Re: [PATCH 4/4] drm/xe: Add watch_queue-based device event notification

[Matthew Brost <matthew.brost@intel.com>]

On Fri, Jun 12, 2026 at 03:53:40PM +0200, Thomas Hellström wrote:
> Add a watch_queue notification channel tied to struct xe_file so that
> userspace can subscribe to asynchronous GPU device events via the
> general kernel notification mechanism.
> 
> Introduce DRM_IOCTL_XE_WATCH_QUEUE to let userspace subscribe a
> notification pipe (opened with pipe2(O_NOTIFICATION_PIPE)) to the device
> event stream.  Embed the watch_id field (0-255) in the WATCH_INFO_ID
> field of every notification, allowing multiple watches to share a single
> pipe and be told apart by the reader.
> 
> Deliver notifications as struct drm_xe_watch_notification records, with
> type always set to WATCH_TYPE_DRM_XE_NOTIFY and subtype drawn from enum
> drm_xe_watch_event.  Define DRM_XE_WATCH_EVENT_VM_ERR as the first
> event, posted by the preempt-rebind worker when a VM encounters an
> unrecoverable error.  Expose xe_watch_queue_post_vm_err_event() as the
> in-kernel posting API.
> 
> Add event definitions in a separate uapi header, <drm/xe_drm_events.h>.
> The main reason is that the header needs to include <linux/watch_queue.h>
> which in turn includes <linux/fcntl.h> which may conflict with the
> system <fcntl.h>. Hence user-space must pay special attention when
> including this file.
> 
> Assisted-by: GitHub_Copilot:claude-sonnet-4.6
> Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> ---
>  MAINTAINERS                          |   1 +
>  drivers/gpu/drm/xe/Kconfig           |   1 +
>  drivers/gpu/drm/xe/Makefile          |   1 +
>  drivers/gpu/drm/xe/xe_device.c       |   7 ++
>  drivers/gpu/drm/xe/xe_device_types.h |   6 ++
>  drivers/gpu/drm/xe/xe_vm.c           |   4 +
>  drivers/gpu/drm/xe/xe_vm_types.h     |   2 +
>  drivers/gpu/drm/xe/xe_watch_queue.c  | 111 +++++++++++++++++++++++++++
>  drivers/gpu/drm/xe/xe_watch_queue.h  |  20 +++++
>  include/uapi/drm/xe_drm.h            |  45 +++++++++++
>  include/uapi/drm/xe_drm_events.h     |  62 +++++++++++++++
>  11 files changed, 260 insertions(+)
>  create mode 100644 drivers/gpu/drm/xe/xe_watch_queue.c
>  create mode 100644 drivers/gpu/drm/xe/xe_watch_queue.h
>  create mode 100644 include/uapi/drm/xe_drm_events.h
> 
> diff --git a/MAINTAINERS b/MAINTAINERS
> index 8c0d9965c636..b7e02cfa692b 100644
> --- a/MAINTAINERS
> +++ b/MAINTAINERS
> @@ -12900,6 +12900,7 @@ F:	Documentation/gpu/xe/
>  F:	drivers/gpu/drm/xe/
>  F:	include/drm/intel/
>  F:	include/uapi/drm/xe_drm.h
> +F:	include/uapi/drm/xe_drm_events.h
>  
>  INTEL ELKHART LAKE PSE I/O DRIVER
>  M:	Raag Jadav <raag.jadav@intel.com>
> diff --git a/drivers/gpu/drm/xe/Kconfig b/drivers/gpu/drm/xe/Kconfig
> index 4d7dcaff2b91..dbdc2fb49c53 100644
> --- a/drivers/gpu/drm/xe/Kconfig
> +++ b/drivers/gpu/drm/xe/Kconfig
> @@ -25,6 +25,7 @@ config DRM_XE
>  	select DRM_MIPI_DSI
>  	select RELAY
>  	select IRQ_WORK
> +	select WATCH_QUEUE
>  	# xe depends on ACPI_VIDEO when ACPI is enabled
>  	# but for select to work, need to select ACPI_VIDEO's dependencies, ick
>  	select BACKLIGHT_CLASS_DEVICE if ACPI
> diff --git a/drivers/gpu/drm/xe/Makefile b/drivers/gpu/drm/xe/Makefile
> index 8e7b146880f4..fc8b4023a044 100644
> --- a/drivers/gpu/drm/xe/Makefile
> +++ b/drivers/gpu/drm/xe/Makefile
> @@ -150,6 +150,7 @@ xe-y += xe_bb.o \
>  	xe_vsec.o \
>  	xe_wa.o \
>  	xe_wait_user_fence.o \
> +	xe_watch_queue.o \
>  	xe_wopcm.o
>  
>  xe-$(CONFIG_I2C)	+= xe_i2c.o
> diff --git a/drivers/gpu/drm/xe/xe_device.c b/drivers/gpu/drm/xe/xe_device.c
> index 867d7c55dc03..788ef2fbd6e5 100644
> --- a/drivers/gpu/drm/xe/xe_device.c
> +++ b/drivers/gpu/drm/xe/xe_device.c
> @@ -9,6 +9,7 @@
>  #include <linux/delay.h>
>  #include <linux/fault-inject.h>
>  #include <linux/units.h>
> +#include <linux/watch_queue.h>
>  
>  #include <drm/drm_client.h>
>  #include <drm/drm_gem_ttm_helper.h>
> @@ -77,6 +78,7 @@
>  #include "xe_vsec.h"
>  #include "xe_wait_user_fence.h"
>  #include "xe_wa.h"
> +#include "xe_watch_queue.h"
>  
>  #include <generated/xe_device_wa_oob.h>
>  #include <generated/xe_wa_oob.h>
> @@ -112,6 +114,8 @@ static int xe_file_open(struct drm_device *dev, struct drm_file *file)
>  	file->driver_priv = xef;
>  	kref_init(&xef->refcount);
>  
> +	init_watch_list(&xef->watch_list, NULL);
> +
>  	task = get_pid_task(rcu_access_pointer(file->pid), PIDTYPE_PID);
>  	if (task) {
>  		xef->process_name = kstrdup(task->comm, GFP_KERNEL);
> @@ -126,6 +130,8 @@ static void xe_file_destroy(struct kref *ref)
>  {
>  	struct xe_file *xef = container_of(ref, struct xe_file, refcount);
>  
> +	remove_watch_from_object(&xef->watch_list, NULL, 0, true);
> +
>  	xa_destroy(&xef->exec_queue.xa);
>  	mutex_destroy(&xef->exec_queue.lock);
>  	xa_destroy(&xef->vm.xa);
> @@ -216,6 +222,7 @@ static const struct drm_ioctl_desc xe_ioctls[] = {
>  	DRM_IOCTL_DEF_DRV(XE_VM_GET_PROPERTY, xe_vm_get_property_ioctl,
>  			  DRM_RENDER_ALLOW),
>  	DRM_IOCTL_DEF_DRV(XE_VM_RESTART, xe_vm_restart_ioctl, DRM_RENDER_ALLOW),
> +	DRM_IOCTL_DEF_DRV(XE_WATCH_QUEUE, xe_watch_queue_ioctl, DRM_RENDER_ALLOW),
>  };
>  
>  static long xe_drm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
> diff --git a/drivers/gpu/drm/xe/xe_device_types.h b/drivers/gpu/drm/xe/xe_device_types.h
> index 32dd2ffbc796..ca726ada30a7 100644
> --- a/drivers/gpu/drm/xe/xe_device_types.h
> +++ b/drivers/gpu/drm/xe/xe_device_types.h
> @@ -11,6 +11,7 @@
>  #include <drm/drm_device.h>
>  #include <drm/drm_file.h>
>  #include <drm/ttm/ttm_device.h>
> +#include <linux/watch_queue.h>
>  
>  #include "xe_devcoredump_types.h"
>  #include "xe_drm_ras_types.h"
> @@ -632,6 +633,11 @@ struct xe_file {
>  
>  	/** @refcount: ref count of this xe file */
>  	struct kref refcount;
> +
> +#ifdef CONFIG_WATCH_QUEUE
> +	/** @watch_list: per-file notification source for device events */
> +	struct watch_list watch_list;
> +#endif
>  };
>  
>  #endif
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index b69a2e5bd9c9..232de0d948d2 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -13,6 +13,7 @@
>  #include <drm/drm_print.h>
>  #include <drm/ttm/ttm_tt.h>
>  #include <uapi/drm/xe_drm.h>
> +#include <uapi/drm/xe_drm_events.h>
>  #include <linux/ascii85.h>
>  #include <linux/delay.h>
>  #include <linux/kthread.h>
> @@ -45,6 +46,7 @@
>  #include "xe_trace_bo.h"
>  #include "xe_vm_madvise.h"
>  #include "xe_wa.h"
> +#include "xe_watch_queue.h"
>  
>  #ifdef CONFIG_FAULT_INJECTION
>  static DECLARE_FAULT_ATTR(rebind_enospc);
> @@ -584,6 +586,7 @@ static void preempt_rebind_work_func(struct work_struct *w)
>  	}
>  
>  	if (err) {
> +		xe_watch_queue_post_vm_err_event(vm->xef, vm->id, err);
>  		if ((err == -ENOMEM || err == -ENOSPC) && xe_vm_is_restartable(vm)) {
>  			vm->preempt.rebind_deactivated = true;
>  			drm_dbg(&vm->xe->drm, "Rebind deactivated VM on error %pe\n",
> @@ -2229,6 +2232,7 @@ int xe_vm_create_ioctl(struct drm_device *dev, void *data,
>  	if (err)
>  		goto err_close_and_put;
>  
> +	vm->id = id;
>  	args->vm_id = id;
>  
>  	return 0;
> diff --git a/drivers/gpu/drm/xe/xe_vm_types.h b/drivers/gpu/drm/xe/xe_vm_types.h
> index 7d295c3b8456..19a673099588 100644
> --- a/drivers/gpu/drm/xe/xe_vm_types.h
> +++ b/drivers/gpu/drm/xe/xe_vm_types.h
> @@ -407,6 +407,8 @@ struct xe_vm {
>  	bool batch_invalidate_tlb;
>  	/** @xef: Xe file handle for tracking this VM's drm client */
>  	struct xe_file *xef;
> +	/** @id: The id of the VM in the VM table of @xef. */
> +	u32 id;
>  };
>  
>  /** struct xe_vma_op_map - VMA map operation */
> diff --git a/drivers/gpu/drm/xe/xe_watch_queue.c b/drivers/gpu/drm/xe/xe_watch_queue.c
> new file mode 100644
> index 000000000000..32763591075b
> --- /dev/null
> +++ b/drivers/gpu/drm/xe/xe_watch_queue.c
> @@ -0,0 +1,111 @@
> +// SPDX-License-Identifier: MIT
> +/*
> + * Copyright © 2026 Intel Corporation
> + */
> +
> +#include <linux/slab.h>
> +#include <linux/timekeeping.h>
> +#include <linux/watch_queue.h>
> +
> +#include <uapi/drm/xe_drm.h>
> +#include <uapi/drm/xe_drm_events.h>
> +
> +#include "xe_device.h"
> +#include "xe_device_types.h"
> +#include "xe_macros.h"
> +#include "xe_watch_queue.h"
> +
> +/**
> + * struct xe_watch_notification_vm_err - kernel-side VM error event notification
> + * @base: common watch notification header; type is %WATCH_TYPE_DRM_XE_NOTIFY,
> + *        subtype is %DRM_XE_WATCH_EVENT_VM_ERR
> + * @vm_id: ID of the VM that hit error
> + * @error_code: error code describing the error condition (negative errno)
> + * @timestamp_ns: CLOCK_MONOTONIC timestamp in nanoseconds at the point the
> + *                error was detected
> + *
> + * Layout mirrors &struct drm_xe_watch_notification_vm_err.
> + */
> +struct xe_watch_notification_vm_err {
> +	struct watch_notification base;
> +	u32 vm_id;
> +	s32 error_code;
> +	u64 timestamp_ns;
> +};
> +
> +/**
> + * xe_watch_queue_ioctl() - Subscribe a pipe to per-file device event notifications
> + * @dev: DRM device
> + * @data: pointer to &struct drm_xe_watch_queue from userspace
> + * @file: DRM file handle of the subscribing process
> + *
> + * Subscribes a notification pipe to receive Xe device events for the calling
> + * process's file handle.  Only events scoped to this file (e.g. VM error on a
> + * VM owned by this file) are delivered.  The pipe must have been opened with
> + * O_NOTIFICATION_PIPE and sized with %IOC_WATCH_QUEUE_SET_SIZE before calling
> + * this IOCTL.
> + *
> + * Return: 0 on success, negative errno on failure.
> + */
> +int xe_watch_queue_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
> +{
> +	struct xe_file *xef = file->driver_priv;
> +	struct xe_device *xe = to_xe_device(dev);
> +	struct drm_xe_watch_queue *args = data;
> +	struct watch_queue *wqueue;
> +	struct watch *watch;
> +	int ret;
> +
> +	if (XE_IOCTL_DBG(xe, args->flags || args->pad))
> +		return -EINVAL;
> +	if (XE_IOCTL_DBG(xe, args->watch_id > 0xff))
> +		return -EINVAL;
> +
> +	wqueue = get_watch_queue(args->fd);
> +	if (XE_IOCTL_DBG(xe, IS_ERR(wqueue)))
> +		return PTR_ERR(wqueue);
> +
> +	watch = kzalloc_obj(*watch, GFP_KERNEL | __GFP_ACCOUNT);
> +	if (XE_IOCTL_DBG(xe, !watch)) {
> +		ret = -ENOMEM;
> +		goto out_put_queue;
> +	}
> +
> +	init_watch(watch, wqueue);
> +	watch->id = 0;
> +	watch->info_id = (u32)args->watch_id << WATCH_INFO_ID__SHIFT;
> +
> +	ret = add_watch_to_object(watch, &xef->watch_list);
> +	if (XE_IOCTL_DBG(xe, ret))
> +		kfree(watch);
> +
> +out_put_queue:
> +	put_watch_queue(wqueue);
> +	return ret;
> +}
> +
> +/**
> + * xe_watch_queue_post_vm_err_event() - Post a VM error event
> + * @xef: xe file handle that owns the VM
> + * @vm_id: userspace ID of the VM that hit error
> + * @error_code: error code describing the error condition (negative errno)
> + *
> + * Posts a %DRM_XE_WATCH_EVENT_VM_ERR notification carrying @vm_id and
> + * @error_code to every pipe that @xef has subscribed via
> + * %DRM_IOCTL_XE_WATCH_QUEUE.  Only the owning process is notified,
> + * preventing information leaks to other clients.
> + */
> +void xe_watch_queue_post_vm_err_event(struct xe_file *xef, u32 vm_id,
> +				      int error_code)
> +{
> +	struct xe_watch_notification_vm_err n = {};
> +
> +	n.base.type    = WATCH_TYPE_DRM_XE_NOTIFY;
> +	n.base.subtype = DRM_XE_WATCH_EVENT_VM_ERR;
> +	n.base.info    = watch_sizeof(struct xe_watch_notification_vm_err);
> +	n.vm_id        = vm_id;
> +	n.error_code   = error_code;
> +	n.timestamp_ns = ktime_get_ns();
> +
> +	post_watch_notification(&xef->watch_list, &n.base, current_cred(), 0);

Is this reclaim safe (i.e., could we post a watch notification in direct
reclaim from our shrinker)? From my look, I think the answer is yes. I
ask because whatever we land here, I think the answer must be KMD->UMD
events are reclaim safe.

Matt 

> +}
> diff --git a/drivers/gpu/drm/xe/xe_watch_queue.h b/drivers/gpu/drm/xe/xe_watch_queue.h
> new file mode 100644
> index 000000000000..ad199ee68205
> --- /dev/null
> +++ b/drivers/gpu/drm/xe/xe_watch_queue.h
> @@ -0,0 +1,20 @@
> +/* SPDX-License-Identifier: MIT */
> +/*
> + * Copyright © 2026 Intel Corporation
> + */
> +
> +#ifndef _XE_WATCH_QUEUE_H_
> +#define _XE_WATCH_QUEUE_H_
> +
> +#include <linux/types.h>
> +
> +struct drm_device;
> +struct drm_file;
> +struct xe_file;
> +
> +int xe_watch_queue_ioctl(struct drm_device *dev, void *data,
> +			 struct drm_file *file);
> +void xe_watch_queue_post_vm_err_event(struct xe_file *xef, u32 vm_id,
> +				      int error_code);
> +
> +#endif /* _XE_WATCH_QUEUE_H_ */
> diff --git a/include/uapi/drm/xe_drm.h b/include/uapi/drm/xe_drm.h
> index 8d5e3f06b8d4..0083dd712f7e 100644
> --- a/include/uapi/drm/xe_drm.h
> +++ b/include/uapi/drm/xe_drm.h
> @@ -86,6 +86,7 @@ extern "C" {
>   *  - &DRM_IOCTL_XE_EXEC_QUEUE_SET_PROPERTY
>   *  - &DRM_IOCTL_XE_VM_GET_PROPERTY
>   *  - &DRM_IOCTL_XE_VM_RESTART
> + *  - &DRM_IOCTL_XE_WATCH_QUEUE
>   */
>  
>  /*
> @@ -112,6 +113,7 @@ extern "C" {
>  #define DRM_XE_EXEC_QUEUE_SET_PROPERTY	0x0e
>  #define DRM_XE_VM_GET_PROPERTY		0x0f
>  #define DRM_XE_VM_RESTART		0x10
> +#define DRM_XE_WATCH_QUEUE		0x11
>  
>  /* Must be kept compact -- no holes */
>  
> @@ -132,6 +134,7 @@ extern "C" {
>  #define DRM_IOCTL_XE_EXEC_QUEUE_SET_PROPERTY	DRM_IOW(DRM_COMMAND_BASE + DRM_XE_EXEC_QUEUE_SET_PROPERTY, struct drm_xe_exec_queue_set_property)
>  #define DRM_IOCTL_XE_VM_GET_PROPERTY		DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_VM_GET_PROPERTY, struct drm_xe_vm_get_property)
>  #define DRM_IOCTL_XE_VM_RESTART			DRM_IOW(DRM_COMMAND_BASE + DRM_XE_VM_RESTART, struct drm_xe_vm_restart)
> +#define DRM_IOCTL_XE_WATCH_QUEUE		DRM_IOW(DRM_COMMAND_BASE + DRM_XE_WATCH_QUEUE, struct drm_xe_watch_queue)
>  
>  /**
>   * DOC: Xe IOCTL Extensions
> @@ -2653,6 +2656,48 @@ enum drm_xe_ras_error_component {
>  	[DRM_XE_RAS_ERR_COMP_SOC_INTERNAL] = "soc-internal"		\
>  }
>  
> +/**
> + * DOC: DRM_XE_WATCH_QUEUE
> + *
> + * Subscribe a notification pipe to receive device events for the calling
> + * process's DRM file handle.  Events are scoped to the subscribing file:
> + * only events that belong to that file (for example, VM error on a VM created
> + * through the same file) are delivered, preventing information leaks between
> + * processes sharing the same GPU device.
> + *
> + * The pipe must first be opened with O_NOTIFICATION_PIPE (i.e. O_EXCL passed
> + * to pipe2()) and sized via %IOC_WATCH_QUEUE_SET_SIZE before subscribing.
> + *
> + * Events are delivered as notification records read from the pipe.  The
> + * @watch_id field is embedded in the notification info field and can be used
> + * to distinguish multiple watches sharing a pipe.
> + *
> + * Currently defined event subtypes:
> + *  - %DRM_XE_WATCH_EVENT_VM_ERR - a VM owned by this file has encountered an error
> + */
> +
> +/**
> + * struct drm_xe_watch_queue - subscribe to device event notifications
> + *
> + * Used with %DRM_IOCTL_XE_WATCH_QUEUE.  Notifications are scoped to the
> + * DRM file handle used to issue this IOCTL.
> + */
> +struct drm_xe_watch_queue {
> +	/** @fd: file descriptor of pipe opened with O_NOTIFICATION_PIPE */
> +	__u32 fd;
> +
> +	/**
> +	 * @watch_id: identifier (0–255) embedded in the watch notification
> +	 * info field; allows multiplexing several watches on one pipe
> +	 */
> +	__u32 watch_id;
> +
> +	/** @flags: must be zero */
> +	__u32 flags;
> +
> +	/** @pad: reserved, must be zero */
> +	__u32 pad;
> +};
>  #if defined(__cplusplus)
>  }
>  #endif
> diff --git a/include/uapi/drm/xe_drm_events.h b/include/uapi/drm/xe_drm_events.h
> new file mode 100644
> index 000000000000..6cc7528bfb9b
> --- /dev/null
> +++ b/include/uapi/drm/xe_drm_events.h
> @@ -0,0 +1,62 @@
> +/* SPDX-License-Identifier: MIT */
> +/*
> + * Copyright © 2026 Intel Corporation
> + */
> +
> +#ifndef _UAPI_XE_DRM_EVENTS_H_
> +#define _UAPI_XE_DRM_EVENTS_H_
> +
> +#include <linux/types.h>
> +#include <linux/watch_queue.h>
> +
> +#if defined(__cplusplus)
> +extern "C" {
> +#endif
> +
> +/**
> + * enum drm_xe_watch_event - Xe device watch event subtypes
> + *
> + * Subtypes for notifications delivered via %WATCH_TYPE_DRM_XE_NOTIFY when
> + * reading from a pipe subscribed with %DRM_IOCTL_XE_WATCH_QUEUE.
> + */
> +enum drm_xe_watch_event {
> +	/**
> +	 * @DRM_XE_WATCH_EVENT_VM_ERR: a VM has encountered an error.
> +	 *
> +	 * Indicates that a fatal or resource error occurred within the given
> +	 * VM.  The vm_id of the affected VM is carried in the
> +	 * @drm_xe_watch_notification_vm_err::vm_id field of the extended
> +	 * notification record.
> +	 */
> +	DRM_XE_WATCH_EVENT_VM_ERR = 0,
> +};
> +
> +/**
> + * struct drm_xe_watch_notification_vm_err - VM error event notification
> + *
> + * Notification record delivered for %DRM_XE_WATCH_EVENT_VM_ERR.
> + * The record type is always %WATCH_TYPE_DRM_XE_NOTIFY and the subtype is
> + * %DRM_XE_WATCH_EVENT_VM_ERR.
> + */
> +struct drm_xe_watch_notification_vm_err {
> +	/** @base: common watch notification header */
> +	struct watch_notification base;
> +
> +	/** @vm_id: ID of the VM that encountered an error */
> +	__u32 vm_id;
> +
> +	/** @error_code: error code describing the error condition (negative errno) */
> +	__s32 error_code;
> +
> +	/**
> +	 * @timestamp_ns: CLOCK_MONOTONIC timestamp in nanoseconds at the
> +	 * point the error was detected
> +	 */
> +	__u64 timestamp_ns;
> +};
> +
> +#if defined(__cplusplus)
> +}
> +#endif
> +
> +#endif /* _UAPI_XE_DRM_EVENTS_H_ */
> -- 
> 2.54.0
>