Re: [PATCH] drm/xe/uapi: Introduce a flag to disallow vm overcommit in fault mode

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

On Thu, Feb 05, 2026 at 09:06:24AM +0100, Thomas Hellström wrote:
> On Wed, 2026-02-04 at 10:56 -0800, Matthew Brost wrote:
> > On Wed, Feb 04, 2026 at 04:33:20PM +0100, Thomas Hellström wrote:
> > > Some compute applications may try to allocate device memory to
> > > probe
> > > how much device memory is actually available, assuming that the
> > > application will be the only one running on the particular GPU.
> > > 
> > > That strategy fails in fault mode since it allows VM overcommit.
> > > 
> > > While this could be resolved in user-space it's further complicated
> > > by cgroups potentially restricting the amount of memory available
> > > to the application.
> > > 
> > > Introduce a vm create flag, DRM_XE_VM_CREATE_NO_VM_OVERCOMMIT, that
> > > allows fault mode to mimic the behaviour of !fault mode WRT this.
> > > It
> > > blocks evicting same vm bos during VM_BIND processing. However,
> > > it does *not* block evicting same-vm bos during pagefault
> > > processing, preferring eviction rather than VM banning in
> > > OOM situations.
> > > 
> > > Cc: John Falkowski <john.falkowski@intel.com>
> > > Cc: Michal Mrozek <michal.mrozek@intel.com>
> > > Cc: Matthew Brost <matthew.brost@intel.com>
> > > Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> > > ---
> > >  drivers/gpu/drm/xe/xe_vm.c       | 11 +++++++++--
> > >  drivers/gpu/drm/xe/xe_vm.h       |  7 +++++++
> > >  drivers/gpu/drm/xe/xe_vm_types.h |  1 +
> > >  include/uapi/drm/xe_drm.h        |  6 ++++++
> > >  4 files changed, 23 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/xe/xe_vm.c
> > > b/drivers/gpu/drm/xe/xe_vm.c
> > > index 8fe54a998385..cf92b6e13a16 100644
> > > --- a/drivers/gpu/drm/xe/xe_vm.c
> > > +++ b/drivers/gpu/drm/xe/xe_vm.c
> > > @@ -1938,7 +1938,8 @@ find_ufence_get(struct xe_sync_entry *syncs,
> > > u32 num_syncs)
> > >  
> > >  #define ALL_DRM_XE_VM_CREATE_FLAGS
> > > (DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE | \
> > >  				    DRM_XE_VM_CREATE_FLAG_LR_MODE
> > > | \
> > > -				   
> > > DRM_XE_VM_CREATE_FLAG_FAULT_MODE)
> > > +				   
> > > DRM_XE_VM_CREATE_FLAG_FAULT_MODE | \
> > > +				   
> > > DRM_XE_VM_CREATE_FLAG_NO_VM_OVERCOMMIT)
> > >  
> > >  int xe_vm_create_ioctl(struct drm_device *dev, void *data,
> > >  		       struct drm_file *file)
> > > @@ -1977,12 +1978,18 @@ int xe_vm_create_ioctl(struct drm_device
> > > *dev, void *data,
> > >  			 args->flags &
> > > DRM_XE_VM_CREATE_FLAG_FAULT_MODE))
> > >  		return -EINVAL;
> > >  
> > > +	if (XE_IOCTL_DBG(xe, !(args->flags &
> > > DRM_XE_VM_CREATE_FLAG_FAULT_MODE) &&
> > > +			 args->flags &
> > > DRM_XE_VM_CREATE_FLAG_NO_VM_OVERCOMMIT))
> > > +		return -EINVAL;
> > > +
> > >  	if (args->flags & DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE)
> > >  		flags |= XE_VM_FLAG_SCRATCH_PAGE;
> > >  	if (args->flags & DRM_XE_VM_CREATE_FLAG_LR_MODE)
> > >  		flags |= XE_VM_FLAG_LR_MODE;
> > >  	if (args->flags & DRM_XE_VM_CREATE_FLAG_FAULT_MODE)
> > >  		flags |= XE_VM_FLAG_FAULT_MODE;
> > > +	if (args->flags & DRM_XE_VM_CREATE_FLAG_NO_VM_OVERCOMMIT)
> > > +		flags |= XE_VM_FLAG_NO_VM_OVERCOMMIT;
> > >  
> > >  	vm = xe_vm_create(xe, flags, xef);
> > >  	if (IS_ERR(vm))
> > > @@ -2903,7 +2910,7 @@ static int vma_lock_and_validate(struct
> > > drm_exec *exec, struct xe_vma *vma,
> > >  			err = drm_exec_lock_obj(exec, &bo-
> > > >ttm.base);
> > >  		if (!err && validate)
> > >  			err = xe_bo_validate(bo, vm,
> > > -					    
> > > !xe_vm_in_preempt_fence_mode(vm) &&
> > > +					    
> > > xe_vm_allow_vm_eviction(vm) &&
> > 
> > One question. This is existing code but can you refresh my memory why
> > we
> > allow overcommit on dma-fencing VMs? Wouldn't the next exec IOCTL
> > immediately fail?
> 
> Yes, but mesa can handle that by user-space eviction. In practice IIRC
> they unbind everything and re-bind on demand. No care for performance.
> 

Right, no I recall the background.

This patch LGTM:
Reviewed-by: Matthew Brost <matthew.brost@intel.com>

> For preempt-fence mode we have the dreaded oom-in-the-rebind-worker
> which we still haven't fixed with a UMD notification, so we're forced
> to avoid that if at all possible.
> 
> /Thomas
> 
> 
> > 
> > Matt
> > 
> > >  					     res_evict, exec);
> > >  	}
> > >  
> > > diff --git a/drivers/gpu/drm/xe/xe_vm.h
> > > b/drivers/gpu/drm/xe/xe_vm.h
> > > index 288115c7844a..f849e369432b 100644
> > > --- a/drivers/gpu/drm/xe/xe_vm.h
> > > +++ b/drivers/gpu/drm/xe/xe_vm.h
> > > @@ -220,6 +220,13 @@ static inline bool
> > > xe_vm_in_preempt_fence_mode(struct xe_vm *vm)
> > >  	return xe_vm_in_lr_mode(vm) && !xe_vm_in_fault_mode(vm);
> > >  }
> > >  
> > > +static inline bool xe_vm_allow_vm_eviction(struct xe_vm *vm)
> > > +{
> > > +	return !xe_vm_in_lr_mode(vm) ||
> > > +		(xe_vm_in_fault_mode(vm) &&
> > > +		 !(vm->flags & XE_VM_FLAG_NO_VM_OVERCOMMIT));
> > > +}
> > > +
> > >  int xe_vm_add_compute_exec_queue(struct xe_vm *vm, struct
> > > xe_exec_queue *q);
> > >  void xe_vm_remove_compute_exec_queue(struct xe_vm *vm, struct
> > > xe_exec_queue *q);
> > >  
> > > diff --git a/drivers/gpu/drm/xe/xe_vm_types.h
> > > b/drivers/gpu/drm/xe/xe_vm_types.h
> > > index 43203e90ee3e..1f6f7e30e751 100644
> > > --- a/drivers/gpu/drm/xe/xe_vm_types.h
> > > +++ b/drivers/gpu/drm/xe/xe_vm_types.h
> > > @@ -232,6 +232,7 @@ struct xe_vm {
> > >  #define XE_VM_FLAG_TILE_ID(flags)	FIELD_GET(GENMASK(7, 6),
> > > flags)
> > >  #define XE_VM_FLAG_SET_TILE_ID(tile)	FIELD_PREP(GENMASK(7, 6),
> > > (tile)->id)
> > >  #define XE_VM_FLAG_GSC			BIT(8)
> > > +#define XE_VM_FLAG_NO_VM_OVERCOMMIT     BIT(9)
> > >  	unsigned long flags;
> > >  
> > >  	/**
> > > diff --git a/include/uapi/drm/xe_drm.h b/include/uapi/drm/xe_drm.h
> > > index 077e66a682e2..e54f8e12acd9 100644
> > > --- a/include/uapi/drm/xe_drm.h
> > > +++ b/include/uapi/drm/xe_drm.h
> > > @@ -975,6 +975,11 @@ struct drm_xe_gem_mmap_offset {
> > >   *    demand when accessed, and also allows per-VM overcommit of
> > > memory.
> > >   *    The xe driver internally uses recoverable pagefaults to
> > > implement
> > >   *    this.
> > > + *  - %DRM_XE_VM_CREATE_FLAG_NO_VM_OVERCOMMIT - Requires also
> > > + *    DRM_XE_VM_CREATE_FLAG_FAULT_MODE. This disallows per-VM
> > > overcommit
> > > + *    but only during a &DRM_IOCTL_XE_VM_BIND operation with the
> > > + *    %DRM_XE_VM_BIND_FLAG_IMMEDIATE flag set. This may be useful
> > > for
> > > + *    user-space naively probing the amount of available memory.
> > >   */
> > >  struct drm_xe_vm_create {
> > >  	/** @extensions: Pointer to the first extension struct, if
> > > any */
> > > @@ -983,6 +988,7 @@ struct drm_xe_vm_create {
> > >  #define DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE	(1 << 0)
> > >  #define DRM_XE_VM_CREATE_FLAG_LR_MODE	        (1 << 1)
> > >  #define DRM_XE_VM_CREATE_FLAG_FAULT_MODE	(1 << 2)
> > > +#define DRM_XE_VM_CREATE_FLAG_NO_VM_OVERCOMMIT  (1 << 3)
> > >  	/** @flags: Flags */
> > >  	__u32 flags;
> > >  
> > > -- 
> > > 2.52.0
> > >