Re: [PATCH v2] drm/xe: Rework eviction rejection of bound external bos

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On Tue, 2025-05-27 at 11:57 -0700, Matthew Brost wrote:
> On Mon, May 26, 2025 at 06:16:32PM +0200, Thomas Hellström wrote:
> > For preempt_fence mode VM's we're rejecting eviction of
> > shared bos during VM_BIND. However, since we do this in the
> > move() callback, we're getting an eviction failure warning from
> > TTM. The TTM callback intended for these things is
> > eviction_valuable().
> > 
> > However, the latter doesn't pass in the struct ttm_operation_ctx
> > needed to determine whether the caller needs this.
> > 
> > Instead, attach the needed information to the vm under the
> > vm->resv, until we've been able to update TTM to provide the
> > needed information. And add sufficient lockdep checks to prevent
> > misuse and races.
> > 
> > v2:
> > - Fix a copy-paste error in xe_vm_clear_validating()
> > 
> > Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> > Fixes: 0af944f0e308 ("drm/xe: Reject BO eviction if BO is bound to
> > current VM")
> > Cc: Oak Zeng <oak.zeng@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_bo.c       | 46 +++++++++++++---------
> >  drivers/gpu/drm/xe/xe_vm.h       | 65
> > ++++++++++++++++++++++++++++++++
> >  drivers/gpu/drm/xe/xe_vm_types.h |  8 ++++
> >  3 files changed, 101 insertions(+), 18 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_bo.c
> > b/drivers/gpu/drm/xe/xe_bo.c
> > index d99d91fe8aa9..3c48a8c5f439 100644
> > --- a/drivers/gpu/drm/xe/xe_bo.c
> > +++ b/drivers/gpu/drm/xe/xe_bo.c
> > @@ -841,21 +841,6 @@ static int xe_bo_move(struct ttm_buffer_object
> > *ttm_bo, bool evict,
> >  		goto out;
> >  	}
> >  
> > -	/* Reject BO eviction if BO is bound to current VM. */
> > -	if (evict && ctx->resv) {
> > -		struct drm_gpuvm_bo *vm_bo;
> > -
> > -		drm_gem_for_each_gpuvm_bo(vm_bo, &bo->ttm.base) {
> > -			struct xe_vm *vm = gpuvm_to_vm(vm_bo->vm);
> > -
> > -			if (xe_vm_resv(vm) == ctx->resv &&
> > -			    xe_vm_in_preempt_fence_mode(vm)) {
> > -				ret = -EBUSY;
> > -				goto out;
> > -			}
> > -		}
> > -	}
> > -
> >  	/*
> >  	 * Failed multi-hop where the old_mem is still marked as
> >  	 * TTM_PL_FLAG_TEMPORARY, should just be a dummy move.
> > @@ -1013,6 +998,25 @@ static long xe_bo_shrink_purge(struct
> > ttm_operation_ctx *ctx,
> >  	return lret;
> >  }
> >  
> > +static bool
> > +xe_bo_eviction_valuable(struct ttm_buffer_object *bo, const struct
> > ttm_place *place)
> > +{
> > +	struct drm_gpuvm_bo *vm_bo;
> > +
> > +	if (!ttm_bo_eviction_valuable(bo, place))
> > +		return false;
> > +
> > +	if (!xe_bo_is_xe_bo(bo))
> > +		return true;
> > +
> > +	drm_gem_for_each_gpuvm_bo(vm_bo, &bo->base) {
> > +		if (xe_vm_is_validating(gpuvm_to_vm(vm_bo->vm)))
> > +			return false;
> > +	}
> > +
> > +	return true;
> > +}
> > +
> >  /**
> >   * xe_bo_shrink() - Try to shrink an xe bo.
> >   * @ctx: The struct ttm_operation_ctx used for shrinking.
> > @@ -1047,7 +1051,7 @@ long xe_bo_shrink(struct ttm_operation_ctx
> > *ctx, struct ttm_buffer_object *bo,
> >  	    (flags.purge && !xe_tt->purgeable))
> >  		return -EBUSY;
> >  
> > -	if (!ttm_bo_eviction_valuable(bo, &place))
> > +	if (!xe_bo_eviction_valuable(bo, &place))
> >  		return -EBUSY;
> >  
> >  	if (!xe_bo_is_xe_bo(bo) || !xe_bo_get_unless_zero(xe_bo))
> > @@ -1588,7 +1592,7 @@ const struct ttm_device_funcs xe_ttm_funcs =
> > {
> >  	.io_mem_pfn = xe_ttm_io_mem_pfn,
> >  	.access_memory = xe_ttm_access_memory,
> >  	.release_notify = xe_ttm_bo_release_notify,
> > -	.eviction_valuable = ttm_bo_eviction_valuable,
> > +	.eviction_valuable = xe_bo_eviction_valuable,
> >  	.delete_mem_notify = xe_ttm_bo_delete_mem_notify,
> >  	.swap_notify = xe_ttm_bo_swap_notify,
> >  };
> > @@ -2431,6 +2435,8 @@ int xe_bo_validate(struct xe_bo *bo, struct
> > xe_vm *vm, bool allow_res_evict)
> >  		.no_wait_gpu = false,
> >  		.gfp_retry_mayfail = true,
> >  	};
> > +	struct pin_cookie cookie;
> > +	int ret;
> >  
> >  	if (vm) {
> >  		lockdep_assert_held(&vm->lock);
> > @@ -2440,8 +2446,12 @@ int xe_bo_validate(struct xe_bo *bo, struct
> > xe_vm *vm, bool allow_res_evict)
> >  		ctx.resv = xe_vm_resv(vm);
> >  	}
> >  
> > +	cookie = xe_vm_set_validating(vm, allow_res_evict);
> >  	trace_xe_bo_validate(bo);
> > -	return ttm_bo_validate(&bo->ttm, &bo->placement, &ctx);
> > +	ret = ttm_bo_validate(&bo->ttm, &bo->placement, &ctx);
> > +	xe_vm_clear_validating(vm, allow_res_evict, cookie);
> > +
> > +	return ret;
> >  }
> >  
> >  bool xe_bo_is_xe_bo(struct ttm_buffer_object *bo)
> > diff --git a/drivers/gpu/drm/xe/xe_vm.h
> > b/drivers/gpu/drm/xe/xe_vm.h
> > index 99e164852f63..fe05c4f7f425 100644
> > --- a/drivers/gpu/drm/xe/xe_vm.h
> > +++ b/drivers/gpu/drm/xe/xe_vm.h
> > @@ -303,6 +303,71 @@ void xe_vm_snapshot_capture_delayed(struct
> > xe_vm_snapshot *snap);
> >  void xe_vm_snapshot_print(struct xe_vm_snapshot *snap, struct
> > drm_printer *p);
> >  void xe_vm_snapshot_free(struct xe_vm_snapshot *snap);
> >  
> > +/**
> > + * xe_vm_set_validating() - Register this task as currently making
> > bos resident
> > + * @vm: Pointer to the vm or NULL.
> > + *
> > + * Register this task as currently making bos resident for the vm.
> > Intended
> > + * to avoid eviction by the same task of shared bos bound to the
> > vm.
> > + * Call with the vm's resv lock held.
> > + *
> > + * Return: A pin cookie that should be used for
> > xe_vm_clear_validating().
> > + */
> > +static inline struct pin_cookie xe_vm_set_validating(struct xe_vm
> > *vm,
> > +						     bool
> > allow_res_evict)
> > +{
> > +	struct pin_cookie cookie = {};
> > +
> > +	if (vm && !allow_res_evict) {
> > +		xe_vm_assert_held(vm);
> > +		cookie = lockdep_pin_lock(&xe_vm_resv(vm)-
> > >lock.base);
> 
> I'm little unclear on why the cookie is needed / what
> lockdep_pin_lock
> does. I looked implementation for lockdep_pin_lock and it completely
> undocumented - could you explain?

lockdep_pin_lock() warns if someone somewhere deeper in the call chain
for some reason unlocks the specific lock while the lock is pinned.

In this case that would mean that vm->validating would be set without
the vm->resv was held.

Mostly paranoia.

/Thomas

> 
> Everything else makes sense and LGTM.
> 
> Matt
> 
> > +		/* Pairs with READ_ONCE in xe_vm_is_validating()
> > */
> > +		WRITE_ONCE(vm->validating, current);
> > +	}
> > +
> > +	return cookie;
> > +}
> > +
> > +/**
> > + * xe_vm_clear_validating() - Unregister this task as currently
> > making bos resident
> > + * @vm: Pointer to the vm or NULL
> > + * @cookie: Cookie obtained from xe_vm_set_validating().
> > + *
> > + * Register this task as currently making bos resident for the vm.
> > Intended
> > + * to avoid eviction by the same task of shared bos bound to the
> > vm.
> > + * Call with the vm's resv lock held.
> > + */
> > +static inline void xe_vm_clear_validating(struct xe_vm *vm, bool
> > allow_res_evict,
> > +					  struct pin_cookie
> > cookie)
> > +{
> > +	if (vm && !allow_res_evict) {
> > +		lockdep_unpin_lock(&xe_vm_resv(vm)->lock.base,
> > cookie);
> > +		/* Pairs with READ_ONCE in xe_vm_is_validating()
> > */
> > +		WRITE_ONCE(vm->validating, NULL);
> > +	}
> > +}
> > +
> > +/**
> > + * xe_vm_is_validating() - Whether bos bound to the vm are
> > currently being made resident
> > + * by the current task.
> > + * @vm: Pointer to the vm.
> > + *
> > + * If this function returns %true, we should be in a vm resv
> > locked region, since
> > + * the current process is the same task that called
> > xe_vm_set_validating().
> > + * The function asserts that that's indeed the case.
> > + *
> > + * Return: %true if the task is currently making bos resident,
> > %false otherwise.
> > + */
> > +static inline bool xe_vm_is_validating(struct xe_vm *vm)
> > +{
> > +	/* Pairs with WRITE_ONCE in xe_vm_is_validating() */
> > +	if (READ_ONCE(vm->validating) == current) {
> > +		xe_vm_assert_held(vm);
> > +		return true;
> > +	}
> > +	return false;
> > +}
> > +
> >  #if IS_ENABLED(CONFIG_DRM_XE_USERPTR_INVAL_INJECT)
> >  void xe_vma_userptr_force_invalidate(struct xe_userptr_vma *uvma);
> >  #else
> > diff --git a/drivers/gpu/drm/xe/xe_vm_types.h
> > b/drivers/gpu/drm/xe/xe_vm_types.h
> > index bfc145baad49..0e1318a15c9e 100644
> > --- a/drivers/gpu/drm/xe/xe_vm_types.h
> > +++ b/drivers/gpu/drm/xe/xe_vm_types.h
> > @@ -310,6 +310,14 @@ struct xe_vm {
> >  	 * protected by the vm resv.
> >  	 */
> >  	u64 tlb_flush_seqno;
> > +	/**
> > +	 * @validating: The task that is currently making bos
> > resident for this vm.
> > +	 * Protected by the VM's resv for writing. Opportunistic
> > reading can be done
> > +	 * using READ_ONCE. Note: This is a workaround for the
> > +	 * TTM eviction_valuable() callback not being passed a
> > struct
> > +	 * ttm_operation_context(). Future work might want to
> > address this.
> > +	 */
> > +	struct task_struct *validating;
> >  	/** @batch_invalidate_tlb: Always invalidate TLB before
> > batch start */
> >  	bool batch_invalidate_tlb;
> >  	/** @xef: XE file handle for tracking this VM's drm client
> > */
> > -- 
> > 2.49.0
> >