Re: [PATCH v2 2/5] drm/xe: Strict migration policy for atomic SVM faults

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

On Mon, Apr 21, 2025 at 12:09:15PM +0530, Ghimiray, Himal Prasad wrote:
> 
> 
> On 17-04-2025 09:43, Matthew Brost wrote:
> > Mixing GPU and CPU atomics does not work unless a strict migration
> > policy of GPU atomics must be device memory. Enforce a policy of must be
> > in VRAM with a retry loop of 2 attempts, if retry loop fails abort
> > fault.
> > 
> > v2:
> >   - Only retry migration on atomics
> >   - Drop alway migrate modparam
> > 
> > Signed-off-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
> > Signed-off-by: Matthew Brost <matthew.brost@intel.com>
> > ---
> >   drivers/gpu/drm/xe/xe_module.c |  3 --
> >   drivers/gpu/drm/xe/xe_module.h |  1 -
> >   drivers/gpu/drm/xe/xe_svm.c    | 57 ++++++++++++++++++++++++++--------
> >   drivers/gpu/drm/xe/xe_svm.h    |  5 ---
> >   4 files changed, 44 insertions(+), 22 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_module.c b/drivers/gpu/drm/xe/xe_module.c
> > index 05c7d0ae6d83..1c4dfafbcd0b 100644
> > --- a/drivers/gpu/drm/xe/xe_module.c
> > +++ b/drivers/gpu/drm/xe/xe_module.c
> > @@ -33,9 +33,6 @@ struct xe_modparam xe_modparam = {
> >   module_param_named(svm_notifier_size, xe_modparam.svm_notifier_size, uint, 0600);
> >   MODULE_PARM_DESC(svm_notifier_size, "Set the svm notifier size(in MiB), must be power of 2");
> > -module_param_named(always_migrate_to_vram, xe_modparam.always_migrate_to_vram, bool, 0444);
> > -MODULE_PARM_DESC(always_migrate_to_vram, "Always migrate to VRAM on GPU fault");
> > -
> >   module_param_named_unsafe(force_execlist, xe_modparam.force_execlist, bool, 0444);
> >   MODULE_PARM_DESC(force_execlist, "Force Execlist submission");
> > diff --git a/drivers/gpu/drm/xe/xe_module.h b/drivers/gpu/drm/xe/xe_module.h
> > index 84339e509c80..5a3bfea8b7b4 100644
> > --- a/drivers/gpu/drm/xe/xe_module.h
> > +++ b/drivers/gpu/drm/xe/xe_module.h
> > @@ -12,7 +12,6 @@
> >   struct xe_modparam {
> >   	bool force_execlist;
> >   	bool probe_display;
> > -	bool always_migrate_to_vram;
> >   	u32 force_vram_bar_size;
> >   	int guc_log_level;
> >   	char *guc_firmware_path;
> > diff --git a/drivers/gpu/drm/xe/xe_svm.c b/drivers/gpu/drm/xe/xe_svm.c
> > index 56b18a293bbc..1cc41ce7b684 100644
> > --- a/drivers/gpu/drm/xe/xe_svm.c
> > +++ b/drivers/gpu/drm/xe/xe_svm.c
> > @@ -726,6 +726,35 @@ static int xe_svm_alloc_vram(struct xe_vm *vm, struct xe_tile *tile,
> >   }
> >   #endif
> > +static bool supports_4K_migration(struct xe_device *xe)
> > +{
> > +	if (xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K)
> > +		return false;
> > +
> > +	return true;
> > +}
> > +
> > +static bool xe_svm_range_needs_migrate_to_vram(struct xe_svm_range *range,
> > +					       struct xe_vma *vma)
> > +{
> > +	struct xe_vm *vm = range_to_vm(&range->base);
> > +	u64 range_size = xe_svm_range_size(range);
> > +
> > +	if (!range->base.flags.migrate_devmem)
> > +		return false;
> > +
> > +	if (xe_svm_range_in_vram(range)) {
> > +		drm_dbg(&vm->xe->drm, "Range is already in VRAM\n");
> > +		return false;
> > +	}
> > +
> > +	if (range_size <= SZ_64K && !supports_4K_migration(vm->xe)) {
> > +		drm_dbg(&vm->xe->drm, "Platform doesn't support SZ_4K range migration\n");
> > +		return false;
> > +	}
> > +
> > +	return true;
> > +}
> >   /**
> >    * xe_svm_handle_pagefault() - SVM handle page fault
> > @@ -750,12 +779,14 @@ int xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma,
> >   			IS_ENABLED(CONFIG_DRM_XE_DEVMEM_MIRROR),
> >   		.check_pages_threshold = IS_DGFX(vm->xe) &&
> >   			IS_ENABLED(CONFIG_DRM_XE_DEVMEM_MIRROR) ? SZ_64K : 0,
> > +		.vram_only = atomic,
> 
> atomic && is_dgfx.

Yes, indeed.

Matt

>  >   	};
> >   	struct xe_svm_range *range;
> >   	struct drm_gpusvm_range *r;
> >   	struct drm_exec exec;
> >   	struct dma_fence *fence;
> >   	struct xe_tile *tile = gt_to_tile(gt);
> > +	int migrate_try_count = atomic ? 3 : 1;
> >   	ktime_t end = 0;
> >   	int err;
> > @@ -782,18 +813,21 @@ int xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma,
> >   	range_debug(range, "PAGE FAULT");
> > -	/* XXX: Add migration policy, for now migrate range once */
> > -	if (!range->skip_migrate && range->base.flags.migrate_devmem &&
> > -	    xe_svm_range_size(range) >= SZ_64K) {
> > -		range->skip_migrate = true;
> > -
> > +	if (--migrate_try_count >= 0 &&
> > +	    xe_svm_range_needs_migrate_to_vram(range, vma)) {
> >   		err = xe_svm_alloc_vram(vm, tile, range, &ctx);
> >   		if (err) {
> > -			drm_dbg(&vm->xe->drm,
> > -				"VRAM allocation failed, falling back to "
> > -				"retrying fault, asid=%u, errno=%pe\n",
> > -				vm->usm.asid, ERR_PTR(err));
> > -			goto retry;
> > +			if (migrate_try_count || !ctx.vram_only) {
> > +				drm_dbg(&vm->xe->drm,
> > +					"VRAM allocation failed, falling back to retrying fault, asid=%u, errno=%pe\n",
> > +					vm->usm.asid, ERR_PTR(err));
> > +				goto retry;
> > +			} else {
> > +				drm_err(&vm->xe->drm,
> > +					"VRAM allocation failed, retry count exceeded, asid=%u, errno=%pe\n",
> > +					vm->usm.asid, ERR_PTR(err));
> > +				return err;
> > +			}
> >   		}
> >   	}
> > @@ -843,9 +877,6 @@ int xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma,
> >   	}
> >   	drm_exec_fini(&exec);
> > -	if (xe_modparam.always_migrate_to_vram)
> > -		range->skip_migrate = false;
> > -
> >   	dma_fence_wait(fence, false);
> >   	dma_fence_put(fence);
> > diff --git a/drivers/gpu/drm/xe/xe_svm.h b/drivers/gpu/drm/xe/xe_svm.h
> > index 3d441eb1f7ea..0e1f376a7471 100644
> > --- a/drivers/gpu/drm/xe/xe_svm.h
> > +++ b/drivers/gpu/drm/xe/xe_svm.h
> > @@ -39,11 +39,6 @@ struct xe_svm_range {
> >   	 * range. Protected by GPU SVM notifier lock.
> >   	 */
> >   	u8 tile_invalidated;
> > -	/**
> > -	 * @skip_migrate: Skip migration to VRAM, protected by GPU fault handler
> > -	 * locking.
> > -	 */
> > -	u8 skip_migrate	:1;
> >   };
> >   /**
>