drm/xe/xe_sync: avoid race during ufence signaling

Marking ufence as signalled after copy_to_user() is too late.
Worker thread which signals ufence by memory write might be raced
with another userspace vm-bind call. In map/unmap scenario unmap
may still see ufence is not signalled causing -EBUSY. Change the
order of marking / write to user-fence fixes this issue.

Fixes: 977e5b82e090 ("drm/xe: Expose user fence from xe_sync_entry")
Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/5536
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Link: https://lore.kernel.org/r/20250820083903.2109891-2-zbigniew.kempczynski@intel.com
(cherry picked from commit 8ae04fe9ffc93d6bc3bc63ac08375427d69cee06)
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>

diff --git a/drivers/gpu/drm/xe/xe_sync.c b/drivers/gpu/drm/xe/xe_sync.c
index f87276df18f280ea0dacc875952a75caf405841f..82872a51f0983aa9654a0276c285a64094dcaed1 100644 (file)
--- a/drivers/gpu/drm/xe/xe_sync.c
+++ b/drivers/gpu/drm/xe/xe_sync.c
@@ -77,6 +77,7 @@ static void user_fence_worker(struct work_struct *w)
 {
        struct xe_user_fence *ufence = container_of(w, struct xe_user_fence, worker);
 
+       WRITE_ONCE(ufence->signalled, 1);
        if (mmget_not_zero(ufence->mm)) {
                kthread_use_mm(ufence->mm);
                if (copy_to_user(ufence->addr, &ufence->value, sizeof(ufence->value)))
@@ -91,7 +92,6 @@ static void user_fence_worker(struct work_struct *w)
         * Wake up waiters only after updating the ufence state, allowing the UMD
         * to safely reuse the same ufence without encountering -EBUSY errors.
         */
-       WRITE_ONCE(ufence->signalled, 1);
        wake_up_all(&ufence->xe->ufence_wq);
        user_fence_put(ufence);
 }