We need to set it to the most recent completed fence, not the most
recent submitted.  Otherwise we have races where we think we can retire
submits that the GPU is not finished with, if the GPU doesn't manage to
overwrite the seqno before we look at it.

This can show up with hang recovery if one of the submits after the
crashing submit also hangs after it is replayed.

Fixes: f97decac ("drm/msm: Support multiple ringbuffers")
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@chromium.org>