amdgpu only uses shared-fences internally, but dmabuf importers rely on
implicit write hazard tracking via the reservation_object.fence_excl.
For example, the importer use the write hazard for timing a page flip to
only occur after the exporter has finished flushing its write into the
surface. As such, on exporting a dmabuf, we must either flush all
outstanding fences (for we do not know which are writes and should have
been exclusive) or alternatively create a new exclusive fence that is
the composite of all the existing shared fences, and so will only be
signaled when all earlier fences are signaled (ensuring that we can not
be signaled before the completion of any earlier write).

v2: reservation_object is already locked by amdgpu_bo_reserve()
v3: Replace looping with get_fences_rcu and special case the promotion
of a single shared fence directly to an exclusive fence, bypassing the
fence array.
v4: Drop the fence array ref after assigning to reservation_object

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=107341
Testcase: igt/amd_prime/amd-to-i915
References: 8e94a46c ("drm/amdgpu: Attach exclusive fence to prime exported bo's. (v5)")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: "Christian König" <christian.koenig@amd.com>
Reviewed-by: "Christian König" <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>