Created attachment 142039 [details]
sample code to reproduce the issue

amdgpu_bo is reference counted, libdrm expects the number of calls to
amdgpu_bo_import must be equal to amdgpu_bo_free.

When importing prime fd into mesa, amdgpu_bo_from_handle will be called, which
calls amdgpu_bo_import to get the unique handle (which in fact is a pointer to
amdgpu_bo), then wrapped into amdgpu_winsys_bo object.

The problem is amdgpu_winsys_bo is also reference counted, if one tries to
import the same prime fd for multiple times, amdgpu_bo_import will be called
for multiple times, but only one amdgpu_winsys_bo will be created, so
amdgpu_bo_free will be called for only once, eventually leaks memory.

I have attached a sample code to reproduce this issue.

a possible solution is to call amdgpu_bo_free after amdgpu_bo_import if
result.buf_handle can be found in ws->bo_export_table, which will balance the
number of calls to amdgpu_bo_import and amdgpu_bo_free.