Re: [PATCH 1/2] cxl/core: Set cxlmd->endpoint to NULL by default

[dan.j.williams]

Li Ming wrote:
[..]
> Hi Dan,
> 
> Thanks for your proposal, I think your change can solve this problem too.
> But the change is a lot, and need more time to review all driver code
> to confirm if there are other places needed such checking. (I found
> that cxl_reset_done also needs some changes like you mentioned) Maybe
> we can consider my change as a quick fix? Then I can prepare a new
> patchset for the consolidation.

I am not convinced that it is a fix. The fact that you say the bug
disappears when patch2 is applied leads me to believe that is
potentially the only bug.

I.e. it may be the case that cxl_dpa_to_region() is safe to assume that
a valid ->endpoint pointer will remain valid once the port
bus_add_device() vs bus_probe_device() hole is plugged.

I would say start with patch2 by itself. Then circle back to prove
that a mere NULL check is a fix or just makes the vulnerability window
smaller and the locking rework is needed.