Re: [PATCH] cxl/mbox: Store event records for EDAC repair regardless of tracing

[Dave Jiang <dave.jiang@intel.com>]

On 9/25/26 4:50 PM, Jonathan Cameron wrote:
> On Fri, 25 Sep 2026 15:53:23 -0700
> Dave Jiang <dave.jiang@intel.com> wrote:
> 
>> On 9/25/26 3:44 PM, Jonathan Cameron wrote:
>>> On Thu, 24 Sep 2026 14:23:20 -0700
>>> Dave Jiang <dave.jiang@intel.com> wrote:
>>>   
>>>> The driver throws away media error records unless a tracepoint is enabled.
>>>>
>>>> cxl_event_trace_record() stores general media and DRAM records for EDAC
>>>> repair, but both calls sit inside a block gated on the cxl_general_media
>>>> and cxl_dram tracepoints. That gate only exists to skip the DPA-to-HPA
>>>> lookup used to annotate those tracepoints. Storing a record does not need
>>>> the lookup.
>>>>
>>>> With both tracepoints off, two things break. The kernel clears the device
>>>> event log either way, so the records are destroyed in hardware and kept
>>>> nowhere else. And live repair stops working altogether: sPPR and memory
>>>> sparing require the target DPA to have reported an error, no DPA ever has
>>>> a record, so every request fails with -EINVAL.  
>>>
>>> Hi Dave,
>>>
>>> Given userspace is always in that path (I think), without the
>>> tracepoints being enabled it will be one impressive guess to
>>> successfully make repair happen - you'll have to magically know
>>> what address will see a hit in the stored error records.
>>>
>>> Maybe that happens in some test case, but it doesn't feel real
>>> to me.  
>>
>> If you don't think that's possible then I'll drop the patch.
>>
> 
> It is possible if you injected the error and so know where it is.
> We may also get other paths by which people obtain that info such
> as anyone who is reporting via fabric management paths.  I just
> don't think it is a fix and perhaps we should wait for a user
> to come along who needs it.

Sure we can just wait.