Borislav Petkov wrote:
> On Fri, May 10, 2024 at 03:12:36PM -0700, Dan Williams wrote:
> > I had asked Fabio to take a look at whether it made sense to continue
> > with the concept of ras_userspace_consumers() especially since it seems
> > limited to the EXTLOG case.
> > 
> > In general I am finding that between OS Native and Firmware First error
> > reporting the logging approaches are inconsistent.
> > 
> > As far I can see rasdaemon would not even notice is the "daemon_active"
> > debugfs file went away [1],
> 
> It tells the kernel that it is consuming the error info from the
> tracepoints.

Yes, my point though was that if it got deleted I doubt anyone would
notice. rasdaemon explicitly does not check the return from
open("daemon_active").

I am also curious about the history here. This "daemon_active" scheme is
an awkward way to detect that something is consuming the tracepoint. It
was added on v4.0, but Steven had added "tracepoint_enabled()" back in
v3.17:

7c65bbc7dcfa tracing: Add trace_<tracepoint>_enabled() function

So even if non-rasdaemon userspace was watching the extlog tracepoints
they would not fire because ras_userspace_consumers() prevents it.

I am finding it difficult to see why ras_userspace_consumers() needs to
continue to be maintained.

> > and it should be the case that the tracepoints always fire whether
> > daemon_active is open or not.
> >
> > So I was expecting this removal to be a conversation starter on the
> > wider topic of error reporting consistency.
> 
> Yeah, and then they'll come and say: ew, we're getting error duplicates
> - once logged in dmesg and once through the tracepoints.

That would be odd since there is no ras_userspace_consumers() in the
ACPI GHES path, so it is already the case that you can get duplicate
error information depending on which path triggers the error.

Tracepoints are individually configurable. 

> So just like with the other thread, we have to figure out what our
> scheme will be wrt hw error logging, agree on it and then make it
> consistent.

From my perspective I want alignement between "firmware first" and "OS
Native" events and I think any movement away from kernel log messages as
a hardware error mechanism towards tracepoints is a good thing.

Recall that tracepoints can also be configured to emit to the kernel
log, so that might be a way to keep legacy kernel log message parsing
environments happy.

> Do we want to have both? Should it be configurable? Probably...

Would be great to hear from folks that have a reasons for kernel log
message error reporting to continue.

> Anything else...?

Uniformity of error response to "fatal" events, but that is mainly a
PCIe error handling concern not  CPU errors.