Luck, Tony wrote: > > Given that is has not kept pace the next question is whether the kernel > > should bother to maintain the contract => "if nothing is watching > > tracepoints some subset (all?) hardware error messages will be reflected > > to the kernel log". > > I'm with you for corrected & recoverable errors. The console is a terrible > place for those logs. > > But it's handy to have fatal errors go to the console. They may make it out to a > serial port, or be stashed in pstore for later retrieval. Trace event logs that > are handled by some user level daemon are just going to disappear in a > puff of reset smoke. That is true, and arguably that fatal event logging to the kernel log should be independent of whether userspace is watching. I.e. ras_userpsace_consumers() gating kernel logging is actively getting in the way when the error severity implies userspace may not get a chance to see or process the event. I would need help from Steven on whether ftrace would be amenable to have a localized tracepoint_printk context to trigger emission of fatal events to the kernel log in addition to the trace buffer.
