[ltt-dev] Crash recovery of traces and final TSC

[Mathieu Desnoyers]

* Michel Dagenais (michel.dagenais at polymtl.ca) wrote:
> 
> > > When the system crashes, the end timestamp is not written in the
> > > subbuffer header. Instead, the value 0 is written. This tends to make
> > > LTTV crash.
> 
> Obviously LTTV should not "crash", whatever the input :-).
> 
> > > - "repair" the trace a posteriori, writing in the header the timestamp
> > > of the last event found in the subbuffer?
> > 
> > Yep, seems like an appropriate solution.
> > 
> > > - change lttv not to not need the "end timestamp" header field?
> > 
> > Hrm, it needs it for quickly seeking to specific sub-buffers.
> 
> People may expect the trace to be readonly, so relying on a posteriori
> fixing is kind of unexpected. You could have a special case for the last
> subbuffer, then not looking at the end time.
> 

When we will create an index for variable-length sub-buffers, the fixup
could be computed and added to this index, if needed.

I agree that the trace should stay read-only.

Mathieu


-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68