Hi.
At the point of a DPDK app crash, or some other abnormal event, you may
want to save the trace buffer, especially if a complete core file is too
large to fit on the disk.
This would be straight-forward to implement on top of DPDK trace.
However, you may also want to save data which is referred to
by-reference in the traces, like for example mbuf pointers. The reason
for not have the data by-value may be that it would be too expensive for
always-on type trace points.
How could one go about solving the latter case with DPDK trace?
Decoding CTF and de-referencing the appropriate data (e.g., in the
signal handler) seems technically possible, but very complicated without
"framework" (i.e., DPDK trace) support. Then you would also need to
regenerate CTF (including new meta data), but now with the by-reference
fields as by-value fields.
Ideally, as far as I see, one would like the app to mark a certain field
as a pointer to X bytes of data, which could potentially be dereferenced
at a point of a crash.
Obviously, mbufs can be recycled and other buffers may change state
compared to what they were during at the time of the trace entry being
created. Nevertheless, such data, although inaccurate/best effort, may
be worth a lot in a postmortem debugging scenario.
You could have something on the side of DPDK trace, which maintained the
buffers and pointers and could be dereferenced/copied around at the
point of the "abnormal event", but I'm curious if there is a solution
where you rely fully on an extended version of DPDK tracing.
/M
