On Fri, Jun 15, 2012 at 3:19 PM, Luck, Tony tony.luck@intel.com wrote:
milli-seconds for recording? This would cripple the kernel. On slow machines, incorporating lockdep into function tracing (and other debug options) causes the system to live lock. Tracing the timer interrupt took so long that by the time it finished, the next timer triggered again.
Ah - I think I misunderstood how this works ... I thought the function trace was taken at some point in the crash/shutdown path. Now I see that it is being generated repeatedly as the kernel runs ... with the idea that if a hang happens, you can see the last call trace that was successful.
Ok then - clearly this is only useful with a ram backend for pstore.
Binary makes sense too - but I'll stick with my comment that a different kernel must be able to do the decode. Or we need to spell out clearly that you must have the same, or compatible kernel booted to make any sense of the saved ftrace record.
The decode also converts PCs to symbols, so for all the data to be useful you have to boot the exact same kernel. You could theoretically boot a different kernel with a compatible format, but you would have to manually extract the PCs from the dump and addr2line them against the correct vmlinux, which is probably more trouble than it's worth. Also, the binary data is stored inside a pstore_ram ringbuffer that may change formats between kernels. I would suggest documenting that this is only supposed to work booting back to the same kernel.