Re: Error removing module: Device or resource busy

[Josh Stone <jistone@redhat.com>]

On 01/25/2012 04:29 PM, Chris Dunlop wrote:
>> Could you save/compare the same script compiled on one machine
>> vs. the other, so we can try to figure out what makes the two
>> gcc versions incompatible in this way?
> 
> See below. Nothing too interesting in the text differences to my gcc-illiterate eye.
[...]
> +-DCC_HAVE_ASM_GOTO

This is it.  With that defined, jump_label.h defines HAVE_JUMP_LABEL,
which enables this part of struct module in module.h:

#ifdef HAVE_JUMP_LABEL
	struct jump_entry *jump_entries;
	unsigned int num_jump_entries;
#endif

Compiling with/without that means everything after is shifted, including
the CONFIG_MODULE_UNLOAD stuff.  You can also compare this with the dump
given by dwarves pahole.

It would probably work to just turn off CONFIG_JUMP_LABEL altogether, so
it's not attempted on either compiler.  I'd be worried about other
hidden incompatibilities cropping up though.  I don't see any other
struct differences with pahole except jump_label stuff, but that doesn't
comfort me much.

So the easiest thing would be for you to compile the kernel with the
target system's own compiler, so everything matches. :)

If you keep the mismatch, we do have a couple of options to make it
easier to compile stap modules on one and run on the other.  One is the
--use-server mode, where stap running on the target will query a network
service running on the compile machine to build for it.  The other is
the --remote mode, where stap running on the compile machine will build
the module locally, then send it over ssh to run on the target and
collect the output.

Josh