[Rd] Stack checking, core dumps, and embedding R

[Thomas Friedrichsmeier]

Dear R developers,

it seems the stack checking issue with embedded applications is not fully 
resolved, yet. The problem arises, when multiple threads are involved. I.e. 
the case, where R is run in a separate thread of the main application. In 
this case, the call to (unix/system.c)

    R_CStackStart = (uintptr_t) __libc_stack_end;

will set the stack start to the start of the main threads stack, while the 
stack of the thread in which R is running may be entirely different. This 
leads to a bad value of R_CStackStart, and subsequently R_CheckStack () is 
likely to fail for no good reason.
I believe, a more robust solution would be to use

	pthread_attr_t stack_addr;

	pthread_attr_getstackaddr (pthread_self (), &stack_addr);
	R_CStackStart = (uintptr_t) stack_addr;

probably with some #ifdefs around it. Of course that's somewhat ugly as well, 
having to pull in the pthread-library for this purpose.
An easier solution might be to explicitely set

	if (R_running_as_main_program) {
		[...]
	} else {
		R_CStackStart = -1;
	}

and thereby disable stack checking for all embedded cases.
Yet another solution, which I do not understand, why it is not being used 
would be to always use the 
	R_CStackStart = (uintptr_t) &i + (6000 * R_CStackDir);
logic, not just when running as the main program, maybe with a slightly larger 
margin for safety.
Finally, if all these are not desirable, I propose a heuristic check somewhat 
like
	if (abs (R_CStackStart - &i) > 20000) {
		/* Very unlikely we're this far away from the stack start at this
			point in the code. Disable checking */
		R_CStackStart = -1;
	}

Regards
Thomas
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : https://stat.ethz.ch/pipermail/r-devel/attachments/20060417/c54a4a5d/attachment.bin