The following problem was first pointed out (in private mail) by Michael
Howard from Microsoft. His writeup is now available at
http://msdn.microsoft.com/library/en-us/dncode/html/secure10102002.asp. From
a representative check of a few widely-used open source crypto programs, this
affects quite a bit of software.

The problem he points out is that clearing sensitive information such as
encryption keys from memory may not work as expected because an optimising
compiler removes the memset() if it decides it's redundant. Consider for
example the following:

int encrypt( const void *key )
{
puts( key ); /* Normally we'd encrypt here */
}

void main( void ) /* Because we can */
{
char key[ 16 ];

strcpy( key, "secretkey" );
encrypt( key );
memset( key, 0, 16 );
}

When compiled with any level of optimisation using gcc, the key clearing call
goes away because of dead code elimination (see the MSDN article for more
details on this, which uses VC++ to get the same effect). While you can
kludge enough stuff around a custom memory-clear call to fool the optimiser
(hacks with 'volatile', touching the memory after it's cleared and hoping the
optimiser is fooled, etc etc) there's no guarantee that it'll work for
anything but the compiler(s) you happen to test it with - any future
enhancement to the optimiser may turn it back into a nop. What it really
needs is the addition of a #pragma dont_remove_this_code_you_bastard in the
compiler. Until then, a lot of security code will be affected by this
problem.

Peter.

[ reply ]