On Wednesday, Aug 13, 2003, at 03:28 US/Pacific, Eygene A. Ryabinkin wrote:
Hi!
I have an idea on buffer overflow prevention. I doubt that it's new, but I
haven't seen an implementation of it in any freely distributable Un*x system.
So, I hardly need your comments on it.
Preliminary: I'm talking about Intel x86 architecture, but maybe it will be
applicable to others as well.
The idea itself: all (correct me if I'm wrong) buffer overflows are based on
the fact that we're using the stack, referenced by SS:ESP pair, both for
procedure return address and for local variables. It seems to me, that would we
have two stacks -- one for real stack and one for variables -- it will solve
a bunch of problems. So, my suggestion: let us organise two segments: one for
normal stack, growing downwards, referenced by SS:ESP pair and the second one,
for local variables, referenced by GS:EBP pair, with either upwards or
downwards growing. Now, if we use first segment for passing variables and
procedure return addresses (normal stack usage), and second segment only for
local procedure variables, we will have the following advantages:
1) Local variables and return address will be physically (by means of CPU)
divided and it will not be possible to touch the return address by
overflowing local buffer.
2) The procedure introduces only one extra register -- GS, since EBP is
very often used for the stack frame.
Of course, this two segments can be made non-executable, just in case.
It's definitely good to be thinking of novel approaches to securing code. Machines in the old days used to have memory partitioned in similar fashions. But realize that overwriting return addresses via stack trashing is only one way a program can be compromised.
C++ programs, for instance, have vtables, 'this' pointers, and all sorts of other fun stuff that can be tinkered with - not to mention the heap.
Approaches that involve partitioning executable code and data are probably the easiest route. It becomes a major challenge to do something interesting with return address manipulation when the memory map is setup with execute-only in one range (for the code) and non-executable (for stack & heap) in another. If you can't inject a return address to the code of your choosing, than what's the point?
Dealing with dynamic code gets tricky (dynamic libs, Java, CLR, and interpreters) in this environment. But it's quite do-able.
What we need to implement the idea: first, rewrite kernel to organise two
segments for every process and to place proper values into the segment
registers upon the program startup. Second, rewrite the compiler to support
the new scheme of local variables addresation. So, the changes are minimal,
in some sence.
As I said, I hardly need your criticism, suggestions, etc. of any type.
rea
What fun is that? ;^)
Craig
-- Craig Pratt Strongbox Network Services Inc. mailto:craig .AT. strong-box.net
-- This message checked for dangerous content by MailScanner on StrongBox.
