Some security people have scoffed at me when I say that security
problems are primarily "just bugs".
Those security people are f*cking morons.
Because honestly, the kind of security person who doesn't accept that
security problems are primarily just bugs, I don't want to work with.
I thought I'd explain why Linus is right.
Linus has an unwritten manifesto of how the Linux kernel should be maintained. It's not written down in one place, instead we are supposed to reverse engineer it from his scathing emails, where he calls people morons for not understanding it. This is one such scathing email. The rules he's expressing here are:
- Large changes to the kernel should happen in small iterative steps, each one thoroughly debugged.
- Minor security concerns aren't major emergencies; they don't allow bypassing the rules more than any other bug/feature.
Last year, some security "hardening" code was added to the kernel to prevent a class of buffer-overflow/out-of-bounds issues. This code didn't address any particular 0day vulnerability, but was designed to prevent a class of future potential exploits from being exploited. This is reasonable.
This code had bugs, but that's no sin. All code has bugs.
The sin, from Linus's point of view, is that when an overflow/out-of-bounds access was detected, the code would kill the user-mode process or kernel. Linus thinks it should have only generated warnings, and let the offending code continue to run.
Of course, that would in theory make the change of little benefit, because it would no longer prevent 0days from being exploited.
But warnings would only be temporary, the first step. There's likely to be be bugs in the large code change, and it would probably uncover bugs in other code. While bounds-checking is a security issue, its first implementation will always find existing code having latent bounds bugs. Or, it'll have "false-positives" triggering on things that aren't actually the flaws its looking for. Killing things made these bugs worse, causing catastrophic failures in the latest kernel that didn't exist before. Warnings, however, would have equally highlighted the bugs, but without causing catastrophic failures. My car runs multiple copies of Linux -- such catastrophic failures would risk my life.
Only after a year, when the bugs have been fixed, would the default behavior of the code be changed to kill buggy code, thus preventing exploitation.
In other words, large changes to the kernel should happen in small, manageable steps. This hardening hasn't existed for 25 years of the Linux kernel, so there's no emergency requiring it be added immediately rather than conservatively, no reason to bypass Linus's development processes. There's no reason it couldn't have been warnings for a year while working out problems, followed by killing buggy code later.
Linus was correct here. No vuln has appeared in the last year that this code would've stopped, so the fact that it killed processes/kernels rather than generated warnings was unnecessary. Conversely, because it killed things, bugs in the kernel code were costly, and required emergency patches.
Despite his unreasonable tone, Linus is a hugely reasonable person. He's not trying to stop changes to the kernel. He's not trying to stop security improvements. He's not even trying to stop processes from getting killed That's not why people are moronic. Instead, they are moronic for not understanding that large changes need to made conservatively, and security issues are no more important than any other feature/bug.
Update: Also, since most security people aren't developers, they are also a bit clueless how things actually work. Bounds-checking, which they define as purely a security feature to stop buffer-overflows is actually overwhelmingly a debugging feature. When you turn on bounds-checking for the first time, it'll trigger on a lot of latent bugs in the code -- things that never caused a problem in the past (like reading past ends of buffers) but cause trouble now. Developers know this, security "experts" tend not to. These kernel changes were made by security people who failed to understand this, who failed to realize that their changes would uncover lots of bugs in existing code, and that killing buggy code was hugely inappropriate.
Update: Another flaw developers are intimately familiar with is how "hardening" code can cause false-positives, triggering on non-buggy code. A good example is where the BIND9 code crashed on an improper assert(). This hardening code designed to prevent exploitation made things worse by triggering on valid input/code.
Update: No, it's probably not okay to call people "morons" as Linus does. They may be wrong, but they usually are reasonable people. On the other hand, security people tend to be sanctimonious bastards with rigid thinking, so after he has dealt with that minority, I can see why Linus treats all security people that way.