codeblog code is freedom — patching my itch

August 20, 2018

security things in Linux v4.18

Filed under: Chrome OS,Debian,Kernel,Security,Ubuntu,Ubuntu-Server — kees @ 11:29 am

Previously: v4.17.

Linux kernel v4.18 was released last week. Here are details on some of the security things I found interesting:

allocation overflow detection helpers

One of the many ways C can be dangerous to use is that it lacks strong primitives to deal with arithmetic overflow. A developer can’t just wrap a series of calculations in a try/catch block to trap any calculations that might overflow (or underflow). Instead, C will happily wrap values back around, causing all kinds of flaws. Some time ago GCC added a set of single-operation helpers that will efficiently detect overflow, so Rasmus Villemoes suggested implementing these (with fallbacks) in the kernel. While it still requires explicit use by developers, it’s much more fool-proof than doing open-coded type-sensitive bounds checking before every calculation. As a first-use of these routines, Matthew Wilcox created wrappers for common size calculations, mainly for use during memory allocations.

removing open-coded multiplication from memory allocation arguments

A common flaw in the kernel is integer overflow during memory allocation size calculations. As mentioned above, C doesn’t provide much in the way of protection, so it’s on the developer to get it right. In an effort to reduce the frequency of these bugs, and inspired by a couple flaws found by Silvio Cesare, I did a first-pass sweep of the kernel to move from open-coded multiplications during memory allocations into either their 2-factor API counterparts (e.g. kmalloc(a * b, GFP...) -> kmalloc_array(a, b, GFP...)), or to use the new overflow-checking helpers (e.g. vmalloc(a * b) -> vmalloc(array_size(a, b))). There’s still lots more work to be done here, since frequently an allocation size will be calculated earlier in a variable rather than in the allocation arguments, and overflows happen in way more places than just memory allocation. Better yet would be to have exceptions raised on overflows where no wrap-around was expected (e.g. Emese Revfy’s size_overflow GCC plugin).

Variable Length Array removals, part 2

As discussed previously, VLAs continue to get removed from the kernel. For v4.18, we continued to get help from a bunch of lovely folks: Andreas Christoforou, Antoine Tenart, Chris Wilson, Gustavo A. R. Silva, Kyle Spiers, Laura Abbott, Salvatore Mesoraca, Stephan Wahren, Thomas Gleixner, Tobin C. Harding, and Tycho Andersen. Almost all the rest of the VLA removals have been queued for v4.19, but it looks like the very last of them (deep in the crypto subsystem) won’t land until v4.20. I’m so looking forward to being able to add -Wvla globally to the kernel build so we can be free from the classes of flaws that VLAs enable, like stack exhaustion and stack guard page jumping. Eliminating VLAs also simplifies the porting work of the stackleak GCC plugin from grsecurity, since it no longer has to hook and check VLA creation.

Kconfig compiler detection

While not strictly a security thing, Masahiro Yamada made giant improvements to the kernel’s Kconfig subsystem so that kernel build configuration now knows what compiler you’re using (among other things) so that configuration is no longer separate from the compiler features. For example, in the past, one could select CONFIG_CC_STACKPROTECTOR_STRONG even if the compiler didn’t support it, and later the build would fail. Or in other cases, configurations would silently down-grade to what was available, potentially leading to confusing kernel images where the compiler would change the meaning of a configuration. Going forward now, configurations that aren’t available to the compiler will simply be unselectable in Kconfig. This makes configuration much more consistent, though in some cases, it makes it harder to discover why some configuration is missing (e.g. CONFIG_GCC_PLUGINS no longer gives you a hint about needing to install the plugin development packages).

That’s it for now! Please let me know if you think I missed anything. Stay tuned for v4.19; the merge window is open. :)

© 2018, Kees Cook. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 License.
CC BY-SA 4.0

2 Comments

  1. Variable Length Array removals, part 2 -What would be impact of KPI for this Kernel hardening feature i.e kmalloc(m*n,GFP_XXX )assign memory of m items of n size or directly calling kmalloc_array(m, bn, GFP…)) as API.

    Comment by Raju Tiwari — August 28, 2018 @ 4:04 am

  2. I’m not sure I know what you mean. Do you mean Kernel Page Table Isolation? That would be entirely unrelated. Normal (non-buggy) callers of kmalloc() and kmalloc_array() will get the same memory allocated — it’s just that multiplicative overflows will be caught in the latter case.

    Comment by kees — October 18, 2018 @ 1:35 pm

Powered by WordPress