7 ms·

Low-Level Software Security for Compiler Developers

128 points by struct 4y ago

JonChesterfield 4y ago
There's a tension here. Emitting 'secure' code almost always means emitting 'slower' code, and one of the few things compilers are assessed on is the performance of the code they generate. Compilers are built as a series of transformation passes. Normalisation is a big deal - if you can simplify N different patterns to the same thing, you only have to match that one canonical form later in the pipeline. So if one pass makes code slower/secure, later passes are apt to undo that transform and/or to miss other optimisations because the code no longer looks as expected. So while it is useful to know various make-it-secure transforms, which this book seems to cover, it's not at all obvious how to implement them without collateral damage. On a final note, compiler transforms are really easy to get wrong, so one should expect the implementation of these guards to be somewhat buggy, and those bugs themselves may introduce vulnerabilities.
- pjmlp 4y ago
  1960's systems were already taking security first approach and the industry would have kept down that route if it wasn't for UNIX and C adoption. IBM even did their RISC research in PL.8 taking into consideration safety and pluggable compiler infrastructure, similar to what people nowadays know from LLVM approach. Some would say that security measures in the car industry also slow drivers down and are a nuisance. https://en.m.wikipedia.org/wiki/Unsafe_at_Any_Speed https://en.m.wikipedia.org/wiki/Unsafe_at_Any_Speed
  - loup-vaillant 4y ago
    > Some would say that security measures in the car industry also slow drivers down and are a nuisance. Not sure about that: what are brakes for? Slowing down & stop, right? But then I ask, how fast would you drive if your car had no brakes? I would guess not very fast at all. Thus, one important role of breaks is to allow you to drive faster. In practice, the more safety measures you put, the more confident people grow and the faster they drive. To a point, of course. Same with programming: I prototype faster with a good static type system than I do with a dynamic one. One reason is that I just write fewer tests (including those one-off verifications in the REPL).
    - MobiusHorizons 4y ago
      I think you are thinking about the wrong type of security measures. I believe the op is talking about features like traction control, stability control, and ecu features that prevent engine power and braking at the same time. In performance driving situations (eg track driving) it is standard practice to disable these for the best track times. As safety features on the road they make a lot of sense, but can get in the way during high performance driving.
    - wizzard0 4y ago
      A fire alarm at home is important, a fire alarm in the chimney or the engine cylinder makes it unusable. Exploit mitigations do work, but a) compiler /does not/ know what are you building and what are your requirements b) they only protect from /specific, known/ threats the same way a generic fire alarm won't protect you from CO leak or an electric shock. c) but they waste time, energy and RAM whether they are relevant or not The only way to get systems that are secure, performant and easy to maintain is to invest in tools that make it easier for developers and users (!) to analyze what the system actually does. Not trying to make everything "magically secure". Pretending a microwave with Super Safety Cat Detector is a Magic Pasta Heater will only end up with lawsuits from owners of dead hamsters - and rightfully so, because it's trying to defraud and dumb down users instead of educating them.
staunton 4y ago
The vast majority of businesses choose speed over security and avoid investing in security since they can offload the cost of incidents to their users. One of the main reasons such "more secure tools" projects are interesting for users is that they provide an easy and cheap avenue towards claiming an effort towards security was made and avoiding liability. On one hand, such tools actually help make things secure, on the other hand, speed and ease of use (not security) being the top priorities, the effect is probably limited. People who care much more about security than average would not start a new project in C/C++ to begin with and where legacy code is involved, dealing with it is hard enough already without trying to "make it secure". The only way to really improve the level of security in the industry is to assign responsibility and damages to those who fail to implement it. So far, it seems all market participants are content with 90% of security concerns being addressed by security theater.
- pjmlp 4y ago
  This is thankfully changing. Returns in digital stores, increasing visibility of how it actually costs in real money to fix those issues, warranty clauses in consulting gigs (usually free of charge), and introduction of cyber security laws like in Germany [0]. [0] - https://www.bsi.bund.de/EN/Das-BSI/Auftrag/Gesetze-und-Verordungen/IT-SiG/2-0/it_sig_2-0.html https://www.bsi.bund.de/EN/Das-BSI/Auftrag/Gesetze-und-Veror...
- loup-vaillant 4y ago
  > The vast majority of businesses choose speed over security I would add that the vast majority of businesses also choose features over speed. In some cases they pay lip service to speed, for instance by choosing C++, but pay zero attention to actual speed, because they end up writing in a pointer fest RAII style that destroys memory locality and miss the cache all the time. Compared to that, even Electron doesn’t look too unreasonable.
  - rightbyte 4y ago
    If they write bad and slow Cpp code surely their Electron code will be even slower?
_a_a_a_ 4y ago
Compilers are an interest of mine so I'll read the article later, but I'm curious whether this is talking about C variety compilers which are generally unsafe, or compilers for managed languages which should never emit code allowing attacks (for some definition of 'never'). Which of these is this article/book discussing?
- eimrine 4y ago
  As far as I've understood the article is about how to fool a CPU.
  - _a_a_a_ 4y ago
    Hardly that, is about corrupting a state to take control. The CPU just does what it's told (not true, halfway down it starts on Covert channels and side-channels which are a different thing entirely). On the 1st half, I'm uncertain how it relates to compilers.
- tptacek 4y ago
  It's a book (in progress), not an article, and it is talking in large part about compilers for unmanaged languages.
- hummus_bae 4y ago
  From skimming a few pages, it seems to be C-family compilers in general.
vonimo 4y ago
That was an excellent read. I look forward to enjoying their section on JIT compiler vulnerabilities, a whole fascinating topic in itself, when it is completed.
tptacek 4y ago
This is pretty great; they've done the work to produce useful capsule summaries of a bunch of memory safety topics (like forward- and backward-edge CFI, JOP, and PAC). Looking forward to seeing how far they can take it. The assembler snippets are useful and could be fleshed out more.

128 points by struct 4y ago

There's a tension here. Emitting 'secure' code almost always means emitting 'slower' code, and one of the few things compilers are assessed on is the performance of the code they generate. Compilers are built as a series of transformation passes. Normalisation is a big deal - if you can simplify N different patterns to the same thing, you only have to match that one canonical form later in the pipeline. So if one pass makes code slower/secure, later passes are apt to undo that transform and/or to miss other optimisations because the code no longer looks as expected. So while it is useful to know various make-it-secure transforms, which this book seems to cover, it's not at all obvious how to implement them without collateral damage. On a final note, compiler transforms are really easy to get wrong, so one should expect the implementation of these guards to be somewhat buggy, and those bugs themselves may introduce vulnerabilities.

pjmlp 4y ago
1960's systems were already taking security first approach and the industry would have kept down that route if it wasn't for UNIX and C adoption. IBM even did their RISC research in PL.8 taking into consideration safety and pluggable compiler infrastructure, similar to what people nowadays know from LLVM approach. Some would say that security measures in the car industry also slow drivers down and are a nuisance. https://en.m.wikipedia.org/wiki/Unsafe_at_Any_Speed https://en.m.wikipedia.org/wiki/Unsafe_at_Any_Speed
- loup-vaillant 4y ago
  > Some would say that security measures in the car industry also slow drivers down and are a nuisance. Not sure about that: what are brakes for? Slowing down & stop, right? But then I ask, how fast would you drive if your car had no brakes? I would guess not very fast at all. Thus, one important role of breaks is to allow you to drive faster. In practice, the more safety measures you put, the more confident people grow and the faster they drive. To a point, of course. Same with programming: I prototype faster with a good static type system than I do with a dynamic one. One reason is that I just write fewer tests (including those one-off verifications in the REPL).
  - MobiusHorizons 4y ago
    I think you are thinking about the wrong type of security measures. I believe the op is talking about features like traction control, stability control, and ecu features that prevent engine power and braking at the same time. In performance driving situations (eg track driving) it is standard practice to disable these for the best track times. As safety features on the road they make a lot of sense, but can get in the way during high performance driving.
  - wizzard0 4y ago
    A fire alarm at home is important, a fire alarm in the chimney or the engine cylinder makes it unusable. Exploit mitigations do work, but a) compiler /does not/ know what are you building and what are your requirements b) they only protect from /specific, known/ threats the same way a generic fire alarm won't protect you from CO leak or an electric shock. c) but they waste time, energy and RAM whether they are relevant or not The only way to get systems that are secure, performant and easy to maintain is to invest in tools that make it easier for developers and users (!) to analyze what the system actually does. Not trying to make everything "magically secure". Pretending a microwave with Super Safety Cat Detector is a Magic Pasta Heater will only end up with lawsuits from owners of dead hamsters - and rightfully so, because it's trying to defraud and dumb down users instead of educating them.

staunton 4y ago

The vast majority of businesses choose speed over security and avoid investing in security since they can offload the cost of incidents to their users. One of the main reasons such "more secure tools" projects are interesting for users is that they provide an easy and cheap avenue towards claiming an effort towards security was made and avoiding liability. On one hand, such tools actually help make things secure, on the other hand, speed and ease of use (not security) being the top priorities, the effect is probably limited. People who care much more about security than average would not start a new project in C/C++ to begin with and where legacy code is involved, dealing with it is hard enough already without trying to "make it secure". The only way to really improve the level of security in the industry is to assign responsibility and damages to those who fail to implement it. So far, it seems all market participants are content with 90% of security concerns being addressed by security theater.

pjmlp 4y ago
This is thankfully changing. Returns in digital stores, increasing visibility of how it actually costs in real money to fix those issues, warranty clauses in consulting gigs (usually free of charge), and introduction of cyber security laws like in Germany [0]. [0] - https://www.bsi.bund.de/EN/Das-BSI/Auftrag/Gesetze-und-Verordungen/IT-SiG/2-0/it_sig_2-0.html https://www.bsi.bund.de/EN/Das-BSI/Auftrag/Gesetze-und-Veror...
loup-vaillant 4y ago
> The vast majority of businesses choose speed over security I would add that the vast majority of businesses also choose features over speed. In some cases they pay lip service to speed, for instance by choosing C++, but pay zero attention to actual speed, because they end up writing in a pointer fest RAII style that destroys memory locality and miss the cache all the time. Compared to that, even Electron doesn’t look too unreasonable.
- rightbyte 4y ago
  If they write bad and slow Cpp code surely their Electron code will be even slower?

_a_a_a_ 4y ago

Compilers are an interest of mine so I'll read the article later, but I'm curious whether this is talking about C variety compilers which are generally unsafe, or compilers for managed languages which should never emit code allowing attacks (for some definition of 'never'). Which of these is this article/book discussing?

eimrine 4y ago
As far as I've understood the article is about how to fool a CPU.
- _a_a_a_ 4y ago
  Hardly that, is about corrupting a state to take control. The CPU just does what it's told (not true, halfway down it starts on Covert channels and side-channels which are a different thing entirely). On the 1st half, I'm uncertain how it relates to compilers.
tptacek 4y ago
It's a book (in progress), not an article, and it is talking in large part about compilers for unmanaged languages.
hummus_bae 4y ago
From skimming a few pages, it seems to be C-family compilers in general.

vonimo 4y ago

That was an excellent read. I look forward to enjoying their section on JIT compiler vulnerabilities, a whole fascinating topic in itself, when it is completed.

tptacek 4y ago

This is pretty great; they've done the work to produce useful capsule summaries of a bunch of memory safety topics (like forward- and backward-edge CFI, JOP, and PAC). Looking forward to seeing how far they can take it. The assembler snippets are useful and could be fleshed out more.