//GCC: -O0
//MSVC: /Od
#include <memory.h>
#include <stdio.h>
struct S { unsigned char x; int y; };
int f(int f) { S a; memset(&a, f, sizeof(a)); unsigned char y[2]; memcpy(&y, &a, sizeof(y)); return y[1]; }
int g() { S a = { 1, 2 }; unsigned char y[2]; memcpy(&y, &a, sizeof(y)); return y[1]; }
int main(int argc, char *argv[]) { f(0xDD); printf("%#x\n", g()); f(0xFF); printf("%#x\n", g()); }
But yeah, I should have mentioned the caveats too. I will update the post once I'm back at my computer.
So if you're not working on an OS kernel somewhere in the syscall path - really not a common activity even among C programmers - it's not a problem. If you are passing a struct from one part of your program to another part, or to a library you use which is not isolated or sandboxed in some way, it's not a problem. If you pass a struct to a _more_ privileged context like a syscall, it's not a problem. If you are passing structs over a network connection or writing to a file, you'll pack them and be aware of every byte anyway, in order to be compatible with another implementation reading them.
If you _are_ passing a struct from a more privileged context to a less privileged context, then yeah you have to memset().
Simply not true; that context could pass a copy of that structure to a less privileged context other than the original one, without doing anything about the padding.
(Well, you could argue that it's the privileged kernel's fault: it should meticulously floss the structure between the members to clear the padding to zero while preserving the values.)
Someone upthread mentioned "sending over a network": that's an example. Sending over a network (or local socket/pipe) begins with pushing it down to a more privileged context, which then assiduously places all the bytes into a buffer that is blasted out on the wire. In this case, the more privileged context has no idea what the structure even is; it's just a blob of bytes.
_6 When a value is stored in an object of structure or union type, including in a member object, the bytes of the object representation that correspond to any padding bytes take unspecified values. 51)_
So sure, a memcpy after a memset aught to preserve all the bytes, but if they are updated through a struct which contains padding then the values in the padding bytes will be left unspecified.
Even if memset in a vacuum is guaranteed, look at the quote by anyfoo. "When a value is stored in an object of structure or union type, including in a member object, the bytes of the object representation that correspond to any padding bytes take unspecified values."
If the compiler knows you're going to write to the members right after the memset, it can put the padding back to its previous values. And by that I mean as far as your code can tell it put it back, but in actuality it never zeroed the padding to begin with.
And since there are performance benefits to pretending, now you have a situation where a "helpful" compiler and a malicious compiler have the same effect: data can get leaked.
Also, type-punning is a thing. memset is byte-oriented/memory-oriented. Just because you're using it to zero a struct of a particular kind doesn't mean that's the only way the memory will accessed. Just because reading the padding of some struct is undefined behaviour doesn't mean accessing those bits by some other means is also undefined.
A compiler usually can't eliminate a call to memset() in most practical cases of initialization (where memory leakage is also a concern) because they almost always pass a reference to a routine in another translation unit. Something like memzero_explicit() can be used anyway -- but you're massively overstating the relevance of compilers eliminating dead stores done via memset(). It's much more of an issue for post-destruction memory sanitizing (which is the primary use case for memzero_explicit) than it is for compiler f*ckery when memset() is used for initialization.
You would have to call such a function between the memset and the first time you write to a member. Otherwise the compiler is allowed to say "I put the padding back, and you can't prove otherwise".
> Just because reading the padding of some struct is undefined behaviour doesn't mean accessing those bits by some other means is also undefined.
It's not always undefined, but it says very clearly that the value of padding becomes unspecified.
> they almost always pass a reference to a routine in another translation unit
> you're massively overstating the relevance of compilers eliminating dead stores done via memset
Unless inlining happened, or link-time optimization, or, or...
If the compiler zeroes the memory most of the time, that makes it even scarier. Because all your tests come back clean and safe, then four years later a macro changes and suddenly you're leaking data all over the place.
I don't think I'm overstating the relevance at all. Any security feature that could disappear because of a reasonable, trying-to-help optimization is one that should have a bright red warning label. And this is such a feature. It doesn't require a "sufficiently smart" compiler, and it doesn't require a malicious compiler. This is the kind of thing that can break by accident and ruin everyone's month.
Not in practice. Compilers make use of undefined behaviour to optimize things that are widely applicable and profitable. No real compiler does what you're saying and no future compiler is likely to without explicitly being asked to.
I agree that, by the letter of the spec, you're right, but you're still most certainly overstating the relevance.
I'm not arguing that this isn't a real problem or that people shouldn't use memzero_explicit() (or similar) where security is on the line, as I already said several times in another sub-thread -- I'm just saying that this kind of thing is extreme language-lawyering beyond the realms of probability. It's still not an excuse to be lax, but let's be realistic about the actual likelihood of it happening.
It's not just any, typical kind of dead store elimination. It also would require other optimization passes that I can assure you no mainstream compiler actually does. You can disagree with me if you want, but you're simply wrong.
That step is 90% of the work. Once you can do that, it's straightforward to assess that every field is assigned after a memset, with no intervening reads, and then remove the memset.
I spent several years working on a production grade compiler and I can assure you it's not. But keep just making things up off the top of your head if it makes you feel smart.
Let's look at a basic common case, as a checklist.
1. All fields are definitely assigned.
2. No functions are called except intrisics.
3. Nothing reads from the object on any control path.
4. Memset happens between creation and first assignment.
We agreed that step 1 is a solved problem, right? Are any of 2-4 difficult? Did I miss any prerequisites for the optimization?
Once you can prove 1-4, isn't the optimization pass as simple as looking for memsets applied to structs, checking 1-4, then deleting the call?
All "undefined behaviour" means is that the spec can no longer guarantee anything about the execution of a program once a constraint is violated. It doesn't say anything about the practicality of actually doing it on any given implementation.
FWIW I didn't downvote you, but your reasoning about undefined behaviour seems overly simplistic. It should never be used to explain away security concerns.
You can memset to 0, and on your compiler it might be secure, but it's not enough to keep your security guaranteed against future compilers.
There's a reason functions like SecureZeroMemory exist. In a similar situation, you try to prevent leaking secure secrets by zeroing memory before releasing it. But the compiler sees you never use the variable again, and optimizes away the zeroing.
The message to take away is not "be less paranoid". It's "be more paranoid".
But the problem is that the security concerns and fixes are all undefined here.
So the initial comment is still very right. If you're relying on the padding to be anything in particular, you're in trouble.
No one in this sub-thread has mentioned (or implied) any "fix". You appear to be putting words in my mouth.
> So the initial comment is still very right.
Of course it's right, but it's also misleading when read in the context of the parent comment.
> If you're relying on the padding to be anything in particular, you're in trouble.
No shit.
I'm talking about the comments hermitdev was replying to that were treating memset as a 'fix'.
And 'fix' is shorthand for the opposite of "open[ing] you up to all kinds of info leaks". I don't think that's putting words in anyone's mouth.
To be fair though, memset() usually IS a fix. As mentioned by the kernel memzero_explicit() docs:
> usually using memset is just fine (!)
-- https://www.kernel.org/doc/htmldocs/kernel-api/API-memzero-e...
A conforming C compiler can't just remove memset() as it pleases. The case that most often requires memzero_explicit() is when zeroing an object after destruction, because the compiler thinks it can statically determine that it's a dead store. It very rarely happens that a compiler elides a memset() used for initialization.
I'm not sure why you seem to think that memset() can just be dropped at will for no reason whatsoever or that it's somehow always undefined behaviour.
For the sake of argument, can you show me some example code where it would be conforming to remove a memset() call? Preferably a realistic example and not a Google'd copypasta. Because it's all too easy to just regurgitate things you heard and think you understood, but no so easy to demonstrate it yourself.
Dead Store Elimination (Still) Considered Harmful
The compiler could also determine that &a.x)[1] is undefined behavior in C (reading past the end of an object), and take advantage of that.
The way to check that is by disassembling the code, and even that doesn’t say what a C compiler should do, only what these particular compilers do.
Code doing that may break with a compiler change or a compilation flag change on one end of the connection, even if you can guarantee both ends use the same CPU.
Not sending the structure over the network is insufficient, and rigorous memset is a clear solution to this problem.
Also, C structure layout for a specific ABI is strictly determined; if that weren’t the case you couldn’t compile two C compilation units with different flags/compilers and expect them to work.
And there exists the very few really secure memset_s variants with a real memory barrier which guarantees memset to be executed in load and store order, flushing the cache, so that Spectre attacks are mitigated. security-relevant crypto libraries, like openssl and variants don't care about that, because memory barriers are "too slow". only the Linux kernel does it properly. It's still a mess.
That's true, but when sending stuff over the network you don't know if the receiver uses the same ABI.
You can't be sure you can send such structures to different OSes or CPUs however.
You cannot do this unless the compilers have the same ABI, which is not part of the standard.
What if you do know the exact layout and it happens to have a gap between fields?
So...you're not wrong, but it's basically the same reason people use C over something with more ergonomics.
Edit: the last time I wrote a networking library, I also just cast to raw structures, mainly because it was less work (didn't want to integrate a serialization library on both sides of the link). This was fine for a while (and very fast), but also hid some bugs in the protocol, and made sending arrays of data very ugly. Eventually we put protobuf on top, and things ended up much nicer.
(Though I should mention that macros actually do make this doable, but they're hardly pretty, and I have yet to see many people embrace this approach, if they realize it's possible at all.)
Yes? Of course you do that? Or pick one of a hundred libraries that do that for you - protobuf, cap'n proto, flatbuffers, etc...
Protobuf serialization functions can turn into a lot of generated code too (read: big binaries) and stomp all over your icache, because they're different for each type. More "data-driven" approaches using type-generic functions with some runtime type-specific data (like proto descriptors&reflection) can be a good tradeoff for code that isn't as "hot". Typically more branchy than a memcpy or a type-specific function though.
And by sticking with the serialization libraries with higher performance, like flatbuffers and capnproto, you also lose a little bit of future flexibility, since its typically harder to change the format of these structures compared to higher level libraries.
> None of them are going to compare to something like sending a raw struct on the wire, if you need absolute performance.
I don't know about that.
For simple flat structs containing integers only (no pointers), Cap'n Proto is nearly identical to raw structs.
For complex structures with pointers, sending raw structs doesn't work.
Admittedly you may suffer unwanted code bloat linking in the Cap'n Proto library if all you really want to do is send a flat, raw struct.
> And by sticking with the serialization libraries with higher performance, like flatbuffers and capnproto, you also lose a little bit of future flexibility, since its typically harder to change the format of these structures compared to higher level libraries.
I disagree with this. Cap'n Proto's compatibility story is essentially the same as Protobuf and JSON: You can add new fields and old programs will ignore them.
(I suspect Flatbuffers is also similar but I haven't looked at it in a long time.)
If you're doing a large project on supported platforms it's probably worth using a pre-existing serialisation library like protobuffers or something.
Not sure what's the appropriate mitigation. Having to memset is a drag, and it's easy to forget to do it in some place or another. I'm tempted to say that you should only use packed structs for shared memory communication across trust boundaries.
Edit: Also, lots of simulation codes I am aware of tend to send flat arrays anyway. So I'd say that your concerns are mostly theoretical in nature.
C is not a very high-level language. Apart from where undefined behavior is called out, you can and should assume how data is represented on a low level. It was made for writing operating systems and their kernels after all, where interfacing in a bit-accurate way is common.
There are platforms on which NULL, integer 0 and floating point 0 is not the same. Hence memsetting won't do the right thing. On modern platforms they are, but still, you don't want to get the Standard C Weenies on your back - they're even more annoying than the Rust Evangelism Strike Force. You don't want demons coming out of your nose, do you? :)
They do have a point, though. Suppose you are handling a struct you thought were memsetted but was not (your stupid team mate used those damn initializers!), then when serializing it you will get undefined bytes in the stream, possibly triggering hash values to mismatch causing mysterious rebuilds and other stuff.
Wrt to field accesses, yes those can overwrite padding bytes. Suppose a is an 8 bit char field, but padded to 4 bytes. Then "foo->a = 22" could be translated to "MOV [RSP+16], EAX" which overwrites the 3 padding bytes with undefined values. The solution, as you mention, is to use packed structs but in those field accesses are much less efficient.
Bottom line: code whichever way you want. :) Me, I like to live in the fast lane from time to time and those memset/cpy/cmp functions are sooo handy. :)
If you said memset(&p, 0, sizeof(p)) then that's not guaranteed to work on all implementations.
> [...] The technicality of the memset example is that it does not set the bits of the pointer by referring to it as a pointer, so the requirement that 0 behave as if it was a null pointer does not apply.
In short, there may exist counter-intuitive but nonetheless standards-compliant platforms where tests such as "assert((uintptr_t)(void * )0 == 0);" or "int x = 0; assert((void * )x == (void * )0);" would fail.
struct S {char data, char pad0, char pad1, char pad2};
int main() {S dataStruct; printf("%s", dataStruct.data); return 0;}
Well, I chose this example specifically to ask whether the resulting memory structure, completely independent of calling a sizeof(), would still be as if I were to call a sizeof() (which then I believe would be 4 bytes large) if not accessing the total size of the struct or any of the padding members.
It makes a difference in the memory footprint, which can become important, if you're programming a device with just 64-bytes total RAM.
This isn't true of modern processors.
struct st s;
bzero(&s, sizeof(s));
// ...
write(fd, &s, sizeof(s));
Having said that, it is possible to get this sort of thing right with C++ and templates (thanks to inlining).
Higher-level languages without compile time code generation (like Java and Python) usually make serialization a train wreck in comparison with “send raw structs” or “make the [C++] compiler generate and typecheck the machine code that does the serialization”
You can get away with directly writing a struct for simple structs, but eventually it will bite you.
The remaining issue is endianness, but you can configure ARM to match x86 at boot so shrug.
Also, if I tell any reasonably experienced developer my protocol sends a packed struct with these fields:
uint8_t version
uint16_t reserved
uint8_t op
int32_t status
uint32_t data_length
// void data[]
I agree that packed structs can be tricky, but they’re much better than the vast majority of serialization libraries I’ve dealt with (including all of the open source ones, and I’ve used all the usual suspects).
ARM in BE mode is pretty rare / nonexistent. Small MIPS chips still float around in mostly low-end networking gear, though, and those do tend to run in BE.
The big annoyance is network people all use big endian probably just to spite everyone else.
The people that wrote that spec knew that and went with big endian anyways.
Also I work with people that do network hardware. They can't care less about little or big endian. Makes no difference to them.
Whether you should do this in 2019 is another question, but sometimes it's still the easiest way to implement a binary protocol.
It's easy to write binary protocol code that way but it isn't portable.
EDIT: Actually I have to apologize, because I’m not certain what the C standard says about whether the padding bytes can change when changing members of the structure (6.2.6.1):
6 When a value is stored in an object of structure or union type, including in a member object, the bytes of the object representation that correspond to any padding bytes take unspecified values. 51)
If that means that the padding can actually also change when just assigning individual members (which I think it does, but footnote 51 only calls out copying the whole struct by assignment, and I’m not 100% sure that “including a member object” means what I think in this context), then that means that copying an unpacked structure across information boundaries is technically always unsafe, because any change of a value in the struct might leak information by changing the padding to undefined values. It’s safe with explicitly packed structs, though (which should be the majority of the cases, as you’d have a representation independent of the architecture), but then so is what the author describes. I’d appreciate if someone could clear this up! memset’ing can still help you getting a structure quickly into a well-defined state, though.