We expected this to hurt performance, but we were unable to measure any impact in practice.
Everyone still working in memory-unsafe languages should really just do this IMO. It would have mitigated this Mongo bug.
Looks like this is the default in OpenBSD.
Note that many malloc implementations will do this for you given an appropriate environment, e.g. setting MALLOC_CONF to opt.junk=free will do this on FreeBSD.
see here: https://godbolt.org/z/rMa8MbYox
I did, of course, test it, and anyway we now run into the "freed memory" pattern regularly when debugging (yes including optimized builds), so it's definitely working.
#include <stdlib.h>
#include <string.h>
void free(void* ptr);
void not_free(void* ptr);
void test_with_free(char* ptr) {
ptr[5] = 6;
void *(* volatile memset_v)(void *s, int c, size_t n) = memset;
memset_v(ptr + 2, 3, 4);
free(ptr);
}
void test_with_other_func(char* ptr) {
ptr[5] = 6;
void *(* volatile memset_v)(void *s, int c, size_t n) = memset;
memset_v(ptr + 2, 3, 4);
not_free(ptr);
}[0] https://llvm.org/docs/LangRef.html#llvm-memset-intrinsics
[1] https://gitweb.git.savannah.gnu.org/gitweb/?p=gnulib.git;a=b...
The Linux kernel extensively uses gcc extensions. That doesn't inherently make it insecure.
The C committee gave you memset_explicit. But note that there is still no guarantee that information can not leak. This is generally a very hard problem as information can leak in many different ways as it may have been copied by the compiler. Fully memory safe languages (so "Safe Rust" but not necessarily real-word Rust) would offer a bit more protection by default, but then there are still side-channel issues.
Creating memset_explicit won't fix existing code. "Oh but what if maybe" is just cope.
If I do memset then free then that's what I want to do
And the way things go I won't be surprised if they break memset_explicit for some other BS reason and then make you use memset_explicit_you_really_mean_it_this_time
Once you accept that optimizing compilers do, well, optimizations, the question is what should be allowed and what not. Both inlining "memset" and eliminating dead stores are both simply optimizations which people generally want.
If you want a store not to be eliminated by a compiler, you can make it volatile. The C standard says this can not be deleted by optimizations. The criticism with this was that later undefined behavior could "undo" this by "travelling in time". We made it clear in ISO C23 that this not allowed (and I believe it never was) - against protests from some compiler folks. Compilers still do not fully conform to this, which shows the limited power WG14 has to change reality.
> Once you accept that optimizing compilers do, well, optimizations
Why in tarnation it is optimizing out a write to a pointer out before a function that takes said pointer? Imagine it is any other function besides free, see how ridiculous that sounds?
Many of us who don't like working under such conditions have just moved on to other languages.
Optimizing out a function call to a heap pointer (especially memset) seems wrong to me. You called the function, it should call the function!
But it's again the C language saving time not wearing a seatbelt or checking the tire pressure for saving 10s on a 2h trip
There are still ways to obtain the desired behavior. Just put a call to a DLL or SO that implements what you need. The compiler cannot inspect the behavior of functions across module boundaries, so it cannot tell whether removing the call preserves semantics or not (for example, it could be that the external function sends the contents of the buffer to a file), so it will not remove it.
When you "call" a "function" in the source you're not specifying to the compiler that you want a specific opcode in the generated executable, you're merely specifying a particular observable behavior. This is why optimizations such as inlining and TCO are valid. If the compiler can prove that a heap allocation can be turned into a stack allocation, or even removed altogether (e.g. free(malloc(1ULL << 50))), the fact that these are exposed to the programmer as "functions" he can "call" poses no obstacle.
In the abstract machine, all expressions are evaluated as specified by the semantics. An actual
implementation need not evaluate part of an expression if it can deduce that its value is not used
and that no needed side effects are produced (including any caused by calling a function or through
volatile access to an object)
Again, could you please explain how compiler can decide to remove call to a function in an external dynamically loaded library, that is not known at compile time, simply based on the name of the function (i.e. not because the call is unreachable)? I do not see any such language in the standard.
And yes, calling unknown function from a dynamically loaded library totally is a side effect.
> And yes, calling unknown function from a dynamically loaded library totally is a side effect.
The thing is that malloc/free aren't "unknown function[s]". From the C89 standard:
> All external identifiers declared in any of the headers are reserved, whether or not the associated header is included.
And from the C23 standard:
> All identifiers with external linkage in any of the following subclauses (including the future library directions) and errno are always reserved for use as identifiers with external linkage
malloc/free are defined in <stdlib.h> and so are reserved names, so compilers are able to optimize under the assumption that malloc/free will have the semantics dictated by the standard.
In fact, the C23 standard explicitly provides an example of this kind of thing:
> Because external identifiers and some macro names beginning with an underscore are reserved, implementations can provide special semantics for such names. For example, the identifier _BUILTIN_abs could be used to indicate generation of in-line code for the abs function. Thus, the appropriate header could specify
#define abs(x) _BUILTIN_abs(x)
What a side effect is, is explained in "5.1.2.3". Calling function is only a side effect when the function contains a side effect, such as modifying an object, or a volatile access, or I/O.
As they're freely replaceable through loading, and designed for that, I would strongly suggest that are among the most magical areas of the C standard.
We get a whole section for those in the standard: 7.24.3 Memory management functions
Hell, malloc is allowed to return you _less than you asked for_:
> The pointer returned if the allocation succeeds is suitably aligned so that it may be assigned to a pointer to any type of object with a fundamental alignment requirement and size less than or equal to the size requested
That was one of the first optimisations we had, back with Fortran and COBOL. Before C existed - and as B started life as a stripped down Fortran compiler, the history carried through.
The K&R book describes the buddy system for malloc, and how its design makes it suitable for compiler optimisations - including ignoring a write to a pointer that does nothing, because the pointer will no longer be valid.
I'd suspect that eliding suitable malloc/free pairs would not break most existing code because most existing code simply does not depend on malloc/free doing anything other than and/or beyond what the C standard requires.
How would you propose that eliding free(malloc(x)) would break "most" existing code, anyways?
Or somebody would try to plug in mimalloc/jemalloc or a debug allocator and wonder what's going on.
Such a program would continue to function as normal; the dirty data would just be left on the stack. If the developer wants to clear that data too, they'd just have to modify the compiler to overwrite the stack just before (or just after) moving the stack pointer.
>Or somebody would try to plug in mimalloc/jemalloc or a debug allocator and wonder what's going on.
Again, that wouldn't be broken. They would see that no dynamic allocations were performed during that particular section. Which would be correct.
> As an example, user kentonv wrote: "I patched the memory allocator used by the Cloudflare Workers runtime to overwrite all memory with a static byte pattern on free". And compiler would, like, "nah, let's leave all that data on stack".
Strictly speaking, I don't think eliding malloc/free would "break" those programs because that behavior is there for security if/when something else goes wrong, not as part of the software's regular intended functionality (or at least I sure hope nothing relies on that behavior for proper functioning!).
> Or somebody would try to plug in mimalloc/jemalloc [] and wonder what's going on.
Why would mimalloc/jemalloc/some other general-purpose allocator care that it doesn't have to execute a matching malloc/free pair any more than the default allocator?
I'm not sure debug allocators would care either? If you're trying to debug mismatched malloc/free pairs then the ones the compiler elides are the ones you don't care about anyways since those are the ones that can be statically proven to be "self-contained" and/or correct. If you're gathering statistics then you probably care more about the malloc/free calls that do occur (i.e., the ones that can't be elided), not those that don't.
In any case, if you want to use a malloc/free implementation that promises more than the C standard does (e.g., special byte pattern on free, statistics/debug info tracking, etc.) there's always -fno-builtin-malloc (or memset_explicit if you're lucky enough to be using C23). Of course, the tradeoff is that you give up some potential performance.
I think that's an overly narrow reading of the footnote. I don't see an obvious reason why "such names" in the footnote should only cover "some macro names beginning with an underscore" and not also "external identifiers". And if implementations are allowed to define special semantics for "external identifiers", then... well, that's exactly what they did!
In addition, there's still the as-if rule. The semantics of malloc/free are defined by the C standard; if the compiler can deduce that there is no observable difference between a version of the program that calls those and a version that does not, why does it matter that the call is emitted? A function call in and of itself is not a side effect, and since the C standard dictates what malloc/free do the compiler knows their possible side effects.
Furthermore, the addition of memset_explicit and its footnote ("The intention is that the memory store is always performed (i.e. never elided), regardless of optimizations. This is in contrast to calls to the memset function (7.26.6.1)") implies that eliding calls is in fact acceptable behavior when optimizations are enabled. If eliding calls were not permissible when optimizing then what's the point of memset_explicit?
> There should be no case where compiler could assume anything about a function it does not see based simply on it's name.
Again, external identifiers defined by the C standard are reserved. Reserved external identifiers aren't just for show. From the C89 standard:
> If the program defines an external identifier with the same name as a reserved external identifier, even in a semantically equivalent form, the behavior is undefined.
And from C23:
> If the program declares or defines an identifier in a context in which it is reserved (other than as allowed by 7.1.4), the behavior is undefined.
This means that yes, under modern compilers' interpretation of UB compilers can assume things about functions based on their names because modern compilers generally optimize assuming UB does not happen. The compiler does not need to see the function's implementation because it is the function's implementation as far as it is concerned.
> 7.1.3 Reserved Identifiers
> [snip]
> Macro names and identifiers with external linkage that are specified in the C standard library clauses.
> This proposal does not propose any changes to these reserved identifiers.
Furthermore, that paper doesn't make the use of reserved external identifiers not UB, so there's no change there either.
If you need better performance, write your own allocator optimized for your specific use case — it's not that hard.
Besides, you if you don't need to clear old allocations, there are likely other optimizations you'll be able to find which would never fly in a system allocator.