errno not always set correctly on failure

[ronorton]

The malloc test contains some calls that are expected to fail such as attempts to allocate very large chunks of memory. Currently these return nullptr but do not set errno to ENOMEM as required by the SUSv2 standard (at least according to the Linux manpage for malloc). Unfortunately check_result notices the incorrect errno but does not fail the test due to the return here so the test emits an error message but does not fail, for example:

realloc(0x40e60000(16384), 9223372036854775807)
Expected error: 12 but got 0
(test continues)

Note that there is a suspicious check for realloc(-1) in override here but I'm sure if this is related to the failure.

[ronorton]

I suppose this is tangentially related to #423 .

[mjp41]

So I am thinking of adding

errno = ENOMEM;

on line 321:

snmalloc/src/pal/pal_posix.h

Lines 298 to 323 in af8ab2d

	static void* reserve(size_t size) noexcept
	{
	// If enforcing access, map pages initially as None, and then
	// add permissions as required. Otherwise, immediately give all
	// access as this is the most efficient to implement.
	auto prot = PalEnforceAccess ? PROT_NONE : PROT_READ | PROT_WRITE;
	void* p = mmap(
	nullptr,
	size,
	prot,
	MAP_PRIVATE | MAP_ANONYMOUS | DefaultMMAPFlags<OS>::flags,
	AnonFD<OS>::fd,
	0);
	if (p != MAP_FAILED)
	{
	#ifdef SNMALLOC_TRACING
	std::cout << "Pal_posix reserved: " << p << " (" << size << ")"
	<< std::endl;
	#endif
	return p;
	}
	return nullptr;
	}

We should then review other reserve operations for specialised Pals such as

snmalloc/src/pal/pal_bsd_aligned.h

Lines 32 to 55 in af8ab2d

	static void* reserve_aligned(size_t size) noexcept
	{
	// Alignment must be a power of 2.
	SNMALLOC_ASSERT(bits::is_pow2(size));
	SNMALLOC_ASSERT(size >= minimum_alloc_size);
	int log2align = static_cast<int>(bits::next_pow2_bits(size));
	auto prot =
	state_using || !PalEnforceAccess ? PROT_READ | PROT_WRITE : PROT_NONE;
	void* p = mmap(
	nullptr,
	size,
	prot,
	MAP_PRIVATE | MAP_ANONYMOUS | MAP_ALIGNED(log2align),
	-1,
	0);
	if (p == MAP_FAILED)
	return nullptr;
	return p;
	}

and add the same to the MAP_FAILED case.

[mjp41]

We probably need to do something here too

snmalloc/src/mem/chunkallocator.h

Lines 194 to 198 in af8ab2d

	if (slab_sizeclass >= NUM_SLAB_SIZES)
	{
	// Your address space is not big enough for this allocation!
	return {nullptr, nullptr};
	}

[rmn30]

mmap should set errno on failure so setting it in pal_posix.h should not be required. Maybe the test is hitting the chunk allocator case?

[rmn30]

I can confirm that setting errno in chunkallocator.h fixes the test failure, but is that portable to non-posix PALs? Do we need a new PAL method?

[rmn30]

I think errno is specified by C / C++ standard so probably safe to just set it?