Skip to content

mm: Default the heap alignment to 16 on 64-bit architectures. - #21961

Open
casaroli wants to merge 2 commits into
apache:masterfrom
casaroli:n222-malloc-align16-up
Open

casaroli wants to merge 2 commits into
apache:masterfrom
casaroli:n222-malloc-align16-up

Conversation

@casaroli

@casaroli casaroli commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

depends-on: /pull/21975

Summary

malloc(), calloc() and realloc() must return memory that is aligned for every type (C11 7.22.3), that is to alignof(max_align_t). On arm64, x86_64 and 64-bit RISC-V, alignof(max_align_t) is 16 (long double). CONFIG_MM_DEFAULT_ALIGNMENT has the default 8, so about half of the heap blocks are 8 modulo 16. On x86_64, an aligned SSE access (movaps, movdqa) to such a block gives a general protection fault. GCC also assumes the alignment of 16 for malloc() on these targets: at -O2 it folds (uintptr_t)malloc(n) & 15 to 0, so code that checks the alignment gets a wrong answer.

This PR makes the default 16 when CONFIG_ARCH_64BIT is set (arm64, x86_64, RV64, and sim on a 64-bit host). The other architectures keep the default 8. A configuration that sets the value keeps it. The second commit updates Documentation/os/memory/index.rst, which said that all allocations are aligned to 8 or 4 bytes.

Impact

64-bit configurations that do not set CONFIG_MM_DEFAULT_ALIGNMENT get 16-byte heap blocks. The code does not change (the same text size). The heap uses a few bytes more: at boot, +144 bytes for 58 allocations on qemu-armv8a:nsh and +232 bytes on rv-virt:nsh64; an allocation uses at most 8 bytes more. The API does not change, in flat, protected and kernel builds. No defconfig changes: no 64-bit defconfig sets the value to 8 or 16, and qemu-intel64:nxterm keeps its 64. savedefconfig gives the in-tree defconfig for qemu-armv8a nsh/knsh, rv-virt nsh64/knsh64/nsh, qemu-intel64 nsh/nxterm, mps2-an500:nsh, esp32s3-devkit:nsh and linum-stm32h753bi:littlefs. The TLSF manager (CONFIG_MM_TLSF_MANAGER) uses its own alignment for malloc() and does not change.

Testing

A new case of the mm cmocka suite checks the alignment of malloc(), calloc() and realloc() memory against alignof(max_align_t): apache/nuttx-apps#4079 (test_nuttx_mm09).

configuration master this PR
qemu-armv8a:nsh (flat, arm64, QEMU HVF) test_nuttx_mm09 FAIL: 1023 pointers not aligned 9 of 9 PASS
rv-virt:nsh64 (flat, RV64, QEMU TCG) test_nuttx_mm09 FAIL: 997 pointers not aligned 9 of 9 PASS

The user heap of the kernel build (qemu-armv8a:knsh) has the same problem on master: two malloc(45) blocks in a row are at 0xc00077e0 and 0xc0007818. ostest passes (exit status 0) with this PR on qemu-armv8a:nsh and qemu-armv8a:knsh. x86_64 (qemu-intel64) was not built.

tools/checkpatch.sh passes. The Documentation build (Sphinx) gives no warning for the changed page.

malloc(), calloc() and realloc() must return memory that is aligned
for every type (C11 7.22.3), that is to alignof(max_align_t).  On
arm64, x86_64 and 64-bit RISC-V, alignof(max_align_t) is 16 (long
double), but CONFIG_MM_DEFAULT_ALIGNMENT is 8, so half of the heap
blocks are 8 modulo 16.  On x86_64 an aligned SSE access to such a
block (movaps, movdqa) gives a general protection fault.  GCC also
assumes the alignment of 16: at -O2 it folds (uintptr_t)malloc(n) & 15
to 0.

Make the default 16 when ARCH_64BIT is set.  The default of the other
architectures and the value of configurations that set it do not
change.

Assisted-by: Claude Code:claude-opus-5-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
The memory management page says that all allocations are aligned to
8 or 4 bytes.  The alignment is CONFIG_MM_DEFAULT_ALIGNMENT: 16 by
default on 64-bit architectures and 8 on the others.

Assisted-by: Claude Code:claude-opus-5-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
@github-actions github-actions Bot added Area: Documentation Improvements or additions to documentation Size: XS The size of the change in this PR is very small labels Nov 17, 2026
@github-actions

Copy link
Copy Markdown

MemBrowse Memory Report · Explore what changed →

No memory changes detected for:


Catch memory regressions in CI with MemBrowse. Free for open source.

@casaroli
casaroli marked this pull request as draft November 18, 2026 00:40
@casaroli
casaroli marked this pull request as ready for review November 18, 2026 00:40
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Area: Documentation Improvements or additions to documentation Size: XS The size of the change in this PR is very small

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants