GrapheneOS

Recent Commits

• jvoisin(01 Oct 26)

  Remove the todo item about cyclic MTE pattern See https://dustri.org/b/probabilistic-analysis-of-mte-tagging-schemes.html for analysis.

• Thomas(25 Sept 26)

  fix 0 size allocations with >PAGE_SIZE alignment for !CONFIG_LARGE_SIZE_CLASSES

• Daniel Micay(25 Sept 26)

  raise bound for slab quarantine get_random_u16 optimization This was 1 lower than it needed to be due to the get_random_uniform API not including the upper bound.

• Daniel Micay(25 Sept 26)

  optimize slab quarantine indexing via power-of-two length

• Daniel Micay(25 Sept 26)

  require power-of-two slab quarantine length multipliers The base quarantine size for each size class is a power of two so this results in the overall length always being a power of two. This will be useful to optimize indexing for both types of slab quarantine. The quarantines for each size class scale the multipliers by a power of two so this results in the overall length always being a power of two.

• Daniel Micay(25 Sept 26)

  reorder size_class struct layout

• Daniel Micay(20 Sept 26)

  add missing "cc" to clobber list for MTE assembly

• jvoisin(05 Aug 26)

  test: add coverage for untested allocator behavior Add small self-contained test cases and wire them into the Makefile and test_smc.py harness: - invalid_free_aligned_sized_small: free_aligned_sized() with a non-power-of-two alignment, covering the "invalid sized deallocation alignment (small)" fatal path. - invalid_malloc_object_size_small_canary: malloc_object_size() queried at an offset past the usable region, covering the "invalid malloc_object_size (canary)" fatal path. - pvalloc: regression test for pvalloc(0) returning a valid allocation. - reallocarray_overflow: reallocarray() returns NULL/ENOMEM on overflow. - posix_memalign_einval: posix_memalign() returns EINVAL for an invalid alignment and leaves the output pointer untouched. - aligned_alloc_einval: aligned_alloc() returns NULL/EINVAL for a non-power-of-two alignment.

• jvoisin(05 Aug 26)

  Compute aligned size class arithmetically get_size_info_align previously scanned the size class table linearly from the class matching the requested size until it found one whose size was a multiple of the alignment, costing up to ~28 iterations (e.g. a small allocation with page-sized alignment). Rounding the size up to a multiple of the (power of 2) alignment and then to a size class always lands on the smallest class whose size is a multiple of the alignment, since get_size_info rounds up to a power-of-2 spacing which preserves divisibility by the alignment. This makes the lookup constant time, which should improve memory-aligned (de)llocations. Verified equivalent to the previous scan for every size and alignment in both the default and extended size class configurations.

• Daniel Micay(06 Sept 26)

  update libdivide to 5.4.0

• Daniel Micay(02 Sept 26)

  resolve out-of-memory edge cases with MAP_FIXED MAP_FIXED can unmap the destination without creating a new mapping in out-of-memory conditions.

• Daniel Micay(04 Sept 26)

  limit size of purging for in-place shrink path It falls back to memset on out-of-memory which would crash for the old guard region.

• Daniel Micay(04 Sept 26)

  always set VMA name for quarantined large allocation

• Daniel Micay(04 Sept 26)

  enforce success for MADV_DONTNEED_LOCKED This should not return ENOMEM.

• Daniel Micay(04 Sept 26)

  replace MADV_DONTNEED with MADV_DONTNEED_LOCKED This works on locked memory instead of returning EINVAL which improves compatibility with mlockall(...) usage.

• Daniel Micay(03 Sept 26)

  remove unnecessary page_align call Large allocation sizes are always page aligned via get_large_size_class in the allocation code paths.

• Daniel Micay(08 Aug 26)

  use pinned commit hashes for GitHub actions

• rdevshp(10 Jul 26)

  Update README.md to properly use CONFIG_N_ARENA for make invocations

• rdevshp(06 Jul 26)

  Fix handling of pvalloc(0)

• rdevshp(07 Jul 26)

  Update test/.gitignore to include additional test binaries

• rdevshp(05 Jul 26)

  Fix C++ exception throwing behavior for the new operator when the new handler throws an exception with a type derived from std::bad_alloc

• Daniel Micay(01 Jul 26)

  replace aligned attribute with alignas

• rdevshp(27 Jun 26)

  github CI: add arm64 tests

• rdevshp(26 Jun 26)

  fix typos and grammar errors in memtag_test.cc/README.md

• rdevshp(26 Jun 26)

  Fix stale mallinfo link in README.md

• rdevshp(26 Jun 26)

  Use __clang__ preprocessor macro to determine the compiler specified by CC Fixes a bug in Makefile where CC=/usr/bin/clang make does not work because it does not start with clang

• Daniel Micay(21 Jun 26)

  clang-tidy: disable bugprone-casting-through-void lint This lint triggers for a void cast used to silence a cast-align warning for write_after_free_check. We aren't violating strict aliasing because the memory was always zeroed with memset prior to reading it as u64 and we don't write to another way.

• Daniel Micay(21 Jun 26)

  clang-tidy: disable useless readability lints The implicit bool conversion lint only works properly for C++ and isn't the code style we want. The math missing parentheses lint expects to use parentheses for basic math operations such as `a - b * c` which isn't the code style we want.

• Daniel Micay(21 Jun 26)

  use #ifdef for non-compound checks

• Daniel Micay(21 Jun 26)

  clang-tidy: filter third_party/ header warnings