mirror of
https://github.com/GrapheneOS/hardened_malloc.git
synced 2024-10-01 01:36:01 -04:00
171 lines
8.7 KiB
Markdown
171 lines
8.7 KiB
Markdown
Debian stable determines the most ancient set of supported dependencies:
|
|
|
|
* glibc 2.24
|
|
* Linux 4.9
|
|
* Clang 3.8 or GCC 6.3
|
|
|
|
However, using more recent releases is highly recommended. Older versions of
|
|
the dependencies may be compatible at the moment but are not tested and will
|
|
explicitly not be supported.
|
|
|
|
Ports to Android (Bionic libc) and musl libc will be created later. The initial
|
|
target is glibc for ease of development, since it supports replacing the malloc
|
|
implementation at runtime using dynamic linking including via `LD_PRELOAD`.
|
|
|
|
# Basic design
|
|
|
|
The current design is very simple and will become a bit more sophisticated as
|
|
the basic features are completed and the implementation is hardened and
|
|
optimized. The allocator is exclusive to 64-bit platforms in order to take full
|
|
advantage of the abundant address space without being constrained by needing to
|
|
keep the design compatible with 32-bit.
|
|
|
|
Small allocations are always located in a large memory region reserved for slab
|
|
allocations. It can be determined that an allocation is one of the small size
|
|
classes from the address range. Each small size class has a separate reserved
|
|
region within the larger region, and the size of a small allocation can simply
|
|
be determined from the range. Each small size class has a separate out-of-line
|
|
metadata array outside of the overall allocation region, with the index of the
|
|
metadata struct within the array mapping to the index of the slab within the
|
|
dedicated size class region. Slabs are a multiple of the page size and are
|
|
paged aligned. The entire small size class region starts out memory protected
|
|
and becomes readable / writable as it gets allocated, with idle slabs beyond
|
|
the cache limit having their pages dropped and the memory protected again.
|
|
|
|
Large allocations are tracked via a global hash table mapping their address to
|
|
their size and guard size. They're simply memory mappings and get mapped on
|
|
allocation and then unmapped on free.
|
|
|
|
# Security properties
|
|
|
|
* Fully out-of-line metadata
|
|
* Deterministic detection of any invalid free (unallocated, unaligned, etc.)
|
|
* Isolated memory region for slab allocations
|
|
* Divided up into isolated inner regions for each size class
|
|
* High entropy random base for each size class region
|
|
* No deterministic / low entropy offsets between allocations with
|
|
different size classes
|
|
* Metadata is completely outside the slab allocation region
|
|
* No references to metadata within the slab allocation region
|
|
* No deterministic / low entropy offsets to metadata
|
|
* Entire slab region starts out non-readable and non-writable
|
|
* Slabs beyond the cache limit are purged and become non-readable and
|
|
non-writable memory again
|
|
* Fine-grained randomization within memory regions
|
|
* Randomly sized guard regions for large allocations
|
|
* Random slot selection within slabs
|
|
* [in-progress] Randomized delayed free for slab allocations
|
|
* [in-progress] Randomized allocation of slabs
|
|
* [more randomization coming as the implementation is matured]
|
|
* Slab allocations are zeroed on free and large allocations are unmapped
|
|
* [in-progress] Detection of write-after-free by verifying zero filling is
|
|
intact
|
|
* Memory in fresh allocations is consistently zeroed due to it either being
|
|
fresh pages or zeroed on free after previous usage
|
|
* [in-progress] Delayed free via a combination of FIFO and randomization for
|
|
slab allocations
|
|
* [in-progress] Random canaries placed after each slab allocation to *absorb*
|
|
and then later detect overflows/underflows
|
|
* High entropy per-slab random values
|
|
* Mangled into a unique value per slab slot (although not with a strong
|
|
keyed hash due to performance limitations)
|
|
* [in-progress] Some slab locations are skipped and remain memory protected,
|
|
leaving slab size class regions interspersed with guard pages
|
|
* Zero size allocations are memory protected
|
|
* Protected allocator metadata
|
|
* Address space for metadata is never used for allocations and vice versa
|
|
* [implementing stronger protection is in-progress]
|
|
* Extension for retrieving the size of allocations with fallback
|
|
to a sentinel for pointers not managed by the allocator
|
|
* Can also return accurate values for pointers *within* small allocations
|
|
* The same applies to pointers within the first page of large allocations,
|
|
otherwise it currently has to return a sentinel
|
|
* No alignment tricks interfering with ASLR like jemalloc, PartitionAlloc, etc.
|
|
* No usage of the legacy brk heap
|
|
* Aggressive sanity checks
|
|
* Errors other than ENOMEM from mmap, munmap, mprotect and mremap treated
|
|
as fatal, which can help to decent memory management gone wrong elsewhere
|
|
in the program
|
|
|
|
# Randomness
|
|
|
|
The current implementation of random number generation for randomization-based
|
|
mitigations is based on generating a keystream from a stream cipher (ChaCha8)
|
|
in small chunks. A separate CSPRNG is used for each small size class, large
|
|
allocations, etc. in order to fit into the existing fine-grained locking model
|
|
without needing to waste memory per thread by having the CSPRNG state in Thread
|
|
Local Storage. Similarly, it's protected via the same approach taken for the
|
|
rest of the metadata. The stream cipher is regularly reseeded from the OS to
|
|
provide backtracking and prediction resistance with a negligible cost. The
|
|
reseed interval simply needs to be adjusted to the point that it stops
|
|
registering as having any significant performance impact. The performance
|
|
impact on recent Linux kernels is primarily from the high cost of system calls
|
|
and locking since the implementation is quite efficient (ChaCha20), especially
|
|
for just generating the key and nonce for another stream cipher (ChaCha8).
|
|
|
|
ChaCha8 is a great fit because it's extremely fast across platforms without
|
|
relying on hardware support or complex platform-specific code. The security
|
|
margins of ChaCha20 would be completely overkill for the use case. Using
|
|
ChaCha8 avoids needing to resort to a non-cryptographically secure PRNG or
|
|
something without a lot of scrunity. The current implementation is simply the
|
|
reference implementation of ChaCha8 converted into a pure keystream by ripping
|
|
out the XOR of the message into the keystream.
|
|
|
|
The random range generation functions are a highly optimized implementation
|
|
too. Traditional uniform random number generation within a range is very high
|
|
overhead and can easily dwarf the cost of an efficient CSPRNG.
|
|
|
|
# Size classes
|
|
|
|
The zero byte size class is a special case of the smallest regular size class. It's allocated in a
|
|
separate region with the memory left non-readable and non-writable.
|
|
|
|
The slab slot count for each size class is not yet finely tuned beyond choosing values avoiding
|
|
internal fragmentation for slabs (i.e. avoiding wasted space due to page size rounding).
|
|
|
|
The choice of size classes is the same as jemalloc, but with a much different approach to the
|
|
slabs containing them:
|
|
|
|
> size classes are multiples of the quantum [16], spaced such that there are four size classes for
|
|
> each doubling in size, which limits internal fragmentation to approximately 20% for all but the
|
|
> smallest size classes
|
|
|
|
| size class | worst case internal fragmentation | slab slots | slab size | worst case internal fragmentation for slabs |
|
|
| - | - | - | - | - |
|
|
| 16 | 100% | 256 | 4096 | 0.0% |
|
|
| 32 | 46.875% | 128 | 4096 | 0.0% |
|
|
| 48 | 31.25% | 85 | 4096 | 0.390625% |
|
|
| 64 | 23.4375% | 64 | 4096 | 0.0% |
|
|
| 80 | 18.75% | 51 | 4096 | 0.390625% |
|
|
| 96 | 15.625% | 42 | 4096 | 1.5625% |
|
|
| 112 | 13.392857142857139% | 36 | 4096 | 1.5625% |
|
|
| 128 | 11.71875% | 64 | 8192 | 0.0% |
|
|
| 160 | 19.375% | 51 | 8192 | 0.390625% |
|
|
| 192 | 16.145833333333343% | 64 | 12288 | 0.0% |
|
|
| 224 | 13.839285714285708% | 54 | 12288 | 1.5625% |
|
|
| 256 | 12.109375% | 64 | 16384 | 0.0% |
|
|
| 320 | 19.6875% | 64 | 20480 | 0.0% |
|
|
| 384 | 16.40625% | 64 | 24576 | 0.0% |
|
|
| 448 | 14.0625% | 64 | 28672 | 0.0% |
|
|
| 512 | 12.3046875% | 64 | 32768 | 0.0% |
|
|
| 640 | 19.84375% | 64 | 40960 | 0.0% |
|
|
| 768 | 16.536458333333343% | 64 | 49152 | 0.0% |
|
|
| 896 | 14.174107142857139% | 64 | 57344 | 0.0% |
|
|
| 1024 | 12.40234375% | 64 | 65536 | 0.0% |
|
|
| 1280 | 19.921875% | 16 | 20480 | 0.0% |
|
|
| 1536 | 16.6015625% | 16 | 24576 | 0.0% |
|
|
| 1792 | 14.229910714285708% | 16 | 28672 | 0.0% |
|
|
| 2048 | 12.451171875% | 16 | 32768 | 0.0% |
|
|
| 2560 | 19.9609375% | 8 | 20480 | 0.0% |
|
|
| 3072 | 16.634114583333343% | 8 | 24576 | 0.0% |
|
|
| 3584 | 14.2578125% | 8 | 28672 | 0.0% |
|
|
| 4096 | 12.4755859375% | 8 | 32768 | 0.0% |
|
|
| 5120 | 19.98046875% | 8 | 40960 | 0.0% |
|
|
| 6144 | 16.650390625% | 8 | 49152 | 0.0% |
|
|
| 7168 | 14.271763392857139% | 8 | 57344 | 0.0% |
|
|
| 8192 | 12.48779296875% | 8 | 65536 | 0.0% |
|
|
| 10240 | 19.990234375% | 6 | 61440 | 0.0% |
|
|
| 12288 | 16.658528645833343% | 5 | 61440 | 0.0% |
|
|
| 14336 | 14.278738839285708% | 4 | 57344 | 0.0% |
|
|
| 16384 | 12.493896484375% | 4 | 65536 | 0.0% |
|