Hardened allocator designed for modern systems. It has integration into Android's Bionic libc and can be used externally with musl and glibc as a dynamic library for use on other Linux-based platforms. It will gain more portability / integration over time.
Go to file
2019-04-07 00:50:26 -04:00
test add real mallinfo implementation for Android 2019-04-06 20:54:26 -04:00
third_party move libdivide.h to third_party/libdivide.h 2018-08-30 03:03:19 -04:00
.gitignore initial commit 2018-08-23 15:42:41 -04:00
Android.bp add real mallinfo implementation for Android 2019-04-06 20:54:26 -04:00
calculate_waste.py slab internal fragmentation is always the same 2018-11-04 03:59:54 -05:00
chacha.c use consistent code style in chacha.c 2018-10-19 17:09:41 -04:00
chacha.h use a consistent style for fixed-size int types 2018-10-04 15:17:11 -04:00
CREDITS reference paper rather than fastrand for Python 2019-01-10 17:12:53 -05:00
h_malloc.c move thread sealing implementation 2019-04-07 00:50:26 -04:00
h_malloc.h clarify hardened_malloc extensions section 2019-04-07 00:34:13 -04:00
KERNEL_FEATURE_WISHLIST.md add RLIMIT_AS alternative to wishlist 2018-12-09 12:48:07 -05:00
LICENSE add basic documentation on memory tagging approach 2019-02-04 11:58:28 -05:00
Makefile add real mallinfo implementation for Android 2019-04-06 20:54:26 -04:00
memory.c add support for labelling memory regions 2019-02-13 13:34:33 -05:00
memory.h add support for labelling memory regions 2019-02-13 13:34:33 -05:00
mutex.h wrap mutex type to enable future optimization 2018-09-07 01:13:08 -04:00
new.cc remove redundant extern "C" block 2019-03-23 13:41:12 -04:00
pages.c label allocate_aligned_pages mappings 2019-03-23 22:29:04 -04:00
pages.h move hash_page to pages.h 2019-03-25 14:54:22 -04:00
preload.sh add lib prefix to the shared object name 2018-12-05 02:05:37 -05:00
random.c use smaller integers for random cache state 2018-11-04 14:24:21 -05:00
random.h use smaller integers for random cache state 2018-11-04 14:24:21 -05:00
README.md document CONFIG_STATS 2019-04-07 00:10:22 -04:00
util.c make error reporting more robust 2018-10-03 16:58:49 -04:00
util.h overhaul glibc compatibility workarounds 2019-03-26 01:45:57 -04:00

Hardened malloc

This is a security-focused general purpose memory allocator providing the malloc API along with various extensions. It provides substantial hardening against heap corruption vulnerabilities. The security-focused design also leads to much less metadata overhead and memory waste from fragmentation than a more traditional allocator design. It aims to provide decent overall performance with a focus on long-term performance and memory usage rather than allocator micro-benchmarks. It offers scalability via a configurable number of entirely independently arenas, with the internal locking within arenas further divided up per size class.

This project currently aims to support Android, musl and glibc. It may support other non-Linux operating systems in the future. For Android and musl, there will be custom integration and other hardening features. The glibc support will be limited to replacing the malloc implementation because musl is a much more robust and cleaner base to build on and can cover the same use cases.

This allocator is intended as a successor to a previous implementation based on extending OpenBSD malloc with various additional security features. It's still heavily based on the OpenBSD malloc design, albeit not on the existing code other than reusing the hash table implementation for the time being. The main differences in the design are that it is solely focused on hardening rather than finding bugs, uses finer-grained size classes along with slab sizes going beyond 4k to reduce internal fragmentation, doesn't rely on the kernel having fine-grained mmap randomization and only targets 64-bit to make aggressive use of the large address space. There are lots of smaller differences in the implementation approach. It incorporates the previous extensions made to OpenBSD malloc including adding padding to allocations for canaries (distinct from the current OpenBSD malloc canaries), write-after-free detection tied to the existing clearing on free, queues alongside the existing randomized arrays for quarantining allocations and proper double-free detection for quarantined allocations. The per-size-class memory regions with their own random bases were loosely inspired by the size and type-based partitioning in PartitionAlloc. The planned changes to OpenBSD malloc ended up being too extensive and invasive so this project was started as a fresh implementation better able to accomplish the goals. For 32-bit, a port of OpenBSD malloc with small extensions can be used instead as this allocator fundamentally doesn't support that environment.

Dependencies

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.

For external malloc replacement with musl, musl 1.1.20 is required. However, there will be custom integration offering better performance in the future along with other hardening for the C standard library implementation.

For Android, only current generation Android Open Source Project branches will be supported, which currently means pie-qpr2-release.

Testing

The preload.sh script can be used for testing with dynamically linked executables using glibc or musl:

./preload.sh krita --new-image RGBA,U8,500,500

It can be necessary to substantially increase the vm.max_map_count sysctl to accomodate the large number of mappings caused by guard slabs and large allocation guard regions. The number of mappings can also be drastically reduced via a significant increase to CONFIG_GUARD_SLABS_INTERVAL but the feature has a low performance and memory usage cost so that isn't recommended.

It can offer slightly better performance when integrated into the C standard library and there are other opportunities for similar hardening within C standard library and dynamic linker implementations. For example, a library region can be implemented to offer similar isolation for dynamic libraries as this allocator offers across different size classes. The intention is that this will be offered as part of hardened variants of the Bionic and musl C standard libraries.

Configuration

You can set some configuration options at compile-time via arguments to the make command as follows:

make CONFIG_EXAMPLE=false

Configuration options are provided when there are significant compromises between portability, performance, memory usage or security. The core design choices are not configurable and the allocator remains very security-focused even with all the optional features disabled.

The following boolean configuration options are available:

  • CONFIG_NATIVE: true (default) or false to control whether the code is optimized for the detected CPU on the host. If this is disabled, setting up a custom -march higher than the baseline architecture is highly recommended due to substantial performance benefits for this code.
  • CONFIG_CXX_ALLOCATOR: true (default) or false to control whether the C++ allocator is replaced for slightly improved performance and detection of mismatched sizes for sized deallocation (often type confusion bugs). This will result in linking against the C++ standard library.
  • CONFIG_ZERO_ON_FREE: true (default) or false to control whether small allocations are zeroed on free, to mitigate use-after-free and uninitialized use vulnerabilities along with purging lots of potentially sensitive data from the process as soon as possible. This has a performance cost scaling to the size of the allocation, which is usually acceptable.
  • CONFIG_WRITE_AFTER_FREE_CHECK: true (default) or false to control sanity checking that new allocations contain zeroed memory. This can detect writes caused by a write-after-free vulnerability and mixes well with the features for making memory reuse randomized / delayed. This has a performance cost scaling to the size of the allocation, which is usually acceptable.
  • CONFIG_SLOT_RANDOMIZE: true (default) or false to randomize selection of free slots within slabs. This has a measurable performance cost and isn't one of the important security features, but the cost has been deemed more than acceptable to be enabled by default.
  • CONFIG_SLAB_CANARY: true (default) or false to enable support for adding 8 byte canaries to the end of memory allocations. The primary purpose of the canaries is to render small fixed size buffer overflows harmless by absorbing them. The first byte of the canary is always zero, containing overflows caused by a missing C string NUL terminator. The other 7 bytes are a per-slab random value. On free, integrity of the canary is checked to detect attacks like linear overflows or other forms of heap corruption caused by imprecise exploit primitives. However, checking on free will often be too late to prevent exploitation so it's not the main purpose of the canaries.
  • CONFIG_SEAL_METADATA: true or false (default) to control whether Memory Protection Keys are used to disable access to all writable allocator state outside of the memory allocator code. It's currently disabled by default due to being extremely experimental and a significant performance cost for this use case on current generation hardware, which may become drastically lower in the future. Whether or not this feature is enabled, the metadata is all contained within an isolated memory region with high entropy random guard regions around it.

The following integer configuration options are available. Proper sanity checks for the chosen values are not written yet, so use them at your own peril:

  • CONFIG_SLAB_QUARANTINE_RANDOM_LENGTH: 1 (default) to control the number of slots in the random array used to randomize reuse for small memory allocations. This sets the length for the largest size class (currently 16384) and the quarantine length for smaller size classes is scaled to match the total memory of the quarantined allocations (1 becomes 1024 for 16 byte allocations).
  • CONFIG_SLAB_QUARANTINE_QUEUE_LENGTH: 1 (default) to control the number of slots in the queue used to delay reuse for small memory allocations. This sets the length for the largest size class (currently 16384) and the quarantine length for smaller size classes is scaled to match the total memory of the quarantined allocations (1 becomes 1024 for 16 byte allocations).
  • CONFIG_GUARD_SLABS_INTERVAL: 1 (default) to control the number of slabs before a slab is skipped and left as an unused memory protected guard slab
  • CONFIG_GUARD_SIZE_DIVISOR: 2 (default) to control the maximum size of the guard regions placed on both sides of large memory allocations, relative to the usable size of the memory allocation
  • CONFIG_REGION_QUARANTINE_RANDOM_LENGTH: 128 (default) to control the number of slots in the random array used to randomize region reuse for large memory allocations
  • CONFIG_REGION_QUARANTINE_QUEUE_LENGTH: 1024 (default) to control the number of slots in the queue used to delay region reuse for large memory allocations
  • CONFIG_REGION_QUARANTINE_SKIP_THRESHOLD: 33554432 (default) to control the size threshold where large allocations will not be quarantined
  • CONFIG_FREE_SLABS_QUARANTINE_RANDOM_LENGTH: 32 (default) to control the number of slots in the random array used to randomize free slab reuse
  • CONFIG_CLASS_REGION_SIZE: 34359738368 (default) to control the size of the size class regions
  • CONFIG_N_ARENA: 1 (default) to control the number of arenas
  • CONFIG_STATS: false (default) to control whether stats on allocation / deallocation count and active allocations are tracked. This is currently only exposed via the mallinfo APIs on Android.

There will be more control over enabled features in the future along with control over fairly arbitrarily chosen values like the size of empty slab caches (making them smaller improves security and reduces memory usage while larger caches can substantially improves performance).

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 page 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.

This allocator is aimed at production usage, not aiding with finding and fixing memory corruption bugs for software development. It does find many latent bugs but won't include features like the option of generating and storing stack traces for each allocation to include the allocation site in related error messages. The design choices are based around minimizing overhead and maximizing security which often leads to different decisions than a tool attempting to find bugs. For example, it uses zero-based sanitization on free and doesn't minimize slack space from size class rounding between the end of an allocation and the canary / guard region. Zero-based filling has the least chance of uncovering latent bugs, but also the best chance of mitigating vulnerabilities. The canary feature is primarily meant to act as padding absorbing small overflows to render them harmless, so slack space is helpful rather than harmful despite not detecting the corruption on free. The canary needs detection on free in order to have any hope of stopping other kinds of issues like a sequential overflow, which is why it's included. It's assumed that an attacker can figure out the allocator is in use so the focus is explicitly not on detecting bugs that are impossible to exploit with it in use like an 8 byte overflow. The design choices would be different if performance was a bit less important and if a core goal was finding latent bugs.

Security properties

  • Fully out-of-line metadata
  • Deterministic detection of any invalid free (unallocated, unaligned, etc.)
    • Validation of the size passed for C++14 sized deallocation by delete even for code compiled with earlier standards (detects type confusion if the size is different) and by various containers using the allocator API directly
  • 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
      • Placed into a queue for reuse in FIFO order to maximize the time spent memory protected
      • Randomized array is used to add a random delay for reuse
  • Fine-grained randomization within memory regions
    • Randomly sized guard regions for large allocations
    • Random slot selection within slabs
    • Randomized delayed free for slab allocations
    • [in-progress] Randomized choice of slabs
    • [in-progress] Randomized allocation of slabs
  • Slab allocations are zeroed on free
  • Detection of write-after-free for slab allocations by verifying zero filling is intact at allocation time
  • Large allocations are purged and memory protected on free with the memory mapping kept reserved in a quarantine to detect use-after-free
    • The quarantine is primarily based on a FIFO ring buffer, with the oldest mapping in the quarantine being unmapped to make room for the most recently freed mapping
    • Another layer of the quarantine swaps with a random slot in an array to randomize the number of large deallocations required to push mappings out of the quarantine
  • Memory in fresh allocations is consistently zeroed due to it either being fresh pages or zeroed on free after previous usage
  • Delayed free via a combination of FIFO and randomization for slab allocations
  • Random canaries placed after each slab allocation to absorb and then later detect overflows/underflows
    • High entropy per-slab random values
    • Leading byte is zeroed to contain C string overflows
  • Possible slab locations are skipped and remain memory protected, leaving slab size class regions interspersed with guard pages
  • Zero size allocations are a dedicated size class with the entire region remaining non-readable and non-writable
  • Protected allocator state (including all metadata)
    • Address space for state is entirely reserved during initialization and never reused for allocations or anything else
    • State within global variables is entirely read-only after initialization with pointers to the isolated allocator state so leaking the address of the library doesn't leak the address of writable state
    • Allocator state is located within a dedicated region with high entropy randomly sized guard regions around it
    • Protection via Memory Protection Keys (MPK) on x86_64 (disabled by default due to low benefit-cost ratio on top of baseline protections)
    • [future] Protection via MTE on ARMv8.5+
  • 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 detect memory management gone wrong elsewhere in the process.
  • [future] Memory tagging for slab allocations via MTE on ARMv8.5+
    • random memory tags as the baseline, providing probabilistic protection against various forms of memory corruption
    • dedicated tag for free slots, set on free, for deterministic protection against accessing freed memory
    • store previous random tag within freed slab allocations, and increment it to get the next tag for that slot to provide deterministic use-after-free detection through multiple cycles of memory reuse
    • guarantee distinct tags for adjacent memory allocations by incrementing past matching values for deterministic detection of linear overflows

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 dedicated region like other size classes but with the slabs never being made readable and writable so the only memory usage is for the slab metadata.

The choice of size classes for slab allocation is the same as jemalloc, which is a careful balance between minimizing internal and external fragmentation. If there are more size classes, more memory is wasted on free slots available only to allocation requests of those sizes (external fragmentation). If there are fewer size classes, the spacing between them is larger and more memory is wasted due to rounding up to the size classes (internal fragmentation). There are 4 special size classes for the smallest sizes (16, 32, 48, 64) that are simply spaced out by the minimum spacing (16). Afterwards, there are four size classes for every power of two spacing which results in bounding the internal fragmentation below 20% for each size class. This also means there are 4 size classes for each doubling in size.

The slot counts tied to the size classes are specific to this allocator rather than being taken from jemalloc. Slabs are always a span of pages so the slot count needs to be tuned to minimize waste due to rounding to the page size. For now, this allocator is set up only for 4096 byte pages as a small page size is desirable for finer-grained memory protection and randomization. It could be ported to larger page sizes in the future. The current slot counts are only a preliminary set of values.

size class worst case internal fragmentation slab slots slab size internal fragmentation for slabs
16 93.75% 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%

The slab allocation size classes currently end at 16384 since that's the final size for 2048 byte spacing and the next spacing class matches the page size of 4096 bytes on the target platforms. This is the minimum set of small size classes required to avoid substantial waste from rounding. Further slab allocation size classes may be offered as an option in the future.

Scalability

Small (slab) allocations

As a baseline form of fine-grained locking, the slab allocator has entirely separate allocators for each size class. Each size class has a dedicated lock, CSPRNG and other state.

The slab allocator's scalability primarily comes from dividing up the slab allocation region into independent arenas assigned to threads. The arenas are just entirely separate slab allocators with their own sub-regions for each size class. Using 4 arenas reserves a region 4 times as large and the relevant slab allocator metadata is determined based on address, as part of the same approach to finding the per-size-class metadata. The part that's still open to different design choices is how arenas are assigned to threads. One approach is statically assigning arenas via round-robin like the standard jemalloc implementation, or statically assigning to a random arena which is essentially the current implementation. Another option is dynamic load balancing via a heuristic like sched_getcpu for per-CPU arenas, which would offer better performance than randomly choosing an arena each time while being more predictable for an attacker. There are actually some security benefits from this assignment being completely static, since it isolates threads from each other. Static assignment can also reduce memory usage since threads may have varying usage of size classes.

When there's substantial allocation or deallocation pressure, the allocator does end up calling into the kernel to purge / protect unused slabs by replacing them with fresh PROT_NONE regions along with unprotecting slabs when partially filled and cached empty slabs are depleted. There will be configuration over the amount of cached empty slabs, but it's not entirely a performance vs. memory trade-off since memory protecting unused slabs is a nice opportunistic boost to security. However, it's not really part of the core security model or features so it's quite reasonable to use much larger empty slab caches when the memory usage is acceptable. It would also be reasonable to attempt to use heuristics for dynamically tuning the size, but there's not a great one size fits all approach so it isn't currently part of this allocator implementation.

Thread caching (or lack thereof)

Thread caches are a commonly implemented optimization in modern allocators but aren't very suitable for a hardened allocator even when implemented via arrays like jemalloc rather than free lists. They would prevent the allocator from having perfect knowledge about which memory is free in a way that's both race free and works with fully out-of-line metadata. It would also interfere with the quality of fine-grained randomization even with randomization support in the thread caches. The caches would also end up with much weaker protection than the dedicated metadata region. Potentially worst of all, it's inherently incompatible with the important quarantine feature.

The primary benefit from a thread cache is performing batches of allocations and batches of deallocations to amortize the cost of the synchronization used by locking. The issue is not contention but rather the cost of synchronization itself. Performing operations in large batches isn't necessarily a good thing in terms of reducing contention to improve scalability. Large thread caches like TCMalloc are a legacy design choice and aren't a good approach for a modern allocator. In jemalloc, thread caches are fairly small and have a form of garbage collection to clear them out when they aren't being heavily used. Since this is a hardened allocator with a bunch of small costs for the security features, the synchronization is already a smaller percentage of the overall time compared to a much leaner performance-oriented allocator. These benefits could be obtained via allocation queues and deallocation queues which would avoid bypassing the quarantine and wouldn't have as much of an impact on randomization. However, deallocation queues would also interfere with having global knowledge about what is free. An allocation queue alone wouldn't have many drawbacks, but it isn't currently planned even as an optional feature since it probably wouldn't be enabled by default and isn't worth the added complexity.

The secondary benefit of thread caches is being able to avoid the underlying allocator implementation entirely for some allocations and deallocations when they're mixed together rather than many allocations being done together or many frees being done together. The value of this depends a lot on the application and it's entirely unsuitable / incompatible with a hardened allocator since it bypasses all of the underlying security and would destroy much of the security value.

Large allocations

The expectation is that the allocator does not need to perform well for large allocations, especially in terms of scalability. When the performance for large allocations isn't good enough, the approach will be to enable more slab allocation size classes. Doubling the maximum size of slab allocations only requires adding 4 size classes while keeping internal waste bounded below 20%.

Large allocations are implemented as a wrapper on top of the kernel memory mapping API. The addresses and sizes are tracked in a global data structure with a global lock. The current implementation is a hash table and could easily use fine-grained locking, but it would have little benefit since most of the locking is in the kernel. Most of the contention will be on the mmap_sem lock for the process in the kernel. Ideally, it could simply map memory when allocating and unmap memory when freeing. However, this is a hardened allocator and the security features require extra system calls due to lack of direct support for this kind of hardening in the kernel. Randomly sized guard regions are placed around each allocation which requires mapping a PROT_NONE region including the guard regions and then unprotecting the usable area between them. The quarantine implementation requires clobbering the mapping with a fresh PROT_NONE mapping using MAP_FIXED on free to hold onto the region while it's in the quarantine, until it's eventually unmapped when it's pushed out of the quarantine. This means there are 2x as many system calls for allocating and freeing as there would be if the kernel supported these features directly.

Memory tagging

Integrating extensive support for ARMv8.5 memory tagging is planned and this section will be expanded cover the details on the chosen design. The approach for slab allocations is currently covered, but it can also be used for the allocator metadata region and large allocations.

Memory allocations are already always multiples of naturally aligned 16 byte units, so memory tags are a natural fit into a malloc implementation due to the 16 byte alignment requirement. The only extra memory consumption will come from the hardware supported storage for the tag values (4 bits per 16 bytes).

The baseline policy will be to generate random tags for each slab allocation slot on first use. The highest value will be reserved for marking freed memory allocations to detect any accesses to freed memory so it won't be part of the generated range. Adjacent slots will be guaranteed to have distinct memory tags in order to guarantee that linear overflows are detected. There are a few ways of implementing this and it will end up depending on the performance costs of different approaches. If there's an efficient way to fetch the adjacent tag values without wasting extra memory, it will be possible to check for them and skip them either by generating a new random value in a loop or incrementing past them since the tiny bit of bias wouldn't matter. Another approach would be alternating odd and even tag values but that would substantially reduce the overall randomness of the tags and there's very little entropy from the start.

Once a slab allocation has been freed, the tag will be set to the reserved value for free memory and the previous tag value will be stored inside the allocation itself. The next time the slot is allocated, the chosen tag value will be the previous value incremented by one to provide use-after-free detection between generations of allocations. The stored tag will be wiped before retagging the memory, to avoid leaking it and as part of preserving the security property of newly allocated memory being zeroed due to zero-on-free. It will eventually wrap all the way around, but this ends up providing a strong guarantee for many allocation cycles due to the combination of 4 bit tags with the FIFO quarantine feature providing delayed free. It also benefits from random slot allocation and the randomized portion of delayed free, which result in a further delay along with preventing a deterministic bypass by forcing a reuse after a certain number of allocation cycles. Similarly to the initial tag generation, tag values for adjacent allocations will be skipped by incrementing past them.

For example, consider this slab of allocations that are not yet used with 16 representing the tag for free memory. For the sake of simplicity, there will be no quarantine or other slabs for this example:

| 16 | 16 | 16 | 16 | 16 | 16 |

Three slots are randomly chosen for allocations, with random tags assigned (2, 15, 7) since these slots haven't ever been used and don't have saved values:

| 16 | 2  | 16 | 15 |  7 | 16 |

The 2nd allocation slot is freed, and is set back to the tag for free memory (16), but with the previous tag value stored in the freed space:

| 16 | 16 | 16 | 7  | 15 | 16 |

The first slot is allocated for the first time, receiving the random value 3:

| 3  | 16 | 16 | 7  | 15 | 16 |

The 2nd slot is randomly chosen again, so the previous tag (2) is retrieved and incremented to 3 as part of the use-after-free mitigation. An adjacent allocation already uses the tag 3, so the tag is further incremented to 4 (it would be incremented to 5 if one of the adjacent tags was 4):

| 3  | 4  | 16 | 7  | 15 | 16 |

The last slot is randomly chosen for the next alocation, and is assigned the random value 15. However, it's placed next to an allocation with the tag 15 so the tag is incremented and wraps around to 0:

| 3  | 4  | 16 | 7  | 15 | 0 |

API extensions

The void free_sized(void *ptr, size_t expected_size) function exposes the sized deallocation sanity checks for C. A performance-oriented allocator could use the same API as an optimization to avoid a potential cache miss from reading the size from metadata.

The size_t malloc_object_size(void *ptr) function returns an upper bound on the accessible size of the relevant object (if any) by querying the malloc implementation. It's similar to the __builtin_object_size intrinsic used by _FORTIFY_SOURCE but via dynamically querying the malloc implementation rather than determining constant sizes at compile-time. The current implementation is just a naive placeholder returning much looser upper bounds than the intended implementation. It's a valid implementation of the API already, but it will become fully accurate once it's finished. This function is not currently safe to call from signal handlers, but another API will be provided to make that possible with a compile-time configuration option to avoid the necessary overhead if the functionality isn't being used (in a way that doesn't change break API compatibility based on the configuration).

The size_t malloc_object_size_fast(void *ptr) is comparable, but avoids expensive operations like locking or even atomics. It provides significantly less useful results falling back to higher upper bounds, but is very fast. In this implementation, it retrieves an upper bound on the size for small memory allocations based on calculating the size class region. This function is safe to use from signal handlers already.

System calls

This is intended to aid with creating system call whitelists via seccomp-bpf and will change over time.

System calls used by all build configurations:

  • futex(uaddr, FUTEX_WAIT_PRIVATE, val, NULL) (via pthread_mutex_lock)
  • futex(uaddr, FUTEX_WAKE_PRIVATE, val) (via pthread_mutex_unlock)
  • getrandom(buf, buflen, 0) (to seed and regularly reseed the CSPRNG)
  • mmap(NULL, size, PROT_NONE, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0)
  • mmap(ptr, size, PROT_NONE, MAP_ANONYMOUS|MAP_PRIVATE|MAP_FIXED, -1, 0)
  • mprotect(ptr, size, PROT_READ)
  • mprotect(ptr, size, PROT_READ|PROT_WRITE)
  • mremap(old, old_size, new_size, 0)
  • mremap(old, old_size, new_size, MREMAP_MAYMOVE|MREMAP_FIXED, new)
  • munmap
  • write(STDERR_FILENO, buf, len) (before aborting due to memory corruption)

Additional system calls when CONFIG_SEAL_METADATA=true is set:

  • pkey_alloc
  • pkey_mprotect instead of mprotect with an additional pkey parameter, but otherwise the same (regular mprotect is never called)
  • uname (to detect old buggy kernel versions)

Additional system calls for Android builds with LABEL_MEMORY:

  • prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, ptr, size, name)