Awesome
Isolation Alloc
Isolation Alloc (or IsoAlloc) is a secure and fast(ish) memory allocator written in C11. It is a drop in replacement for malloc
on Linux / Mac OS using LD_PRELOAD
or DYLD_INSERT_LIBRARIES
respectively. Its security strategy is originally inspired by Chrome's PartitionAlloc. A memory allocation isolation security strategy is best summed up as maintaining spatial separation, or isolation between objects of different sizes or types. While IsoAlloc wraps malloc
and enforces naive isolation by default very strict isolation of allocations can be achieved using the APIs directly.
IsoAlloc is designed and tested for 64 bit Linux and MacOS. The space afforded by a 64 bit process makes this possible, therefore Isolation Alloc does not support 32 bit targets. The number of bits of entropy provided to mmap
based page allocations is far too low in a 32 bit process to provide much security value. It may work on operating systems other than Linux/MacOS but that is also untested at this time. There is partial FreeBSD support but CI is often flakey. A minimal Solaris/Illumos support is available, LTO not supported by the compilers backends.
Design
At a high level IsoAlloc creates zones which are used to manage regions of memory that hold individual allocations of a specific size. If you are familiar with the implementation of arenas in other heap allocators then the concepts here will be familiar to you.
There is one iso_alloc_root
structure which contains a pointer to a fixed number of iso_alloc_zone
structures. These iso_alloc_zone
structures are referred to as zones. Zones point to user chunks and a bitmap that is used to manage those chunks. The translation between bitmap and user chunks is referred to as bit slots. The pages that back both the user chunks and the bitmap are allocated separately. The pointers that reference these in the zone meta data are masked in between allocation and free operations. The bitmap contains 2 bits of state per user chunk. The current bit value specification is as follows:
00
free chunk10
currently in use01
was used but is now free11
canary chunk
All user chunk pages and bitmap pages are surrounded by guard page allocations with the PROT_NONE
permission. Zones are created for specific sizes, or manually created through the exposed API for a particular size or object type. Internally managed zones will live for the entire lifetime of the process, but zones created via the API can be destroyed at any time.
If DEBUG
, LEAK_DETECTOR
, or MEM_USAGE
are specified during compilation a memory leak and memory usage routine will be called from the destructor which will print useful information about the state of the heap at that time. These can also be invoked via the API, which is documented further below.
- All chunk sizes are a multiple of 32 and are always 8 byte aligned.
- The
iso_alloc_root
structure is thread safe and guarded by a mutex or spinlock whenTHREAD_SUPPORT
is enabled. - Each zone bitmap contains 2 bits per chunk.
- All zones are 4 MB in size regardless of the chunk sizes they manage.
- Default zones are created in the constructor for sizes: 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192 bytes.
- Zones are created on demand for larger allocations or when these default zones are exhausted.
- The free bit slot cache is 255 entries, it helps speed up allocations.
- All allocations larger than
SMALL_SIZE_MAX
live in big zones which have a size limitation of 4 GB. See PERFORMANCE.
There is support for Address Sanitizer, Memory Sanitizer, and Undefined Behavior Sanitizer. If you want to enable it just uncomment the ENABLE_ASAN
, ENABLE_MSAN
, or ENABLE_UBSAN
flags in the Makefile. Like any other usage of Address Sanitizer these are mutually exclusive. IsoAlloc will use Address Sanitizer macros to poison and unpoison user chunks appropriately. IsoAlloc still catches a number of issues Address Sanitizer does not, including double/unaligned/wild free's.
A feature similar to GWP-ASAN can be enabled with ALLOC_SANITY
in the Makefile. It samples calls to iso_alloc/malloc
and allocates a page of memory surrounded by guard pages in order to detect Use-After-Free and linear heap overflows. All sampled sanity allocations are verified with canaries to detect over/underflows into the surrounding bytes of the page. A percentage of sanity allocations are allocated at end of the page to detect linear overflows. This feature works on all supported platforms.
You can also enable UNINIT_READ_SANITY
for detecting uninitialized read vulnerabilities using the userfaultfd
syscall. You can read more about that feature here. This feature is only available on Linux and requires ALLOC_SANITY
and THREAD_SUPPORT
to be enabled.
See the PERFORMANCE documentation for more information.
Thread Safety
IsoAlloc is thread safe by way of protecting the root structure with a global lock built with either a pthread mutex, or a C11 atomic_flag
when USE_SPINLOCK
is enabled. This means every thread that wants to allocate or free a chunk needs to wait until it can take ownership of the lock. This design choice has some tradeoffs. It can negatively impact performance of multi threaded programs that perform a lot of allocations. This is because every thread shares the same set of global zones. The benefit of this is that you can allocate and free any chunk from any thread with no additional complexity required. In order to help alleviate contention on this lock each thread has a zone cache built using thread local storage (TLS). This is implemented as a simple FILO cache of the most recently used zones by that thread. It's size is 8 by default but can be increased modifying the ZONE_CACHE_SZ
define in the internal header file. Making this cache too large can lead to negative performance implications for certain allocation patterns. For example, if a thread allocates multiple 32 byte chunks in a row then the cache may be populated entirely by the same zone that holds 32 byte chunks. Now when the thread goes to allocate a 64 byte chunk it iterates through the entire cache, does not find a usable zone, and then has to take the slow path which iterates through all zones again. This cache is also used when thread support is disabled but it does not live in TLS and is instead allocated on its own set of pages. See the PERFORMANCE documentation for more information on the various caches in use in IsoAlloc.
When enabled, the CPU_PIN
feature will restrict allocations from a given zone to the CPU core that created that zone. Free operations are not restricted in this way. This mode is compatible with and without thread support, but is only available on Linux, and will introduce a negative performance hit to the hot path and may increase memory usage. The benefit of this mode is that it introduces an isolation mechanism based on CPU core with no configuration beyond enabling the CPU_PIN
define in the Makefile.
Security Properties
- Zones cannot overflow or underflow into one another.
- All user pages are surrounded by guard pages including big zones.
- All bitmap pages are surrounded by guard pages.
- Double free's are checked for on every call to
iso_free
. - For zones managing allocations 8192 bytes or smaller around %1 of their chunks are permanent canaries.
- All free'd chunks get a canary written to them and verified upon reallocation.
- The state of all zones can be verified at any anytime using
iso_verify_zones
oriso_verify_zone(zone)
. - Canaries are unique and are composed of a 64 bit secret value xor'd by the address of the chunk itself.
- A reused chunk will always have its canary checked before its returned by
iso_alloc
. - The top byte of user chunk canaries is
0x00
to prevent unbounded C string reads from leaking it. - A chunk can be permanently free'd with a call to
iso_free_permanently
. - If
SANITIZE_CHUNKS
is set all user chunks are cleared when passed toiso_free
with the constant0xde
. - When freeing a chunk the canary in adjacent chunks above/below are verified.
- Some important zone metadata pointers are masked in-between
iso_alloc
andiso_free
operations. - Passing a pointer to
iso_free
that was not allocated withiso_alloc
will abort. - Pointers passed to
iso_free
must be 8 byte aligned, and a multiple of the zone chunk size. - The free bit slot cache provides a chunk quarantine or delayed free mechanism.
- When private zones are destroyed they are overwritten and marked
PROT_NONE
to prevent use-after-free. - Big zone meta data lives at a random offset from its base page.
- A call to
realloc
will always return a new chunk. UsePERM_FREE_REALLOC
to make these free's permanent. - Enable
FUZZ_MODE
in the Makefile to verify all zones upon alloc/free, and never reuse private zones. - When
CPU_PIN
is enabled allocation from a zone will be restricted to the CPU core that created it. - When
UAF_PTR_PAGE
is enabled calls toiso_free
will be sampled to search for dangling references. - Enable
VERIFY_FREE_BIT_SLOTS
to verify there are no duplicates in the bit slot cache upon free. - When
ALLOC_SANITY
is enabled a percentage of allocations will be sampled to detect UAF/overflows, see above. - Randomized hints are passed to
mmap
to ensure contiguous page ranges are not allocated. - When
ABORT_ON_NULL
is enabled IsoAlloc will abort instead of returningNULL
. - By default
NO_ZERO_ALLOCATIONS
will return a pointer to a page markedPROT_NONE
for all0
sized allocations. - When
ABORT_NO_ENTROPY
is enabled IsoAlloc will abort when it can't gather enough entropy. - When
RANDOMIZE_FREELIST
is enabled IsoAlloc will randomize the free list upon creation. May have a perf hit. - Zones are retired and replaced after they've allocated and freed a specific number of chunks. This is calculated as
ZONE_ALLOC_RETIRE * max_chunk_count_for_zone
. - Big zones are retired after being allocated and freed a total of
BIG_ZONE_ALLOC_RETIRE
times. MEMORY_TAGGING
When enabled IsoAlloc will create a 1 byte tag for each chunk in private zones. See the MEMORY_TAGGING.md documentation, or this test for an example of how to use it.MEMCPY_SANITY
andMEMSET_SANITY
Configures the allocator will hook all calls tomemcpy
/memset
and check for out of bounds r/w operations when either src or dst points to a chunk allocated by IsoAllocSTRONG_SIZE_ISOLATION
Enables a policy that enforces stronger memory isolation by sizePROTECT_FREE_BIG_ZONES
Marks big zones on the free list asPROT_NONE
when not in useBIG_ZONE_META_DATA_GUARD
Enables guard pages for big zone meta dataBIG_ZONE_GUARD
Enables guard pages for big zone user pagesARM_MTE
Enables support for the ARM v8.5a Memory Tagging Extension
Building
The Makefile targets are very simple:
make library
- Builds a release version of the library without C++ support
make library_debug
- Builds a debug version of the library
make library_debug_no_output
- Builds a debug version of the library with no logging output
make analyze_library_debug
- Builds the library with clang's scan-build if installed
make tests
- Builds and runs all tests
make tagging_tests
- Builds and runs the memory tagging tests
make libc_sanity_tests
- Builds the memcpy/memset libc hook sanity tests
make perf_tests
- Builds and runs a simple performance test that uses gprof. Linux only
make malloc_cmp_test
- Builds and runs a test that uses both iso_alloc and malloc for comparison
make c_library_objects
- Builds .o files to be linked in another compilation step
make c_library_objects_debug
- Builds debug .o files to be linked in another compilation step
make cpp_library
- Builds the library with a C++ interface that overloads operators new
and delete
make cpp_library_debug
- Builds a debug version of the library with a C++ interface that overloads operators new
and delete
make cpp_tests
- Builds and runs the C++ tests
make format
- Runs clang formatter according to the specification in .clang-format
make clean
- Cleans up the root directory
Android
To build Android libraries for the ARM64 architecture just cd
into the android/jni
directory and run ndk-build
.
For those of you on an M1 based Mac you can still build IsoAlloc with the following command:
arch -x86_64 /bin/bash -c $ANDROID_NDK_HOME/build/ndk-build
or
~/Library/Android/sdk/ndk/26.1.10909125/ndk-build
Linking With C++
If you want to use IsoAlloc with a C++ program you can use the c_library_objects
Makefile target. This will produce .o object files you can pass to your compiler. These targets are used internally to build the library with new
and delete
support.
Debugging
If you try to use Isolation Alloc in an existing program then and you are getting crashes here are some tips to help you get started. If you aren't using LD_PRELOAD
then first make sure you actually replaced all malloc, calloc, realloc
and free
calls to their iso_alloc
equivalents. Don't forget things like strdup
that return a pointer from malloc
.
If you are getting consistent crashes you can build a debug version of the library with make library_debug
and then catch the crash in GDB with a command similar to this gdb -q -command=misc/commands.gdb <your_binary>
.
If all else fails please file an issue on the github project page.
API
void iso_alloc_initialize()
- Initializes the IsoAlloc root. Only needed if AUTO_CTOR_DTOR
is disabled. See Makefile.
void iso_alloc_destroy()
- Destroys the IsoAlloc root. Only needed if AUTO_CTOR_DTOR
is disabled. See Makefile.
void *iso_alloc(size_t size)
- Equivalent to malloc
. Returns a pointer to a chunk of memory that is size bytes in size. To free this chunk just pass it to iso_free
.
void *iso_calloc(size_t nmemb, size_t size)
- Equivalent to calloc
. Allocates a chunk big enough for an array of nmemb elements of size bytes. The array is zeroized.
void *iso_realloc(void *p, size_t size)
- Equivalent to realloc
. Reallocates a new chunk, if necessary, to be size bytes big and copies the contents of p to it.
void *iso_reallocarray(void *p, size_t nmemb, size_t size)
- Equivalent to reallocarray
. In the same principles as iso_realloc
, reallocates a new chunk but for an array of nmemb elements of size bytes and in addition check for possible size overflow.
void iso_free(void *p)
- Frees any chunk allocated and returned by any API call (e.g. iso_alloc, iso_calloc, iso_realloc, iso_strdup, iso_strndup
).
void iso_free_size(void *p, size_t size)
- The same as iso_free
but requires a size argument so a strict size check can be performed
void iso_free_permanently(void *p)
- Same as iso_free
but marks the chunk in such a way that it will not be reallocated
void iso_free_from_zone(void *p, iso_alloc_zone_handle *zone)
- Free's a chunk from a private zone. Can take a tagged or untagged pointer if MEMORY_TAGGING
is enabled. These chunks are not quarantined.
void iso_free_from_zone_permanently(void *p, iso_alloc_zone_handle *zone)
- Permanently free's a chunk from a zone
size_t iso_chunksz(void *p)
- Returns the size of the chunk returned by iso_alloc
char *iso_strdup(const char *str)
- Equivalent to strdup
. Returned pointer must be free'd by iso_free
.
char *iso_strndup(const char *str, size_t n)
- Equivalent to strndup
. Returned pointer must be free'd by iso_free.
iso_alloc_zone_handle *iso_alloc_new_zone(size_t size)
- Allocates a new private zone for allocations up to size bytes. Returns a handle to that zone.
char *iso_strdup_from_zone(iso_alloc_zone_handle *zone, const char *str)
- Equivalent to iso_strdup
except string is duplicated in specified zone.
char *iso_strndup_from_zone(iso_alloc_zone_handle *zone, const char *str, size_t n)
- Equivalent to iso_strndup
except string is duplicated in specified zone.
void *iso_alloc_from_zone(iso_alloc_zone_handle *zone)
- Equivalent to iso_alloc
except allocation is done in specified zone.
void *iso_alloc_from_zone_tagged(iso_alloc_zone_handle *zone)
- Same as iso_alloc_from_zone
but returns a tagged pointer if MEMORY_TAGGING
is enabled.
void iso_alloc_verify_ptr_tag(void *p, iso_alloc_zone_handle *zone)
- Verifies the tag for a pointer is correct, aborts if not. Requires MEMORY_TAGGING
.
void iso_alloc_destroy_zone(iso_alloc_zone_handle *zone)
- Destroy a zone created with iso_alloc_from_zone
.
void *iso_alloc_tag_ptr(void *p, iso_alloc_zone_handle *zone)
- Tags a pointer from a private zone if MEMORY_TAGGING
is enabled.
void *iso_alloc_untag_ptr(void *p, iso_alloc_zone_handle *zone)
- Untags a pointer from a private zone if MEMORY_TAGGING
is enabled.
uint8_t iso_alloc_get_mem_tag(void *p, iso_alloc_zone_handle *zone)
- Retrieves the 1 byte tag for an untagged pointer if MEMORY_TAGGING
is enabled.
void iso_alloc_protect_root()
- Temporarily protects the iso_alloc
root structure by marking it unreadable.
void iso_alloc_unprotect_root()
- Undoes the operation performed by iso_alloc_protect_root
.
uint64_t iso_alloc_detect_leaks()
- Returns the total number of leaks detected for all zones. Will print debug logs when compiled with -DDEBUG
uint64_t iso_alloc_detect_zone_leaks(iso_alloc_zone_handle *zone)
- Returns the total number of leaks detected for specified zone. Will print debug logs when compiled with -DDEBUG
uint64_t iso_alloc_mem_usage()
- Returns the total memory usage for all zones. Will print debug logs when compiled with -DDEBUG
uint64_t iso_alloc_zone_mem_usage(iso_alloc_zone_handle *zone)
- Returns the total memory usage for a specified zone. Will print debug logs when compiled with -DDEBUG
void iso_verify_zones()
- Verifies the state of all zones. Will abort if inconsistencies are found.
void iso_verify_zone(iso_alloc_zone_handle *zone)
- Verifies the state of specified zone. Will abort if inconsistencies are found.
int32_t iso_alloc_name_zone(iso_alloc_zone_handle *zone, char *name)
- Allows naming of private zones via prctl on Android.
void iso_flush_caches()
- Flushes all thread specific caches. Intended to be used upon thread destruction.
size_t iso_zone_chunk_count(iso_alloc_zone_handle *zone)
- Returns the total number of chunks a private zone can hold not including canary chunks. If canaries are disabled this number is absolute, otherwise it is a safe lower bound and actual number may be higher due to canary creation random seed.
Experimental APIs
These APIs are exposed via the public header iso_alloc.h
but are subject to backward breaking changes at any time.
int32_t iso_get_alloc_traces(iso_alloc_traces_t *traces_out)
- Retrieves the current global iso_alloc_traces_t
structure from the allocator
int32_t iso_get_free_traces(iso_free_traces_t *traces_out)
- Retrieves the current global iso_free_traces_t
structure from the allocator
void iso_alloc_search_stack(void *p)
- Searches from p
until the current stack frame in iso_alloc_search_stack
for any pointers into IsoAlloc user pages. Any pointers found are logged to stdout. If p
is NULL
then the entire stack is searched.
Data Structures
typedef struct {
/* The address of the last ALLOC_BTS_DEPTH (8) callers as referenced by stack frames */
uint64_t callers[ALLOC_BTS_DEPTH];
/* The smallest allocation size requested by this call path */
size_t lower_bound_size;
/* The largest allocation size requested by this call path */
size_t upper_bound_size;
/* A 16 bit hash of the back trace */
uint16_t backtrace_hash;
} iso_alloc_traces_t;
typedef struct {
/* The address of the last 8 callers as referenced by stack frames */
uint64_t callers[BACKTRACE_DEPTH];
/* A 16 bit hash of the back trace */
uint16_t backtrace_hash;
/* Call count */
size_t call_count;
} iso_free_traces_t;
When HEAP_PROFILER
is enabled these structure will contain information collected by the allocator by sampling malloc
and free
calls. This data structure is experimental and is subject to change. See interfaces_test.c
file for an example of how to retrieve and inspect these structures. Note: On MacOS you may need to use gcc/g++ to compile this functionality without error.