struct / isoalloc

A general purpose memory allocator that implements an isolation security strategy to mitigate memory safety issues while maintaining good performance
https://secure.dev
Apache License 2.0
363 stars 25 forks source link
exploit exploitation heap memory-allocator memory-management secure security

unit tests

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:

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.

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

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.

Memory Isolation Visualized

Design of IsoAlloc schema