ISO C++ allocator The mt allocator [hereinafter referred to simply as "the allocator"] is a fixed size (power of two) allocator that was initially developed specifically to suit the needs of multi threaded applications [hereinafter referred to as an MT application]. Over time the allocator has evolved and been improved in many ways, in particular it now also does a good job in single threaded applications [hereinafter referred to as a ST application]. (Note: In this document, when referring to single threaded applications this also includes applications that are compiled with gcc without thread support enabled. This is accomplished using ifdef's on __GTHREADS). This allocator is tunable, very flexible, and capable of high-performance. The aim of this document is to describe - from an application point of view - the "inner workings" of the allocator. There are three general components to the allocator: a datum describing the characteristics of the memory pool, a policy class containing this pool that links instantiation types to common or individual pools, and a class inheriting from the policy class that is the actual allocator. The datum describing pools characteristics is template<bool _Thread> class __pool This class is parametrized on thread support, and is explicitly specialized for both multiple threads (with bool==true) and single threads (via bool==false.) It is possible to use a custom pool datum instead of the default class that is provided. There are two distinct policy classes, each of which can be used with either type of underlying pool datum. template<bool _Thread> struct __common_pool_policy template<typename _Tp, bool _Thread> struct __per_type_pool_policy The first policy, __common_pool_policy, implements a common pool. This means that allocators that are instantiated with different types, say char and long will both use the same pool. This is the default policy. The second policy, __per_type_pool_policy, implements a separate pool for each instantiating type. Thus, char and long will use separate pools. This allows per-type tuning, for instance. Putting this all together, the actual allocator class is template<typename _Tp, typename _Poolp = __default_policy> class __mt_alloc : public __mt_alloc_base<_Tp>, _Poolp This class has the interface required for standard library allocator classes, namely member functions allocate and deallocate, plus others. Certain allocation parameters can be modified, or tuned. There exists a nested struct __pool_base::_Tune that contains all these parameters, which include settings for Alignment Maximum bytes before calling ::operator new directly Minimum bytes Size of underlying global allocations Maximum number of supported threads Migration of deallocations to the global free list Shunt for global new and delete Adjusting parameters for a given instance of an allocator can only happen before any allocations take place, when the allocator itself is initialized. For instance: #include <ext/mt_allocator.h> struct pod { int i; int j; }; int main() { typedef pod value_type; typedef __gnu_cxx::__mt_alloc<value_type> allocator_type; typedef __gnu_cxx::__pool_base::_Tune tune_type; tune_type t_default; tune_type t_opt(16, 5120, 32, 5120, 20, 10, false); tune_type t_single(16, 5120, 32, 5120, 1, 10, false); tune_type t; t = allocator_type::_M_get_options(); allocator_type::_M_set_options(t_opt); t = allocator_type::_M_get_options(); allocator_type a; allocator_type::pointer p1 = a.allocate(128); allocator_type::pointer p2 = a.allocate(5128); a.deallocate(p1, 128); a.deallocate(p2, 5128); return 0; } The static variables (pointers to freelists, tuning parameters etc) are initialized as above, or are set to the global defaults. The very first allocate() call will always call the _S_initialize_once() function. In order to make sure that this function is called exactly once we make use of a __gthread_once call in MT applications and check a static bool (_S_init) in ST applications. The _S_initialize() function: - If the GLIBCXX_FORCE_NEW environment variable is set, it sets the bool _S_force_new to true and then returns. This will cause subsequent calls to allocate() to return memory directly from a new() call, and deallocate will only do a delete() call. - If the GLIBCXX_FORCE_NEW environment variable is not set, both ST and MT applications will: - Calculate the number of bins needed. A bin is a specific power of two size of bytes. I.e., by default the allocator will deal with requests of up to 128 bytes (or whatever the value of _S_max_bytes is when _S_init() is called). This means that there will be bins of the following sizes (in bytes): 1, 2, 4, 8, 16, 32, 64, 128. - Create the _S_binmap array. All requests are rounded up to the next "large enough" bin. I.e., a request for 29 bytes will cause a block from the "32 byte bin" to be returned to the application. The purpose of _S_binmap is to speed up the process of finding out which bin to use. I.e., the value of _S_binmap[ 29 ] is initialized to 5 (bin 5 = 32 bytes). - Create the _S_bin array. This array consists of bin_records. There will be as many bin_records in this array as the number of bins that we calculated earlier. I.e., if _S_max_bytes = 128 there will be 8 entries. Each bin_record is then initialized: - bin_record->first = An array of pointers to block_records. There will be as many block_records pointers as there are maximum number of threads (in a ST application there is only 1 thread, in a MT application there are _S_max_threads). This holds the pointer to the first free block for each thread in this bin. I.e., if we would like to know where the first free block of size 32 for thread number 3 is we would look this up by: _S_bin[ 5 ].first[ 3 ] The above created block_record pointers members are now initialized to their initial values. I.e. _S_bin[ n ].first[ n ] = NULL; - Additionally a MT application will: - Create a list of free thread id's. The pointer to the first entry is stored in _S_thread_freelist_first. The reason for this approach is that the __gthread_self() call will not return a value that corresponds to the maximum number of threads allowed but rather a process id number or something else. So what we do is that we create a list of thread_records. This list is _S_max_threads long and each entry holds a size_t thread_id which is initialized to 1, 2, 3, 4, 5 and so on up to _S_max_threads. Each time a thread calls allocate() or deallocate() we call _S_get_thread_id() which looks at the value of _S_thread_key which is a thread local storage pointer. If this is NULL we know that this is a newly created thread and we pop the first entry from this list and saves the pointer to this record in the _S_thread_key variable. The next time we will get the pointer to the thread_record back and we use the thread_record->thread_id as identification. I.e., the first thread that calls allocate will get the first record in this list and thus be thread number 1 and will then find the pointer to its first free 32 byte block in _S_bin[ 5 ].first[ 1 ] When we create the _S_thread_key we also define a destructor (_S_thread_key_destr) which means that when the thread dies, this thread_record is returned to the front of this list and the thread id can then be reused if a new thread is created. This list is protected by a mutex (_S_thread_freelist_mutex) which is only locked when records are removed or added to the list. - Initialize the free and used counters of each bin_record: - bin_record->free = An array of size_t. This keeps track of the number of blocks on a specific thread's freelist in each bin. I.e., if a thread has 12 32-byte blocks on it's freelists and allocates one of these, this counter would be decreased to 11. - bin_record->used = An array of size_t. This keeps track of the number of blocks currently in use of this size by this thread. I.e., if a thread has made 678 requests (and no deallocations...) of 32-byte blocks this counter will read 678. The above created arrays are now initialized with their initial values. I.e. _S_bin[ n ].free[ n ] = 0; - Initialize the mutex of each bin_record: The bin_record->mutex is used to protect the global freelist. This concept of a global freelist is explained in more detail in the section "A multi threaded example", but basically this mutex is locked whenever a block of memory is retrieved or returned to the global freelist for this specific bin. This only occurs when a number of blocks are grabbed from the global list to a thread specific list or when a thread decides to return some blocks to the global freelist. Notes about deallocation. This allocator does not explicitly release memory. Because of this, memory debugging programs like valgrind or purify may notice leaks: sorry about this inconvenience. Operating systems will reclaim allocated memory at program termination anyway. If sidestepping this kind of noise is desired, there are three options: use an allocator, like new_allocator that releases memory while debugging, use GLIBCXX_FORCE_NEW to bypass the allocator's internal pools, or use a custom pool datum that releases resources on destruction. On systems with the function __cxa_atexit, the allocator can be forced to free all memory allocated before program termination with the member function __pool_type::_M_destroy. However, because this member function relies on the precise and exactly-conforming ordering of static destructors, including those of a static local __pool object, it should not be used, ever, on systems that don't have the necessary underlying support. In addition, in practice, forcing deallocation can be tricky, as it requires the __pool object to be fully-constructed before the object that uses it is fully constructed. For most (but not all) STL containers, this works, as an instance of the allocator is constructed as part of a container's constructor. However, this assumption is implementation-specific, and subject to change. For an example of a pool that frees memory, see the following example. Let's start by describing how the data on a freelist is laid out in memory. This is the first two blocks in freelist for thread id 3 in bin 3 (8 bytes): +----------------+ | next* ---------|--+ (_S_bin[ 3 ].first[ 3 ] points here) | | | | | | | | | +----------------+ | | thread_id = 3 | | | | | | | | | | | +----------------+ | | DATA | | (A pointer to here is what is returned to the | | | the application when needed) | | | | | | | | | | | | | | | | | | +----------------+ | +----------------+ | | next* |<-+ (If next == NULL it's the last one on the list) | | | | | | +----------------+ | thread_id = 3 | | | | | | | +----------------+ | DATA | | | | | | | | | | | | | | | +----------------+ With this in mind we simplify things a bit for a while and say that there is only one thread (a ST application). In this case all operations are made to what is referred to as the global pool - thread id 0 (No thread may be assigned this id since they span from 1 to _S_max_threads in a MT application). When the application requests memory (calling allocate()) we first look at the requested size and if this is > _S_max_bytes we call new() directly and return. If the requested size is within limits we start by finding out from which bin we should serve this request by looking in _S_binmap. A quick look at _S_bin[ bin ].first[ 0 ] tells us if there are any blocks of this size on the freelist (0). If this is not NULL - fine, just remove the block that _S_bin[ bin ].first[ 0 ] points to from the list, update _S_bin[ bin ].first[ 0 ] and return a pointer to that blocks data. If the freelist is empty (the pointer is NULL) we must get memory from the system and build us a freelist within this memory. All requests for new memory is made in chunks of _S_chunk_size. Knowing the size of a block_record and the bytes that this bin stores we then calculate how many blocks we can create within this chunk, build the list, remove the first block, update the pointer (_S_bin[ bin ].first[ 0 ]) and return a pointer to that blocks data. Deallocation is equally simple; the pointer is casted back to a block_record pointer, lookup which bin to use based on the size, add the block to the front of the global freelist and update the pointer as needed (_S_bin[ bin ].first[ 0 ]). The decision to add deallocated blocks to the front of the freelist was made after a set of performance measurements that showed that this is roughly 10% faster than maintaining a set of "last pointers" as well. In the ST example we never used the thread_id variable present in each block. Let's start by explaining the purpose of this in a MT application. The concept of "ownership" was introduced since many MT applications allocate and deallocate memory to shared containers from different threads (such as a cache shared amongst all threads). This introduces a problem if the allocator only returns memory to the current threads freelist (I.e., there might be one thread doing all the allocation and thus obtaining ever more memory from the system and another thread that is getting a longer and longer freelist - this will in the end consume all available memory). Each time a block is moved from the global list (where ownership is irrelevant), to a threads freelist (or when a new freelist is built from a chunk directly onto a threads freelist or when a deallocation occurs on a block which was not allocated by the same thread id as the one doing the deallocation) the thread id is set to the current one. What's the use? Well, when a deallocation occurs we can now look at the thread id and find out if it was allocated by another thread id and decrease the used counter of that thread instead, thus keeping the free and used counters correct. And keeping the free and used counters corrects is very important since the relationship between these two variables decides if memory should be returned to the global pool or not when a deallocation occurs. When the application requests memory (calling allocate()) we first look at the requested size and if this is >_S_max_bytes we call new() directly and return. If the requested size is within limits we start by finding out from which bin we should serve this request by looking in _S_binmap. A call to _S_get_thread_id() returns the thread id for the calling thread (and if no value has been set in _S_thread_key, a new id is assigned and returned). A quick look at _S_bin[ bin ].first[ thread_id ] tells us if there are any blocks of this size on the current threads freelist. If this is not NULL - fine, just remove the block that _S_bin[ bin ].first[ thread_id ] points to from the list, update _S_bin[ bin ].first[ thread_id ], update the free and used counters and return a pointer to that blocks data. If the freelist is empty (the pointer is NULL) we start by looking at the global freelist (0). If there are blocks available on the global freelist we lock this bins mutex and move up to block_count (the number of blocks of this bins size that will fit into a _S_chunk_size) or until end of list - whatever comes first - to the current threads freelist and at the same time change the thread_id ownership and update the counters and pointers. When the bins mutex has been unlocked, we remove the block that _S_bin[ bin ].first[ thread_id ] points to from the list, update _S_bin[ bin ].first[ thread_id ], update the free and used counters, and return a pointer to that blocks data. The reason that the number of blocks moved to the current threads freelist is limited to block_count is to minimize the chance that a subsequent deallocate() call will return the excess blocks to the global freelist (based on the _S_freelist_headroom calculation, see below). However if there isn't any memory on the global pool we need to get memory from the system - this is done in exactly the same way as in a single threaded application with one major difference; the list built in the newly allocated memory (of _S_chunk_size size) is added to the current threads freelist instead of to the global. The basic process of a deallocation call is simple: always add the block to the front of the current threads freelist and update the counters and pointers (as described earlier with the specific check of ownership that causes the used counter of the thread that originally allocated the block to be decreased instead of the current threads counter). And here comes the free and used counters to service. Each time a deallocation() call is made, the length of the current threads freelist is compared to the amount memory in use by this thread. Let's go back to the example of an application that has one thread that does all the allocations and one that deallocates. Both these threads use say 516 32-byte blocks that was allocated during thread creation for example. Their used counters will both say 516 at this point. The allocation thread now grabs 1000 32-byte blocks and puts them in a shared container. The used counter for this thread is now 1516. The deallocation thread now deallocates 500 of these blocks. For each deallocation made the used counter of the allocating thread is decreased and the freelist of the deallocation thread gets longer and longer. But the calculation made in deallocate() will limit the length of the freelist in the deallocation thread to _S_freelist_headroom % of it's used counter. In this case, when the freelist (given that the _S_freelist_headroom is at it's default value of 10%) exceeds 52 (516/10) blocks will be returned to the global pool where the allocating thread may pick them up and reuse them. In order to reduce lock contention (since this requires this bins mutex to be locked) this operation is also made in chunks of blocks (just like when chunks of blocks are moved from the global freelist to a threads freelist mentioned above). The "formula" used can probably be improved to further reduce the risk of blocks being "bounced back and forth" between freelists.