On the iPhone platform, memory is a very precious resources. If they are not handled properly, the application will receive one or two memory warning, and then your application will be killed by the OS. So I decided to write my own memory allocator to preallocate a large chunk of memory so that my game will not be killed by the OS when running, it will either start or not start. This is my first time to write a memory allocator, and it is not as sophisticated as
"Ready, Set, Allocate!", but it just works fine enough for me.
In my little engine, a pool allocator is written for memory allocation, with different pre-defined pool size ranging from 8, 16, 32, 64bytes to 1048546bytes. As my target platform is iPhone, the maximum pool size is 1048546bytes which is used only for a few high resolution textures, most of the memory is spent on the the smaller pool size. During the program starts, a large chunk of memory is created and it is divided into different smaller chunks for different pool size as follows:
Notice that the large byte chuck is located in the smaller memory address for proper byte alignment. And within each chunk, it is divided into equally sized block for each particular size allocation:
The memory blocks within each chunk are maintained as a linked list so that when the pool memory allocator need to allocate/deallocate memory, it just need to return a free memory block from the list/add it back to the linked list. The memory within each memory blocks is used to store the 'next pointer' for the next free memory block in the linked list so that we do not need to allocate extra memory to keep track of the linked list (this approach is learnt from
Game Engine Architecture):
For each allocation, the allocator need to decided which pool chunk need to be used depends on the size of the allocation. Then within that chunk, a free memory block is returned which is just the head of the linked list within that chunk. For each deallocation, as we divide the memory into different chunk, we know the boundary address of each pool chunk, so by checking the deallocated pointer address, we can determine which pool chunk it belongs to, then we can just add the memory back to that chunk's free memory block linked list. One drawback of this approach is it does not verify whether the deallocated pointer is actually allocated by the user nor it is double freed. To 'partly overcome' this problem, I added limited check to verify the input to each deallocation. First, I would check whether the input is byte aligned, for example if the deallocated pointer is within the 1048546 bytes chunk, then the pointer address must be 1048546 byte aligned. Second, as we partition the memory chunk, we know how many memory blocks is within each memory chunk, we can maintain a current free blocks number which will increase and decrease for each allocation and deallocation. When the program exits and this free block number does not match with the total number of blocks, then memory may either be leaked or double freed. But this only solve the problem partially.
To actually solve the problem and also check for memory leaks. I need to log every allocation and deallocation. Originally, for each allocation, I just store the returned pointer address and location of each allocation(which source file and line number) using the macro
__FILE__, __LINE__. But this does not do well enough to track down all memory leak as some files are templated such as btAlignedAllocator.h in the bullet physics library(yes, I use bullet physics in my engine). Using the macro __FILE__, __LINE__ will only log down the allocation in these header file which does not help much for tracking memory leak. Therefore, I also log the callstack of each allocation using the system call
backtrace() and backtrace_symbols()(which is available in Unix-like platform). Then I can track down all memory leak easily. However, logging every allocation is a very slow process and it is only enabled in debug build/ enabled when necessary.
In conclusion, my memory allocator still have different things to improve such as verifying the user deallocation; adding some meta-data within each memory block such as the allocation size; And for the thread safety, currently I only use a mutex to protected the memory, I may switch to a lock-free version in the future. Despite these short comings, this allocator works well enough for me as it avoid receiving memory warning from the iOS, avoiding memory fragmentation and help me track down memory leaks.
Reference:
[1]
http://g.oswego.edu/dl/html/malloc.html
[2] Ready, Set, Allocate!:
http://altdevblogaday.com/2011/04/11/ready-set-allocate-part-1/
[3] Game Engine Architecture:
http://www.gameenginebook.com/
[4]
http://gcc.gnu.org/onlinedocs/cpp/Standard-Predefined-Macros.html
[5]
http://developer.apple.com/library/mac/#documentation/Darwin/Reference/ManPages/man3/backtrace_symbols.3.html
