Reminder

Work on and test your code on GitHub Codespaces, the Docker container, your local Linux machine, or login.khoury.northeastern.edu.

For this assignment, we will write a safer and more efficient version of the memory allocator from Assignment 5. Our goal is to get performance closer to that of the actual malloc, while making it safe for use by multiple threads.

Task 1: Use mmap to Allocate Page-sized Blocks

Understanding mmap

For the basic allocator implemented in Assignment 5, you were asked to use the sbrk system call to change the size of the heap. This call is not the only way to request memory for our process. In fact, the sbrk syscall is considered deprecated and its use is discouraged. We have used it in our first allocator because it provides a simple and easy-to-use interface for requesting heap memory from the OS.

So what does a modern Unix/Linux want us to use for memory allocation? The modern way to allocate memory is to directly request pages for a process using the mmap system call. In general, mmap allows us to map a file or a device into memory, meaning that any reads/writes within the memory region returned from a successful mmap call are reflected in the file (more on this in the file system project). In other words, mmap allows us to allocate memory from the operating system, backed by a particular file. On Linux and some other operating systems, we can even request “anonymous” mappings, meaning that the returned region is not backed by any file. In this mode, mmap can behave like malloc.

An “mmapped” region’s size is always aligned along the boundary of a page. This means that one constraint is that allocations with mmap should be made in multiples the size of a page. Even if you specify a size that is not a multiple of the system’s page size, mmap will round up to the nearest page. For example, if our page size is 4KB (4096 bytes), and we only use 12 bytes in total during our program’s execution, we have quite a bit of waste! In practice, for large desktop applications this is not a major issue.

Task

The first task is to update our allocator implementation to use mmap instead of sbrk. Remember that, unlike sbrk, mmap will return a pointer to the beginning of a page of memory. Here is the overall updated strategy for mymalloc and myfree:

mymalloc

1. For requests of size ≤ PAGE_SIZE - sizeof(block_t) (aka “small block”):
   • Use your free list to see if there is a large enough block to accommodate the request.
   • If a block is found, select it and remove it from the list (marking it as “not free”)
   • If a large enough block is not found, use mmap to request a new page and set it up as a block.
   • If the block is larger than the requested memory size, determine whether it is feasible to split it: after “shrinking” the original block, is the “leftover” big enough include a block header and at least one byte of user memory.
   • If splitting, set up the leftover as a new memory block and add it to your free list.
   • Return the selected block’s user memory pointer.
2. For requests of size > PAGE_SIZE - sizeof(block_t) (aka “large block”):
   • Compute the number of pages needed to satisfy the request (including a block header) and allocate that many pages with mmap.
   • Set up the block header and set the size to the no. of pages × PAGE_SIZE.
   • Return the block’s user memory pointer.

myfree

1. For small blocks, i.e., blocks of memory of size ≤ PAGE_SIZE - sizeof(block_t): add the block to your free list and coalesce (see below) if needed.
2. For large blocks, i.e., blocks of size > PAGE_SIZE - sizeof(block_t): use munmap to unmap it.

Coalescing Free Blocks

When adding a block into your free list, keep the list sorted by the memory address of the blocks. This will allow coalescing: whenever two blocks in the free list form a continuous area of memory, they should be merged into one block (coalesced)

Since you insert into the free list and need to handle this in two different places, a helper function is a good idea.

Tracing Operations

In addition to the original debug messages, use debug_printf to log the following messages from mymalloc, mycalloc, and myfree.

Function	small / large	Condition / Event	Printout
mymalloc	small	exact block is found in free list	"malloc: block of size %zu found\n"
mymalloc	small	no block found	"malloc: block of size %zu not found - calling mmap\n"
mymalloc	small	splitting block	"malloc: splitting - blocks of size %zu and %zu created\n"
mymalloc	large	mmap request	"malloc: large block - mmap region of size %zu\n"
myfree	small	coalescing	"free: coalesce blocks of size %zu and %zu to new block of size %zu\n"
myfree	large	unmaping	"free: munmap region of size %zu"

For example, let’s say that we are servicing the first malloc request after the program was started and the request size is 1024 bytes, header size is 16 bytes, the system page size is 4096 bytes. Then the following log trace should be visible:

DEBUG: Malloc 1024 bytes
DEBUG: malloc: block of size 1024 not found - calling mmap
DEBUG: malloc: splitting - blocks of size 1024 and 3040 created
DEBUG: malloc: block of size 1024 found

If free was called immediately after, we get the following messages:

DEBUG: free: coalesce blocks of size 1024 and 3040 to new block of size 4080
DEBUG: Freed 1024 bytes

Here’s a large block example with malloc followed by a free of the same block. Request is for 12000 bytes, header is 16 bytes, page size is 4096:

DEBUG: Malloc 12000 bytes
DEBUG: malloc: large block - mmap region of size 12288
...
DEBUG: free: munmap region of size 12288
DEBUG: Freed 12288 bytes

Task 2: A Thread-safe Allocator

Data races can affect memory allocators too. In a multi-threaded environment, we cannot simply make requests to our malloc and free functions based on our previous implementation. We could have a scenario where two or more threads request memory at the same time, and potentially all allocate the same block of memory in the heap. This would certainly be unlucky!

Luckily, we have mutexes to enforce mutual exclusion and help protect against data races. Remember, when we use pthread_mutex_lock and pthread_mutex_unlock, this creates a critical section where only one thread that has acquired the lock can execute a section of code, thus enforcing sequential execution over shared data.

Task

Implement locking mechanisms such that, whenever there is an allocation (malloc or calloc) or deallocation (free), a lock protects that section of code from being run by another thread.

Assignment Strategy

1. You have a previous implementation of an allocator. Structurally many pieces will look the same, so you do not need to build things from scratch.
2. You do not have to work on this assignment in the order of the tasks. Task 1 and Task 2 are independent of each other. Start with Task 2 if you have no idea where to begin with Task 1.
3. One way to approach the assignment:
   1. Start single threaded, with your code from Assignment 5
   2. Initially, ignore splitting and coalescing blocks, and just accept wasted memory when using mmap
   3. Move to multiple threads, adding locks around all allocations and frees
   4. Add splitting of blocks
   5. Add coalescing of blocks

Deliverables

All Tasks

Implement your memory allocator in mymalloc.c and include any additional .c and .h files your implementation relies on. For example, you might want to compile your helper data structure separately.

Commit the code to your repository. Do not include any executables, .o files, or other binary, temporary, or hidden files.

Once you are done, remember to submit your solution to Gradescope and do not forget to include your partner.

Hints & Tips

• Check out the man pages of mmap, munmap, malloc, calloc, free, realloc, …

• mmap arguments. Request multiples of PAGE_SIZE. You’ll want an anonymous private mapping, that is both readable and writable. The flags you’re looking for are MAP_PRIVATE and MAP_ANONYMOUS, and the protection PROT_READ and PROT_WRITE. For anonymous mappings, use -1 as the file descriptor. Offset should be 0.

• Compile and test often.

• Use assert to check that your assumptions about state are valid.

• Write your own (unit) tests. Doing so will save you time in the long run, especially in conjunction with the debugger. In office hours, the instructors or the TAs may ask you to show how you tested code that fails.

• Unit tests can be just a bunch of functions and a main, with asserts to check expected results. Use our tests for queue/vector from Assignment 4 as an example.

• Follow good coding practices. Make sure your function prototypes (signatures) are correct and always provide purpose statements. Add comments where appropriate to document your thinking, although strive to write self-documenting code.

• Split code into short functions. Avoid producing “spaghetti code”. A multi-branch if-else if-else or a multi-case switch should be the only reason to go beyond 40-50 lines per function. Even so, the body of each branch/case should be at most 3-5 lines long.

• It is ok to use global variables, but be judicious with their use. Give them meaningful names and explain their purpose.

• If multiple threads are reading and writing the same variable, consider using pthread_mutex_lock and pthread_mutex_unlock to ensure consistency.

• Have one global lock for initializing your free list.

• Be conservative with your locks. That is, don’t try to optimize too much by moving stuff inside and outside of locked sections–take the conservative approach first, then optimize later.

• You probably will need locks for your myfree function and mycalloc as well.

• Make sure when you memset (if you are using memset in calloc) that you are not memset’ing over your block header.

• Write additional tests to examine high volumes of mallocs and frees, mallocs of a wide range of sizes to exercise the two block size strategies, in particular the edge cases.

• If your program hangs up at some point and does not continue running, it is likely a deadlock. Check the source that every time you lock, you unlock too.

• Use sysconf(_SC_PAGE_SIZE) to get the OS’s page size. Check the manpage for sysconf to see which header you need to include.

• On Linux, anonymously mapped pages are guaranteed to be initialized to 0, meaning a freshly mmapped page does not need to be memset to initialize it.

• The first five tests from Assignment 5 might be still useful, however, the assumptions about sbrk calls are no longer valid. Because we are releasing large blocks of memory to the OS and coalescing smaller blocks, tests 6 and 7 might be prohibitively slow.

F.A.Q.

Q: If you need to split a block but the amount of remaining memory in the block is less than the amount of memory of a new block header (which we need to split the remaining memory into a new block), what should we do?
A: In this case you do not need to do anything, that is an acceptable amount of fragmentation to live with.

Going Further and Additional Resources

• A nice survey of dynamic memory allocation strategies: http://www.cs.northwestern.edu/~pdinda/ics-s05/doc/dsa.pdf
• Investigate the ‘buddy allocator’ and ‘arena allocators’
• John Lakos Memory Allocator talks
  • part 1: https://www.youtube.com/watch?v=nZNd5FjSquk
  • part 2: https://www.youtube.com/watch?v=CFzuFNSpycI
• Some high performance memory allocators are
  • tcmalloc
  • jemalloc
• Pool Allocators: https://blog.molecular-matters.com/2012/09/17/memory-allocation-strategies-a-pool-allocator/
• Our allocator works on the ‘heap’ working with dynamic memory. Other memory may work on the ‘stack’ which is for really hot code. Read up on alloca (in alloca.h) and understand that we could also write a custom stack allocator.