| .. | ||
| README.md | ||
Homework 3 Dynamic Memory Allocator - CSE 320 - Spring 2022
Professor Eugene Stark
Due Date: Friday 3/25/2022 @ 11:59pm
We HIGHLY suggest that you read this entire document, the book chapter, and examine the base code prior to beginning. If you do not read the entire document before beginning, you may find yourself doing extra work.
😱 Start early so that you have an adequate amount of time to test your program!
😱 The functions
malloc,free,realloc,memalign,calloc, etc., are NOT ALLOWED in your implementation. If any of these functions, or any other function with similar functionality is found in your program, you will receive a ZERO.
NOTE: In this document, we refer to a word as 2 bytes (16 bits) and a memory row as 4 words (64 bits). We consider a page of memory to be 1024 bytes (1 KB)
Introduction
You must read Chapter 9.9 Dynamic Memory Allocation Page 839 before starting this assignment. This chapter contains all the theoretical information needed to complete this assignment. Since the textbook has sufficient information about the different design strategies and implementation details of an allocator, this document will not cover this information. Instead, it will refer you to the necessary sections and pages in the textbook.
Takeaways
After completing this assignment, you will have a better understanding of:
- The inner workings of a dynamic memory allocator
- Memory padding and alignment
- Structs and linked lists in C
- errno numbers in C
- Unit testing in C
Overview
You will create an allocator for the x86-64 architecture with the following features:
- Free lists segregated by size class, using first-fit policy within each size class, augmented with a set of "quick lists" holding small blocks segregated by size.
- Immediate coalescing of large blocks on free with adjacent free blocks; delayed coalescing on free of small blocks.
- Boundary tags to support efficient coalescing, with footer optimization that allows footers to be omitted from allocated blocks.
- Block splitting without creating splinters.
- Allocated blocks aligned to "double memory row" (16-byte) boundaries.
- Free lists maintained using last in first out (LIFO) discipline.
- Obfuscation of block headers and footers to detect heap corruption and attempts to free blocks not previously obtained via allocation.
You will implement your own versions of the malloc, realloc, and free functions.
You will use existing Criterion unit tests and write your own to help debug your implementation.
Free List Management Policy
Your allocator MUST use the following scheme to manage free blocks:
Free blocks will be stored in a fixed array of NUM_FREE_LISTS free lists,
segregated by size class (see Chapter 9.9.14 Page 863 for a discussion
of segregated free lists).
Each individual free list will be organized as a circular, doubly linked list
(more information below).
The size classes are based on a power-of-two geometric sequence (1, 2, 4, 8, 16, ...),
according to the following scheme:
The first free list (at index 0) holds blocks of the minimum size M
(where M = 32 for this assignment).
The second list (at index 1) holds blocks of size (M, 2M].
The third list (at index 2) holds blocks of size (2M, 4M].
The fourth list holds blocks whose size is in the interval (4M, 8M].
The fifth list holds blocks whose size is in the interval (8M, 16M],
and so on. This pattern continues up to the interval (128M, 256M],
and then the last list (at index NUM_FREE_LISTS-1; i.e. 9)
holds blocks of size greater than 256M.
Allocation requests will be satisfied by searching the free lists in increasing
order of size class.
Block Placement Policy
When allocating memory, use a segregated fits policy, modified by the use of quick lists as follows. When an allocation request is received, the quick list containing blocks of the appropriate size is first checked to try to quickly obtain a block of exactly the right size. If there is no quick list of that size (quick lists are only maintained for a fixed set of the smallest block sizes), or if there is a quick list but it is empty, then the request will be satisfied from the main free lists.
Satisfying a request from the main free lists is accomplished as follows: First, the smallest size class that is sufficiently large to satisfy the request is determined. The free lists are then searched, starting from the list for the determined size class and continuing in increasing order of size, until a nonempty list is found. The request is then satisfied by the first block in that list that is sufficiently large; i.e. a first-fit policy (discussed in Chapter 9.9.7 Page 849) is applied within each individual free list.
If there is no exact match for an allocation request in the quick lists, and there
is no block in the main free lists that is large enough to satisfy the allocation request,
sf_mem_grow should be called to extend the heap by an additional page of memory.
After coalescing this page with any free block that immediately precedes it, you should
attempt to use the resulting block of memory to satisfy the allocation request;
splitting it if it is too large and no "splinter" (i.e. a remainder smaller than the
minimum block size) would result. If the block of memory is still not large enough,
another call to sf_mem_grow should be made; continuing to grow the heap until either
a large enough block is obtained or the return value from sf_mem_grow indicates that
there is no more memory.
As discussed in the book, segregated free lists allow the allocator to approximate a best-fit policy, with lower overhead than would be the case if an exact best-fit policy were implemented. The rationale for the use of quick lists is that when a small block is freed, it is likely that there will soon be another allocation request for a block of that same size. By putting the block in a quick list, it can be re-used for such a request without the overhead of coalescing and/or splitting that would be required if the block were inserted back into the main pool.
Splitting Blocks & Splinters
Your allocator must split blocks at allocation time to reduce the amount of internal fragmentation. Details about this feature can be found in Chapter 9.9.8 Page 849. Due to alignment and overhead constraints, there will be a minimum useful block size that the allocator can support. For this assignment, pointers returned by the allocator in response to allocation requests are required to be aligned to 16-byte boundaries; i.e. the pointers returned will be addresses that are multiples of 2^4. The 16-byte alignment requirement implies that the minimum block size for your allocator will be 32 bytes. No "splinters" of smaller size than this are ever to be created. If splitting a block to be allocated would result in a splinter, then the block should not be split; rather, the block should be used as-is to satisfy the allocation request (i.e., you will "over-allocate" by issuing a block slightly larger than that required).
🤔 How do the alignment and overhead requirements constrain the minimum block size? As you read more details about the format of a block header, block footer, and alignment requirements, you should try to answer this question.
Freeing a Block
When a block is freed, if it is a small block it is inserted at the front of the quick list of the
appropriate size. Blocks in the quick lists are free, but the allocation bit remains set in
the header to prevent them from being coalesced with adjacent blocks. In addition, there is a
separate "in quick list" bit in the block header that is set for blocks in the quick lists,
to allow them to be readily distinguished from blocks that are actually allocated.
To avoid arbitrary growth of the quick lists, the capacity of each is limited to QUICK_LIST_MAX blocks.
If an attempt is made to insert a block into a quick list that is already at capacity,
the quick list is flushed by removing each of the blocks it currently contains and adding
them back into the main free lists, coalescing them with any adjacent free blocks as described
below. After flushing the quick list, the block currently being freed is inserted into the
now-empty list, leaving just one block in that list.
When a block is added into the main free lists, an attempt should first be made to coalesce the block with any free block that immediately precedes or follows it in the heap. (See Chapter 9.9.10 Page 850 for a discussion of the coalescing procedure.) Once the block has been coalesced, it should be inserted at the front of the free list for the appropriate size class (based on the size after coalescing). The reason for performing coalescing is to combat the external fragmentation that would otherwise result due to the splitting of blocks upon allocation. Note that blocks inserted into quick lists are not immediately coalesced; they are only coalesced at such later time as the quick list is flushed and the blocks are moved into the main free lists. This is an example of a "deferred coalescing" strategy.
Block Headers & Footers
In Chapter 9.9.6 Page 847 Figure 9.35, a block header is defined as 2 words (32 bits) to hold the block size and allocated bit. In this assignment, the header will be 4 words (i.e. 64 bits or 1 memory row). The header fields will be similar to those in the textbook but with some differences.
Block Header Format:
+----------------------------+----------------------+--------+--------+---------+---------+ <- header
| payload size | block_size | unused | alloc |prv alloc|in qklst |
| (0/1) |(4 LSB's implicitly 0)| (0) | (1) | (0/1) | (0) |
| (32 bits) | (28 bits) | 1 bit | 1 bit | 1 bit | 1 bit |
+---------------------------------------------------+--------+--------+---------+---------+ <- (aligned)
- The
payload_sizefield, which occupies the four most-significant bytes of the header of an allocated block, will be used to store the payload size that was requested by the client for that block. In a free block (including a block in a quick list) this field be zero. - The
block_sizefield gives the number of bytes for the entire block (including header/footer, payload, and padding). It occupies the four least-significant bytes of the block header or footer, except that three of the four least-significant bits of the block size, which would normally always be zero due to alignment requirements, are used to store additional information. This means that these bits have to be masked when retrieving the block size from the header and when the block size is stored in the header the previously existing values of these bits have to be preserved. - The
allocbit (bit 2, mask 0x4) is a boolean. It is 1 if the block is allocated and 0 if it is free. - The
prev_alloc(bit 1, mask 0x2) is also a boolean. It is 1 if the immediately preceding block in the heap is allocated and 0 if it is not. - The
in_qklst(bit 0, mask 0x1) is also a boolean. It is 1 if the block is currently in a quick list, and 0 if it is not. Note that if this bit is a 1, then theallocbit will also be a 1. - The remaining bit out of the four least-significant bits (bit 3, mask 0x8) is not used, and should always be 0.
😱 Note that the fact that only four bytes are available for storing the block size and payload size in a block header means that the maximum size of a block will be at most 2^32 bytes. The types of the arguments to the
sf_malloc()andsf_realloc()functions reflect this limitation.
Each free block will also have a footer, which occupies the last memory row of the block. The footer of a free block (including a block in a quick list) must contain exactly the same information as the header. In an allocated block, the footer will not be present, and the space that it would otherwise occupy may be used for payload.
🤔 Here is an example of determining the block size required to satisfy a particular requested payload size. Suppose the requested size is 25 bytes. An additional 8 bytes will be required to store the block header, which must always be present. This means that a block of at least 33 bytes must be used, however due to alignment requirements this has to be rounded up to the next multiple of the alignment size. If the alignment size were 16 bytes (which would be just large enough to enable the memory returned by the allocator to store in an aligned fashion any of the basic data types supported by the x86-64 architecture), then a block of at least 48 bytes would have to be used. As a result, there would be 15 bytes of "padding" at the end of the payload area, which contributes to internal fragmentation. Besides the header, when the block is free it is also necessary to store a footer, as well and next and previous links for the freelist. These will take an additional 24 bytes of space, however when the block is free there is no payload so the payload area can be used to store this information, assuming that the payload area is big enough in the first place. But the payload area is 40 bytes (25 bytes plus 15 bytes of padding), which is certainly bigger than 24 bytes, so a block of total size 48 would be fine. Note that a block cannot be smaller than 32 bytes, as there there would not then be enough space to store the header, footer, and freelist links when the block is free.
Obfuscation of Headers and Footers
Your allocator has to satisfy one further requirement as regards the storage of the
block headers and footers. The headers and footers will not be stored directly in
memory; rather their contents will first be obfuscated by performing a bitwise XOR
(C operator ^) with a "magic" value that is obtained by referencing the preprocessor
symbol MAGIC. This value is set randomly (by the utility code provided for you)
when the heap is first initialized. When a header or footer is read from memory,
it must again be XOR'ed with the magic value to expose the true contents.
The purpose of obfuscating the headers and footers in this way is to help detect attempts
to free pointers that were not obtained from a previous call to malloc, and also to make
it possible to detect some situations in which the heap has been corrupted by overwriting
of headers and/or footers.
In the initial stages of debugging, you might find it helpful to turn off the header
and footer obfuscation. This can be accomplished by making an initial call of
sf_set_magic(0x0). The effect of this is that the magic value will then always be 0x0,
rather than a randomly chosen value. Once you have your code working with obfuscation
turned off in this way, don't forget to turn it back on again to test your code in the
correct configuration, because the sf_set_magic() function will be replaced by a dummy
version during grading.
Getting Started
Fetch and merge the base code for hw3 as described in hw0 from the
following link: https://gitlab02.cs.stonybrook.edu/cse320/hw3
Remember to use the --strategy-option=theirs flag with the git merge
command as described in the hw1 doc to avoid merge conflicts in the Gitlab
CI file.
Directory Structure
.
├── .gitignore
├── .gitlab-ci.yml
└── hw3
├── hw3.sublime-project
├── include
│ ├── debug.h
│ └── sfmm.h
├── lib
│ └── sfutil.o
├── Makefile
├── src
│ ├── main.c
│ └── sfmm.c
└── tests
└── sfmm_tests.c
The lib folder contains the object file for the sfutil library. This
library provides you with several functions to aid you with the implementation
of your allocator. Do NOT delete this file as it
is an essential part of your homework assignment.
The provided Makefile creates object files from the .c files in the src
directory, places the object files inside the build directory, and then links
the object files together, including lib/sfutil.o, to make executables that
are stored to the bin directory.
Note: make clean will not delete sfutil.o or the lib folder, but it
will delete all other contained .o files.
The sfmm.h header file contains function prototypes and defines the format
of the various data structures that you are to use.
😱 DO NOT modify
sfmm.hor the Makefile. Both will be replaced when we run tests for grading. If you wish to add things to a header file, please create a new header file in theincludefolder
All functions for your allocator (sf_malloc, sf_free, sf_realloc,
sf_internal_fragmentation, and sf_peak_utilization)
must be implemented in src/sfmm.c.
The program in src/main.c contains a basic example of using the allocation functions.
Running make will create a sfmm executable in the bin directory. This can be run
using the command bin/sfmm.
Allocation Functions
You will implement the three functions (sf_malloc, sf_realloc, and sf_free)
in the file src/sfmm.c. The file include/sfmm.h contains the prototypes and
documentation shown below.
Note: Standard C library functions set errno when there is an error.
To avoid conflicts with these functions, your allocation functions will set sf_errno,
a variable declared as extern in sfmm.h.
/*
* This is your implementation of sf_malloc. It acquires uninitialized memory that
* is aligned and padded properly for the underlying system.
*
* @param size The number of bytes requested to be allocated.
*
* @return If size is 0, then NULL is returned without setting sf_errno.
* If size is nonzero, then if the allocation is successful a pointer to a valid region of
* memory of the requested size is returned. If the allocation is not successful, then
* NULL is returned and sf_errno is set to ENOMEM.
*/
void *sf_malloc(sf_size_t size);
/*
* Resizes the memory pointed to by ptr to size bytes.
*
* @param ptr Address of the memory region to resize.
* @param size The minimum size to resize the memory to.
*
* @return If successful, the pointer to a valid region of memory is
* returned, else NULL is returned and sf_errno is set appropriately.
*
* If sf_realloc is called with an invalid pointer sf_errno should be set to EINVAL.
* If there is no memory available sf_realloc should set sf_errno to ENOMEM.
*
* If sf_realloc is called with a valid pointer and a size of 0 it should free
* the allocated block and return NULL without setting sf_errno.
*/
void *sf_realloc(void *ptr, sf_size_t size);
/*
* Marks a dynamically allocated region as no longer in use.
* Adds the newly freed block to the free list.
*
* @param ptr Address of memory returned by the function sf_malloc.
*
* If ptr is invalid, the function calls abort() to exit the program.
*/
void sf_free(void *ptr);
😱 Make sure these functions have these exact names and arguments. They must also appear in the correct file. If you do not name the functions correctly with the correct arguments, your program will not compile when we test it. YOU WILL GET A ZERO
Statistics Functions
Besides the allocation functions discussed above, you are to implement the following two functions that return statistics about the memory utilization of the allocator:
/*
* Get the current amount of internal fragmentation of the heap.
*
* @return the current amount of internal fragmentation, defined to be the
* ratio of the total amount of payload to the total size of allocated blocks.
* If there are no allocated blocks, then the returned value should be 0.0.
*/
double sf_internal_fragmentation();
/*
* Get the peak memory utilization for the heap.
*
* @return the peak memory utilization over the interval starting from the
* time the heap was initialized, up to the current time. The peak memory
* utilization at a given time, as defined in the lecture and textbook,
* is the ratio of the maximum aggregate payload up to that time, divided
* by the current heap size. If the heap has not yet been initialized,
* this function should return 0.0.
*/
double sf_peak_utilization();
These functions are also to be implemented in sfmm.c.
Any functions other than
sf_malloc,sf_free,sf_realloc,sf_internal_fragmentation, andsf_peak_utilizationWILL NOT be graded.
Initialization Functions
In the lib directory, we have provided you with the sfutil.o object file.
When linked with your program, this object file allows you to access the
sfutil library, which contains the following functions:
/*
* @return The starting address of the heap for your allocator.
*/
void *sf_mem_start();
/*
* @return The ending address of the heap for your allocator.
*/
void *sf_mem_end();
/*
* This function increases the size of your heap by adding one page of
* memory to the end.
*
* @return On success, this function returns a pointer to the start of the
* additional page, which is the same as the value that would have been returned
* by get_heap_end() before the size increase. On error, NULL is returned.
*/
void *sf_mem_grow();
/* The size of a page of memory returned by sf_mem_grow(). */
#define PAGE_SZ ((sf_size_t)1024)
/*
* @return The "magic number" used to obfuscate header and footer contents
* to make it difficult to free a block without having first succesfully
* malloc'ed one. To obtain the ability to turn off obfuscation using the
* -DWEAK_MAGIC compilation flag, you should not call this function directly
* but rather use the preprocessor symbol MAGIC where the magic number is
* required.
*/
sf_header sf_magic();
/* Define WEAK_MAGIC during compilation to use MAGIC of 0x0 for debugging purposes. */
#ifndef WEAK_MAGIC
#define MAGIC (sf_magic())
#else
#define MAGIC ((sf_header)0x0)
#endif
😱 As these functions are provided in a pre-built .o file, the source is not available to you. You will not be able to debug these using gdb. You must treat them as black boxes.
sf_mem_grow
The function sf_mem_grow is to be invoked by sf_malloc, at the time of the
first allocation request to obtain an initial free block, and on subsequent allocations
when a large enough block to satisfy the request is not found.
For this assignment, your implementation MUST ONLY use sf_mem_grow to
extend the heap. DO NOT use any system calls such as brk or sbrk
to do this.
Function sf_mem_grow returns memory to your allocator in pages.
Each page is 1024 bytes (1 KB) and there are a limited, small number of pages
available (the actual number may vary, so do not hard-code any particular limit
into your program). Each call to sf_mem_grow extends the heap by one page and
returns a pointer to the new page (this will be the same pointer as would have
been obtained from sf_mem_end before the call to sf_mem_grow.
The sf_mem_grow function also keeps track of the starting and ending addresses
of the heap for you. You can get these addresses through the sf_mem_start and
sf_mem_end functions.
😄 A real allocator would typically use the brk/sbrk system calls calls for small memory allocations and the mmap/munmap system calls for large allocations. To allow your program to use other functions provided by glibc, which rely on glibc's allocator (i.e.
malloc), we have providedsf_mem_growas a safe wrapper around sbrk. This makes it so your heap and the one managed by glibc do not interfere with each other.
sf_magic and MAGIC
The sf_magic function returns the random bit pattern that is being used
to obfuscate header and footer contents. This bit pattern must be XOR'ed
with header or footer contents to obfuscate them before storing them in memory,
and it must also be XOR'ed with header or footer contents to de-obfuscate them
after reading them from memory.
Instead of calling the sf_magic function directly, you should use the
preprocessor symbol MAGIC to obtain the magic bit pattern.
The definition of MAGIC is affected by whether or not the WEAK_MAGIC
C preprocessor symbol is defined during compilation. If WEAK_MAGIC
was not defined, then MAGIC calls sf_magic() to obtain the magic bit_pattern.
If WEAK_MAGIC was defined, then MAGIC is defined to be 0x0.
The latter is useful during debugging, so you can examine the contents of headers
and footers in gdb without being confused by the obfuscation.
You can arrange for WEAK_MAGIC to be defined during compilation by uncommenting
the -D WEAK_MAGIC in the DFLAGS setting in the Makefile.
Make sure that you test your code with -D WEAK_MAGIC commented out (as it is
in the basecode distribution), because this is how your code will be compiled
during grading.
Implementation Details
Memory Row Size
The table below lists the sizes of data types (following Intel standard terminlogy) on x86-64 Linux Mint:
| C declaration | Data type | x86-64 Size (Bytes) |
|---|---|---|
| char | Byte | 1 |
| short | Word | 2 |
| int | Double word | 4 |
| long int | Quadword | 8 |
| unsigned long | Quadword | 8 |
| pointer | Quadword | 8 |
| float | Single precision | 4 |
| double | Double precision | 8 |
| long double | Extended precision | 16 |
:nerd: You can find these sizes yourself using the sizeof operator. For example,
printf("%lu\n", sizeof(int))prints 4.
In this assignment we will assume that each "memory row" is 8 bytes (64 bits) in size.
All pointers returned by your sf_malloc are to be 16-byte aligned; that is, they will be
addresses that are multiples of 16. This requirement permits such pointers to be used to
store any of the basic machine data types in a "naturally aligned" fashion.
A value stored in memory is said to be naturally aligned if the address at which it
is stored is a multiple of the size of the value. For example, an int value is
naturally aligned when stored at an address that is a multiple of 4. A long double value
is naturally aligned when stored at an address that is a multiple of 16.
Keeping values naturally aligned in memory is a hardware-imposed requirement for some
architectures, and improves the efficiency of memory access in other architectures.
Block Header & Footer Fields
The various header and footer formats are specified in include/sfmm.h:
Format of an allocated memory block
+-----------------------------------------------------------------------------------------+
| 64-bit-wide row |
+-----------------------------------------------------------------------------------------+
+----------------------------+----------------------+--------+--------+---------+---------+ <- header
| payload size | block_size | unused | alloc |prv alloc|in qklst |
| (0/1) |(4 LSB's implicitly 0)| (0) | (1) | (0/1) | (0) |
| (32 bits) | (28 bits) | 1 bit | 1 bit | 1 bit | 1 bit |
+---------------------------------------------------+--------+--------+---------+---------+ <- (aligned)
| |
| Payload and Padding |
| (N rows) |
| |
| |
+-----------------------------------------------------------------------------------------+
NOTE: For an allocated block, there is no footer (it is used for payload).
NOTE: The actual stored header is obfuscated by bitwise XOR'ing with MAGIC.
The above diagram shows the un-obfuscated contents.
Format of a memory block in a quick list
+-----------------------------------------------------------------------------------------+
| 64-bit-wide row |
+-----------------------------------------------------------------------------------------+
+----------------------------+----------------------+--------+--------+---------+---------+ <- header
| unused | block_size | unused | alloc |prv alloc|in qklst |
| (0) |(4 LSB's implicitly 0)| (0) | (1) | (0/1) | (1) |
| (32 bits) | (28 bits) | 1 bit | 1 bit | 1 bit | 1 bit |
+---------------------------------------------------+--------+--------+---------+---------+ <- (aligned)
| |
| Payload and Padding |
| (N rows) |
| |
| |
+-----------------------------------------------------------------------------------------+
NOTE: For a block in a quick list, there is no footer.
Format of a free memory block
+----------------------------+----------------------+--------+--------+---------+---------+ <- header
| unused | block_size | unused | alloc |prv alloc|in qklst |
| (0) |(4 LSB's implicitly 0)| (0) | (0) | (0/1) | (0) |
| (32 bits) | (28 bits) | 1 bit | 1 bit | 1 bit | 1 bit |
+------------------------------------------------------------+--------+---------+---------+ <- (aligned)
| |
| Pointer to next free block |
| (1 row) |
+-----------------------------------------------------------------------------------------+
| |
| Pointer to previous free block |
| (1 row) |
+-----------------------------------------------------------------------------------------+
| |
| Unused |
| (N rows) |
| |
| |
+------------------------------------------------------------+--------+---------+---------+ <- footer
| unused | block_size | unused | alloc |prv alloc|in qklst |
| (0) |(4 LSB's implicitly 0)| (0) | (0) | (0/1) | (0) |
| (32 bits) | (28 bits) | 1 bit | 1 bit | 1 bit | 1 bit |
+------------------------------------------------------------+--------+---------+---------+
NOTE: For a free block, footer contents must always be identical to header contents.
NOTE: The actual stored footer is obfuscated by bitwise XOR'ing with MAGIC.
The above diagram shows the un-obfuscated contents.
The sfmm.h header file contains C structure definitions corresponding to the above diagrams:
#define IN_QUICK_LIST 0x1
#define PREV_BLOCK_ALLOCATED 0x2
#define THIS_BLOCK_ALLOCATED 0x4
typedef uint32_t sf_size_t;
typedef uint64_t sf_header;
typedef sf_header sf_footer;
/*
* Structure of a block.
* The first field of this structure is actually the footer of the *previous* block.
* This must be taken into account when creating sf_block pointers from memory addresses.
*/
typedef struct sf_block {
sf_footer prev_footer; // NOTE: This actually belongs to the *previous* block.
sf_header header; // This is where the current block really starts.
union {
/* A free block contains links to other blocks in a free list. */
struct {
struct sf_block *next;
struct sf_block *prev;
} links;
/* An allocated block contains a payload (aligned), starting here. */
char payload[0]; // Length varies according to block size.
} body;
} sf_block;
For sf_block, the body field is a union, which has been used to emphasize
the difference between the information contained in a free block and that contained
in an allocated block. If the block is free, then its body has a links field,
which is a struct containing next and prev pointers. If the block is
allocated, then its body does not have a links field, but rather has a payload,
which starts at the same address that the links field would have started if the
block were free. The size of the payload is obviously not zero, but as it is
variable and only determined at run time, the payload field has been declared
to be an array of length 0 just to enable the use of bp->body.payload to obtain
a pointer to the payload area, if bp is a pointer to sf_block.
👍 You can use casts to convert a generic pointer value to one of type
sf_block *orsf_header *, in order to make use of the above structure definitions to easily access the various fields. You can even cast an integer value to these pointer types; this is sometimes required when calculating the locations of blocks in the heap.
When a block is free, it must have a valid footer whose contents are identical to the header contents. We will use a "footer optimization" technique that permits a footer to be omitted from allocated blocks; thereby making the space that would otherwise be occupied by the footer available for use by payload. The footer optimization technique involves maintaining a bit in the header of each block that can be checked to find out if the immediately preceding block is allocated or free. If the preceding block is free, then its footer can be examined to find out its size and then the size can be used to calculate the block's starting address for the purpose of performing coalescing. If the preceding block is not free, then it has no footer, but as we can only coalesce with a free block there is no need for the information that we would have found in the footer, anyway.
😱 Note that the
prev_footerfield in thesf_blockstructure is actually part of the previous block in the heap. In order to initialize ansf_blockpointer to correctly access the fields of a block, it is necessary to compute the address of the footer of the immediately preceding block in the heap and then cast that address to typesf_block *. The footer of a particular block can be obtained by first getting ansf_block *pointer for that block and then using the contained information (i.e. the block size) to obtain theprev_footerfield of the next block in the heap. Thesf_blockstructure has been specified this way so as to permit it to be defined with a fixed size, even though the payload size is unknown and will vary.
Quick List and Free List Heads
In the file include/sfmm.h, you will see the following declaration:
#define NUM_QUICK_LISTS 10 /* Number of quick lists. */
#define QUICK_LIST_MAX 5 /* Maximum number of blocks permitted on a single quick list. */
struct {
int length; // Number of blocks currently in the list.
struct sf_block *first; // Pointer to first block in the list.
} sf_quick_lists[NUM_QUICK_LISTS];
#define NUM_FREE_LISTS 10
struct sf_block sf_free_list_heads[NUM_FREE_LISTS];
The array sf_quick_lists contains the heads of the quick lists,
which are maintained as singly linked lists accessed in LIFO fashion
(i.e. like stacks). The capacity of each quick list is limited to
a maximum of QUICK_LIST_MAX blocks. Inserting into a quick list that is
at capacity causes the quick list to be flushed as discussed elsewhere.
The array sf_free_list_heads contains the heads of the main free lists,
which are maintained as circular, doubly linked lists.
Each node in a free list contains a next pointer that points to the next
node in the list, and a prev pointer that points the previous node.
For each index i with 0 <= i < NUM_FREE_LISTS the variable sf_free_list_head[i]
is a dummy, "sentinel" node, which is used to connect the beginning and the end of
the list at index i. This sentinel node is always present and (aside from its next
and free pointers) does not contain any other data. If the list is empty,
then the fields sf_freelist_heads[i].body.links.next and sf_freelist_heads[i].body.links.prev
both contain &sf_freelist_heads[i] (i.e. the sentinel node points back to itself).
If the list is nonempty, then sf_freelist_heads[i].body.links.next points to the
first node in the list and sf_freelist_heads[i].body.links.prev points to the
last node in the list.
Inserting into and deleting from a circular doubly linked list is done
in the usual way, except that, owing to the use of the sentinel, there
are no edge cases for inserting or removing at the beginning or the end
of the list.
If you need a further introduction to this data structure, you can readily
find information on it by googling ("circular doubly linked lists with sentinel").
😱 You MUST use the
sf_free_list_headsarray for the heads of your free lists and you MUST maintain these lists as circular, doubly linked lists. The helper functions discussed later, as well as the unit tests, will assume that you have done this when accessing your free lists.
😱 Note that the head of a freelist must be initialized before the list can be used. The initialization is accomplished by setting the
nextandprevpointers of the sentinel node to point back to the node itself.
Overall Structure of the Heap
The overall structure of the allocatable area of your heap will be a sequence of allocated and free blocks. Your heap should also contain a prologue and epilogue (as described in the book, page 855) to arrange for the proper block alignment and to avoid edge cases when coalescing blocks. The overall organization of the heap is as shown below:
Format of the heap
+-----------------------------------------------------------------------------------------+
| 64-bit-wide row |
+-----------------------------------------------------------------------------------------+
+-----------------------------------------------------------------------------------------+ <- heap start
| | (aligned)
| Unused |
| (1 row) |
+----------------------------+----------------------+--------+--------+---------+---------+ <- header
| payload size |minimum block_size(32)| unused | alloc |prv alloc|in qklst |
| (0) |(4 LSB's implicitly 0)| (0) | (1) | (0/1) | (0) | prologue block
| (32 bits) | (28 bits) | 1 bit | 1 bit | 1 bit | 1 bit |
+------------------------------------------------------------+--------+---------+---------+ <- (aligned)
| |
| Unused Payload Area |
| (3 rows) |
+------------------------------------------------------------+--------+---------+---------+ <- header
| payload size | block_size | unused | alloc |prv alloc|in qklst |
| (0/1) |(4 LSB's implicitly 0)| (0) | (0/1) | (0/1) | (0/1) | first block
| (32 bits) | (28 bits) | 1 bit | 1 bit | 1 bit | 1 bit |
+------------------------------------------------------------+--------+---------+---------+ <- (aligned)
| |
| Payload and Padding |
| (N rows) |
| |
| |
+--------------------------------------------+------------------------+---------+---------+
| |
| |
| |
| |
| Additional allocated and free blocks |
| |
| |
| |
+-----------------------------------------------------------------------------------------+
| payload size | block_size | unused | alloc |prv alloc|in qklst |
| (0) | (0) | (0) | (1) | (0/1) | (0) | epilogue
| (32 bits) | (28 bits) | 1 bit | 1 bit | 1 bit | 1 bit |
+------------------------------------------------------------+--------+---------+---------+ <- heap_end
(aligned)
NOTE: The actual stored epilogue is obfuscated by bitwise XOR'ing with MAGIC.
The above diagram shows the un-obfuscated contents.
The heap begins with unused "padding", so that the header of each block will start
sizeof(sf_header) bytes before an alignment boundary.
The first block of the heap is the "prologue", which is an allocated block of minimum
size with an unused payload area.
At the end of the heap is an "epilogue", which consists only of an allocated header, with block size set to 0. The prologue and epilogue are never used to satisfy allocation requests and they are never freed. Whenever the heap is extended, a new epilogue is created at the end of the newly added region and the old epilogue becomes the header of the new block. This is as described in the book.
We do not make any separate C structure definitions for the prologue and epilogue.
They can be manipulated using the existing sf_block structure, though care must be taken
not to access fields that are not valid for these special blocks
(i.e. prev_footer for the prologue and anything other than header and prev_footer
for the epilogue).
As your heap is initially empty, at the time of the first call to sf_malloc
you will need to make one call to sf_mem_grow to obtain a page of memory
within which to set up the prologue and initial epilogue.
The remainder of the memory in this first page should then be inserted into
the free list as a single block.
Notes on sf_malloc
When implementing your sf_malloc function, first determine if the request size
is 0. If so, then return NULL without setting sf_errno.
If the request size is non-zero, then you should determine the size of the
block to be allocated by adding the header size and the size of any necessary
padding to reach a size that is a multiple of 16 to maintain proper alignment.
Remember also that the block has to be big enough to store the footer
as well as the next and prev pointers when the block is free.
As these fields are not present in an allocated block this space can (and should)
be overlapped with the payload area.
As has already been discussed, the above constraints lead to a minimum block size
of 32 bytes, so you should not attempt to allocate any block smaller than this.
After having determined the required block size, you should first check the
quick lists to see if they contain a block of that size.
If they do not, you should determine the index of the first main free list
that would be able to satisfy a request of that size.
Search that free list from the beginning until the first sufficiently large
block is found. If there is no such block, continue with the next larger
size class.
If a big enough block is found, then after splitting it (if it will not leave
a splinter), you should insert the remainder part back into the appropriate
freelist. When splitting a block, the "lower part" should be used to
satisfy the allocation request and the "upper part" should become the remainder.
Do not insert this remainder portion into any quick list; it should be put
directly into the main free lists.
If a big enough block is not found in any of the freelists, then you
must use sf_mem_grow to request more memory
(for requests larger than a page, more than one such call might be required).
If your allocator ultimately cannot satisfy the request, your sf_malloc function
must set sf_errno to ENOMEM and return NULL.
Notes on sf_mem_grow
After each call to sf_mem_grow, you must attempt to coalesce the newly
allocated page with any free block immediately preceding it, in order to build
blocks larger than one page. Insert the new block at the beginning of
the appropriate main freelist (not a quick list).
Note: Do not coalesce past the beginning or end of the heap.
Notes on sf_free
When implementing sf_free, you must first verify that the pointer being
passed to your function belongs to an allocated block. This can be done by
examining the fields in the block header. In this assignment, we will consider
the following cases to be invalid pointers:
- The pointer is
NULL. - The pointer is not 16-byte aligned.
- After XOR'ing the stored header with
MAGIC:- The block size is less than the minimum block size of 32.
- The block size is not a multiple of 16
- The header of the block is before the start of the first block of the heap, or the footer of the block is after the end of the last block in the heap.
- The
allocatedbit in the header is 0. - The
prev_allocfield in the header is 0, indicating that the previous block is free, but theallocfield of the previous block header is not 0.
If an invalid pointer is passed to your function, you must call abort to exit
the program. Use the man page for the abort function to learn more about this.
After confirming that a valid pointer was given, you must free the block. If the block size matches the size of one of the quick lists, it should be inserted into that quick list, flushing the quick list first if it is already at capacity. Otherwise, the block is inserted at the front of the appropriate main free list, after coalescing with any adjacent free block.
Note that blocks in a main free list must not be marked as allocated,
and they must have a valid footer with contents identical to the block header.
In contrast, blocks in a quick list are marked as allocated, and they
do not have any footer. In addition, blocks in a quick list have the
IN_QUICK_LIST bit set in their header.
Notes on sf_realloc
When implementing your sf_realloc function, you must first verify that the
pointer passed to your function is valid. The criteria for pointer validity
are the same as those described in the 'Notes on sf_free' section above.
If the pointer is valid but the size parameter is 0, free the block and return NULL.
After verifying the parameters, consider the cases described below.
Note that in some cases, sf_realloc is more complicated than calling sf_malloc
to allocate more memory, memcpy to move the old memory to the new memory, and
sf_free to free the old memory.
Reallocating to a Larger Size
When reallocating to a larger size, always follow these three steps:
-
Call
sf_mallocto obtain a larger block. -
Call
memcpyto copy the data in the block given by the client to the block returned bysf_malloc. Be sure to copy the entire payload area, but no more. -
Call
sf_freeon the block given by the client (inserting into a quick list or main freelist and coalescing if required). -
Return the block given to you by
sf_mallocto the client.
If sf_malloc returns NULL, sf_realloc must also return NULL. Note that
you do not need to set sf_errno in sf_realloc because sf_malloc should
take care of this.
Reallocating to a Smaller Size
When reallocating to a smaller size, your allocator must use the block that was passed by the caller. You must attempt to split the returned block. There are two cases for splitting:
- Splitting the returned block results in a splinter. In this case, do not split the block. Leave the splinter in the block, update the header field if necessary, and return the same block back to the caller.
Example:
b b +----------------------+ +------------------------+ | allocated | | allocated. | | Blocksize: 64 bytes | sf_realloc(b, 32) | Block size: 64 bytes | | payload: 48 bytes | | payload: 32 bytes | | | | | | | | | +----------------------+ +------------------------+
In the example above, splitting the block would have caused a 24-byte splinter. Therefore, the block is not split.
- The block can be split without creating a splinter. In this case, split the block and update the block size fields in both headers. Free the remainder block by inserting it into the appropriate free list (after coalescing, if possible -- do not insert the remainder block into a quick list). Return a pointer to the payload of the now-smaller block to the caller.
Note that in both of these sub-cases, you return a pointer to the same block that was given to you.
Example:
b b +----------------------+ +------------------------+ | allocated | | allocated | free | | Blocksize: 128 bytes | sf_realloc(b, 50) | 64 bytes | 64 bytes. | | payload: 80 bytes | | payload: | | | | | 50 bytes | goes into | | | | | free list | +----------------------+ +------------------------+
Helper Functions
The sfutil library additionally contains the following helper functions,
which should be self explanatory. They all output to stderr.
void sf_show_block(sf_block *bp);
void sf_show_blocks();
void sf_show_free_list(int index);
void sf_show_free_lists();
void sf_show_quick_list(int index);
void sf_show_quick_lists();
void sf_show_heap();
We have provided these functions to help you visualize your free lists and allocated blocks.
Unit Testing
For this assignment, we will use Criterion to test your allocator. We have provided a basic set of test cases and you will have to write your own as well.
You will use the Criterion framework alongside the provided helper functions to ensure your allocator works exactly as specified.
In the tests/sfmm_tests.c file, there are ten unit test examples. These tests
check for the correctness of sf_malloc, sf_realloc, and sf_free.
We provide some basic assertions, but by no means are they exhaustive. It is your
job to ensure that your header/footer bits are set correctly and that blocks are
allocated/freed as specified.
Compiling and Running Tests
When you compile your program with make, a sfmm_tests executable will be
created in the bin folder alongside the main executable. This can be run
with bin/sfmm_tests. To obtain more information about each test run, you can
use the verbose print option: bin/sfmm_tests --verbose.
You might also find it helpful to suppress the running of tests concurrently
by giving the --j1 option.
It is also possible to restrict the set of tests that are run. For example,
using --filter suite_name/test_name will only run the test named test_name
in test suite suite_name (if there is such a test, otherwise it will run
no tests).
Writing Criterion Tests
The first test malloc_an_int tests sf_malloc.
It allocates space for an integer and assigns a value to that space.
It then runs an assertion to make sure that the space returned by sf_malloc
was properly assigned.
cr_assert(*x == 4, "sf_malloc failed to give proper space for an int!");
The string after the assertion only gets printed to the screen if the assertion
failed (i.e. *x != 4). However, if there is a problem before the assertion,
such as a SEGFAULT, the unit test will print the error to the screen and
continue to run the rest of the unit tests.
For this assignment you must write 5 additional unit tests
which test new functionality and add them to sfmm_tests.c below the following
comment:
😱 You should definitely not regard the style in which the given tests have been written as an example of the correct way to write such tests. These handout tests have been deliberately coded in such a way as to to avoid giving away too much information about how you might write the allocator code. The tests contain many hard-coded numeric values and intentionally somewhat obscure pointer manipulations. You would do well not to follow this example, but rather to devise functions and macros that make your own code easier to write and to read. Exactly how you might do this has been left for you to work out!
//############################################
//STUDENT UNIT TESTS SHOULD BE WRITTEN BELOW
//DO NOT DELETE THESE COMMENTS
//############################################
For additional information on Criterion library, take a look at the official documentation located here! This documentation is VERY GOOD.
Hand-in instructions
Make sure your directory tree looks like it did originally after merging the basecode, and and that your homework compiles.
This homework's tag is: hw3
$ git submit hw3
A Word to the Wise
This program will be very difficult to get working unless you are extremely disciplined about your coding style. Think carefully about how to modularize your code in a way that makes it easier to understand and avoid mistakes. Verbose, repetitive code is error-prone and evil! When writing your program try to comment as much as possible. Format the code consistently. It is much easier for your TA and the professor to help you if we can quickly figure out what your code does.