vgc is an implementation of a conservative, thread-local, mark-and-sweep
garbage collector. The implementation provides a fully functional replacement
for the standard POSIX malloc(), calloc(), realloc(), and free() calls.
The focus of vgc is to provide a conceptually clean implementation of
a mark-and-sweep GC, without delving into the depths of architecture-specific
optimization (see e.g. the Boehm GC for such an undertaking). It
should be particularly suitable for learning purposes and is open for all kinds
of optimization (PRs welcome!).
The original motivation for gc (the parent fork) was the original author's desire to write their own LISP implementation in C, entirely from scratch - and that required garbage collection.
Ironically enough, my original motivation for vgc (this fork) is my desire to write my own programming language
in C, entirely from scratch - and that also required garbage collection.
This work would not have been possible without the ability to read the work of others,
most notably the Boehm GC,
orangeduck's tgc (which also follows the ideals of being tiny and simple),
The Garbage Collection Handbook,
mkirchner, and
the many other contributors who worked on the original gc.
- Read the quickstart below to see how to get started quickly
- The concepts section describes the basic concepts and design
decisions that went into the implementation of
vgc. - Interleaved with the concepts, there are implementation sections that detail the implementation of the core components, see hash map implementation, dumping registers on the stack, finding roots, and depth-first, recursive marking.
$ git clone [email protected]:voidvoxel/gc.git
$ cd gc/src/voidvoxel/garbage_collection
To compile using the clang compiler:
$ make test
To use the GNU Compiler Collection (GCC):
$ make test CC=gcc
The tests should complete successfully. To create the current coverage report:
$ make coverage
struct Vector3 {
float x;
float y;
float z;
};
typedef struct Vector3 Vector3;
struct String {
size_t length;
char *data;
};
typedef struct String String;
struct Entity {
String *name;
Vector3 position;
};
typedef struct Entity Entity;
void do_something()
{
vgcx_var(Entity, x);
x->name = vgcx_new(String);
}
void do_lots_of_things()
{
int total_iterations = 1000000;
for (int i = 0; i < total_iterations; i++)
{
do_something();
}
}
int main(int argc, char **argv) {
vgcx_start();
do_lots_of_things();
vgcx_stop();
}This describes the core API, see gc.h for more details and the low-level API.
In order to initialize and start garbage collection, use the vgc_start()
function and pass a bottom-of-stack address:
void vgc_start(vgc_GC* gc, void* stack_bp);The bottom-of-stack parameter stack_bp needs to point to a stack-allocated
variable and marks the low end of the stack from where root
finding (scanning) starts.
Garbage collection can be stopped, disabled and resumed with
void vgc_stop(vgc_GC* gc);
void vgc_pause(vgc_GC* gc);
void vgc_resume(vgc_GC* gc);and manual garbage collection can be triggered with
size_t vgc_collect(vgc_GC* gc);vgc supports malloc(), calloc()and realloc()-style memory allocation.
The respective function signatures mimick the POSIX functions (with the
exception that we need to pass the garbage collector along as the first
argument):
void* vgc_malloc(vgc_GC* gc, size_t size);
void* vgc_calloc(vgc_GC* gc, size_t count, size_t size);
void* vgc_realloc(vgc_GC* gc, void* ptr, size_t size);It is possible to pass a pointer to a destructor function through the extended interface:
void* dtor(void* obj) {
// do some cleanup work
obj->parent->deregister();
obj->db->disconnect()
...
// no need to free obj
}
...
SomeObject* obj = vgc_malloc_ext(gc, sizeof(SomeObject), dtor);
...vgc supports static allocations that are garbage collected only when the
GC shuts down via vgc_stop(). Just use the appropriate helper function:
void* vgc_malloc_static(vgc_GC* gc, size_t size, void (*dtor)(void*));Static allocation expects a pointer to a finalization function; just set to
NULL if finalization is not required.
Note that vgc currently does not guarantee a specific ordering when it
collects static variables, If static vars need to be deallocated in a
particular order, the user should call vgc_free() on them in the desired
sequence prior to calling vgc_stop(), see below.
It is also possible to trigger explicit memory deallocation using
void vgc_free(vgc_GC* gc, void* ptr);Calling vgc_free() is guaranteed to (a) finalize/destruct on the object
pointed to by ptr if applicable and (b) to free the memory that ptr points to
irrespective of the current scheduling for garbage collection and will also
work if GC has been disabled using vgc_pause() above.
vgc also offers a strdup() implementation that returns a garbage-collected
copy:
char* vgc_strdup (vgc_GC* gc, const char* s);The fundamental idea behind garbage collection is to automate the memory allocation/deallocation cycle. This is accomplished by keeping track of all allocated memory and periodically triggering deallocation for memory that is still allocated but unreachable.
Many advanced garbage collectors also implement their own approach to memory
allocation (i.e. replace malloc()). This often enables them to layout memory
in a more space-efficient manner or for faster access but comes at the price of
architecture-specific implementations and increased complexity. vgc sidesteps
these issues by falling back on the POSIX *alloc() implementations and keeping
memory management and garbage collection metadata separate. This makes vgc
much simpler to understand but, of course, also less space- and time-efficient
than more optimized approaches.
The core data structure inside vgc is a hash map that maps the address of
allocated memory to the garbage collection metadata of that memory:
The items in the hash map are allocations, modeled with the Allocation
struct:
typedef struct Allocation {
void* ptr; // mem pointer
size_t size; // allocated size in bytes
char tag; // the tag for mark-and-sweep
void (*dtor)(void*); // destructor
struct Allocation* next; // separate chaining
} Allocation;Each Allocation instance holds a pointer to the allocated memory, the size of
the allocated memory at that location, a tag for mark-and-sweep (see below), an
optional pointer to the destructor function and a pointer to the next
Allocation instance (for separate chaining, see below).
The allocations are collected in an AllocationMap
typedef struct AllocationMap {
size_t capacity;
size_t min_capacity;
double downsize_factor;
double upsize_factor;
double sweep_factor;
size_t sweep_limit;
size_t size;
Allocation** allocs;
} AllocationMap;that, together with a set of static functions inside gc.c, provides hash
map semantics for the implementation of the public API.
The AllocationMap is the central data structure in the vgc_GC
struct which is part of the public API:
typedef struct vgc_GC {
struct AllocationMap* allocs;
bool disabled;
void *stack_bp;
size_t min_size;
} vgc_GC;With the basic data structures in place, any vgc_*alloc() memory allocation
request is a two-step procedure: first, allocate the memory through system (i.e.
standard malloc()) functionality and second, add or update the associated
metadata to the hash map.
For vgc_free(), use the pointer to locate the metadata in the hash map,
determine if the deallocation requires a destructor call, call if required,
free the managed memory and delete the metadata entry from the hash map.
These data structures and the associated interfaces enable the management of the metadata required to build a garbage collector.
vgc triggers collection under two circumstances: (a) when any of the calls to
the system allocation fail (in the hope to deallocate sufficient memory to
fulfill the current request); and (b) when the number of entries in the hash
map passes a dynamically adjusted high water mark.
If either of these cases occurs, vgc stops the world and starts a
mark-and-sweep garbage collection run over all current allocations. This
functionality is implemented in the vgc_collect() function which is part of the
public API and delegates all work to the vgc_mark() and vgc_sweep() functions
that are part of the private API.
vgc_mark() has the task of finding roots and tagging all
known allocations that are referenced from a root (or from an allocation that
is referenced from a root, i.e. transitively) as "used". Once the marking of
is completed, vgc_sweep() iterates over all known allocations and
deallocates all unused (i.e. unmarked) allocations, returns to vgc_collect() and
the world continues to run.
vgc will keep memory allocations that are reachable and collect everything
else. An allocation is considered reachable if any of the following is true:
- There is a pointer on the stack that points to the allocation content.
The pointer must reside in a stack frame that is at least as deep in the call
stack as the bottom-of-stack variable passed to
vgc_start()(i.e.stack_bpis the smallest stack address considered during the mark phase). - There is a pointer inside
vgc_*alloc()-allocated content that points to the allocation content. - The allocation is tagged with
VGC_TAG_ROOT.
The naïve mark-and-sweep algorithm runs in two stages. First, in a mark stage, the algorithm finds and marks all root allocations and all allocations that are reachable from the roots. Second, in the sweep stage, the algorithm passes over all known allocations, collecting all allocations that were not marked and are therefore deemed unreachable.
At the beginning of the mark stage, we first sweep across all known
allocations and find explicit roots with the VGC_TAG_ROOT tag set.
Each of these roots is a starting point for depth-first recursive
marking.
vgc subsequently detects all roots in the stack (starting from the bottom-of-stack
pointer stack_bp that is passed to vgc_start()) and the registers (by dumping them
on the stack prior to the mark phase) and
uses these as starting points for marking as well.
Given a root allocation, marking consists of (1) setting the tag field in an
Allocation object to VGC_TAG_MARK and (2) scanning the allocated memory for
pointers to known allocations, recursively repeating the process.
The underlying implementation is a simple, recursive depth-first search that scans over all memory content to find potential references:
void vgc_mark_alloc(vgc_GC* gc, void* ptr)
{
Allocation* alloc = vgc_allocation_map_get(gc->allocs, ptr);
if (alloc && !(alloc->tag & VGC_TAG_MARK)) {
alloc->tag |= VGC_TAG_MARK;
for (char* p = (char*) alloc->ptr;
p < (char*) alloc->ptr + alloc->size;
++p) {
vgc_mark_alloc(gc, *(void**)p);
}
}
}In gc.c, vgc_mark() starts the marking process by marking the
known roots on the stack via a call to vgc_mark_roots(). To mark the roots we
do one full pass through all known allocations. We then proceed to dump the
registers on the stack.
In order to make the CPU register contents available for root finding, vgc
dumps them on the stack. This is implemented in a somewhat portable way using
setjmp(), which stores them in a jmp_buf variable right before we mark the
stack:
...
/* Dump registers onto stack and scan the stack */
void (*volatile _mark_stack)(vgc_GC*) = vgc_mark_stack;
jmp_buf ctx;
memset(&ctx, 0, sizeof(jmp_buf));
setjmp(ctx);
_mark_stack(gc);
...The detour using the volatile function pointer _mark_stack to the
vgc_mark_stack() function is necessary to avoid the inlining of the call to
vgc_mark_stack().
After marking all memory that is reachable and therefore potentially still in
use, collecting the unreachable allocations is trivial. Here is the
implementation from vgc_sweep():
size_t vgc_sweep(vgc_GC* gc)
{
size_t total = 0;
for (size_t i = 0; i < gc->allocs->capacity; ++i) {
Allocation* chunk = gc->allocs->allocs[i];
Allocation* next = NULL;
while (chunk) {
if (chunk->tag & VGC_TAG_MARK) {
/* unmark */
chunk->tag &= ~VGC_TAG_MARK;
chunk = chunk->next;
} else {
total += chunk->size;
if (chunk->dtor) {
chunk->dtor(chunk->ptr);
}
free(chunk->ptr);
next = chunk->next;
vgc_allocation_map_remove(gc->allocs, chunk->ptr, false);
chunk = next;
}
}
}
vgc_allocation_map_resize_to_fit(gc->allocs);
return total;
}We iterate over all allocations in the hash map (the for loop), following every
chain (the while loop with the chunk = chunk->next update) and either (1)
unmark the chunk if it was marked; or (2) call the destructor on the chunk and
free the memory if it was not marked, keeping a running total of the amount of
memory we free.
That concludes the mark & sweep run. The stopped world is resumed and we're ready for the next run!
