3. Garbage collecting a language with the Memory Pool System

Have you written the lexer, parser, code generator and the runtime system for your programming language, and come to the realization that you are going to need a memory manager too? If so, you’ve come to the right place.

In this guide, I’ll explain how to use the MPS to add incremental, moving, generational garbage collection to the runtime system for a programming language.

I’m assuming that you are familiar with the overall architecture of the MPS (see the chapter Overview of the Memory Pool System) and that you’ve downloaded and built the MPS (see the chapter Building the Memory Pool System).

3.1. The Scheme interpreter

As a running example throughout this guide, I’ll be using a small interpreter for a subset of the Scheme programming language. I’ll be quoting the relevant sections of code as needed, but you may find it helpful to experiment with this interpreter yourself, in either of its versions:

scheme-malloc.c

The toy Scheme interpreter before integration with the MPS, using malloc and free(2) for memory management.

scheme.c

The toy Scheme interpreter after integration with the MPS.

This simple interpreter allocates two kinds of objects on the heap:

  1. All Scheme objects (there are no unboxed objects).
  2. The global symbol table: a hash table consisting of a vector of pointers to strings.

A Scheme object (whose type is not necessarily known) is represented by an obj_t, which is a pointer to a union of every type in the language:

typedef union obj_u *obj_t;
typedef union obj_u {
    type_s type;
    pair_s pair;
    symbol_s symbol;
    integer_s integer;
    special_s special;
    operator_s operator;
    string_s string;
    port_s port;
    character_s character;
    vector_s vector;
    table_s table;
    buckets_s buckets;
} obj_s;

Each of these types is a structure whose first word is a number specifying the type of the object (TYPE_PAIR for pairs, TYPE_SYMBOL for symbols, and so on). For example, pairs are represented by a pointer to the structure pair_s defined as follows:

typedef struct pair_s {
    type_t type;        /* TYPE_PAIR */
    obj_t car, cdr;     /* first and second projections */
} pair_s;

Because the first word of every object is its type, functions can operate on objects generically, testing TYPE(obj) as necessary (which is a macro for obj->type.type). For example, the print() function is implemented like this:

static void print(obj_t obj, unsigned depth, FILE *stream)
{
    switch (TYPE(obj)) {
    case TYPE_INTEGER:
        fprintf(stream, "%ld", obj->integer.integer);
        break;

    case TYPE_SYMBOL:
        fputs(obj->symbol.string, stream);
        break;

    /* ... and so on for the other types ... */
    }
}

Each constructor allocates memory for the new object by calling malloc. For example, make_pair is the constructor for pairs:

static obj_t make_pair(obj_t car, obj_t cdr)
{
    obj_t obj = (obj_t)malloc(sizeof(pair_s));
    if (obj == NULL) error("out of memory");
    obj->pair.type = TYPE_PAIR;
    CAR(obj) = car;
    CDR(obj) = cdr;
    return obj;
}

Objects are never freed, because it is necessary to prove that they are dead before their memory can be reclaimed. To prove that they are dead, we need a tracing garbage collector, which the MPS will provide.

3.2. Choosing an arena class

You’ll recall from the Overview of the Memory Pool System that the functionality of the MPS is divided between the arenas, which request memory from (and return it to) the operating system, and pools, which allocate blocks of memory for your program.

There are two main classes of arena: the client arena, mps_arena_class_cl(), which gets its memory from your program, and the virtual memory arena, mps_arena_class_vm(), which gets its memory from the operating system’s virtual memory interface.

The client arena is intended for use on embedded systems where there is no virtual memory, and has a couple of disadvantages (you have to decide how much memory you are going to use; and the MPS can’t return memory to the operating system for use by other processes) so for general-purpose programs you’ll want to use the virtual memory arena.

You’ll need a couple of headers: mps.h for the MPS interface, and mpsavm.h for the virtual memory arena class:

#include "mps.h"
#include "mpsavm.h"

There’s only one arena, and many MPS functions take an arena as an argument, so it makes sense for the arena to be a global variable rather than having to pass it around everywhere:

static mps_arena_t arena;

Create an arena by calling mps_arena_create(). This function takes a third argument when creating a virtual memory arena: the size of the amount of virtual virtual address space (not RAM), in bytes, that the arena will reserve initially. The MPS will ask for more address space if it runs out, but the more times it has to extend its address space, the less efficient garbage collection will become. The MPS works best if you reserve an address space that is several times larger than your peak memory usage.

Let’s reserve 32 megabytes:

mps_res_t res;
res = mps_arena_create(&arena,
                       mps_arena_class_vm(),
                       (size_t)(32 * 1024 * 1024));
if (res != MPS_RES_OK) error("Couldn't create arena");

mps_arena_create() is typical of functions in the MPS interface in that it stores its result in a location pointed to by an out parameter (here, &arena) and returns a result code, which is MPS_RES_OK if the function succeeded, or some other value if it failed.

Note

The MPS is designed to co-operate with other memory managers, so when integrating your language with the MPS you need not feel obliged to move all your memory management to the MPS: you can continue to use malloc and free to manage some of your memory, for example, while using the MPS for the rest.

The toy Scheme interpreter illustrates this by continuing to use malloc and free to manage its global symbol table.

3.3. Choosing a pool class

Pool classes come with a policy for how their memory will be managed: some pool classes use automatic memory management and others use manual; some use moving collection and others non-moving.

The section Choosing a pool class in the Pool reference contains a procedure for choosing a pool class. In the case of the toy Scheme interpreter, the answers to the questions are (1) yes, the MPS needs to automatically reclaim unreachable blocks; (2) yes, it’s acceptable for the MPS to move blocks in memory and protect them with barriers(1); and (3) the Scheme objects will contain exact references to other Scheme objects in the same pool.

The recommended class is AMC (Automatic Mostly-Copying). This pool class uses automatic memory management, moving garbage collection, allocation points and formatted objects, so it will provide an introduction to these features of the MPS.

Note

The MPS is designed for pools of different classes to co-exist in the same arena, so that objects requiring different memory management policies can be segregated into pools of suitable classes.

Topic

Pools.

3.4. Describing your objects

In order for the MPS to be able to automatically manage your objects, you need to tell it how to perform various operations on an object (scan it for references; replace it with a forwarding or padding object, and so on). You do this by creating an object format. Here’s the code for creating the object format for the toy Scheme interpreter:

struct mps_fmt_A_s obj_fmt_s = {
    ALIGNMENT,
    obj_scan,
    obj_skip,
    NULL,
    obj_fwd,
    obj_isfwd,
    obj_pad,
};

mps_fmt_t obj_fmt;
res = mps_fmt_create_A(&obj_fmt, arena, &obj_fmt_s);
if (res != MPS_RES_OK) error("Couldn't create obj format");

The structure mps_fmt_A_s is the simplest of several object format variants that are appropriate for moving pools like AMC.

The first element of the structure is the alignment of objects belonging to this format. Determining the alignment is hard to do portably, because it depends on the target architecture and on the way the compiler lays out its structures in memory. Here are some things you might try:

  1. Some modern compilers support the alignof operator:

    #define ALIGNMENT alignof(obj_s)
    
  2. On older compilers you may be able to use this trick:

    #define ALIGNMENT offsetof(struct {char c; obj_s obj;}, obj)
    

    but this is not reliable because some compilers pack structures more tightly than their alignment requirements in some circumstances (for example, GCC if the -fstruct-pack option is specified).

  3. The MPS interface provides the type mps_word_t, which is an unsigned integral type that is the same size as the platform’s object pointer types.

    So if you know that all your objects can be word-aligned, you can use:

    #define ALIGNMENT sizeof(mps_word_t)
    

The other elements of the structure are the format methods, which are described in the following sections. (The NULL in the structure is a placeholder for the copy method, which is now obsolete.)

3.4.1. The scan method

The scan method is a function of type mps_fmt_scan_t. It is called by the MPS to scan a block of memory. Its task is to identify all references within the objects in the block of memory, and “fix” them, by calling the macros MPS_FIX1() and MPS_FIX2() on each reference (possibly via the convenience macro MPS_FIX12()).

“Fixing” is a generic operation whose effect depends on the context in which the scan method was called. The scan method is called to discover references and so determine which objects are alive and which are dead, and also to update references after objects have been moved.

Here’s the scan method for the toy Scheme interpreter:

static mps_res_t obj_scan(mps_ss_t ss, mps_addr_t base, mps_addr_t limit)
{
    MPS_SCAN_BEGIN(ss) {
        while (base < limit) {
            obj_t obj = base;
            switch (TYPE(obj)) {
            case TYPE_PAIR:
                FIX(CAR(obj));
                FIX(CDR(obj));
                base = (char *)base + ALIGN(sizeof(pair_s));
                break;
            case TYPE_INTEGER:
                base = (char *)base + ALIGN(sizeof(integer_s));
                break;
            /* ... and so on for the other types ... */
            default:
                assert(0);
                fprintf(stderr, "Unexpected object on the heap\n");
                abort();
            }
        }
    } MPS_SCAN_END(ss);
    return MPS_RES_OK;
}

The scan method receives a scan state (ss) argument, and the block of memory to scan, from base (inclusive) to limit (exclusive). This block of memory is known to be packed with objects belonging to the object format, and so the scan method loops over the objects in the block, dispatching on the type of each object, and then updating base to point to the next object in the block.

For each reference in an object obj_scan fixes it by calling MPS_FIX12() via the macro FIX, which is defined as follows:

#define FIX(ref)                                                        \
    do {                                                                \
        mps_addr_t _addr = (ref); /* copy to local to avoid type pun */ \
        mps_res_t res = MPS_FIX12(ss, &_addr);                          \
        if (res != MPS_RES_OK) return res;                              \
        (ref) = _addr;                                                  \
    } while (0)

Each call to MPS_FIX12() must appear between calls to the macros MPS_SCAN_BEGIN() and MPS_SCAN_END(). It’s usually most convenient to call MPS_SCAN_BEGIN() at the start of the function and MPS_SCAN_END() at the end, as here.

Notes

  1. When the MPS calls your scan method, it may be part-way through moving your objects. It is therefore essential that the scan method only examine objects in the range of addresses it is given. Objects in other ranges of addresses are not guaranteed to be in a consistent state.
  2. Scanning is an operation on the critical path of the MPS, which means that it is important that it runs as quickly as possible.
  3. If your reference is tagged, you must remove the tag before fixing it. (This is not quite true, but see Tagged references for the full story.)
  4. The “fix” operation may update the reference. So if your reference is tagged, you must make sure that the tag is restored after the reference is updated.
  5. The “fix” operation may fail by returning a result code other than MPS_RES_OK. A scan function must propagate such a result code to the caller, and should do so as soon as practicable.

3.4.2. The skip method

The skip method is a function of type mps_fmt_skip_t. It is called by the MPS to skip over an object belonging to the format, and also to determine its size.

Here’s the skip method for the toy Scheme interpreter:

static mps_addr_t obj_skip(mps_addr_t base)
{
    obj_t obj = base;
    switch (TYPE(obj)) {
    case TYPE_PAIR:
        base = (char *)base + ALIGN(sizeof(pair_s));
        break;
    case TYPE_INTEGER:
        base = (char *)base + ALIGN(sizeof(integer_s));
        break;
    /* ... and so on for the other types ... */
    default:
        assert(0);
        fprintf(stderr, "Unexpected object on the heap\n");
        abort();
    }
    return base;
}

The argument base is the address to the base of the object. The skip method must return the address of the base of the “next object”: in formats of variant A like this one, this is the address just past the end of the object, rounded up to the object format’s alignment.

3.4.3. The forward method

The forward method is a function of type mps_fmt_fwd_t. It is called by the MPS after it has moved an object, and its task is to replace the old object with a forwarding object pointing to the new location of the object.

Diagram: Copying garbage collection.

Copying garbage collection.

The forwarding object must satisfy these properties:

  1. It must be scannable and skippable, and so it will need to have a type field to distinguish it from other Scheme objects.
  2. It must contain a pointer to the new location of the object (a forwarding pointer).
  3. The scan method and the skip method will both need to know the length of the forwarding object. This can be arbitrarily long (in the case of string objects, for example) so it must contain a length field.

This poses a problem, because the above analysis suggests that forwarding objects need to contain at least three words, but Scheme objects might be as small as two words (for example, integers).

This conundrum can be solved by having two types of forwarding object. The first type is suitable for forwarding objects of three words or longer:

typedef struct fwd_s {
    type_t type;                  /* TYPE_FWD */
    obj_t fwd;                    /* forwarded object */
    size_t size;                  /* total size of this object */
} fwd_s;

while the second type is suitable for forwarding objects of two words:

typedef struct fwd2_s {
    type_t type;                  /* TYPE_FWD2 */
    obj_t fwd;                    /* forwarded object */
} fwd2_s;

Here’s the forward method for the toy Scheme interpreter:

static void obj_fwd(mps_addr_t old, mps_addr_t new)
{
    obj_t obj = old;
    mps_addr_t limit = obj_skip(old);
    size_t size = (char *)limit - (char *)old;
    assert(size >= ALIGN_UP(sizeof(fwd2_s)));
    if (size == ALIGN_UP(sizeof(fwd2_s))) {
        TYPE(obj) = TYPE_FWD2;
        obj->fwd2.fwd = new;
    } else {
        TYPE(obj) = TYPE_FWD;
        obj->fwd.fwd = new;
        obj->fwd.size = size;
    }
}

The argument old is the old address of the object, and new is the location to which it has been moved.

The forwarding objects must be scannable and skippable, so the following code must be added to obj_scan and obj_skip:

case TYPE_FWD:
    base = (char *)base + ALIGN_UP(obj->fwd.size);
    break;
case TYPE_FWD2:
    base = (char *)base + ALIGN_UP(sizeof(fwd2_s));
    break;

Note

Objects that consist of a single word present a problem for the design of the forwarding object. In the toy Scheme interpreter, this happens on some 64-bit platforms, where a pointer is 8 bytes long, and a character_s object (which consists of a 4-byte int and a 1-byte char) is also 8 bytes long.

There are a couple of solutions to this problem:

  1. Allocate the small objects with enough padding so that they can be forwarded. (This is how the problem is solved in the toy Scheme interpreter.)
  2. Use a tag to distinguish between the client object and a forwarding object that replaces it. It might help to allocate the small objects in their own pool so that the number of types that the scan method has to distinguish is minimized. Since these objects do not contain references, they could be allocated from the AMCZ (Automatic Mostly-Copying Zero-rank) pool, and so the cost of scanning them could be avoided.

3.4.4. The is-forwarded method

The is-forwarded method is a function of type mps_fmt_isfwd_t. It is called by the MPS to determine if an object is a forwarding object, and if it is, to determine the location where that object was moved.

Here’s the is-forwarded method for the toy Scheme interpreter:

static mps_addr_t obj_isfwd(mps_addr_t addr)
{
    obj_t obj = addr;
    switch (TYPE(obj)) {
    case TYPE_FWD2:
        return obj->fwd2.fwd;
    case TYPE_FWD:
        return obj->fwd.fwd;
    }
    return NULL;
}

It receives the address of an object, and returns the address to which that object was moved, or NULL if the object was not moved.

3.4.5. The padding method

The padding method is a function of type mps_fmt_pad_t. It is called by the MPS to fill a block of memory with a padding object: this is an object that fills gaps in a block of formatted objects, for example to enable the MPS to pack objects into fixed-size units (such as operating system pages).

A padding object must be scannable and skippable, and not confusable with a forwarding object. This means they need a type and a size. However, padding objects might need to be as small as the alignment of the object format, which was specified to be a single word. As with forwarding objects, this can be solved by having two types of padding object. The first type is suitable for padding objects of two words or longer:

typedef struct pad_s {
    type_t type;                  /* TYPE_PAD */
    size_t size;                  /* total size of this object */
} pad_s;

while the second type is suitable for padding objects consisting of a single word:

typedef struct pad1_s {
    type_t type;                  /* TYPE_PAD1 */
} pad1_s;

Here’s the padding method:

static void obj_pad(mps_addr_t addr, size_t size)
{
    obj_t obj = addr;
    assert(size >= ALIGN(sizeof(pad1_s)));
    if (size == ALIGN(sizeof(pad1_s))) {
        TYPE(obj) = TYPE_PAD1;
    } else {
        TYPE(obj) = TYPE_PAD;
        obj->pad.size = size;
    }
}

The argument addr is the address at which the padding object must be created, and size is its size in bytes: this will always be a multiple of the alignment of the object format.

The padding objects must be scannable and skippable, so the following code must be added to obj_scan and obj_skip:

case TYPE_PAD:
    base = (char *)base + ALIGN(obj->pad.size);
    break;
case TYPE_PAD1:
    base = (char *)base + ALIGN(sizeof(pad1_s));
    break;

3.5. Generation chains

The AMC pool requires not only an object format but a generation chain. This specifies the generation structure of the generational garbage collection.

You create a generation chain by constructing an array of structures of type mps_gen_param_s, one for each generation, and passing them to mps_chain_create(). Each of these structures contains two values, the capacity of the generation in kilobytes, and the mortality, the proportion of objects in the generation that you expect to survive a collection of that generation.

These numbers are hints to the MPS that it may use to make decisions about when and what to collect: nothing will go wrong (other than suboptimal performance) if you make poor choices. Making good choices for the capacity and mortality of each generation is not easy, and is postponed to the chapter Tuning the Memory Pool System for performance.

Here’s the code for creating the generation chain for the toy Scheme interpreter:

mps_gen_param_s obj_gen_params[] = {
    { 150, 0.85 },
    { 170, 0.45 },
};

res = mps_chain_create(&obj_chain,
                       arena,
                       LENGTH(obj_gen_params),
                       obj_gen_params);
if (res != MPS_RES_OK) error("Couldn't create obj chain");

Note that these numbers have have been deliberately chosen to be small, so that the MPS is forced to collect often, so that you can see it working. Don’t just copy these numbers unless you also want to see frequent garbage collections!

3.6. Creating the pool

Now you know enough to create an AMC (Automatic Mostly-Copying) pool! Let’s review the pool creation code. First, the header for the AMC pool class:

#include "mpscamc.h"

Second, the object format:

struct mps_fmt_A_s obj_fmt_s = {
    sizeof(mps_word_t),
    obj_scan,
    obj_skip,
    NULL,
    obj_fwd,
    obj_isfwd,
    obj_pad,
};

mps_fmt_t obj_fmt;
res = mps_fmt_create_A(&obj_fmt, arena, &obj_fmt_s);
if (res != MPS_RES_OK) error("Couldn't create obj format");

Third, the generation chain:

mps_gen_param_s obj_gen_params[] = {
    { 150, 0.85 },
    { 170, 0.45 },
};

mps_chain_t obj_chain;
res = mps_chain_create(&obj_chain,
                       arena,
                       LENGTH(obj_gen_params),
                       obj_gen_params);
if (res != MPS_RES_OK) error("Couldn't create obj chain");

And finally the pool:

mps_pool_t obj_pool;
res = mps_pool_create(&obj_pool,
                      arena,
                      mps_class_amc(),
                      obj_fmt,
                      obj_chain);
if (res != MPS_RES_OK) error("Couldn't create obj pool");

3.7. Roots

The object format tells the MPS how to find references from one object to another. This allows the MPS to extrapolate the reachability property: if object A is reachable, and the scan method fixes a reference from A to another object B, then B is reachable too.

But how does this process get started? How does the MPS know which objects are reachable a priori? Such objects are known as roots, and you must register them with the MPS, creating root descriptions of type mps_root_t.

The most important root consists of the contents of the registers and the control stack of each thread in your program: this is covered in Threads, below.

Other roots may be found in static variables in your program, or in memory allocated by other memory managers. For these roots you must describe to the MPS how to scan them for references.

The toy Scheme interpreter has a number of static variables that point to heap-allocated objects. First, the special objects, including:

static obj_t obj_empty;         /* (), the empty list */

Second, the predefined symbols, including:

static obj_t obj_quote;         /* "quote" symbol */

And third, the global symbol table:

static obj_t *symtab;
static size_t symtab_size;

You tell the MPS how to scan these by writing root scanning functions of type mps_reg_scan_t. These functions are similar to the scan method in an object format, described above.

In the case of the toy Scheme interpreter, the root scanning function for the special objects and the predefined symbols could be written like this:

static mps_res_t globals_scan(mps_ss_t ss, void *p, size_t s)
{
    MPS_SCAN_BEGIN(ss) {
        FIX(obj_empty);
        /* ... and so on for the special objects ... */
        FIX(obj_quote);
        /* ... and so on for the predefined symbols ... */
    } MPS_SCAN_END(ss);
    return MPS_RES_OK;
}

but in fact the interpreter already has tables of these global objects, so it’s simpler and more extensible for the root scanning function to iterate over them:

static mps_res_t globals_scan(mps_ss_t ss, void *p, size_t s)
{
    MPS_SCAN_BEGIN(ss) {
        size_t i;
        for (i = 0; i < LENGTH(sptab); ++i)
            FIX(*sptab[i].varp);
        for (i = 0; i < LENGTH(isymtab); ++i)
            FIX(*isymtab[i].varp);
    } MPS_SCAN_END(ss);
    return MPS_RES_OK;
}

Each root scanning function must be registered with the MPS by calling mps_root_create(), like this:

mps_root_t globals_root;
res = mps_root_create(&globals_root, arena, mps_rank_exact(), 0,
                      globals_scan, NULL, 0);
if (res != MPS_RES_OK) error("Couldn't register globals root");

The third argument (here mps_rank_exact()) is the rank of references in the root. “Exact” means that:

  1. each reference in the root is a genuine pointer to another object managed by the MPS, or else a null pointer (unlike ambiguous references); and
  2. each reference keeps the target of the reference alive (unlike weak references(1)).

The fourth argument is the root mode, which tells the MPS whether it is allowed to place a barrier(1) on the root. The root mode 0 means that it is not allowed.

The sixth and seventh arguments (here NULL and 0) are passed to the root scanning function where they are received as the parameters p and s respectively. In this case there was no need to use them.

What about the global symbol table? This is trickier, because it gets rehashed from time to time, and during the rehashing process there are two copies of the symbol table in existence. Because the MPS is asynchronous, it might be scanning, moving, or collecting, at any point in time, and if it is doing so during the rehashing of the symbol table it had better scan both the old and new copies of the table. This is most conveniently done by registering a new root to refer to the new copy, and then after the rehash has completed, de-registering the old root by calling mps_root_destroy().

It would be possible to write a root scanning function of type mps_reg_scan_t, as described above, to fix the references in the global symbol table, but the case of a table of references is sufficiently common that the MPS provides a convenient (and optimized) function, mps_root_create_table(), for registering it:

static mps_root_t symtab_root;

/* ... */

mps_addr_t ref = symtab;
res = mps_root_create_table(&symtab_root, arena, mps_rank_exact(), 0,
                            ref, symtab_size);
if (res != MPS_RES_OK) error("Couldn't register new symtab root");

The root must be re-registered whenever the global symbol table changes size:

static void rehash(void) {
    obj_t *old_symtab = symtab;
    unsigned old_symtab_size = symtab_size;
    mps_root_t old_symtab_root = symtab_root;
    unsigned i;
    mps_addr_t ref;
    mps_res_t res;

    symtab_size *= 2;
    symtab = malloc(sizeof(obj_t) * symtab_size);
    if (symtab == NULL) error("out of memory");

    /* Initialize the new table to NULL so that "find" will work. */
    for (i = 0; i < symtab_size; ++i)
        symtab[i] = NULL;

    ref = symtab;
    res = mps_root_create_table(&symtab_root, arena, mps_rank_exact(), 0,
                                ref, symtab_size);
    if (res != MPS_RES_OK) error("Couldn't register new symtab root");

    for (i = 0; i < old_symtab_size; ++i)
        if (old_symtab[i] != NULL) {
            obj_t *where = find(old_symtab[i]->symbol.string);
            assert(where != NULL);    /* new table shouldn't be full */
            assert(*where == NULL);   /* shouldn't be in new table */
            *where = old_symtab[i];
        }

    mps_root_destroy(old_symtab_root);
    free(old_symtab);
}

Notes

  1. The old root description (referring to the old copy of the symbol table) is not destroyed until after the new root description has been registered. This is because the MPS is asynchronous: it might be scanning, moving, or collecting, at any point in time. If the old root description were destroyed before the new root description was registered, there would be a period during which:
    1. the symbol table was not reachable (at least as far as the MPS was concerned) and so all the objects referenced by it (and all the objects reachable from those objects) might be dead; and
    2. if the MPS moved an object, it would not know that the object was referenced by the symbol table, and so would not update the reference there to point to the new location of the object. This would result in out-of-date references in the old symbol table, and these would be copied into the new symbol table.
  2. The root might be scanned as soon as it is registered, so it is important to fill it with scannable references (NULL in this case) before registering it.
  3. The order of operations at the end is important: the old root must be de-registered before its memory is freed.

Topic

Roots.

3.8. Threads

In a multi-threaded environment where incremental garbage collection is used, you must register each of your threads with the MPS so that the MPS can examine their state.

Even in a single-threaded environment (like the toy Scheme interpreter) it may also be necessary to register the (only) thread if either of these conditions apply:

  1. you are using moving garbage collection (as with the AMC (Automatic Mostly-Copying) pool);
  2. the thread’s registers and control stack constitute a root (that is, objects may be kept alive via references in local variables: this is almost always the case for programs written in C).

You register a thread with an arena by calling mps_thread_reg():

mps_thr_t thread;
res = mps_thread_reg(&thread, arena);
if (res != MPS_RES_OK) error("Couldn't register thread");

You register the thread’s registers and control stack as a root by calling mps_root_create_reg() and passing mps_stack_scan_ambig():

void *marker = &marker;
mps_root_t reg_root;
res = mps_root_create_reg(&reg_root,
                          arena,
                          mps_rank_ambig(),
                          0,
                          thread,
                          mps_stack_scan_ambig,
                          marker,
                          0);
if (res != MPS_RES_OK) error("Couldn't create root");

In order to scan the control stack, the MPS needs to know where the bottom of the stack is, and that’s the role of the marker variable: the compiler places it on the stack, so its address is a position within the stack. As long as you don’t exit from this function while the MPS is running, your program’s active local variables will always be higher up on the stack than marker, and so will be scanned for references by the MPS.

Topic

Threads.

3.9. Allocation

It probably seemed a long journey to get here, but at last we’re ready to start allocating.

Manual pools typically support malloc-like allocation using the function mps_alloc(). But automatic pools cannot, because of the following problem:

static obj_t make_pair(obj_t car, obj_t cdr)
{
    obj_t obj;
    mps_addr_t addr;
    mps_res_t res;
    res = mps_alloc(&addr, pool, sizeof(pair_s));
    if (res != MPS_RES_OK) error("out of memory in make_pair");
    obj = addr;

    /* What happens if the MPS scans obj just now? */

    obj->pair.type = TYPE_PAIR;
    CAR(obj) = car;
    CDR(obj) = cdr;
    return obj;
}

Because the MPS is asynchronous, it might scan any reachable object at any time, including immediately after the object has been allocated. In this case, if the MPS attempts to scan obj at the indicated point, the object’s type field will be uninitialized, and so the scan method may abort.

The MPS solves this problem via the fast, nearly lock-free Allocation point protocol. This needs an additional structure, an allocation point, to be attached to the pool by calling mps_ap_create():

static mps_ap_t obj_ap;

/* ... */

res = mps_ap_create(&obj_ap, obj_pool, mps_rank_exact());
if (res != MPS_RES_OK) error("Couldn't create obj allocation point");

And then the constructor can be implemented like this:

static obj_t make_pair(obj_t car, obj_t cdr)
{
    obj_t obj;
    mps_addr_t addr;
    size_t size = ALIGN(sizeof(pair_s));
    do {
        mps_res_t res = mps_reserve(&addr, obj_ap, size);
        if (res != MPS_RES_OK) error("out of memory in make_pair");
        obj = addr;
        obj->pair.type = TYPE_PAIR;
        CAR(obj) = car;
        CDR(obj) = cdr;
    } while (!mps_commit(obj_ap, addr, size));
    return obj;
}

The function mps_reserve() allocates a block of memory that the MPS knows is uninitialized: the MPS promises not to scan this block or move it until after it is committed(2) by calling mps_commit(). So the new object can be initialized safely.

However, there’s a second problem:

    CAR(obj) = car;
    CDR(obj) = cdr;

    /* What if the MPS moves car or cdr just now? */

} while (!mps_commit(obj_ap, addr, size));

Because obj is not yet committed, the MPS won’t scan it, and that means that it won’t discover that it contains references to car and cdr, and so won’t update these references to point to their new locations.

In such a circumstance (that is, when objects have moved since you called mps_reserve()), mps_commit() returns false, and we have to initialize the object again (most conveniently done via a while loop, as here).

Notes

  1. When using the Allocation point protocol it is up to you to ensure that the requested size is aligned, because mps_reserve() is on the MPS’s critical path, and so it is highly optimized: in nearly all cases it is just an increment to a pointer and a test.
  2. It is very rare for mps_commit() to return false, but in the course of millions of allocations even very rare events occur, so it is important not to do anything you don’t want to repeat between calling mps_reserve() and mps_commit(). Also, the shorter the interval, the less likely mps_commit() is to return false.

Topic

Allocation.

3.10. Maintaining consistency

The MPS is asynchronous: this means that it might be scanning, moving, or collecting, at any point in time (potentially, between any pair of instructions in your program). So you must make sure that your data structures always obey these rules:

  1. A root must be scannable by its root scanning function as soon as it has been registered.

    See the discussion of the global symbol table in the toy Scheme interpreter.

  2. A formatted object must be scannable by the scan method as soon as it has been committed(2) by calling mps_commit().

    See the discussion of the pair constructor in the toy Scheme interpreter.

  3. All objects in automatically managed pools that are reachable by your code must always be provably reachable from a root via a chain of references that are fixed by a scanning function.

    See the discussion of the global symbol table in the toy Scheme interpreter.

  4. Formatted objects must remain scannable throughout their lifetime.

Examples of code that breaks these rules, together with tactics for tracking down the causes, appear in the chapter Debugging with the Memory Pool System.

3.11. Tidying up

When your program is done with the MPS, it’s good practice to tear down all the MPS data structures. This causes the MPS to check the consistency of its data structures and report any problems it detects. It also causes the MPS to flush its telemetry stream.

MPS data structures must be destroyed or deregistered in the reverse order to that in which they were registered or created. So you must destroy all allocation points created in a pool before destroying the pool; destroy all roots and pools, and deregister all threads, that were created in an arena before destroying the arena, and so on.

Here’s the tear-down code from the toy Scheme interpreter:

mps_ap_destroy(obj_ap);
mps_pool_destroy(obj_pool);
mps_chain_destroy(obj_chain);
mps_fmt_destroy(obj_fmt);
mps_root_destroy(reg_root);
mps_thread_dereg(thread);
mps_arena_destroy(arena);

3.12. What next?

This article has covered the basic knowledge needed to add incremental, moving, generational garbage collection to the runtime system for a programming language.

If everything is working for your language, then the next step is the chapter Tuning the Memory Pool System for performance.

But in the more likely event that things don’t work out quite as smoothly for your language as they did in the toy Scheme interpreter, then you’ll be more interested in the chapter Debugging with the Memory Pool System.