NAME ^

src/gc/gc_gms.c - Generational mark and sweep garbage collection

OVERVIEW ^

The following comments describe a generational garbage collection scheme for Parrot.

Keywords:

 - non-copying, mark & sweep
 - generational
 - implicit reclamation, treadmill

DESCRIPTION ^

A plain mark & sweep collector performs work depending on the amount of all allocated objects. The advantage of a generational GC is achieved by not processing all objects. This is based on the weak generational hypothesis, which states that young objects are likely to die early. Old objects, which have survived a few GC cycles tend to be long-lived.

The terms young and old objects imply that there is some ordering in object creation time and the ordering is also followed by object references.

Specifically object references have to follow the marking direction. In pure functional programming languages this can be a very simple scheme:

        +------------+    object references
        v            |
   old   .... young .... youngest
                            |
                   <--------  scan direction

If (simplified) the only reference-like operation of the interpreter is:

   cons = (car, cdr)

and the object references "car" and "cdr" are created prior to the "aggregate" "cons", all object references point always to older objects. By scanning from the youngest to the oldest objects, all non-marked objects can be reclaimed immediately. And the scan can be aborted at any time after some processing, creating a generational GC in a trivial way.

But the programming languages we are serving are working basically the other direction, when it comes to object history:

  @a[$i] = $n

A reference operation like this needs first an aggregate and then the contents of it. So the scan direction is from old objects to younger ones. In such a scheme it's a bit more complicated to skip parts of the objects.

To take advantage of not processing all the objects, these are divided into generations, e.g.:

   old               young := nursery
   generation 0      generation 1

A mark phase now processes the root set and only objects from the young generation. When all objects are either referenced by the root set or only by the young generation, the algorithm is correct and complete.

But there is of course the possibilty that a young object is stored into an aggregate of an older generation. This case is tracked by the write barrier, which remembers all such operations in the IGP (inter generational pointer) list. When now generation 1 is marked, the IGP list can be considered as an extension to the root set, so that again all live objects of the young generation are detected.

Structures ^

typedef struct _gc_gms_gen Gc_gms_gen

Describes the state of one generation for one pool.

typedef struct _gc_gms_hdr Gc_gms_hdr

This header is in front of all Parrot objects. It forms a doubly-linked list of all objects in one pool and points to its generation.

PObj_to_GMSH(o)

GMSH_to_PObj(p)

These two macros convert from and to headers and objects.

typedef struct _gc_gms_hdr_list Gc_gms_hdr_list

A chained list of headers used e.g. for the IGP list.

 * XXX

Main problem TODO 1):

 [ PCont ]       ... continuation object in old generation
    |
    v
 [ Stack chunk ] --> [ e.g. P register frame ]  ... new generation

By pushing a new stack chunk onto the (old) existing stack frame, we'd need a WRITE_BARRIER that promotes the stack chunk to the old generation of the continuation. This would also need an IGP entry for the stack chunk buffer. But - as buffers aren't really containers in Parrot - this isn't possible.

To get that right, the code needs better support by the running interpreter. - never promote continuations (and stacks) in the current stack frame to an old generation - create scope_enter / scope_exit opcodes

A scope_enter happens on a subroutine call *and' with new_pad / push_pad opcodes. Each lexical scope must have its distinct register frame, else timely destruction can't work. If the frame needs active destruction, the old frame should be converted to the (new-1) generation, the inner frame is the nursery. On scope exit the newest (nursery) generation is collected and the current generation number is reset back to (new-1).

If the scope_enter doesn't indicate timely destruction, generation promoting should be done only, if object statistics indicate the presence of a fair amount of live objects.

TODO 2) in lazy sweep If timely destruction didn't find (all) eager objects, go back to older generations, until all these objects have been seen.

TODO 3) interpreter startup After all internal structures are created, promote interpreter state into initial first old generation by running one GC cycle before program execution begins (or just treat all objects as being alive).

Initialization functions ^

static void parrot_gc_gms_deinit

Free used resources.

static void gc_gms_pool_init

Initialize pool variables. This function must set the pool function pointers for add_free_object, get_free_object, alloc_objects, and more_objects.

void Parrot_gc_gms_init

Initialize the state structures of the gc system. Called immediately before creation of memory pools.

Interface functions ^

static void gc_gms_add_free_object

Unused. White (dead) objects are added in a bunch to the free_list.

static void gc_gms_chain_objects

TODO: interfere active_destroy and put these items into a separate white area, so that a sweep has just to run through these objects

Header chain layout: - all objects are chained together forming a circular list - pool->marker is the "anchor" of the circle (shown twice below)

 1) object allocation

 1a) one bunch of allocated objects was consumed: the free ptr did
     hit the marker

 +===+---+---+---+---+---+===+
 I M I w | w | w | w | w I M I
 +   +---+---+---+---+---+   +
       ^                   ^
       |                   |
       white               free == marker

 All these pointer ranges include the first element, but not the last one.

  [white ... free_list)   is the list of all whites

 1b) after allocating another bunch of objects

 +===+---+---+---+---+---+---+---+---+---+---+===+
 I M I w | w | w | w | w | f | f | f | f | f I M I
 +   +---+---+---+---+---+---+---+---+---+---+   +
       ^                   ^                   ^
       |                   |                   |
       white               free                marker
static void gc_gms_alloc_objects

Allocate new objects for the given pool.

static void gc_gms_more_objects

Run a GC cycle or allocate new objects for the given pool.

static PObj *gc_gms_get_free_object

Get a new object off the free_list in the given pool.

2) object consumption the free ptr moves towards the marker

Generation handling functions ^

overall header chain layout

           gen 0         gen 1      ...    gen N
  marker [first last) [first last)  ...   [first last)  marker

The last (youngest) generation N holds these (pool) pointers:

  [ black ... gray )          during marking
  [ gray ... white )          during marking
  [ white ... free_list )     allocated items
  [ free_list ... marker )    free items

The black, white, and generation ranges have additionally (TODO) *fin variants, which refer to PMCs that need destruction/finalization. These are always in front of the ranges to be processed first.

static Gc_gms_gen *gc_gms_create_gen

Create a generation structure for the given generation number.

static void gc_gms_init_gen

Initalize the generation system by creating the first two generations.

static Gc_gms_gen *gc_gms_find_gen

RT #48260: Not yet documented!!!

static void gc_gms_promote

RT #48260: Not yet documented!!!

static void gc_gms_store_hdr_list

RT #48260: Not yet documented!!!

static void gc_gms_clear_hdr_list

RT #48260: Not yet documented!!!

static void gc_gms_store_igp

RT #48260: Not yet documented!!!

static void gc_gms_clear_igp

RT #48260: Not yet documented!!!

void parrot_gc_gms_wb

Called by the write barrier. The aggregate belongs to an older generation then the new value written into it. Put the header of the new value onto the IGP list for the current generation, if it contains pointers to other items, and promote it to the old generation.

void parrot_gc_gms_wb_key

RT #48260: Not yet documented!!!

static void gc_gms_merge_gen

RT #48260: Not yet documented!!!

static void gc_gms_use_gen

RT #48260: Not yet documented!!!

static int set_gen_cb

RT #48260: Not yet documented!!!

static void gc_gms_set_gen

RT #48260: Not yet documented!!!

Marking functions ^

Header chain layout

Init: gray := black := white

 3) marking the root set

 3a) the white 'h' is to be set to gray to be scanned for children

 +---+---+---+---+---+---+->      +---+->
 | b | b | g | g | g | w          | h |
 +---+---+---+---+---+---+      <-+---+
   ^       ^           ^
   |       |           |
   black   gray        white

 3b) DFS if 'h' needs timely destruction

 +---+---+---+---+---+---+---+->
 | b | b | h | g | g | g | w
 +---+---+---+---+---+---+---+
   ^       ^               ^
   |       |               |
   black   gray            white


 3c) BFS in the normal case

 +---+---+---+---+---+---+---+->
 | b | b | g | g | g | h | w
 +---+---+---+---+---+---+---+
   ^       ^               ^
   |       |               |
   black   gray            white

 3d) the white is a scalar and immediately blackened


 +---+---+---+---+---+---+---+->
 | b | b | h | g | g | g | w
 +---+---+---+---+---+---+---+
   ^           ^           ^
   |           |           |
   black       gray        white

 3e) blacken the gray 'h' during trace_children

 +---+---+---+---+---+---+---+->
 | b | b | h | g | g | g | w
 +---+---+---+---+---+---+---+
   ^       ^               ^
   |       |               |
   black   gray            white


 +---+---+---+---+---+---+---+->
 | b | b | h | g | g | g | w
 +---+---+---+---+---+---+---+
   ^           ^           ^
   |           |           |
   black       gray        white
static void gc_gms_setto_gray

Set the white header h to gray.

static void gc_gms_setto_black

Set the white header h to black.

void parrot_gc_gms_pobject_lives

Set the object live - called by the pobject_lives macro

static int init_mark_cb

RT #48260: Not yet documented!!!

static void gc_gms_init_mark

Initialize the mark phase of GC.

static int trace_igp_cb

RT #48260: Not yet documented!!!

static int gc_gms_trace_root

Trace the root set. If trace_stack is true, trace system areas.

static int trace_children_cb

RT #48260: Not yet documented!!!

static int gc_gms_trace_children

Traverse gray objects: mark and blacken. Returns 0 if the trace was aborted lazily.

static int sweep_cb_pmc

move everything from white up to the free_list to the free_list scan for active destroy objects TODO put these in front of the pool at pool->white_fin

static int sweep_cb_buf

RT #48260: Not yet documented!!!

static void gc_gms_sweep

Free unused resources, put white objects onto free_list.

static int end_cycle_cb

RT #48260: Not yet documented!!!

static void gc_gms_end_cycle

RT #48260: Not yet documented!!!

Interface function main entry ^

static void parrot_gc_gms_run

Interface to Parrot_do_dod_run. flags is one of:

  GC_lazy_FLAG   ... timely destruction
  GC_finish_FLAG ... run a final sweep to destruct objects at
                      interpreter shutdown
static void gms_debug_verify

RT #48260: Not yet documented!!!

SEE ALSO ^

src/gc/dod.c, include/parrot/dod.h, include/parrot/pobj.h, src/gc/gc_ims.c

HISTORY ^

Initial version by leo (2005.01.12 - 2005.01.30)


parrot