#include "Trace.h"
#include "RetainerProfile.h"
#include "RaiseAsync.h"
+#include "Sparks.h"
#include "GC.h"
#include "Compact.h"
#include <string.h> // for memset()
+/* -----------------------------------------------------------------------------
+ Global variables
+ -------------------------------------------------------------------------- */
+
/* STATIC OBJECT LIST.
*
* During GC:
*/
StgClosure* static_objects; // live static objects
StgClosure* scavenged_static_objects; // static objects scavenged so far
+#ifdef THREADED_RTS
+SpinLock static_objects_sync;
+#endif
/* N is the oldest generation being collected, where the generations
* are numbered starting at 0. A major GC (indicated by the major_gc
nat N;
rtsBool major_gc;
-/* Youngest generation that objects should be evacuated to in
- * evacuate(). (Logically an argument to evacuate, but it's static
- * a lot of the time so we optimise it into a global variable).
- */
-nat evac_gen;
-
-/* Whether to do eager promotion or not.
- */
-rtsBool eager_promotion;
-
-/* Flag indicating failure to evacuate an object to the desired
- * generation.
- */
-rtsBool failed_to_evac;
-
/* Data used for allocation area sizing.
*/
-lnat new_blocks; // blocks allocated during this GC
-lnat new_scavd_blocks; // ditto, but depth-first blocks
static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
/* Mut-list stats */
mutlist_OTHERS;
#endif
+/* Thread-local data for each GC thread
+ */
+gc_thread *gc_threads = NULL;
+gc_thread *gct = NULL; // this thread's gct TODO: make thread-local
+
+// For stats:
+long copied; // *words* copied & scavenged during this GC
+long scavd_copied; // *words* copied only during this GC
+
/* -----------------------------------------------------------------------------
Static function declarations
-------------------------------------------------------------------------- */
-static void mark_root ( StgClosure **root );
-
-static void zero_static_object_list ( StgClosure* first_static );
+static void mark_root (StgClosure **root);
+static void zero_static_object_list (StgClosure* first_static);
+static void initialise_N (rtsBool force_major_gc);
+static void alloc_gc_threads (void);
+static void init_collected_gen (nat g, nat threads);
+static void init_uncollected_gen (nat g, nat threads);
+static void init_gc_thread (gc_thread *t);
+static void update_task_list (void);
+static void resize_generations (void);
+static void resize_nursery (void);
#if 0 && defined(DEBUG)
-static void gcCAFs ( void );
+static void gcCAFs (void);
#endif
/* -----------------------------------------------------------------------------
- inline functions etc. for dealing with the mark bitmap & stack.
+ The mark bitmap & stack.
-------------------------------------------------------------------------- */
#define MARK_STACK_BLOCKS 4
StgPtr oldgen_scan;
/* -----------------------------------------------------------------------------
- GarbageCollect
-
- Rough outline of the algorithm: for garbage collecting generation N
- (and all younger generations):
-
- - follow all pointers in the root set. the root set includes all
- mutable objects in all generations (mutable_list).
-
- - for each pointer, evacuate the object it points to into either
-
- + to-space of the step given by step->to, which is the next
- highest step in this generation or the first step in the next
- generation if this is the last step.
-
- + to-space of generations[evac_gen]->steps[0], if evac_gen != 0.
- When we evacuate an object we attempt to evacuate
- everything it points to into the same generation - this is
- achieved by setting evac_gen to the desired generation. If
- we can't do this, then an entry in the mut list has to
- be made for the cross-generation pointer.
-
- + if the object is already in a generation > N, then leave
- it alone.
-
- - repeatedly scavenge to-space from each step in each generation
- being collected until no more objects can be evacuated.
-
- - free from-space in each step, and set from-space = to-space.
+ GarbageCollect: the main entry point to the garbage collector.
Locks held: all capabilities are held throughout GarbageCollect().
-
-------------------------------------------------------------------------- */
void
{
bdescr *bd;
step *stp;
- lnat live, allocated, copied = 0, scavd_copied = 0;
+ lnat live, allocated;
lnat oldgen_saved_blocks = 0;
- nat g, s, i;
+ nat n_threads; // number of threads participating in GC
+
+ nat g, s, t;
#ifdef PROFILING
CostCentreStack *prev_CCS;
}
#endif
- // tell the STM to discard any cached closures its hoping to re-use
+ // tell the STM to discard any cached closures it's hoping to re-use
stmPreGCHook();
// tell the stats department that we've started a GC
/* Figure out which generation to collect
*/
- if (force_major_gc) {
- N = RtsFlags.GcFlags.generations - 1;
- major_gc = rtsTrue;
+ initialise_N(force_major_gc);
+
+ /* Allocate + initialise the gc_thread structures.
+ */
+ alloc_gc_threads();
+
+ /* How many threads will be participating in this GC?
+ * We don't try to parallelise minor GC.
+ */
+#if defined(THREADED_RTS)
+ if (N == 0) {
+ n_threads = 1;
} else {
- N = 0;
- for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
- if (generations[g].steps[0].n_blocks +
- generations[g].steps[0].n_large_blocks
- >= generations[g].max_blocks) {
- N = g;
- }
- }
- major_gc = (N == RtsFlags.GcFlags.generations-1);
+ n_threads = RtsFlags.ParFlags.gcThreads;
}
+#else
+ n_threads = 1;
+#endif
#ifdef RTS_GTK_FRONTPANEL
if (RtsFlags.GcFlags.frontpanel) {
*/
static_objects = END_OF_STATIC_LIST;
scavenged_static_objects = END_OF_STATIC_LIST;
+#ifdef THREADED_RTS
+ initSpinLock(&static_objects_sync);
+#endif
- /* Keep a count of how many new blocks we allocated during this GC
- * (used for resizing the allocation area, later).
- */
- new_blocks = 0;
- new_scavd_blocks = 0;
-
- // Initialise to-space in all the generations/steps that we're
- // collecting.
- //
+ // Initialise all the generations/steps that we're collecting.
for (g = 0; g <= N; g++) {
-
- // throw away the mutable list. Invariant: the mutable list
- // always has at least one block; this means we can avoid a check for
- // NULL in recordMutable().
- if (g != 0) {
- freeChain(generations[g].mut_list);
- generations[g].mut_list = allocBlock();
- for (i = 0; i < n_capabilities; i++) {
- freeChain(capabilities[i].mut_lists[g]);
- capabilities[i].mut_lists[g] = allocBlock();
- }
- }
-
- for (s = 0; s < generations[g].n_steps; s++) {
-
- // generation 0, step 0 doesn't need to-space
- if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
- continue;
- }
-
- stp = &generations[g].steps[s];
- ASSERT(stp->gen_no == g);
-
- // start a new to-space for this step.
- stp->old_blocks = stp->blocks;
- stp->n_old_blocks = stp->n_blocks;
-
- // allocate the first to-space block; extra blocks will be
- // chained on as necessary.
- stp->hp_bd = NULL;
- bd = gc_alloc_block(stp);
- stp->blocks = bd;
- stp->n_blocks = 1;
- stp->scan = bd->start;
- stp->scan_bd = bd;
-
- // allocate a block for "already scavenged" objects. This goes
- // on the front of the stp->blocks list, so it won't be
- // traversed by the scavenging sweep.
- gc_alloc_scavd_block(stp);
-
- // initialise the large object queues.
- stp->new_large_objects = NULL;
- stp->scavenged_large_objects = NULL;
- stp->n_scavenged_large_blocks = 0;
-
- // mark the large objects as not evacuated yet
- for (bd = stp->large_objects; bd; bd = bd->link) {
- bd->flags &= ~BF_EVACUATED;
- }
-
- // for a compacted step, we need to allocate the bitmap
- if (stp->is_compacted) {
- nat bitmap_size; // in bytes
- bdescr *bitmap_bdescr;
- StgWord *bitmap;
-
- bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
-
- if (bitmap_size > 0) {
- bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
- / BLOCK_SIZE);
- stp->bitmap = bitmap_bdescr;
- bitmap = bitmap_bdescr->start;
-
- debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
- bitmap_size, bitmap);
-
- // don't forget to fill it with zeros!
- memset(bitmap, 0, bitmap_size);
-
- // For each block in this step, point to its bitmap from the
- // block descriptor.
- for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
- bd->u.bitmap = bitmap;
- bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
-
- // Also at this point we set the BF_COMPACTED flag
- // for this block. The invariant is that
- // BF_COMPACTED is always unset, except during GC
- // when it is set on those blocks which will be
- // compacted.
- bd->flags |= BF_COMPACTED;
- }
- }
- }
- }
+ init_collected_gen(g,n_threads);
}
-
- /* make sure the older generations have at least one block to
- * allocate into (this makes things easier for copy(), see below).
- */
+
+ // Initialise all the generations/steps that we're *not* collecting.
for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
- for (s = 0; s < generations[g].n_steps; s++) {
- stp = &generations[g].steps[s];
- if (stp->hp_bd == NULL) {
- ASSERT(stp->blocks == NULL);
- bd = gc_alloc_block(stp);
- stp->blocks = bd;
- stp->n_blocks = 1;
- }
- if (stp->scavd_hp == NULL) {
- gc_alloc_scavd_block(stp);
- stp->n_blocks++;
- }
- /* Set the scan pointer for older generations: remember we
- * still have to scavenge objects that have been promoted. */
- stp->scan = stp->hp;
- stp->scan_bd = stp->hp_bd;
- stp->new_large_objects = NULL;
- stp->scavenged_large_objects = NULL;
- stp->n_scavenged_large_blocks = 0;
- }
-
- /* Move the private mutable lists from each capability onto the
- * main mutable list for the generation.
- */
- for (i = 0; i < n_capabilities; i++) {
- for (bd = capabilities[i].mut_lists[g];
- bd->link != NULL; bd = bd->link) {
- /* nothing */
- }
- bd->link = generations[g].mut_list;
- generations[g].mut_list = capabilities[i].mut_lists[g];
- capabilities[i].mut_lists[g] = allocBlock();
- }
+ init_uncollected_gen(g,n_threads);
}
/* Allocate a mark stack if we're doing a major collection.
mark_stack_bdescr = NULL;
}
- eager_promotion = rtsTrue; // for now
+ // Initialise all our gc_thread structures
+ for (t = 0; t < n_threads; t++) {
+ init_gc_thread(&gc_threads[t]);
+ }
+
+ // Initialise stats
+ copied = 0;
+ scavd_copied = 0;
+
+ // start threads etc.
+ // For now, we just have one thread, and set gct to gc_threads[0]
+ gct = &gc_threads[0];
/* -----------------------------------------------------------------------
* follow all the roots that we know about:
* - mutable lists from each generation > N
* we want to *scavenge* these roots, not evacuate them: they're not
* going to move in this GC.
- * Also: do them in reverse generation order. This is because we
- * often want to promote objects that are pointed to by older
- * generations early, so we don't have to repeatedly copy them.
- * Doing the generations in reverse order ensures that we don't end
- * up in the situation where we want to evac an object to gen 3 and
- * it has already been evaced to gen 2.
+ * Also do them in reverse generation order, for the usual reason:
+ * namely to reduce the likelihood of spurious old->new pointers.
*/
{
- int st;
for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
generations[g].saved_mut_list = generations[g].mut_list;
generations[g].mut_list = allocBlock();
// mut_list always has at least one block.
}
-
for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
scavenge_mutable_list(&generations[g]);
- evac_gen = g;
- for (st = generations[g].n_steps-1; st >= 0; st--) {
- scavenge(&generations[g].steps[st]);
- }
}
}
- /* follow roots from the CAF list (used by GHCi)
- */
- evac_gen = 0;
+ // follow roots from the CAF list (used by GHCi)
+ gct->evac_gen = 0;
markCAFs(mark_root);
- /* follow all the roots that the application knows about.
- */
- evac_gen = 0;
+ // follow all the roots that the application knows about.
+ gct->evac_gen = 0;
GetRoots(mark_root);
- /* Mark the weak pointer list, and prepare to detect dead weak
- * pointers.
- */
+ // Mark the weak pointer list, and prepare to detect dead weak pointers.
markWeakPtrList();
initWeakForGC();
- /* Mark the stable pointer table.
- */
+ // Mark the stable pointer table.
markStablePtrTable(mark_root);
/* -------------------------------------------------------------------------
loop:
flag = rtsFalse;
- // scavenge static objects
- if (major_gc && static_objects != END_OF_STATIC_LIST) {
- IF_DEBUG(sanity, checkStaticObjects(static_objects));
- scavenge_static();
- }
-
- /* When scavenging the older generations: Objects may have been
- * evacuated from generations <= N into older generations, and we
- * need to scavenge these objects. We're going to try to ensure that
- * any evacuations that occur move the objects into at least the
- * same generation as the object being scavenged, otherwise we
- * have to create new entries on the mutable list for the older
- * generation.
- */
-
- // scavenge each step in generations 0..maxgen
- {
- long gen;
- int st;
-
- loop2:
- // scavenge objects in compacted generation
- if (mark_stack_overflowed || oldgen_scan_bd != NULL ||
- (mark_stack_bdescr != NULL && !mark_stack_empty())) {
- scavenge_mark_stack();
- flag = rtsTrue;
- }
-
- for (gen = RtsFlags.GcFlags.generations; --gen >= 0; ) {
- for (st = generations[gen].n_steps; --st >= 0; ) {
- if (gen == 0 && st == 0 && RtsFlags.GcFlags.generations > 1) {
- continue;
- }
- stp = &generations[gen].steps[st];
- evac_gen = gen;
- if (stp->hp_bd != stp->scan_bd || stp->scan < stp->hp) {
- scavenge(stp);
- flag = rtsTrue;
- goto loop2;
- }
- if (stp->new_large_objects != NULL) {
- scavenge_large(stp);
- flag = rtsTrue;
- goto loop2;
- }
- }
- }
- }
+ scavenge_loop();
// if any blackholes are alive, make the threads that wait on
// them alive too.
}
}
- /* Update the pointers from the task list - these are
- * treated as weak pointers because we want to allow a main thread
- * to get a BlockedOnDeadMVar exception in the same way as any other
- * thread. Note that the threads should all have been retained by
- * GC by virtue of being on the all_threads list, we're just
- * updating pointers here.
- */
- {
- Task *task;
- StgTSO *tso;
- for (task = all_tasks; task != NULL; task = task->all_link) {
- if (!task->stopped && task->tso) {
- ASSERT(task->tso->bound == task);
- tso = (StgTSO *) isAlive((StgClosure *)task->tso);
- if (tso == NULL) {
- barf("task %p: main thread %d has been GC'd",
-#ifdef THREADED_RTS
- (void *)task->id,
-#else
- (void *)task,
-#endif
- task->tso->id);
- }
- task->tso = tso;
- }
- }
- }
+ // Update pointers from the Task list
+ update_task_list();
// Now see which stable names are still alive.
gcStablePtrTable();
- // Tidy the end of the to-space chains
- for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
- for (s = 0; s < generations[g].n_steps; s++) {
- stp = &generations[g].steps[s];
- if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
- ASSERT(Bdescr(stp->hp) == stp->hp_bd);
- stp->hp_bd->free = stp->hp;
- Bdescr(stp->scavd_hp)->free = stp->scavd_hp;
- }
- }
- }
-
#ifdef PROFILING
// We call processHeapClosureForDead() on every closure destroyed during
// the current garbage collection, so we invoke LdvCensusForDead().
IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
+ // Two-space collector: free the old to-space.
+ // g0s0->old_blocks is the old nursery
+ // g0s0->blocks is to-space from the previous GC
+ if (RtsFlags.GcFlags.generations == 1) {
+ if (g0s0->blocks != NULL) {
+ freeChain(g0s0->blocks);
+ g0s0->blocks = NULL;
+ }
+ }
+
+ // For each workspace, in each thread:
+ // * clear the BF_EVACUATED flag from each copied block
+ // * move the copied blocks to the step
+ {
+ gc_thread *thr;
+ step_workspace *ws;
+ bdescr *prev;
+
+ for (t = 0; t < n_threads; t++) {
+ thr = &gc_threads[t];
+
+ for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
+ for (s = 0; s < generations[g].n_steps; s++) {
+ ws = &thr->steps[g][s];
+ if (g==0 && s==0) continue;
+
+ ASSERT( ws->scan_bd == ws->todo_bd );
+ ASSERT( ws->scan_bd ? ws->scan == ws->scan_bd->free : 1 );
+
+ // Push the final block
+ if (ws->scan_bd) { push_scan_block(ws->scan_bd, ws); }
+
+ // update stats: we haven't counted the block at the
+ // front of the scavd_list yet.
+ scavd_copied += ws->scavd_list->free - ws->scavd_list->start;
+
+ ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
+
+ prev = ws->scavd_list;
+ for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
+ bd->flags &= ~BF_EVACUATED; // now from-space
+ prev = bd;
+ }
+ prev->link = ws->stp->blocks;
+ ws->stp->blocks = ws->scavd_list;
+ ws->stp->n_blocks += ws->n_scavd_blocks;
+ ASSERT(countBlocks(ws->stp->blocks) == ws->stp->n_blocks);
+ }
+ }
+ }
+ }
+
+ // Two-space collector: swap the semi-spaces around.
+ // Currently: g0s0->old_blocks is the old nursery
+ // g0s0->blocks is to-space from this GC
+ // We want these the other way around.
+ if (RtsFlags.GcFlags.generations == 1) {
+ bdescr *nursery_blocks = g0s0->old_blocks;
+ nat n_nursery_blocks = g0s0->n_old_blocks;
+ g0s0->old_blocks = g0s0->blocks;
+ g0s0->n_old_blocks = g0s0->n_blocks;
+ g0s0->blocks = nursery_blocks;
+ g0s0->n_blocks = n_nursery_blocks;
+ }
+
/* run through all the generations/steps and tidy up
*/
- copied = new_blocks * BLOCK_SIZE_W;
- scavd_copied = new_scavd_blocks * BLOCK_SIZE_W;
for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
if (g <= N) {
bdescr *next;
stp = &generations[g].steps[s];
- if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
- // stats information: how much we copied
- if (g <= N) {
- copied -= stp->hp_bd->start + BLOCK_SIZE_W -
- stp->hp_bd->free;
- scavd_copied -= stp->scavd_hpLim - stp->scavd_hp;
- }
- }
-
// for generations we collected...
if (g <= N) {
* freed blocks will probaby be quickly recycled.
*/
if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
- if (stp->is_compacted) {
+ if (stp->is_compacted)
+ {
// for a compacted step, just shift the new to-space
// onto the front of the now-compacted existing blocks.
for (bd = stp->blocks; bd != NULL; bd = bd->link) {
// add the new blocks to the block tally
stp->n_blocks += stp->n_old_blocks;
ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
- } else {
+ }
+ else // not copacted
+ {
freeChain(stp->old_blocks);
- for (bd = stp->blocks; bd != NULL; bd = bd->link) {
- bd->flags &= ~BF_EVACUATED; // now from-space
- }
}
stp->old_blocks = NULL;
stp->n_old_blocks = 0;
stp->large_objects = stp->scavenged_large_objects;
stp->n_large_blocks = stp->n_scavenged_large_blocks;
- } else {
- // for older generations...
-
+ }
+ else // for older generations...
+ {
/* For older generations, we need to append the
* scavenged_large_object list (i.e. large objects that have been
* promoted during this GC) to the large_object list for that step.
}
}
- /* Reset the sizes of the older generations when we do a major
- * collection.
- *
- * CURRENT STRATEGY: make all generations except zero the same size.
- * We have to stay within the maximum heap size, and leave a certain
- * percentage of the maximum heap size available to allocate into.
- */
- if (major_gc && RtsFlags.GcFlags.generations > 1) {
- nat live, size, min_alloc;
- nat max = RtsFlags.GcFlags.maxHeapSize;
- nat gens = RtsFlags.GcFlags.generations;
-
- // live in the oldest generations
- live = oldest_gen->steps[0].n_blocks +
- oldest_gen->steps[0].n_large_blocks;
-
- // default max size for all generations except zero
- size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
- RtsFlags.GcFlags.minOldGenSize);
-
- // minimum size for generation zero
- min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
- RtsFlags.GcFlags.minAllocAreaSize);
-
- // Auto-enable compaction when the residency reaches a
- // certain percentage of the maximum heap size (default: 30%).
- if (RtsFlags.GcFlags.generations > 1 &&
- (RtsFlags.GcFlags.compact ||
- (max > 0 &&
- oldest_gen->steps[0].n_blocks >
- (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
- oldest_gen->steps[0].is_compacted = 1;
-// debugBelch("compaction: on\n", live);
- } else {
- oldest_gen->steps[0].is_compacted = 0;
-// debugBelch("compaction: off\n", live);
- }
-
- // if we're going to go over the maximum heap size, reduce the
- // size of the generations accordingly. The calculation is
- // different if compaction is turned on, because we don't need
- // to double the space required to collect the old generation.
- if (max != 0) {
-
- // this test is necessary to ensure that the calculations
- // below don't have any negative results - we're working
- // with unsigned values here.
- if (max < min_alloc) {
- heapOverflow();
- }
-
- if (oldest_gen->steps[0].is_compacted) {
- if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
- size = (max - min_alloc) / ((gens - 1) * 2 - 1);
- }
- } else {
- if ( (size * (gens - 1) * 2) + min_alloc > max ) {
- size = (max - min_alloc) / ((gens - 1) * 2);
- }
- }
-
- if (size < live) {
- heapOverflow();
- }
- }
-
-#if 0
- debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
- min_alloc, size, max);
-#endif
-
- for (g = 0; g < gens; g++) {
- generations[g].max_blocks = size;
- }
- }
-
+ // update the max size of older generations after a major GC
+ resize_generations();
+
// Guess the amount of live data for stats.
live = calcLive();
- /* Free the small objects allocated via allocate(), since this will
- * all have been copied into G0S1 now.
- */
+ // Free the small objects allocated via allocate(), since this will
+ // all have been copied into G0S1 now.
if (RtsFlags.GcFlags.generations > 1) {
if (g0s0->blocks != NULL) {
freeChain(g0s0->blocks);
// Start a new pinned_object_block
pinned_object_block = NULL;
- /* Free the mark stack.
- */
+ // Free the mark stack.
if (mark_stack_bdescr != NULL) {
freeGroup(mark_stack_bdescr);
}
- /* Free any bitmaps.
- */
+ // Free any bitmaps.
for (g = 0; g <= N; g++) {
for (s = 0; s < generations[g].n_steps; s++) {
stp = &generations[g].steps[s];
}
}
- /* Two-space collector:
- * Free the old to-space, and estimate the amount of live data.
- */
- if (RtsFlags.GcFlags.generations == 1) {
- nat blocks;
-
- /* For a two-space collector, we need to resize the nursery. */
-
- /* set up a new nursery. Allocate a nursery size based on a
- * function of the amount of live data (by default a factor of 2)
- * Use the blocks from the old nursery if possible, freeing up any
- * left over blocks.
- *
- * If we get near the maximum heap size, then adjust our nursery
- * size accordingly. If the nursery is the same size as the live
- * data (L), then we need 3L bytes. We can reduce the size of the
- * nursery to bring the required memory down near 2L bytes.
- *
- * A normal 2-space collector would need 4L bytes to give the same
- * performance we get from 3L bytes, reducing to the same
- * performance at 2L bytes.
- */
- blocks = g0s0->n_blocks;
-
- if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
- blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
- RtsFlags.GcFlags.maxHeapSize ) {
- long adjusted_blocks; // signed on purpose
- int pc_free;
-
- adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
-
- debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
- RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
-
- pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
- if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even be < 0 */ {
- heapOverflow();
- }
- blocks = adjusted_blocks;
-
- } else {
- blocks *= RtsFlags.GcFlags.oldGenFactor;
- if (blocks < RtsFlags.GcFlags.minAllocAreaSize) {
- blocks = RtsFlags.GcFlags.minAllocAreaSize;
- }
- }
- resizeNurseries(blocks);
-
- } else {
- /* Generational collector:
- * If the user has given us a suggested heap size, adjust our
- * allocation area to make best use of the memory available.
- */
-
- if (RtsFlags.GcFlags.heapSizeSuggestion) {
- long blocks;
- nat needed = calcNeeded(); // approx blocks needed at next GC
-
- /* Guess how much will be live in generation 0 step 0 next time.
- * A good approximation is obtained by finding the
- * percentage of g0s0 that was live at the last minor GC.
- */
- if (N == 0) {
- g0s0_pcnt_kept = (new_blocks * 100) / countNurseryBlocks();
- }
-
- /* Estimate a size for the allocation area based on the
- * information available. We might end up going slightly under
- * or over the suggested heap size, but we should be pretty
- * close on average.
- *
- * Formula: suggested - needed
- * ----------------------------
- * 1 + g0s0_pcnt_kept/100
- *
- * where 'needed' is the amount of memory needed at the next
- * collection for collecting all steps except g0s0.
- */
- blocks =
- (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
- (100 + (long)g0s0_pcnt_kept);
-
- if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
- blocks = RtsFlags.GcFlags.minAllocAreaSize;
- }
-
- resizeNurseries((nat)blocks);
-
- } else {
- // we might have added extra large blocks to the nursery, so
- // resize back to minAllocAreaSize again.
- resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
- }
- }
+ resize_nursery();
// mark the garbage collected CAFs as dead
#if 0 && defined(DEBUG) // doesn't work at the moment
}
}
+/* -----------------------------------------------------------------------------
+ Figure out which generation to collect, initialise N and major_gc.
+ -------------------------------------------------------------------------- */
+
+static void
+initialise_N (rtsBool force_major_gc)
+{
+ nat g;
+
+ if (force_major_gc) {
+ N = RtsFlags.GcFlags.generations - 1;
+ major_gc = rtsTrue;
+ } else {
+ N = 0;
+ for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
+ if (generations[g].steps[0].n_blocks +
+ generations[g].steps[0].n_large_blocks
+ >= generations[g].max_blocks) {
+ N = g;
+ }
+ }
+ major_gc = (N == RtsFlags.GcFlags.generations-1);
+ }
+}
+
+/* -----------------------------------------------------------------------------
+ Initialise the gc_thread structures.
+ -------------------------------------------------------------------------- */
+
+static void
+alloc_gc_thread (gc_thread *t, int n)
+{
+ nat g, s;
+ step_workspace *ws;
+
+ t->thread_index = n;
+ t->free_blocks = NULL;
+ t->gc_count = 0;
+
+ init_gc_thread(t);
+
+ t->steps = stgMallocBytes(RtsFlags.GcFlags.generations *
+ sizeof(step_workspace *),
+ "initialise_gc_thread");
+
+ for (g = 0; g < RtsFlags.GcFlags.generations; g++)
+ {
+ t->steps[g] = stgMallocBytes(generations[g].n_steps *
+ sizeof(step_workspace),
+ "initialise_gc_thread/2");
+
+ for (s = 0; s < generations[g].n_steps; s++)
+ {
+ ws = &t->steps[g][s];
+ ws->stp = &generations[g].steps[s];
+ ws->gct = t;
+
+ ws->scan_bd = NULL;
+ ws->scan = NULL;
+
+ ws->todo_bd = NULL;
+ ws->buffer_todo_bd = NULL;
+
+ ws->scavd_list = NULL;
+ ws->n_scavd_blocks = 0;
+ }
+ }
+}
+
+
+static void
+alloc_gc_threads (void)
+{
+ if (gc_threads == NULL) {
+#if defined(THREADED_RTS)
+ nat i;
+
+ gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
+ sizeof(gc_thread),
+ "alloc_gc_threads");
+
+ for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
+ alloc_gc_thread(&gc_threads[i], i);
+ }
+#else
+ gc_threads = stgMallocBytes (sizeof(gc_thread),
+ "alloc_gc_threads");
+
+ alloc_gc_thread(gc_threads, 0);
+#endif
+ }
+}
+
+/* ----------------------------------------------------------------------------
+ Initialise a generation that is to be collected
+ ------------------------------------------------------------------------- */
+
+static void
+init_collected_gen (nat g, nat n_threads)
+{
+ nat s, t, i;
+ step_workspace *ws;
+ step *stp;
+ bdescr *bd;
+
+ // Throw away the current mutable list. Invariant: the mutable
+ // list always has at least one block; this means we can avoid a
+ // check for NULL in recordMutable().
+ if (g != 0) {
+ freeChain(generations[g].mut_list);
+ generations[g].mut_list = allocBlock();
+ for (i = 0; i < n_capabilities; i++) {
+ freeChain(capabilities[i].mut_lists[g]);
+ capabilities[i].mut_lists[g] = allocBlock();
+ }
+ }
+
+ for (s = 0; s < generations[g].n_steps; s++) {
+
+ // generation 0, step 0 doesn't need to-space
+ if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
+ continue;
+ }
+
+ stp = &generations[g].steps[s];
+ ASSERT(stp->gen_no == g);
+
+ // deprecate the existing blocks
+ stp->old_blocks = stp->blocks;
+ stp->n_old_blocks = stp->n_blocks;
+ stp->blocks = NULL;
+ stp->n_blocks = 0;
+
+ // we don't have any to-be-scavenged blocks yet
+ stp->todos = NULL;
+ stp->n_todos = 0;
+
+ // initialise the large object queues.
+ stp->scavenged_large_objects = NULL;
+ stp->n_scavenged_large_blocks = 0;
+
+ // mark the large objects as not evacuated yet
+ for (bd = stp->large_objects; bd; bd = bd->link) {
+ bd->flags &= ~BF_EVACUATED;
+ }
+
+ // for a compacted step, we need to allocate the bitmap
+ if (stp->is_compacted) {
+ nat bitmap_size; // in bytes
+ bdescr *bitmap_bdescr;
+ StgWord *bitmap;
+
+ bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
+
+ if (bitmap_size > 0) {
+ bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
+ / BLOCK_SIZE);
+ stp->bitmap = bitmap_bdescr;
+ bitmap = bitmap_bdescr->start;
+
+ debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
+ bitmap_size, bitmap);
+
+ // don't forget to fill it with zeros!
+ memset(bitmap, 0, bitmap_size);
+
+ // For each block in this step, point to its bitmap from the
+ // block descriptor.
+ for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
+ bd->u.bitmap = bitmap;
+ bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
+
+ // Also at this point we set the BF_COMPACTED flag
+ // for this block. The invariant is that
+ // BF_COMPACTED is always unset, except during GC
+ // when it is set on those blocks which will be
+ // compacted.
+ bd->flags |= BF_COMPACTED;
+ }
+ }
+ }
+ }
+
+ // For each GC thread, for each step, allocate a "todo" block to
+ // store evacuated objects to be scavenged, and a block to store
+ // evacuated objects that do not need to be scavenged.
+ for (t = 0; t < n_threads; t++) {
+ for (s = 0; s < generations[g].n_steps; s++) {
+
+ // we don't copy objects into g0s0, unless -G0
+ if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
+
+ ws = &gc_threads[t].steps[g][s];
+
+ ws->scan_bd = NULL;
+ ws->scan = NULL;
+
+ ws->todo_large_objects = NULL;
+
+ // allocate the first to-space block; extra blocks will be
+ // chained on as necessary.
+ ws->todo_bd = NULL;
+ ws->buffer_todo_bd = NULL;
+ gc_alloc_todo_block(ws);
+
+ // allocate a block for "already scavenged" objects. This goes
+ // on the front of the stp->blocks list, so it won't be
+ // traversed by the scavenging sweep.
+ ws->scavd_list = NULL;
+ ws->n_scavd_blocks = 0;
+ gc_alloc_scavd_block(ws);
+ }
+ }
+}
+
+
+/* ----------------------------------------------------------------------------
+ Initialise a generation that is *not* to be collected
+ ------------------------------------------------------------------------- */
+
+static void
+init_uncollected_gen (nat g, nat threads)
+{
+ nat s, t, i;
+ step_workspace *ws;
+ step *stp;
+ bdescr *bd;
+
+ for (s = 0; s < generations[g].n_steps; s++) {
+ stp = &generations[g].steps[s];
+ stp->scavenged_large_objects = NULL;
+ stp->n_scavenged_large_blocks = 0;
+ }
+
+ for (t = 0; t < threads; t++) {
+ for (s = 0; s < generations[g].n_steps; s++) {
+
+ ws = &gc_threads[t].steps[g][s];
+ stp = ws->stp;
+
+ ws->buffer_todo_bd = NULL;
+ ws->todo_large_objects = NULL;
+
+ // If the block at the head of the list in this generation
+ // is less than 3/4 full, then use it as a todo block.
+ if (isPartiallyFull(stp->blocks))
+ {
+ ws->todo_bd = stp->blocks;
+ stp->blocks = stp->blocks->link;
+ stp->n_blocks -= 1;
+ ws->todo_bd->link = NULL;
+
+ // this block is also the scan block; we must scan
+ // from the current end point.
+ ws->scan_bd = ws->todo_bd;
+ ws->scan = ws->scan_bd->free;
+
+ // subtract the contents of this block from the stats,
+ // because we'll count the whole block later.
+ copied -= ws->scan_bd->free - ws->scan_bd->start;
+ }
+ else
+ {
+ ws->scan_bd = NULL;
+ ws->scan = NULL;
+ ws->todo_bd = NULL;
+ gc_alloc_todo_block(ws);
+ }
+
+ // Do the same trick for the scavd block
+ if (isPartiallyFull(stp->blocks))
+ {
+ ws->scavd_list = stp->blocks;
+ stp->blocks = stp->blocks->link;
+ stp->n_blocks -= 1;
+ ws->scavd_list->link = NULL;
+ ws->n_scavd_blocks = 1;
+ // subtract the contents of this block from the stats,
+ // because we'll count the whole block later.
+ scavd_copied -= ws->scavd_list->free - ws->scavd_list->start;
+ }
+ else
+ {
+ ws->scavd_list = NULL;
+ ws->n_scavd_blocks = 0;
+ gc_alloc_scavd_block(ws);
+ }
+ }
+ }
+
+ // Move the private mutable lists from each capability onto the
+ // main mutable list for the generation.
+ for (i = 0; i < n_capabilities; i++) {
+ for (bd = capabilities[i].mut_lists[g];
+ bd->link != NULL; bd = bd->link) {
+ /* nothing */
+ }
+ bd->link = generations[g].mut_list;
+ generations[g].mut_list = capabilities[i].mut_lists[g];
+ capabilities[i].mut_lists[g] = allocBlock();
+ }
+}
+
+/* -----------------------------------------------------------------------------
+ Initialise a gc_thread before GC
+ -------------------------------------------------------------------------- */
+
+static void
+init_gc_thread (gc_thread *t)
+{
+ t->evac_gen = 0;
+ t->failed_to_evac = rtsFalse;
+ t->eager_promotion = rtsTrue;
+ t->thunk_selector_depth = 0;
+}
+
+/* -----------------------------------------------------------------------------
+ Function we pass to GetRoots to evacuate roots.
+ -------------------------------------------------------------------------- */
+
static void
mark_root(StgClosure **root)
{
}
}
+/* ----------------------------------------------------------------------------
+ Update the pointers from the task list
+
+ These are treated as weak pointers because we want to allow a main
+ thread to get a BlockedOnDeadMVar exception in the same way as any
+ other thread. Note that the threads should all have been retained
+ by GC by virtue of being on the all_threads list, we're just
+ updating pointers here.
+ ------------------------------------------------------------------------- */
+
+static void
+update_task_list (void)
+{
+ Task *task;
+ StgTSO *tso;
+ for (task = all_tasks; task != NULL; task = task->all_link) {
+ if (!task->stopped && task->tso) {
+ ASSERT(task->tso->bound == task);
+ tso = (StgTSO *) isAlive((StgClosure *)task->tso);
+ if (tso == NULL) {
+ barf("task %p: main thread %d has been GC'd",
+#ifdef THREADED_RTS
+ (void *)task->id,
+#else
+ (void *)task,
+#endif
+ task->tso->id);
+ }
+ task->tso = tso;
+ }
+ }
+}
+
+/* ----------------------------------------------------------------------------
+ Reset the sizes of the older generations when we do a major
+ collection.
+
+ CURRENT STRATEGY: make all generations except zero the same size.
+ We have to stay within the maximum heap size, and leave a certain
+ percentage of the maximum heap size available to allocate into.
+ ------------------------------------------------------------------------- */
+
+static void
+resize_generations (void)
+{
+ nat g;
+
+ if (major_gc && RtsFlags.GcFlags.generations > 1) {
+ nat live, size, min_alloc;
+ nat max = RtsFlags.GcFlags.maxHeapSize;
+ nat gens = RtsFlags.GcFlags.generations;
+
+ // live in the oldest generations
+ live = oldest_gen->steps[0].n_blocks +
+ oldest_gen->steps[0].n_large_blocks;
+
+ // default max size for all generations except zero
+ size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
+ RtsFlags.GcFlags.minOldGenSize);
+
+ // minimum size for generation zero
+ min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
+ RtsFlags.GcFlags.minAllocAreaSize);
+
+ // Auto-enable compaction when the residency reaches a
+ // certain percentage of the maximum heap size (default: 30%).
+ if (RtsFlags.GcFlags.generations > 1 &&
+ (RtsFlags.GcFlags.compact ||
+ (max > 0 &&
+ oldest_gen->steps[0].n_blocks >
+ (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
+ oldest_gen->steps[0].is_compacted = 1;
+// debugBelch("compaction: on\n", live);
+ } else {
+ oldest_gen->steps[0].is_compacted = 0;
+// debugBelch("compaction: off\n", live);
+ }
+
+ // if we're going to go over the maximum heap size, reduce the
+ // size of the generations accordingly. The calculation is
+ // different if compaction is turned on, because we don't need
+ // to double the space required to collect the old generation.
+ if (max != 0) {
+
+ // this test is necessary to ensure that the calculations
+ // below don't have any negative results - we're working
+ // with unsigned values here.
+ if (max < min_alloc) {
+ heapOverflow();
+ }
+
+ if (oldest_gen->steps[0].is_compacted) {
+ if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
+ size = (max - min_alloc) / ((gens - 1) * 2 - 1);
+ }
+ } else {
+ if ( (size * (gens - 1) * 2) + min_alloc > max ) {
+ size = (max - min_alloc) / ((gens - 1) * 2);
+ }
+ }
+
+ if (size < live) {
+ heapOverflow();
+ }
+ }
+
+#if 0
+ debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
+ min_alloc, size, max);
+#endif
+
+ for (g = 0; g < gens; g++) {
+ generations[g].max_blocks = size;
+ }
+ }
+}
+
+/* -----------------------------------------------------------------------------
+ Calculate the new size of the nursery, and resize it.
+ -------------------------------------------------------------------------- */
+
+static void
+resize_nursery (void)
+{
+ if (RtsFlags.GcFlags.generations == 1)
+ { // Two-space collector:
+ nat blocks;
+
+ /* set up a new nursery. Allocate a nursery size based on a
+ * function of the amount of live data (by default a factor of 2)
+ * Use the blocks from the old nursery if possible, freeing up any
+ * left over blocks.
+ *
+ * If we get near the maximum heap size, then adjust our nursery
+ * size accordingly. If the nursery is the same size as the live
+ * data (L), then we need 3L bytes. We can reduce the size of the
+ * nursery to bring the required memory down near 2L bytes.
+ *
+ * A normal 2-space collector would need 4L bytes to give the same
+ * performance we get from 3L bytes, reducing to the same
+ * performance at 2L bytes.
+ */
+ blocks = g0s0->n_old_blocks;
+
+ if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
+ blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
+ RtsFlags.GcFlags.maxHeapSize )
+ {
+ long adjusted_blocks; // signed on purpose
+ int pc_free;
+
+ adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
+
+ debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
+ RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
+
+ pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
+ if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
+ {
+ heapOverflow();
+ }
+ blocks = adjusted_blocks;
+ }
+ else
+ {
+ blocks *= RtsFlags.GcFlags.oldGenFactor;
+ if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
+ {
+ blocks = RtsFlags.GcFlags.minAllocAreaSize;
+ }
+ }
+ resizeNurseries(blocks);
+ }
+ else // Generational collector
+ {
+ /*
+ * If the user has given us a suggested heap size, adjust our
+ * allocation area to make best use of the memory available.
+ */
+ if (RtsFlags.GcFlags.heapSizeSuggestion)
+ {
+ long blocks;
+ nat needed = calcNeeded(); // approx blocks needed at next GC
+
+ /* Guess how much will be live in generation 0 step 0 next time.
+ * A good approximation is obtained by finding the
+ * percentage of g0s0 that was live at the last minor GC.
+ *
+ * We have an accurate figure for the amount of copied data in
+ * 'copied', but we must convert this to a number of blocks, with
+ * a small adjustment for estimated slop at the end of a block
+ * (- 10 words).
+ */
+ if (N == 0)
+ {
+ g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
+ / countNurseryBlocks();
+ }
+
+ /* Estimate a size for the allocation area based on the
+ * information available. We might end up going slightly under
+ * or over the suggested heap size, but we should be pretty
+ * close on average.
+ *
+ * Formula: suggested - needed
+ * ----------------------------
+ * 1 + g0s0_pcnt_kept/100
+ *
+ * where 'needed' is the amount of memory needed at the next
+ * collection for collecting all steps except g0s0.
+ */
+ blocks =
+ (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
+ (100 + (long)g0s0_pcnt_kept);
+
+ if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
+ blocks = RtsFlags.GcFlags.minAllocAreaSize;
+ }
+
+ resizeNurseries((nat)blocks);
+ }
+ else
+ {
+ // we might have added extra large blocks to the nursery, so
+ // resize back to minAllocAreaSize again.
+ resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
+ }
+ }
+}
+
/* -----------------------------------------------------------------------------
Sanity code for CAF garbage collection.
debugTrace(DEBUG_gccafs, "%d CAFs live", i);
}
#endif
-
-/* -----------------------------------------------------------------------------
- * Debugging
- * -------------------------------------------------------------------------- */
-
-#if DEBUG
-void
-printMutableList(generation *gen)
-{
- bdescr *bd;
- StgPtr p;
-
- debugBelch("mutable list %p: ", gen->mut_list);
-
- for (bd = gen->mut_list; bd != NULL; bd = bd->link) {
- for (p = bd->start; p < bd->free; p++) {
- debugBelch("%p (%s), ", (void *)*p, info_type((StgClosure *)*p));
- }
- }
- debugBelch("\n");
-}
-#endif /* DEBUG */
projects / ghc-hetmet.git / blobdiff