/* -----------------------------------------------------------------------------
- * $Id: Storage.c,v 1.41 2001/07/23 17:23:20 simonmar Exp $
+ * $Id: Storage.c,v 1.58 2002/01/24 01:45:55 sof Exp $
*
* (c) The GHC Team, 1998-1999
*
*
* ---------------------------------------------------------------------------*/
+#include "PosixSource.h"
#include "Rts.h"
#include "RtsUtils.h"
#include "RtsFlags.h"
#include "MBlock.h"
#include "Weak.h"
#include "Sanity.h"
+#include "Arena.h"
#include "Storage.h"
#include "Schedule.h"
#include "StoragePriv.h"
-#ifndef SMP
-nat nursery_blocks; /* number of blocks in the nursery */
-#endif
+#include "RetainerProfile.h" // for counting memory blocks (memInventory)
StgClosure *caf_list = NULL;
bdescr *small_alloc_list; /* allocate()d small objects */
bdescr *large_alloc_list; /* allocate()d large objects */
+bdescr *pinned_object_block; /* allocate pinned objects into this block */
nat alloc_blocks; /* number of allocate()d blocks since GC */
nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
static void stgDeallocForGMP (void *ptr, size_t size);
void
-initStorage (void)
+initStorage( void )
{
nat g, s;
step *stp;
generation *gen;
- /* If we're doing heap profiling, we want a two-space heap with a
- * fixed-size allocation area so that we get roughly even-spaced
- * samples.
- */
-
- /* As an experiment, try a 2 generation collector
- */
-
-#if defined(PROFILING) || defined(DEBUG)
- if (RtsFlags.ProfFlags.doHeapProfile) {
- RtsFlags.GcFlags.generations = 1;
- RtsFlags.GcFlags.steps = 1;
- RtsFlags.GcFlags.oldGenFactor = 0;
- RtsFlags.GcFlags.heapSizeSuggestion = 0;
+ if (RtsFlags.GcFlags.maxHeapSize != 0 &&
+ RtsFlags.GcFlags.heapSizeSuggestion >
+ RtsFlags.GcFlags.maxHeapSize) {
+ RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
}
-#endif
- if (RtsFlags.GcFlags.heapSizeSuggestion >
+ if (RtsFlags.GcFlags.maxHeapSize != 0 &&
+ RtsFlags.GcFlags.minAllocAreaSize >
RtsFlags.GcFlags.maxHeapSize) {
- RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
+ prog_belch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
+ exit(1);
}
initBlockAllocator();
stp->scan = NULL;
stp->scan_bd = NULL;
stp->large_objects = NULL;
+ stp->n_large_blocks = 0;
stp->new_large_objects = NULL;
stp->scavenged_large_objects = NULL;
+ stp->n_scavenged_large_blocks = 0;
stp->is_compacted = 0;
+ stp->bitmap = NULL;
}
}
/* The oldest generation has one step and it is compacted. */
if (RtsFlags.GcFlags.compact) {
- oldest_gen->steps[0].is_compacted = 1;
+ if (RtsFlags.GcFlags.generations == 1) {
+ belch("WARNING: compaction is incompatible with -G1; disabled");
+ } else {
+ oldest_gen->steps[0].is_compacted = 1;
+ }
}
oldest_gen->steps[0].to = &oldest_gen->steps[0];
g0s0->blocks = NULL;
g0s0->n_blocks = 0;
for (cap = free_capabilities; cap != NULL; cap = cap->link) {
- cap->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
- cap->rCurrentNursery = cap->rNursery;
- for (bd = cap->rNursery; bd != NULL; bd = bd->link) {
+ cap->r.rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
+ cap->r.rCurrentNursery = cap->r.rNursery;
+ for (bd = cap->r.rNursery; bd != NULL; bd = bd->link) {
bd->u.back = (bdescr *)cap;
}
}
*/
}
#else /* SMP */
- nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
- g0s0->blocks = allocNursery(NULL, nursery_blocks);
- g0s0->n_blocks = nursery_blocks;
+ g0s0->blocks = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
+ g0s0->n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
g0s0->to_blocks = NULL;
g0s0->n_to_blocks = 0;
- MainRegTable.rNursery = g0s0->blocks;
- MainRegTable.rCurrentNursery = g0s0->blocks;
+ MainCapability.r.rNursery = g0s0->blocks;
+ MainCapability.r.rCurrentNursery = g0s0->blocks;
/* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
#endif
}
ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
for (cap = free_capabilities; cap != NULL; cap = cap->link) {
- for (bd = cap->rNursery; bd; bd = bd->link) {
+ for (bd = cap->r.rNursery; bd; bd = bd->link) {
bd->free = bd->start;
ASSERT(bd->gen_no == 0);
ASSERT(bd->step == g0s0);
IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
}
- cap->rCurrentNursery = cap->rNursery;
+ cap->r.rCurrentNursery = cap->r.rNursery;
}
#else
for (bd = g0s0->blocks; bd; bd = bd->link) {
ASSERT(bd->step == g0s0);
IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
}
- MainRegTable.rNursery = g0s0->blocks;
- MainRegTable.rCurrentNursery = g0s0->blocks;
+ MainCapability.r.rNursery = g0s0->blocks;
+ MainCapability.r.rCurrentNursery = g0s0->blocks;
#endif
}
bdescr *
-allocNursery (bdescr *last_bd, nat blocks)
+allocNursery (bdescr *tail, nat blocks)
{
bdescr *bd;
nat i;
- /* Allocate a nursery */
+ // Allocate a nursery: we allocate fresh blocks one at a time and
+ // cons them on to the front of the list, not forgetting to update
+ // the back pointer on the tail of the list to point to the new block.
for (i=0; i < blocks; i++) {
+ // @LDV profiling
+ /*
+ processNursery() in LdvProfile.c assumes that every block group in
+ the nursery contains only a single block. So, if a block group is
+ given multiple blocks, change processNursery() accordingly.
+ */
bd = allocBlock();
- bd->link = last_bd;
+ bd->link = tail;
+ // double-link the nursery: we might need to insert blocks
+ if (tail != NULL) {
+ tail->u.back = bd;
+ }
bd->step = g0s0;
bd->gen_no = 0;
bd->flags = 0;
bd->free = bd->start;
- last_bd = bd;
+ tail = bd;
}
- return last_bd;
+ tail->u.back = NULL;
+ return tail;
}
void
resizeNursery ( nat blocks )
{
bdescr *bd;
+ nat nursery_blocks;
#ifdef SMP
barf("resizeNursery: can't resize in SMP mode");
#endif
+ nursery_blocks = g0s0->n_blocks;
if (nursery_blocks == blocks) {
- ASSERT(g0s0->n_blocks == blocks);
return;
}
IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
blocks));
- for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) {
- next_bd = bd->link;
- freeGroup(bd);
- bd = next_bd;
+
+ bd = g0s0->blocks;
+ while (nursery_blocks > blocks) {
+ next_bd = bd->link;
+ next_bd->u.back = NULL;
+ nursery_blocks -= bd->blocks; // might be a large block
+ freeGroup(bd);
+ bd = next_bd;
}
g0s0->blocks = bd;
+ // might have gone just under, by freeing a large block, so make
+ // up the difference.
+ if (nursery_blocks < blocks) {
+ g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
+ }
}
- g0s0->n_blocks = nursery_blocks = blocks;
+ g0s0->n_blocks = blocks;
+ ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
}
/* -----------------------------------------------------------------------------
-------------------------------------------------------------------------- */
StgPtr
-allocate(nat n)
+allocate( nat n )
{
bdescr *bd;
StgPtr p;
return p;
}
-lnat allocated_bytes(void)
+lnat
+allocated_bytes( void )
{
return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
}
+/* ---------------------------------------------------------------------------
+ Allocate a fixed/pinned object.
+
+ We allocate small pinned objects into a single block, allocating a
+ new block when the current one overflows. The block is chained
+ onto the large_object_list of generation 0 step 0.
+
+ NOTE: The GC can't in general handle pinned objects. This
+ interface is only safe to use for ByteArrays, which have no
+ pointers and don't require scavenging. It works because the
+ block's descriptor has the BF_LARGE flag set, so the block is
+ treated as a large object and chained onto various lists, rather
+ than the individual objects being copied. However, when it comes
+ to scavenge the block, the GC will only scavenge the first object.
+ The reason is that the GC can't linearly scan a block of pinned
+ objects at the moment (doing so would require using the
+ mostly-copying techniques). But since we're restricting ourselves
+ to pinned ByteArrays, not scavenging is ok.
+
+ This function is called by newPinnedByteArray# which immediately
+ fills the allocated memory with a MutableByteArray#.
+ ------------------------------------------------------------------------- */
+
+StgPtr
+allocatePinned( nat n )
+{
+ StgPtr p;
+ bdescr *bd = pinned_object_block;
+
+ ACQUIRE_LOCK(&sm_mutex);
+
+ TICK_ALLOC_HEAP_NOCTR(n);
+ CCS_ALLOC(CCCS,n);
+
+ // If the request is for a large object, then allocate()
+ // will give us a pinned object anyway.
+ if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
+ RELEASE_LOCK(&sm_mutex);
+ return allocate(n);
+ }
+
+ // If we don't have a block of pinned objects yet, or the current
+ // one isn't large enough to hold the new object, allocate a new one.
+ if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
+ pinned_object_block = bd = allocBlock();
+ dbl_link_onto(bd, &g0s0->large_objects);
+ bd->gen_no = 0;
+ bd->step = g0s0;
+ bd->flags = BF_LARGE;
+ bd->free = bd->start;
+ alloc_blocks++;
+ }
+
+ p = bd->free;
+ bd->free += n;
+ RELEASE_LOCK(&sm_mutex);
+ return p;
+}
+
/* -----------------------------------------------------------------------------
Allocation functions for GMP.
+ allocated_bytes();
for (cap = free_capabilities; cap != NULL; cap = cap->link) {
- for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
+ for ( bd = cap->r.rCurrentNursery->link; bd != NULL; bd = bd->link ) {
allocated -= BLOCK_SIZE_W;
}
- if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
+ if (cap->r.rCurrentNursery->free < cap->r.rCurrentNursery->start
+ BLOCK_SIZE_W) {
- allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
- - cap->rCurrentNursery->free;
+ allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
+ - cap->r.rCurrentNursery->free;
}
}
#else /* !SMP */
- bdescr *current_nursery = MainRegTable.rCurrentNursery;
+ bdescr *current_nursery = MainCapability.r.rCurrentNursery;
- allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
+ allocated = (g0s0->n_blocks * BLOCK_SIZE_W) + allocated_bytes();
for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
allocated -= BLOCK_SIZE_W;
}
continue;
}
stp = &generations[g].steps[s];
- live += (stp->n_blocks - 1) * BLOCK_SIZE_W;
+ live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
if (stp->hp_bd != NULL) {
live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
/ sizeof(W_);
extern lnat
calcNeeded(void)
{
- lnat needed = 0;
- nat g, s;
- step *stp;
-
- for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
- for (s = 0; s < generations[g].n_steps; s++) {
- if (g == 0 && s == 0) { continue; }
- stp = &generations[g].steps[s];
- if (generations[g].steps[0].n_blocks > generations[g].max_blocks
- && stp->is_compacted == 0) {
- needed += 2 * stp->n_blocks;
- } else {
- needed += stp->n_blocks;
- }
+ lnat needed = 0;
+ nat g, s;
+ step *stp;
+
+ for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
+ for (s = 0; s < generations[g].n_steps; s++) {
+ if (g == 0 && s == 0) { continue; }
+ stp = &generations[g].steps[s];
+ if (generations[g].steps[0].n_blocks +
+ generations[g].steps[0].n_large_blocks
+ > generations[g].max_blocks
+ && stp->is_compacted == 0) {
+ needed += 2 * stp->n_blocks;
+ } else {
+ needed += stp->n_blocks;
+ }
+ }
}
- }
- return needed;
+ return needed;
}
/* -----------------------------------------------------------------------------
for (bd = large_alloc_list; bd; bd = bd->link) {
total_blocks += bd->blocks;
}
-
+
+#ifdef PROFILING
+ if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
+ for (bd = firstStack; bd != NULL; bd = bd->link)
+ total_blocks += bd->blocks;
+ }
+#endif
+
+ // count the blocks allocated by the arena allocator
+ total_blocks += arenaBlocks();
+
/* count the blocks on the free list */
free_blocks = countFreeList();
ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
}
-static nat
+
+nat
countBlocks(bdescr *bd)
{
nat n;
for (n=0; bd != NULL; bd=bd->link) {
- n++;
+ n += bd->blocks;
}
return n;
}
for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
for (s = 0; s < generations[g].n_steps; s++) {
- if (g == 0 && s == 0) { continue; }
- checkHeap(generations[g].steps[s].blocks);
ASSERT(countBlocks(generations[g].steps[s].blocks)
== generations[g].steps[s].n_blocks);
+ ASSERT(countBlocks(generations[g].steps[s].large_objects)
+ == generations[g].steps[s].n_large_blocks);
+ if (g == 0 && s == 0) { continue; }
+ checkHeap(generations[g].steps[s].blocks);
checkChain(generations[g].steps[s].large_objects);
if (g > 0) {
checkMutableList(generations[g].mut_list, g);
}
}
+// handy function for use in gdb, because Bdescr() is inlined.
+extern bdescr *_bdescr( StgPtr p );
+
+bdescr *
+_bdescr( StgPtr p )
+{
+ return Bdescr(p);
+}
+
#endif