/* -----------------------------------------------------------------------------
- * $Id: Storage.c,v 1.49 2001/08/14 13:40:09 sewardj Exp $
*
- * (c) The GHC Team, 1998-1999
+ * (c) The GHC Team, 1998-2004
*
* Storage manager front end
*
#include "MBlock.h"
#include "Weak.h"
#include "Sanity.h"
-
+#include "Arena.h"
+#include "OSThreads.h"
+#include "Capability.h"
#include "Storage.h"
#include "Schedule.h"
-#include "StoragePriv.h"
+#include "RetainerProfile.h" // for counting memory blocks (memInventory)
-#ifndef SMP
-nat nursery_blocks; /* number of blocks in the nursery */
-#endif
+#include <stdlib.h>
+#include <string.h>
StgClosure *caf_list = NULL;
+StgClosure *revertible_caf_list = NULL;
+rtsBool keepCAFs;
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 */
StgPtr alloc_Hp = NULL; /* next free byte in small_alloc_list */
StgPtr alloc_HpLim = NULL; /* end of block at small_alloc_list */
-generation *generations; /* all the generations */
-generation *g0; /* generation 0, for convenience */
-generation *oldest_gen; /* oldest generation, for convenience */
-step *g0s0; /* generation 0, step 0, for convenience */
+generation *generations = NULL; /* all the generations */
+generation *g0 = NULL; /* generation 0, for convenience */
+generation *oldest_gen = NULL; /* oldest generation, for convenience */
+step *g0s0 = NULL; /* generation 0, step 0, for convenience */
-lnat total_allocated = 0; /* total memory allocated during run */
+ullong total_allocated = 0; /* total memory allocated during run */
+
+nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
+step *nurseries = NULL; /* array of nurseries, >1 only if SMP */
/*
* Storage manager mutex: protects all the above state from
* simultaneous access by two STG threads.
*/
#ifdef SMP
-pthread_mutex_t sm_mutex = PTHREAD_MUTEX_INITIALIZER;
+Mutex sm_mutex = INIT_MUTEX_VAR;
#endif
/*
static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
static void stgDeallocForGMP (void *ptr, size_t size);
+static void
+initStep (step *stp, int g, int s)
+{
+ stp->no = s;
+ stp->blocks = NULL;
+ stp->n_to_blocks = 0;
+ stp->n_blocks = 0;
+ stp->gen = &generations[g];
+ stp->gen_no = g;
+ stp->hp = NULL;
+ stp->hpLim = NULL;
+ stp->hp_bd = NULL;
+ 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;
+}
+
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 (generations != NULL) {
+ // multi-init protection
+ return;
}
-#endif
+ /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
+ * doing something reasonable.
+ */
+ ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info));
+ ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
+ ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
+
if (RtsFlags.GcFlags.maxHeapSize != 0 &&
RtsFlags.GcFlags.heapSizeSuggestion >
RtsFlags.GcFlags.maxHeapSize) {
RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
}
+ if (RtsFlags.GcFlags.maxHeapSize != 0 &&
+ RtsFlags.GcFlags.minAllocAreaSize >
+ RtsFlags.GcFlags.maxHeapSize) {
+ errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
+ exit(1);
+ }
+
initBlockAllocator();
+#if defined(SMP)
+ initMutex(&sm_mutex);
+#endif
+
/* allocate generation info array */
generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
- * sizeof(struct _generation),
+ * sizeof(struct generation_),
"initStorage: gens");
/* Initialise all generations */
for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
gen = &generations[g];
gen->no = g;
- gen->mut_list = END_MUT_LIST;
- gen->mut_once_list = END_MUT_LIST;
+ gen->mut_list = allocBlock();
gen->collections = 0;
gen->failed_promotions = 0;
gen->max_blocks = 0;
/* Oldest generation: one step */
oldest_gen->n_steps = 1;
oldest_gen->steps =
- stgMallocBytes(1 * sizeof(struct _step), "initStorage: last step");
+ stgMallocBytes(1 * sizeof(struct step_), "initStorage: last step");
/* set up all except the oldest generation with 2 steps */
for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
generations[g].n_steps = RtsFlags.GcFlags.steps;
generations[g].steps =
- stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct _step),
+ stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct step_),
"initStorage: steps");
}
} else {
/* single generation, i.e. a two-space collector */
g0->n_steps = 1;
- g0->steps = stgMallocBytes (sizeof(struct _step), "initStorage: steps");
+ g0->steps = stgMallocBytes (sizeof(struct step_), "initStorage: steps");
}
+#ifdef SMP
+ n_nurseries = RtsFlags.ParFlags.nNodes;
+ nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
+ "initStorage: nurseries");
+#else
+ n_nurseries = 1;
+ nurseries = g0->steps; // just share nurseries[0] with g0s0
+#endif
+
/* Initialise all steps */
for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
for (s = 0; s < generations[g].n_steps; s++) {
- stp = &generations[g].steps[s];
- stp->no = s;
- stp->blocks = NULL;
- stp->n_blocks = 0;
- stp->gen = &generations[g];
- stp->gen_no = g;
- stp->hp = NULL;
- stp->hpLim = NULL;
- stp->hp_bd = NULL;
- 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;
+ initStep(&generations[g].steps[s], g, s);
}
}
+#ifdef SMP
+ for (s = 0; s < n_nurseries; s++) {
+ initStep(&nurseries[s], 0, s);
+ }
+#endif
+
/* Set up the destination pointers in each younger gen. step */
for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
for (s = 0; s < generations[g].n_steps-1; s++) {
}
generations[g].steps[s].to = &generations[g+1].steps[0];
}
+ oldest_gen->steps[0].to = &oldest_gen->steps[0];
- /* The oldest generation has one step and it is compacted. */
+#ifdef SMP
+ for (s = 0; s < n_nurseries; s++) {
+ nurseries[s].to = generations[0].steps[0].to;
+ }
+#endif
+
+ /* The oldest generation has one step. */
if (RtsFlags.GcFlags.compact) {
- oldest_gen->steps[0].is_compacted = 1;
+ if (RtsFlags.GcFlags.generations == 1) {
+ errorBelch("WARNING: compaction is incompatible with -G1; disabled");
+ } else {
+ oldest_gen->steps[0].is_compacted = 1;
+ }
}
- oldest_gen->steps[0].to = &oldest_gen->steps[0];
+
+#ifdef SMP
+ if (RtsFlags.GcFlags.generations == 1) {
+ errorBelch("-G1 is incompatible with SMP");
+ stg_exit(1);
+ }
+ // No -H, for now
+ if (RtsFlags.GcFlags.heapSizeSuggestion > 0) {
+ errorBelch("-H<size> is incompatible with SMP");
+ stg_exit(1);
+ }
+#endif
/* generation 0 is special: that's the nursery */
generations[0].max_blocks = 0;
weak_ptr_list = NULL;
caf_list = NULL;
+ revertible_caf_list = NULL;
/* initialise the allocate() interface */
small_alloc_list = NULL;
- large_alloc_list = NULL;
alloc_blocks = 0;
alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
/* Tell GNU multi-precision pkg about our custom alloc functions */
mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
-#ifdef SMP
- pthread_mutex_init(&sm_mutex, NULL);
-#endif
-
IF_DEBUG(gc, statDescribeGens());
}
void
newCAF(StgClosure* caf)
{
- /* Put this CAF on the mutable list for the old generation.
- * This is a HACK - the IND_STATIC closure doesn't really have
- * a mut_link field, but we pretend it has - in fact we re-use
- * the STATIC_LINK field for the time being, because when we
- * come to do a major GC we won't need the mut_link field
- * any more and can use it as a STATIC_LINK.
- */
- ACQUIRE_LOCK(&sm_mutex);
+ ACQUIRE_SM_LOCK;
- if (is_dynamically_loaded_rwdata_ptr((StgPtr)caf)) {
- ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
- ((StgIndStatic *)caf)->static_link = caf_list;
- caf_list = caf;
- } else {
- ((StgIndStatic *)caf)->saved_info = NULL;
- ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
- oldest_gen->mut_once_list = (StgMutClosure *)caf;
+ if(keepCAFs)
+ {
+ // HACK:
+ // If we are in GHCi _and_ we are using dynamic libraries,
+ // then we can't redirect newCAF calls to newDynCAF (see below),
+ // so we make newCAF behave almost like newDynCAF.
+ // The dynamic libraries might be used by both the interpreted
+ // program and GHCi itself, so they must not be reverted.
+ // This also means that in GHCi with dynamic libraries, CAFs are not
+ // garbage collected. If this turns out to be a problem, we could
+ // do another hack here and do an address range test on caf to figure
+ // out whether it is from a dynamic library.
+ ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
+ ((StgIndStatic *)caf)->static_link = caf_list;
+ caf_list = caf;
}
-
- RELEASE_LOCK(&sm_mutex);
+ else
+ {
+ /* Put this CAF on the mutable list for the old generation.
+ * This is a HACK - the IND_STATIC closure doesn't really have
+ * a mut_link field, but we pretend it has - in fact we re-use
+ * the STATIC_LINK field for the time being, because when we
+ * come to do a major GC we won't need the mut_link field
+ * any more and can use it as a STATIC_LINK.
+ */
+ ((StgIndStatic *)caf)->saved_info = NULL;
+ recordMutableGen(caf, oldest_gen);
+ }
+
+ RELEASE_SM_LOCK;
#ifdef PAR
/* If we are PAR or DIST then we never forget a CAF */
{ globalAddr *newGA;
- //belch("<##> Globalising CAF %08x %s",caf,info_type(caf));
+ //debugBelch("<##> Globalising CAF %08x %s",caf,info_type(caf));
newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
ASSERT(newGA);
}
#endif /* PAR */
}
+// An alternate version of newCaf which is used for dynamically loaded
+// object code in GHCi. In this case we want to retain *all* CAFs in
+// the object code, because they might be demanded at any time from an
+// expression evaluated on the command line.
+// Also, GHCi might want to revert CAFs, so we add these to the
+// revertible_caf_list.
+//
+// The linker hackily arranges that references to newCaf from dynamic
+// code end up pointing to newDynCAF.
+void
+newDynCAF(StgClosure *caf)
+{
+ ACQUIRE_SM_LOCK;
+
+ ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
+ ((StgIndStatic *)caf)->static_link = revertible_caf_list;
+ revertible_caf_list = caf;
+
+ RELEASE_SM_LOCK;
+}
+
/* -----------------------------------------------------------------------------
Nursery management.
-------------------------------------------------------------------------- */
-void
-allocNurseries( void )
-{
-#ifdef SMP
- {
- Capability *cap;
+static bdescr *
+allocNursery (step *stp, bdescr *tail, nat blocks)
+{
bdescr *bd;
+ nat i;
+
+ // 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 = tail;
+ // double-link the nursery: we might need to insert blocks
+ if (tail != NULL) {
+ tail->u.back = bd;
+ }
+ bd->step = stp;
+ bd->gen_no = 0;
+ bd->flags = 0;
+ bd->free = bd->start;
+ tail = bd;
+ }
+ tail->u.back = NULL;
+ return tail;
+}
- 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) {
- bd->u.back = (bdescr *)cap;
- }
+static void
+assignNurseriesToCapabilities (void)
+{
+#ifdef SMP
+ nat i;
+
+ for (i = 0; i < n_nurseries; i++) {
+ capabilities[i].r.rNursery = &nurseries[i];
+ capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
}
- /* Set the back links to be equal to the Capability,
- * so we can do slightly better informed locking.
- */
- }
#else /* SMP */
- nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
- g0s0->blocks = allocNursery(NULL, nursery_blocks);
- g0s0->n_blocks = nursery_blocks;
- g0s0->to_blocks = NULL;
- g0s0->n_to_blocks = 0;
- MainRegTable.rNursery = g0s0->blocks;
- MainRegTable.rCurrentNursery = g0s0->blocks;
- /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
+ MainCapability.r.rNursery = &nurseries[0];
+ MainCapability.r.rCurrentNursery = nurseries[0].blocks;
#endif
}
+
+void
+allocNurseries( void )
+{
+ nat i;
+
+ for (i = 0; i < n_nurseries; i++) {
+ nurseries[i].blocks =
+ allocNursery(&nurseries[i], NULL,
+ RtsFlags.GcFlags.minAllocAreaSize);
+ nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
+ nurseries[i].to_blocks = NULL;
+ nurseries[i].n_to_blocks = 0;
+ /* hp, hpLim, hp_bd, to_space etc. aren't used in the nursery */
+ }
+ assignNurseriesToCapabilities();
+}
void
resetNurseries( void )
{
- bdescr *bd;
-#ifdef SMP
- Capability *cap;
-
- /* All tasks must be stopped */
- ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
-
- for (cap = free_capabilities; cap != NULL; cap = cap->link) {
- for (bd = cap->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));
+ nat i;
+ bdescr *bd;
+ step *stp;
+
+ for (i = 0; i < n_nurseries; i++) {
+ stp = &nurseries[i];
+ for (bd = stp->blocks; bd; bd = bd->link) {
+ bd->free = bd->start;
+ ASSERT(bd->gen_no == 0);
+ ASSERT(bd->step == stp);
+ IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
+ }
}
- cap->rCurrentNursery = cap->rNursery;
- }
-#else
- for (bd = g0s0->blocks; 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));
- }
- MainRegTable.rNursery = g0s0->blocks;
- MainRegTable.rCurrentNursery = g0s0->blocks;
-#endif
+ assignNurseriesToCapabilities();
}
-bdescr *
-allocNursery (bdescr *last_bd, nat blocks)
+lnat
+countNurseryBlocks (void)
{
- bdescr *bd;
- nat i;
+ nat i;
+ lnat blocks = 0;
- /* Allocate a nursery */
- for (i=0; i < blocks; i++) {
- bd = allocBlock();
- bd->link = last_bd;
- bd->step = g0s0;
- bd->gen_no = 0;
- bd->flags = 0;
- bd->free = bd->start;
- last_bd = bd;
- }
- return last_bd;
+ for (i = 0; i < n_nurseries; i++) {
+ blocks += nurseries[i].n_blocks;
+ }
+ return blocks;
}
-void
-resizeNursery ( nat blocks )
+static void
+resizeNursery ( step *stp, nat blocks )
{
bdescr *bd;
+ nat nursery_blocks;
-#ifdef SMP
- barf("resizeNursery: can't resize in SMP mode");
-#endif
-
- if (nursery_blocks == blocks) {
- ASSERT(g0s0->n_blocks == blocks);
- return;
- }
+ nursery_blocks = stp->n_blocks;
+ if (nursery_blocks == blocks) return;
- else if (nursery_blocks < blocks) {
- IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
+ if (nursery_blocks < blocks) {
+ IF_DEBUG(gc, debugBelch("Increasing size of nursery to %d blocks\n",
blocks));
- g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
+ stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
}
-
else {
bdescr *next_bd;
- IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
+ IF_DEBUG(gc, debugBelch("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 = stp->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;
+ }
+ stp->blocks = bd;
+ // might have gone just under, by freeing a large block, so make
+ // up the difference.
+ if (nursery_blocks < blocks) {
+ stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
}
- g0s0->blocks = bd;
}
- g0s0->n_blocks = nursery_blocks = blocks;
+ stp->n_blocks = blocks;
+ ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
+}
+
+//
+// Resize each of the nurseries to the specified size.
+//
+void
+resizeNurseries (nat blocks)
+{
+ nat i;
+ for (i = 0; i < n_nurseries; i++) {
+ resizeNursery(&nurseries[i], blocks);
+ }
}
/* -----------------------------------------------------------------------------
bdescr *bd;
StgPtr p;
- ACQUIRE_LOCK(&sm_mutex);
+ ACQUIRE_SM_LOCK;
TICK_ALLOC_HEAP_NOCTR(n);
CCS_ALLOC(CCCS,n);
nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
bd = allocGroup(req_blocks);
dbl_link_onto(bd, &g0s0->large_objects);
+ g0s0->n_large_blocks += req_blocks;
bd->gen_no = 0;
bd->step = g0s0;
bd->flags = BF_LARGE;
- bd->free = bd->start;
- /* don't add these blocks to alloc_blocks, since we're assuming
- * that large objects are likely to remain live for quite a while
- * (eg. running threads), so garbage collecting early won't make
- * much difference.
- */
+ bd->free = bd->start + n;
alloc_blocks += req_blocks;
- RELEASE_LOCK(&sm_mutex);
+ RELEASE_SM_LOCK;
return bd->start;
/* small allocation (<LARGE_OBJECT_THRESHOLD) */
p = alloc_Hp;
alloc_Hp += n;
- RELEASE_LOCK(&sm_mutex);
+ RELEASE_SM_LOCK;
return p;
}
lnat
allocated_bytes( void )
{
- return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
+ lnat allocated;
+
+ allocated = alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp);
+ if (pinned_object_block != NULL) {
+ allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
+ pinned_object_block->free;
+ }
+
+ return allocated;
+}
+
+void
+tidyAllocateLists (void)
+{
+ if (small_alloc_list != NULL) {
+ ASSERT(alloc_Hp >= small_alloc_list->start &&
+ alloc_Hp <= small_alloc_list->start + BLOCK_SIZE);
+ small_alloc_list->free = alloc_Hp;
+ }
}
/* ---------------------------------------------------------------------------
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);
}
+ ACQUIRE_SM_LOCK;
+
+ TICK_ALLOC_HEAP_NOCTR(n);
+ CCS_ALLOC(CCCS,n);
+
+ // we always return 8-byte aligned memory. bd->free must be
+ // 8-byte aligned to begin with, so we just round up n to
+ // the nearest multiple of 8 bytes.
+ if (sizeof(StgWord) == 4) {
+ n = (n+1) & ~1;
+ }
+
// 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)) {
dbl_link_onto(bd, &g0s0->large_objects);
bd->gen_no = 0;
bd->step = g0s0;
- bd->flags = BF_LARGE;
+ bd->flags = BF_PINNED | BF_LARGE;
bd->free = bd->start;
alloc_blocks++;
}
p = bd->free;
bd->free += n;
- RELEASE_LOCK(&sm_mutex);
+ RELEASE_SM_LOCK;
return p;
}
StgArrWords* arr;
nat data_size_in_words, total_size_in_words;
- /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
- ASSERT(size_in_bytes % sizeof(W_) == 0);
-
- data_size_in_words = size_in_bytes / sizeof(W_);
+ /* round up to a whole number of words */
+ data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
/* allocate and fill it in. */
SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
/* and return a ptr to the goods inside the array */
- return(BYTE_ARR_CTS(arr));
+ return arr->payload;
}
static void *
{
nat allocated;
bdescr *bd;
+ nat i;
+ allocated = allocated_bytes();
+ for (i = 0; i < n_nurseries; i++) {
+ allocated += nurseries[i].n_blocks * BLOCK_SIZE_W;
+ }
+
#ifdef SMP
- Capability *cap;
-
- /* All tasks must be stopped. Can't assert that all the
- capabilities are owned by the scheduler, though: one or more
- tasks might have been stopped while they were running (non-main)
- threads. */
- /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
-
- allocated =
- n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
- + allocated_bytes();
-
- for (cap = free_capabilities; cap != NULL; cap = cap->link) {
- for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
- allocated -= BLOCK_SIZE_W;
- }
- if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
- + BLOCK_SIZE_W) {
- allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
- - cap->rCurrentNursery->free;
- }
+ for (i = 0; i < n_nurseries; i++) {
+ Capability *cap;
+ for ( bd = capabilities[i].r.rCurrentNursery;
+ bd != NULL; bd = bd->link ) {
+ allocated -= BLOCK_SIZE_W;
+ }
+ cap = &capabilities[i];
+ if (cap->r.rCurrentNursery->free <
+ cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
+ allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
+ - cap->r.rCurrentNursery->free;
+ }
}
+#else
+ bdescr *current_nursery = MainCapability.r.rCurrentNursery;
-#else /* !SMP */
- bdescr *current_nursery = MainRegTable.rCurrentNursery;
-
- allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
- allocated -= BLOCK_SIZE_W;
+ allocated -= BLOCK_SIZE_W;
}
if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
- allocated -= (current_nursery->start + BLOCK_SIZE_W)
- - current_nursery->free;
+ allocated -= (current_nursery->start + BLOCK_SIZE_W)
+ - current_nursery->free;
}
#endif
#ifdef DEBUG
+static lnat
+stepBlocks (step *stp)
+{
+ lnat total_blocks;
+ bdescr *bd;
+
+ total_blocks = stp->n_blocks;
+ for (bd = stp->large_objects; bd; bd = bd->link) {
+ total_blocks += bd->blocks;
+ /* hack for megablock groups: they have an extra block or two in
+ the second and subsequent megablocks where the block
+ descriptors would normally go.
+ */
+ if (bd->blocks > BLOCKS_PER_MBLOCK) {
+ total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
+ * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
+ }
+ }
+ return total_blocks;
+}
+
void
memInventory(void)
{
- nat g, s;
+ nat g, s, i;
step *stp;
bdescr *bd;
lnat total_blocks = 0, free_blocks = 0;
/* count the blocks we current have */
for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
- for (s = 0; s < generations[g].n_steps; s++) {
- stp = &generations[g].steps[s];
- total_blocks += stp->n_blocks;
- if (RtsFlags.GcFlags.generations == 1) {
- /* two-space collector has a to-space too :-) */
- total_blocks += g0s0->n_to_blocks;
+ for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
+ total_blocks += bd->blocks;
}
- for (bd = stp->large_objects; bd; bd = bd->link) {
- total_blocks += bd->blocks;
- /* hack for megablock groups: they have an extra block or two in
- the second and subsequent megablocks where the block
- descriptors would normally go.
- */
- if (bd->blocks > BLOCKS_PER_MBLOCK) {
- total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
- * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
- }
+ for (s = 0; s < generations[g].n_steps; s++) {
+ if (g==0 && s==0) continue;
+ stp = &generations[g].steps[s];
+ total_blocks += stepBlocks(stp);
}
- }
+ }
+
+ for (i = 0; i < n_nurseries; i++) {
+ total_blocks += stepBlocks(&nurseries[i]);
+ }
+
+ if (RtsFlags.GcFlags.generations == 1) {
+ /* two-space collector has a to-space too :-) */
+ total_blocks += g0s0->n_to_blocks;
}
/* any blocks held by allocate() */
for (bd = small_alloc_list; bd; bd = bd->link) {
total_blocks += bd->blocks;
}
- for (bd = large_alloc_list; bd; bd = bd->link) {
- total_blocks += bd->blocks;
+
+#ifdef PROFILING
+ if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
+ total_blocks += retainerStackBlocks();
}
-
+#endif
+
+ // count the blocks allocated by the arena allocator
+ total_blocks += arenaBlocks();
+
/* count the blocks on the free list */
free_blocks = countFreeList();
if (total_blocks + free_blocks != mblocks_allocated *
BLOCKS_PER_MBLOCK) {
- fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
+ debugBelch("Blocks: %ld live + %ld free = %ld total (%ld around)\n",
total_blocks, free_blocks, total_blocks + free_blocks,
mblocks_allocated * BLOCKS_PER_MBLOCK);
}
ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
}
-static nat
+
+nat
countBlocks(bdescr *bd)
{
nat 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);
- checkChain(generations[g].steps[s].large_objects);
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);
+ checkHeap(generations[g].steps[s].blocks);
+ checkChain(generations[g].steps[s].large_objects);
if (g > 0) {
checkMutableList(generations[g].mut_list, g);
- checkMutOnceList(generations[g].mut_once_list, g);
}
}
}
+
+ for (s = 0; s < n_nurseries; s++) {
+ ASSERT(countBlocks(nurseries[s].blocks)
+ == nurseries[s].n_blocks);
+ ASSERT(countBlocks(nurseries[s].large_objects)
+ == nurseries[s].n_large_blocks);
+ }
+
checkFreeListSanity();
}
}
+// handy function for use in gdb, because Bdescr() is inlined.
+extern bdescr *_bdescr( StgPtr p );
+
+bdescr *
+_bdescr( StgPtr p )
+{
+ return Bdescr(p);
+}
+
#endif