/* -----------------------------------------------------------------------------
- * $Id: Storage.c,v 1.21 1999/11/09 15:46:59 simonmar Exp $
+ * $Id: Storage.c,v 1.52 2001/10/18 14:41:01 simonmar Exp $
*
* (c) The GHC Team, 1998-1999
*
*
* ---------------------------------------------------------------------------*/
+#include "PosixSource.h"
#include "Rts.h"
#include "RtsUtils.h"
#include "RtsFlags.h"
#include "Hooks.h"
#include "BlockAlloc.h"
#include "MBlock.h"
-#include "gmp.h"
#include "Weak.h"
#include "Sanity.h"
+#include "Arena.h"
#include "Storage.h"
#include "Schedule.h"
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 */
generation *oldest_gen; /* oldest generation, for convenience */
step *g0s0; /* generation 0, step 0, for convenience */
+lnat total_allocated = 0; /* total memory allocated during run */
+
/*
* Storage manager mutex: protects all the above state from
* simultaneous access by two STG threads.
static void stgDeallocForGMP (void *ptr, size_t size);
void
-initStorage (void)
+initStorage( void )
{
nat g, s;
- step *step;
+ 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;
}
#endif
- if (RtsFlags.GcFlags.heapSizeSuggestion >
+ 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) {
+ prog_belch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
+ exit(1);
+ }
+
initBlockAllocator();
/* allocate generation info array */
/* Initialise all steps */
for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
for (s = 0; s < generations[g].n_steps; s++) {
- step = &generations[g].steps[s];
- step->no = s;
- step->blocks = NULL;
- step->n_blocks = 0;
- step->gen = &generations[g];
- step->hp = NULL;
- step->hpLim = NULL;
- step->hp_bd = NULL;
- step->scan = NULL;
- step->scan_bd = NULL;
- step->large_objects = NULL;
- step->new_large_objects = NULL;
- step->scavenged_large_objects = NULL;
+ 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;
}
}
generations[g].steps[s].to = &generations[g+1].steps[0];
}
- /* The oldest generation has one step and its destination is the
- * same step. */
+ /* The oldest generation has one step and it is compacted. */
+ if (RtsFlags.GcFlags.compact) {
+ 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];
/* generation 0 is special: that's the nursery */
alloc_blocks = 0;
alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
-#ifdef COMPILER
/* Tell GNU multi-precision pkg about our custom alloc functions */
mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
-#endif
#ifdef SMP
pthread_mutex_init(&sm_mutex, NULL);
#endif
- IF_DEBUG(gc, stat_describe_gens());
+ IF_DEBUG(gc, statDescribeGens());
}
void
exitStorage (void)
{
- stat_exit(calcAllocated());
+ stat_exit(calcAllocated());
}
+/* -----------------------------------------------------------------------------
+ CAF management.
+
+ The entry code for every CAF does the following:
+
+ - builds a CAF_BLACKHOLE in the heap
+ - pushes an update frame pointing to the CAF_BLACKHOLE
+ - invokes UPD_CAF(), which:
+ - calls newCaf, below
+ - updates the CAF with a static indirection to the CAF_BLACKHOLE
+
+ Why do we build a BLACKHOLE in the heap rather than just updating
+ the thunk directly? It's so that we only need one kind of update
+ frame - otherwise we'd need a static version of the update frame too.
+
+ newCaf() does the following:
+
+ - it puts the CAF on the oldest generation's mut-once list.
+ This is so that we can treat the CAF as a root when collecting
+ younger generations.
+
+ For GHCI, we have additional requirements when dealing with CAFs:
+
+ - we must *retain* all dynamically-loaded CAFs ever entered,
+ just in case we need them again.
+ - we must be able to *revert* CAFs that have been evaluated, to
+ their pre-evaluated form.
+
+ To do this, we use an additional CAF list. When newCaf() is
+ called on a dynamically-loaded CAF, we add it to the CAF list
+ instead of the old-generation mutable list, and save away its
+ old info pointer (in caf->saved_info) for later reversion.
+
+ To revert all the CAFs, we traverse the CAF list and reset the
+ info pointer to caf->saved_info, then throw away the CAF list.
+ (see GC.c:revertCAFs()).
+
+ -- SDM 29/1/01
+
+ -------------------------------------------------------------------------- */
+
void
newCAF(StgClosure* caf)
{
* any more and can use it as a STATIC_LINK.
*/
ACQUIRE_LOCK(&sm_mutex);
- ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
- oldest_gen->mut_once_list = (StgMutClosure *)caf;
-#ifdef DEBUG
- {
- const StgInfoTable *info;
-
- info = get_itbl(caf);
- ASSERT(info->type == IND_STATIC);
-#if 0
- STATIC_LINK2(info,caf) = caf_list;
- caf_list = caf;
-#endif
+ 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;
}
-#endif
+
RELEASE_LOCK(&sm_mutex);
+
+#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));
+ newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
+ ASSERT(newGA);
+ }
+#endif /* PAR */
}
/* -----------------------------------------------------------------------------
cap->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
cap->rCurrentNursery = cap->rNursery;
for (bd = cap->rNursery; bd != NULL; bd = bd->link) {
- bd->back = (bdescr *)cap;
+ bd->u.back = (bdescr *)cap;
}
}
/* Set the back links to be equal to the Capability,
*/
}
#else /* SMP */
- nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
- g0s0->blocks = allocNursery(NULL, nursery_blocks);
- g0s0->n_blocks = nursery_blocks;
- g0s0->to_space = NULL;
+ 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 */
Capability *cap;
/* All tasks must be stopped */
- ASSERT(n_free_capabilities == RtsFlags.ConcFlags.nNodes);
+ 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 == g0);
+ ASSERT(bd->gen_no == 0);
ASSERT(bd->step == g0s0);
IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
}
#else
for (bd = g0s0->blocks; bd; bd = bd->link) {
bd->free = bd->start;
- ASSERT(bd->gen == g0);
+ ASSERT(bd->gen_no == 0);
ASSERT(bd->step == g0s0);
IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
}
bd = allocBlock();
bd->link = last_bd;
bd->step = g0s0;
- bd->gen = g0;
- bd->evacuated = 0;
+ bd->gen_no = 0;
+ bd->flags = 0;
bd->free = bd->start;
last_bd = bd;
}
-------------------------------------------------------------------------- */
StgPtr
-allocate(nat n)
+allocate( nat n )
{
bdescr *bd;
StgPtr p;
nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
bd = allocGroup(req_blocks);
dbl_link_onto(bd, &g0s0->large_objects);
- bd->gen = g0;
+ bd->gen_no = 0;
bd->step = g0s0;
- bd->evacuated = 0;
+ 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.
*/
+ alloc_blocks += req_blocks;
RELEASE_LOCK(&sm_mutex);
return bd->start;
bd = allocBlock();
bd->link = small_alloc_list;
small_alloc_list = bd;
- bd->gen = g0;
+ bd->gen_no = 0;
bd->step = g0s0;
- bd->evacuated = 0;
+ bd->flags = 0;
alloc_Hp = bd->start;
alloc_HpLim = bd->start + BLOCK_SIZE_W;
alloc_blocks++;
}
-
+
p = alloc_Hp;
alloc_Hp += n;
RELEASE_LOCK(&sm_mutex);
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.
/* allocate and fill it in. */
arr = (StgArrWords *)allocate(total_size_in_words);
- SET_ARR_HDR(arr, &ARR_WORDS_info, CCCS, data_size_in_words);
+ 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));
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.ConcFlags.nNodes); */
+ /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
allocated =
n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
}
#endif
+ total_allocated += allocated;
return allocated;
}
{
nat g, s;
lnat live = 0;
- step *step;
+ step *stp;
if (RtsFlags.GcFlags.generations == 1) {
- live = (g0s0->to_blocks - 1) * BLOCK_SIZE_W +
+ live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W +
((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
return live;
}
if (g == 0 && s == 0) {
continue;
}
- step = &generations[g].steps[s];
- live += (step->n_blocks - 1) * BLOCK_SIZE_W +
- ((lnat)step->hp_bd->free - (lnat)step->hp_bd->start) / sizeof(W_);
+ stp = &generations[g].steps[s];
+ 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_);
+ }
}
}
return live;
extern lnat
calcNeeded(void)
{
- lnat needed = 0;
- nat g, s;
- step *step;
-
- for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
- for (s = 0; s < generations[g].n_steps; s++) {
- if (g == 0 && s == 0) { continue; }
- step = &generations[g].steps[s];
- if (generations[g].steps[0].n_blocks > generations[g].max_blocks) {
- needed += 2 * step->n_blocks;
- } else {
- needed += step->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;
}
/* -----------------------------------------------------------------------------
#ifdef DEBUG
-extern void
+void
memInventory(void)
{
nat g, s;
- step *step;
+ step *stp;
bdescr *bd;
lnat total_blocks = 0, free_blocks = 0;
for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
for (s = 0; s < generations[g].n_steps; s++) {
- step = &generations[g].steps[s];
- total_blocks += step->n_blocks;
+ 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->to_blocks;
+ total_blocks += g0s0->n_to_blocks;
}
- for (bd = step->large_objects; bd; bd = bd->link) {
+ 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
total_blocks += bd->blocks;
}
+ // 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);
-
-#if 0
if (total_blocks + free_blocks != mblocks_allocated *
BLOCKS_PER_MBLOCK) {
fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
total_blocks, free_blocks, total_blocks + free_blocks,
mblocks_allocated * BLOCKS_PER_MBLOCK);
}
-#endif
-}
-/* Full heap sanity check. */
+ ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
+}
-extern void
-checkSanity(nat N)
+static nat
+countBlocks(bdescr *bd)
{
- nat g, s;
-
- if (RtsFlags.GcFlags.generations == 1) {
- checkHeap(g0s0->to_space, NULL);
- checkChain(g0s0->large_objects);
- } else {
-
- for (g = 0; g <= N; g++) {
- for (s = 0; s < generations[g].n_steps; s++) {
- if (g == 0 && s == 0) { continue; }
- checkHeap(generations[g].steps[s].blocks, NULL);
- }
+ nat n;
+ for (n=0; bd != NULL; bd=bd->link) {
+ n += bd->blocks;
}
- for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
- for (s = 0; s < generations[g].n_steps; s++) {
- checkHeap(generations[g].steps[s].blocks,
- generations[g].steps[s].blocks->start);
- checkChain(generations[g].steps[s].large_objects);
- }
+ return n;
+}
+
+/* Full heap sanity check. */
+void
+checkSanity( void )
+{
+ nat g, s;
+
+ if (RtsFlags.GcFlags.generations == 1) {
+ checkHeap(g0s0->to_blocks);
+ checkChain(g0s0->large_objects);
+ } else {
+
+ 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);
+ if (g > 0) {
+ checkMutableList(generations[g].mut_list, g);
+ checkMutOnceList(generations[g].mut_once_list, g);
+ }
+ }
+ }
+ checkFreeListSanity();
}
- checkFreeListSanity();
- }
}
#endif
projects / ghc-hetmet.git / blobdiff