static bdescr *
allocGroup_sync(nat n)
{
bdescr *bd;
ACQUIRE_SPIN_LOCK(&gc_alloc_block_sync);
bd = allocGroup(n);
RELEASE_SPIN_LOCK(&gc_alloc_block_sync);
return bd;
}
bdescr *
allocGroup_lock(W_ n)
{
bdescr *bd;
ACQUIRE_SM_LOCK;
bd = allocGroup(n);
RELEASE_SM_LOCK;
return bd;
}
//
// Allocate a chunk of blocks that is at least min and at most max
// blocks in size. This API is used by the nursery allocator that
// wants contiguous memory preferably, but doesn't require it. When
// memory is fragmented we might have lots of chunks that are
// less than a full megablock, so allowing the nursery allocator to
// use these reduces fragmentation considerably. e.g. on a GHC build
// with +RTS -H, I saw fragmentation go from 17MB down to 3MB on a
// single compile.
//
// Further to this: in #7257 there is a program that creates serious
// fragmentation such that the heap is full of tiny <4 block chains.
// The nursery allocator therefore has to use single blocks to avoid
// fragmentation, but we make sure that we allocate large blocks
// preferably if there are any.
//
bdescr *
allocLargeChunk (W_ min, W_ max)
{
bdescr *bd;
StgWord ln, lnmax;
if (min >= BLOCKS_PER_MBLOCK) {
return allocGroup(max);
}
ln = log_2_ceil(min);
lnmax = log_2_ceil(max); // tops out at MAX_FREE_LIST
while (ln < lnmax && free_list[ln] == NULL) {
ln++;
}
if (ln == lnmax) {
return allocGroup(max);
}
bd = free_list[ln];
if (bd->blocks <= max) // exactly the right size!
{
dbl_link_remove(bd, &free_list[ln]);
initGroup(bd);
}
else // block too big...
{
bd = split_free_block(bd, max, ln);
ASSERT(bd->blocks == max);
initGroup(bd);
}
n_alloc_blocks += bd->blocks;
if (n_alloc_blocks > hw_alloc_blocks) hw_alloc_blocks = n_alloc_blocks;
IF_DEBUG(sanity, memset(bd->start, 0xaa, bd->blocks * BLOCK_SIZE));
IF_DEBUG(sanity, checkFreeListSanity());
return bd;
}
static void
allocBlocks_sync(nat n, bdescr **hd, bdescr **tl,
nat gen_no, step *stp,
StgWord32 flags)
{
bdescr *bd;
nat i;
ACQUIRE_SPIN_LOCK(&gc_alloc_block_sync);
bd = allocGroup(n);
for (i = 0; i < n; i++) {
bd[i].blocks = 1;
bd[i].gen_no = gen_no;
bd[i].step = stp;
bd[i].flags = flags;
bd[i].link = &bd[i+1];
bd[i].u.scan = bd[i].free = bd[i].start;
}
*hd = bd;
*tl = &bd[n-1];
RELEASE_SPIN_LOCK(&gc_alloc_block_sync);
}
AdjustorWritable allocateExec (W_ bytes, AdjustorExecutable *exec_ret)
{
void *ret;
W_ n;
ACQUIRE_SM_LOCK;
// round up to words.
n = (bytes + sizeof(W_) + 1) / sizeof(W_);
if (n+1 > BLOCK_SIZE_W) {
barf("allocateExec: can't handle large objects");
}
if (exec_block == NULL ||
exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
bdescr *bd;
W_ pagesize = getPageSize();
bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
bd->gen_no = 0;
bd->flags = BF_EXEC;
bd->link = exec_block;
if (exec_block != NULL) {
exec_block->u.back = bd;
}
bd->u.back = NULL;
setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
exec_block = bd;
}
*(exec_block->free) = n; // store the size of this chunk
exec_block->gen_no += n; // gen_no stores the number of words allocated
ret = exec_block->free + 1;
exec_block->free += n + 1;
RELEASE_SM_LOCK
*exec_ret = ret;
return ret;
}
StgPtr
allocate (Capability *cap, W_ n)
{
bdescr *bd;
StgPtr p;
TICK_ALLOC_HEAP_NOCTR(WDS(n));
CCS_ALLOC(cap->r.rCCCS,n);
if (cap->r.rCurrentTSO != NULL) {
// cap->r.rCurrentTSO->alloc_limit -= n*sizeof(W_)
ASSIGN_Int64((W_*)&(cap->r.rCurrentTSO->alloc_limit),
(PK_Int64((W_*)&(cap->r.rCurrentTSO->alloc_limit))
- n*sizeof(W_)));
}
if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
// The largest number of words such that
// the computation of req_blocks will not overflow.
W_ max_words = (HS_WORD_MAX & ~(BLOCK_SIZE-1)) / sizeof(W_);
W_ req_blocks;
if (n > max_words)
req_blocks = HS_WORD_MAX; // signal overflow below
else
req_blocks = (W_)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
// Attempting to allocate an object larger than maxHeapSize
// should definitely be disallowed. (bug #1791)
if ((RtsFlags.GcFlags.maxHeapSize > 0 &&
req_blocks >= RtsFlags.GcFlags.maxHeapSize) ||
req_blocks >= HS_INT32_MAX) // avoid overflow when
// calling allocGroup() below
{
heapOverflow();
// heapOverflow() doesn't exit (see #2592), but we aren't
// in a position to do a clean shutdown here: we
// either have to allocate the memory or exit now.
// Allocating the memory would be bad, because the user
// has requested that we not exceed maxHeapSize, so we
// just exit.
stg_exit(EXIT_HEAPOVERFLOW);
}
ACQUIRE_SM_LOCK
bd = allocGroup(req_blocks);
dbl_link_onto(bd, &g0->large_objects);
g0->n_large_blocks += bd->blocks; // might be larger than req_blocks
g0->n_new_large_words += n;
RELEASE_SM_LOCK;
initBdescr(bd, g0, g0);
bd->flags = BF_LARGE;
bd->free = bd->start + n;
cap->total_allocated += n;
return bd->start;
}
/* small allocation (<LARGE_OBJECT_THRESHOLD) */
bd = cap->r.rCurrentAlloc;
if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
if (bd) finishedNurseryBlock(cap,bd);
// The CurrentAlloc block is full, we need to find another
// one. First, we try taking the next block from the
// nursery:
bd = cap->r.rCurrentNursery->link;
if (bd == NULL) {
// The nursery is empty: allocate a fresh block (we can't
// fail here).
ACQUIRE_SM_LOCK;
bd = allocBlock();
cap->r.rNursery->n_blocks++;
RELEASE_SM_LOCK;
initBdescr(bd, g0, g0);
bd->flags = 0;
// If we had to allocate a new block, then we'll GC
// pretty quickly now, because MAYBE_GC() will
// notice that CurrentNursery->link is NULL.
} else {
newNurseryBlock(bd);
// we have a block in the nursery: take it and put
// it at the *front* of the nursery list, and use it
// to allocate() from.
//
// Previously the nursery looked like this:
//
// CurrentNursery
// /
// +-+ +-+
// nursery -> ... |A| -> |B| -> ...
// +-+ +-+
//
// After doing this, it looks like this:
//
// CurrentNursery
// /
// +-+ +-+
// nursery -> |B| -> ... -> |A| -> ...
// +-+ +-+
// |
// CurrentAlloc
//
// The point is to get the block out of the way of the
// advancing CurrentNursery pointer, while keeping it
// on the nursery list so we don't lose track of it.
cap->r.rCurrentNursery->link = bd->link;
if (bd->link != NULL) {
bd->link->u.back = cap->r.rCurrentNursery;
}
}
dbl_link_onto(bd, &cap->r.rNursery->blocks);
cap->r.rCurrentAlloc = bd;
IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
}
p = bd->free;
bd->free += n;
IF_DEBUG(sanity, ASSERT(*((StgWord8*)p) == 0xaa));
return p;
}
static StgCompactNFDataBlock *
compactAllocateBlockInternal(Capability *cap,
StgWord aligned_size,
StgCompactNFDataBlock *first,
AllocateOp operation)
{
StgCompactNFDataBlock *self;
bdescr *block, *head;
uint32_t n_blocks;
generation *g;
n_blocks = aligned_size / BLOCK_SIZE;
// Attempting to allocate an object larger than maxHeapSize
// should definitely be disallowed. (bug #1791)
if ((RtsFlags.GcFlags.maxHeapSize > 0 &&
n_blocks >= RtsFlags.GcFlags.maxHeapSize) ||
n_blocks >= HS_INT32_MAX) // avoid overflow when
// calling allocGroup() below
{
reportHeapOverflow();
// reportHeapOverflow() doesn't exit (see #2592), but we aren't
// in a position to do a clean shutdown here: we
// either have to allocate the memory or exit now.
// Allocating the memory would be bad, because the user
// has requested that we not exceed maxHeapSize, so we
// just exit.
stg_exit(EXIT_HEAPOVERFLOW);
}
// It is imperative that first is the first block in the compact
// (or NULL if the compact does not exist yet)
// because the evacuate code does not update the generation of
// blocks other than the first (so we would get the statistics
// wrong and crash in Sanity)
if (first != NULL) {
block = Bdescr((P_)first);
g = block->gen;
} else {
g = g0;
}
ACQUIRE_SM_LOCK;
block = allocGroup(n_blocks);
switch (operation) {
case ALLOCATE_NEW:
ASSERT(first == NULL);
ASSERT(g == g0);
dbl_link_onto(block, &g0->compact_objects);
g->n_compact_blocks += block->blocks;
g->n_new_large_words += aligned_size / sizeof(StgWord);
break;
case ALLOCATE_IMPORT_NEW:
dbl_link_onto(block, &g0->compact_blocks_in_import);
/* fallthrough */
case ALLOCATE_IMPORT_APPEND:
ASSERT(first == NULL);
ASSERT(g == g0);
g->n_compact_blocks_in_import += block->blocks;
g->n_new_large_words += aligned_size / sizeof(StgWord);
break;
case ALLOCATE_APPEND:
g->n_compact_blocks += block->blocks;
if (g == g0)
g->n_new_large_words += aligned_size / sizeof(StgWord);
break;
default:
#if defined(DEBUG)
ASSERT(!"code should not be reached");
#else
RTS_UNREACHABLE;
#endif
}
RELEASE_SM_LOCK;
cap->total_allocated += aligned_size / sizeof(StgWord);
self = (StgCompactNFDataBlock*) block->start;
self->self = self;
self->next = NULL;
head = block;
initBdescr(head, g, g);
head->flags = BF_COMPACT;
for (block = head + 1, n_blocks --; n_blocks > 0; block++, n_blocks--) {
block->link = head;
block->blocks = 0;
block->flags = BF_COMPACT;
}
return self;
}
static bool allocBlocks(int level, int numBlocks, Block* result)
{
//printf("Allocating %d blocks at level %d\n", numBlocks, level);
if(level == 0)
{
//printf(" Will attempt to find top-level group of %d blocks.\n", numBlocks);
return findTopLevelGroup(numBlocks, result);
}
//Need to find subblock(s)
//Best to worst options:
//-Use same parent block from last allocation
//-Allocate new parent block somewhere else
//-Depth-first search subdivided blocks looking for free blocks
//-Return false
//First try lastParents
int bestOvershootParent = -1;
int bestOvershoot = 0xFFFF;
int overshoot = 0;
for(int i = 0; i < PARENT_SAVES; i++)
{
if(validBlock(lastParents[level][i]))
{
if(findGroupInBlock(lastParents[level][i], numBlocks, result, &overshoot))
{
if(overshoot < bestOvershoot)
{
bestOvershoot = overshoot;
bestOvershootParent = i;
}
}
}
}
if(bestOvershootParent != -1)
{
Block use = lastParents[level][bestOvershootParent];
if(!findGroupInBlock(use, numBlocks, result, NULL))
{
//This should never ever happen
//puts("VERY FATAL ERROR: CHECK FOUND LAST_PARENT CASE");
return false;
}
//puts("Allocation successful via lastParents");
//printf("Resulting block group: ");
//printBlock(*result);
//puts("");
allocGroup(*result, numBlocks);
return true;
}
//lastParents must be updated
//If one is not yet valid, set it to a newly allocated block
//Immediately use subblocks of that new block to fulfill current request
//puts("Did not find good cached parent block.");
bool allValid = true;
Block* toReplace;
for(int i = 0; i < PARENT_SAVES; i++)
{
if(!validBlock(lastParents[level][i]))
{
allValid = false;
toReplace = &lastParents[level][i];
break;
}
}
if(!allValid)
{
bool success = replaceParentSave(toReplace, level - 1);
if(success)
{
//use *toReplace (now a fresh subdivded block) as parent
*result = getFirstSubblock(*toReplace);
allocGroup(*result, numBlocks);
//puts("Allocation successful via newParent");
return true;
}
}
else
{
int worstGroup = 100;
int worstBlock = -1;
for(int i = 0; i < PARENT_SAVES; i++)
{
int thisGroup = getMaxGroup(lastParents[level][i]);
if(thisGroup < worstGroup)
{
worstGroup = thisGroup;
worstBlock = i;
}
}
toReplace = &lastParents[level][worstBlock];
//replace worstBlock
bool success = replaceParentSave(toReplace, level - 1);
if(success)
{
*result = getFirstSubblock(*toReplace);
allocGroup(*result, numBlocks);
return true;
//puts("Allocation sucessful via new parent.");
}
}
//If here, then depth-first search for an appropriate parent block
//that isn't necesarily listed in lastParent.
Block first;
for(int i = 0; i < numSubblocks[0]; i++)
{
Block topLevel = {{i, -1, -1, -1}};
if(getBlockStatus(topLevel) == SUBDIV)
{
if(groupSearch(level, numBlocks, topLevel, &first))
{
*result = first;
allocGroup(*result, numBlocks);
return true;
//puts("Allocation successful via worst-case dfs for parent");
}
}
}
//puts("Block group allocation failed!");
return false;
}
bdescr *
allocBlock(void)
{
return allocGroup(1);
}
