Storage.c

/* -----------------------------------------------------------------------------
 *
 * (c) The GHC Team, 1998-2012
 *
 * Storage manager front end
 *
 * Documentation on the architecture of the Storage Manager can be
 * found in the online commentary:
 *
 *   https://gitlab.haskell.org/ghc/ghc/wikis/commentary/rts/storage
 *
 * ---------------------------------------------------------------------------*/

#include "PosixSource.h"
#include "Rts.h"

#include "Storage.h"
#include "GCThread.h"
#include "RtsUtils.h"
#include "Stats.h"
#include "BlockAlloc.h"
#include "Weak.h"
#include "Sanity.h"
#include "Arena.h"
#include "Capability.h"
#include "Schedule.h"
#include "RetainerProfile.h"        // for counting memory blocks (memInventory)
#include "OSMem.h"
#include "Trace.h"
#include "GC.h"
#include "Evac.h"
#include "NonMoving.h"
#if defined(ios_HOST_OS)
#include "Hash.h"
#endif

#include <string.h>

#include "ffi.h"

/*
 * All these globals require sm_mutex to access in THREADED_RTS mode.
 */
StgIndStatic  *dyn_caf_list        = NULL;
StgIndStatic  *debug_caf_list      = NULL;
StgIndStatic  *revertible_caf_list = NULL;
bool           keepCAFs;

W_ large_alloc_lim;    /* GC if n_large_blocks in any nursery
                        * reaches this. */

bdescr *exec_block;

generation *generations = NULL; /* all the generations */
generation *g0          = NULL; /* generation 0, for convenience */
generation *oldest_gen  = NULL; /* oldest generation, for convenience */

/*
 * Array of nurseries, size == n_capabilities
 *
 * nursery[i] belongs to NUMA node (i % n_numa_nodes)
 * This is chosen to be the same convention as capabilities[i], so
 * that when not using nursery chunks (+RTS -n), we just map
 * capabilities to nurseries 1:1.
 */
nursery *nurseries = NULL;
uint32_t n_nurseries;

/*
 * When we are using nursery chunks, we need a separate next_nursery
 * pointer for each NUMA node.
 */
volatile StgWord next_nursery[MAX_NUMA_NODES];

#if defined(THREADED_RTS)
/*
 * Storage manager mutex:  protects all the above state from
 * simultaneous access by two STG threads.
 */
Mutex sm_mutex;
#endif

static void allocNurseries (uint32_t from, uint32_t to);
static void assignNurseriesToCapabilities (uint32_t from, uint32_t to);

void
initGeneration (generation *gen, int g)
{
    gen->no = g;
    gen->collections = 0;
    gen->par_collections = 0;
    gen->failed_promotions = 0;
    gen->max_blocks = 0;
    gen->blocks = NULL;
    gen->n_blocks = 0;
    gen->n_words = 0;
    gen->live_estimate = 0;
    gen->old_blocks = NULL;
    gen->n_old_blocks = 0;
    gen->large_objects = NULL;
    gen->n_large_blocks = 0;
    gen->n_large_words = 0;
    gen->n_new_large_words = 0;
    gen->compact_objects = NULL;
    gen->n_compact_blocks = 0;
    gen->compact_blocks_in_import = NULL;
    gen->n_compact_blocks_in_import = 0;
    gen->scavenged_large_objects = NULL;
    gen->n_scavenged_large_blocks = 0;
    gen->live_compact_objects = NULL;
    gen->n_live_compact_blocks = 0;
    gen->compact_blocks_in_import = NULL;
    gen->n_compact_blocks_in_import = 0;
    gen->mark = 0;
    gen->compact = 0;
    gen->bitmap = NULL;
#if defined(THREADED_RTS)
    initSpinLock(&gen->sync);
#endif
    gen->threads = END_TSO_QUEUE;
    gen->old_threads = END_TSO_QUEUE;
    gen->weak_ptr_list = NULL;
    gen->old_weak_ptr_list = NULL;
}

void
initStorage (void)
{
  uint32_t g, n;

  if (generations != NULL) {
      // multi-init protection
      return;
  }

  initMBlocks();

  /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
   * doing something reasonable.
   */
  /* We use the NOT_NULL variant or gcc warns that the test is always true */
  ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLOCKING_QUEUE_CLEAN_info));
  ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
  ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));

  initBlockAllocator();

#if defined(THREADED_RTS)
  initMutex(&sm_mutex);
#endif

  ACQUIRE_SM_LOCK;

  /* allocate generation info array */
  generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
                                             * sizeof(struct generation_),
                                             "initStorage: gens");

  /* Initialise all generations */
  for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
      initGeneration(&generations[g], g);
  }

  /* A couple of convenience pointers */
  g0 = &generations[0];
  oldest_gen = &generations[RtsFlags.GcFlags.generations-1];

  /* Set up the destination pointers in each younger gen. step */
  for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
      generations[g].to = &generations[g+1];
  }
  oldest_gen->to = oldest_gen;

  // Nonmoving heap uses oldest_gen so initialize it after initializing oldest_gen
  nonmovingInit();

#if defined(THREADED_RTS)
  // nonmovingAddCapabilities allocates segments, which requires taking the gc
  // sync lock, so initialize it before nonmovingAddCapabilities
  initSpinLock(&gc_alloc_block_sync);
#endif

  if (RtsFlags.GcFlags.useNonmoving)
      nonmovingAddCapabilities(n_capabilities);

  /* The oldest generation has one step. */
  if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) {
      if (RtsFlags.GcFlags.generations == 1) {
          errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled");
      } else {
          oldest_gen->mark = 1;
          if (RtsFlags.GcFlags.compact)
              oldest_gen->compact = 1;
      }
  }

  generations[0].max_blocks = 0;

  dyn_caf_list = (StgIndStatic*)END_OF_CAF_LIST;
  debug_caf_list = (StgIndStatic*)END_OF_CAF_LIST;
  revertible_caf_list = (StgIndStatic*)END_OF_CAF_LIST;

  if (RtsFlags.GcFlags.largeAllocLim > 0) {
      large_alloc_lim = RtsFlags.GcFlags.largeAllocLim * BLOCK_SIZE_W;
  } else {
      large_alloc_lim = RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W;
  }

  exec_block = NULL;

  N = 0;

  for (n = 0; n < n_numa_nodes; n++) {
      next_nursery[n] = n;
  }
  storageAddCapabilities(0, n_capabilities);

  IF_DEBUG(gc, statDescribeGens());

  RELEASE_SM_LOCK;

  traceEventHeapInfo(CAPSET_HEAP_DEFAULT,
                     RtsFlags.GcFlags.generations,
                     RtsFlags.GcFlags.maxHeapSize * BLOCK_SIZE,
                     RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE,
                     MBLOCK_SIZE,
                     BLOCK_SIZE);
}

void storageAddCapabilities (uint32_t from, uint32_t to)
{
    uint32_t n, g, i, new_n_nurseries;
    nursery *old_nurseries;

    if (RtsFlags.GcFlags.nurseryChunkSize == 0) {
        new_n_nurseries = to;
    } else {
        memcount total_alloc = to * RtsFlags.GcFlags.minAllocAreaSize;
        new_n_nurseries =
            stg_max(to, total_alloc / RtsFlags.GcFlags.nurseryChunkSize);
    }

    old_nurseries = nurseries;
    if (from > 0) {
        nurseries = stgReallocBytes(nurseries,
                                    new_n_nurseries * sizeof(struct nursery_),
                                    "storageAddCapabilities");
    } else {
        nurseries = stgMallocBytes(new_n_nurseries * sizeof(struct nursery_),
                                   "storageAddCapabilities");
    }

    // we've moved the nurseries, so we have to update the rNursery
    // pointers from the Capabilities.
    for (i = 0; i < from; i++) {
        uint32_t index = capabilities[i]->r.rNursery - old_nurseries;
        capabilities[i]->r.rNursery = &nurseries[index];
    }

    /* The allocation area.  Policy: keep the allocation area
     * small to begin with, even if we have a large suggested heap
     * size.  Reason: we're going to do a major collection first, and we
     * don't want it to be a big one.  This vague idea is borne out by
     * rigorous experimental evidence.
     */
    allocNurseries(n_nurseries, new_n_nurseries);
    n_nurseries = new_n_nurseries;

    /*
     * Assign each of the new capabilities a nursery.  Remember to start from
     * next_nursery, because we may have already consumed some of the earlier
     * nurseries.
     */
    assignNurseriesToCapabilities(from,to);

    // allocate a block for each mut list
    for (n = from; n < to; n++) {
        for (g = 1; g < RtsFlags.GcFlags.generations; g++) {
            capabilities[n]->mut_lists[g] =
                allocBlockOnNode(capNoToNumaNode(n));
        }
    }

    // Initialize NonmovingAllocators and UpdRemSets
    if (RtsFlags.GcFlags.useNonmoving) {
        nonmovingAddCapabilities(to);
        for (i = 0; i < to; ++i) {
            init_upd_rem_set(&capabilities[i]->upd_rem_set);
        }
    }

#if defined(THREADED_RTS) && defined(CC_LLVM_BACKEND) && (CC_SUPPORTS_TLS == 0)
    newThreadLocalKey(&gctKey);
#endif

    initGcThreads(from, to);
}


void
exitStorage (void)
{
    nonmovingExit();
    updateNurseriesStats();
    stat_exit();
}

void
freeStorage (bool free_heap)
{
    stgFree(generations);
    if (free_heap) freeAllMBlocks();
#if defined(THREADED_RTS)
    closeMutex(&sm_mutex);
#endif
    stgFree(nurseries);
#if defined(THREADED_RTS) && defined(CC_LLVM_BACKEND) && (CC_SUPPORTS_TLS == 0)
    freeThreadLocalKey(&gctKey);
#endif
    freeGcThreads();
}

/* -----------------------------------------------------------------------------
   Note [CAF management]
   ~~~~~~~~~~~~~~~~~~~~~

   The entry code for every CAF does the following:

      - calls newCAF, which builds a CAF_BLACKHOLE on the heap and atomically
        updates the CAF with IND_STATIC pointing to the CAF_BLACKHOLE

      - if newCAF returns zero, it re-enters the CAF (see Note [atomic
        CAF entry])

      - pushes an update frame pointing 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, and various other parts of the RTS that deal with update
   frames would also need special cases for static update frames.

   newCAF() does the following:

      - atomically locks the CAF (see [atomic CAF entry])

      - it builds a CAF_BLACKHOLE on the heap

      - it updates the CAF with an IND_STATIC pointing to the
        CAF_BLACKHOLE, atomically.

      - it puts the CAF on the oldest generation's mutable list.
        This is so that we treat the CAF as a root when collecting
        younger generations.

      - links the CAF onto the CAF list (see below)

   ------------------
   Note [atomic CAF entry]
   ~~~~~~~~~~~~~~~~~~~~~~~

   With THREADED_RTS, newCAF() is required to be atomic (see
   #5558). This is because if two threads happened to enter the same
   CAF simultaneously, they would create two distinct CAF_BLACKHOLEs,
   and so the normal threadPaused() machinery for detecting duplicate
   evaluation will not detect this.  Hence in lockCAF() below, we
   atomically lock the CAF with WHITEHOLE before updating it with
   IND_STATIC, and return zero if another thread locked the CAF first.
   In the event that we lost the race, CAF entry code will re-enter
   the CAF and block on the other thread's CAF_BLACKHOLE.

   ------------------
   Note [GHCi CAFs]
   ~~~~~~~~~~~~~~~~

   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

   ------------------
   Note [Static objects under the nonmoving collector]
   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

   Static object management is a bit tricky under the nonmoving collector as we
   need to maintain a bit more state than in the moving collector. In
   particular, the moving collector uses the low bits of the STATIC_LINK field
   to determine whether the object has been moved to the scavenger's work list
   (see Note [STATIC_LINK fields] in Storage.h).

   However, the nonmoving collector also needs a place to keep its mark bit.
   This is problematic as we therefore need at least three bits of state
   but can assume only two bits are available in STATIC_LINK (due to 32-bit
   systems).

   To accommodate this we move handling of static objects entirely to the
   oldest generation when the nonmoving collector is in use. To do this safely
   and efficiently we allocate the blackhole created by lockCAF() directly in
   the non-moving heap. This means that the moving collector can completely
   ignore static objects in minor collections since they are guaranteed not to
   have any references into the moving heap. Of course, the blackhole itself
   likely will contain a reference into the moving heap but this is
   significantly easier to handle, being a heap-allocated object (see Note
   [Aging under the non-moving collector] in NonMoving.c for details).

   During the moving phase of a major collection we treat static objects
   as we do any other reference into the non-moving heap by pushing them
   to the non-moving mark queue (see Note [Aging under the non-moving
   collector]).

   This allows the non-moving collector to have full control over the flags
   in STATIC_LINK, which it uses as described in Note [STATIC_LINK fields]).
   This is implemented by NonMovingMark.c:bump_static_flag.

   In short, the plan is:

     - lockCAF allocates its blackhole in the nonmoving heap. This is important
       to ensure that we do not need to place the static object on the mut_list
       lest we would need somw way to ensure that it evacuate only once during
       a moving collection.

     - evacuate_static_object adds merely pushes objects to the mark queue

     - the nonmoving collector uses the flags in STATIC_LINK as its mark bit.

   -------------------------------------------------------------------------- */

STATIC_INLINE StgInd *
lockCAF (StgRegTable *reg, StgIndStatic *caf)
{
    const StgInfoTable *orig_info;
    Capability *cap = regTableToCapability(reg);
    StgInd *bh;

    orig_info = caf->header.info;

#if defined(THREADED_RTS)
    const StgInfoTable *cur_info;

    if (orig_info == &stg_IND_STATIC_info ||
        orig_info == &stg_WHITEHOLE_info) {
        // already claimed by another thread; re-enter the CAF
        return NULL;
    }

    cur_info = (const StgInfoTable *)
        cas((StgVolatilePtr)&caf->header.info,
            (StgWord)orig_info,
            (StgWord)&stg_WHITEHOLE_info);

    if (cur_info != orig_info) {
        // already claimed by another thread; re-enter the CAF
        return NULL;
    }

    // successfully claimed by us; overwrite with IND_STATIC
#endif

    // Push stuff that will become unreachable after updating to UpdRemSet to
    // maintain snapshot invariant
    const StgInfoTable *orig_info_tbl = INFO_PTR_TO_STRUCT(orig_info);
    // OSA: Assertions to make sure my understanding of static thunks is correct
    ASSERT(orig_info_tbl->type == THUNK_STATIC);
    // Secondly I think static thunks can't have payload: anything that they
    // reference should be in SRTs
    ASSERT(orig_info_tbl->layout.payload.ptrs == 0);
    // Because the payload is empty we just push the SRT
    IF_NONMOVING_WRITE_BARRIER_ENABLED {
        StgThunkInfoTable *thunk_info = itbl_to_thunk_itbl(orig_info_tbl);
        if (thunk_info->i.srt) {
            updateRemembSetPushClosure(cap, GET_SRT(thunk_info));
        }
    }

    // For the benefit of revertCAFs(), save the original info pointer
    caf->saved_info = orig_info;

    // Allocate the blackhole indirection closure
    if (RtsFlags.GcFlags.useNonmoving) {
        // See Note [Static objects under the nonmoving collector].
        ACQUIRE_SM_LOCK;
        bh = (StgInd *)nonmovingAllocate(cap, sizeofW(*bh));
        RELEASE_SM_LOCK;
        recordMutableCap((StgClosure*)bh,
                         regTableToCapability(reg), oldest_gen->no);
    } else {
        bh = (StgInd *)allocate(cap, sizeofW(*bh));
    }
    bh->indirectee = (StgClosure *)cap->r.rCurrentTSO;
    SET_HDR(bh, &stg_CAF_BLACKHOLE_info, caf->header.prof.ccs);
    // Ensure that above writes are visible before we introduce reference as CAF indirectee.
    write_barrier();

    caf->indirectee = (StgClosure *)bh;
    write_barrier();
    SET_INFO((StgClosure*)caf,&stg_IND_STATIC_info);

    return bh;
}

StgInd *
newCAF(StgRegTable *reg, StgIndStatic *caf)
{
    StgInd *bh;

    bh = lockCAF(reg, caf);
    if (!bh) return NULL;

    if(keepCAFs)
    {
        // Note [dyn_caf_list]
        // If we are in GHCi _and_ we are using dynamic libraries,
        // then we can't redirect newCAF calls to newRetainedCAF (see below),
        // so we make newCAF behave almost like newRetainedCAF.
        // 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.

        ACQUIRE_SM_LOCK; // dyn_caf_list is global, locked by sm_mutex
        caf->static_link = (StgClosure*)dyn_caf_list;
        dyn_caf_list = (StgIndStatic*)((StgWord)caf | STATIC_FLAG_LIST);
        RELEASE_SM_LOCK;
    }
    else
    {
        // Put this CAF on the mutable list for the old generation.
        // N.B. the nonmoving collector works a bit differently: see
        // Note [Static objects under the nonmoving collector].
        if (oldest_gen->no != 0 && !RtsFlags.GcFlags.useNonmoving) {
            recordMutableCap((StgClosure*)caf,
                             regTableToCapability(reg), oldest_gen->no);
        }

#if defined(DEBUG)
        // In the DEBUG rts, we keep track of live CAFs by chaining them
        // onto a list debug_caf_list.  This is so that we can tell if we
        // ever enter a GC'd CAF, and emit a suitable barf().
        //
        // The saved_info field of the CAF is used as the link field for
        // debug_caf_list, because this field is only used by newDynCAF
        // for revertible CAFs, and we don't put those on the
        // debug_caf_list.
        ACQUIRE_SM_LOCK; // debug_caf_list is global, locked by sm_mutex
        ((StgIndStatic *)caf)->saved_info = (const StgInfoTable*)debug_caf_list;
        debug_caf_list = (StgIndStatic*)caf;
        RELEASE_SM_LOCK;
#endif
    }

    return bh;
}

// External API for setting the keepCAFs flag. see #3900.
void
setKeepCAFs (void)
{
    keepCAFs = 1;
}

// 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 newRetainedCAF.
//
StgInd* newRetainedCAF (StgRegTable *reg, StgIndStatic *caf)
{
    StgInd *bh;

    bh = lockCAF(reg, caf);
    if (!bh) return NULL;

    ACQUIRE_SM_LOCK;

    caf->static_link = (StgClosure*)revertible_caf_list;
    revertible_caf_list = (StgIndStatic*)((StgWord)caf | STATIC_FLAG_LIST);

    RELEASE_SM_LOCK;

    return bh;
}

// If we are using loadObj/unloadObj in the linker, then we want to
//
//  - retain all CAFs in statically linked code (keepCAFs == true),
//    because we might link a new object that uses any of these CAFs.
//
//  - GC CAFs in dynamically-linked code, so that we can detect when
//    a dynamically-linked object is unloadable.
//
// So for this case, we set keepCAFs to true, and link newCAF to newGCdCAF
// for dynamically-linked code.
//
StgInd* newGCdCAF (StgRegTable *reg, StgIndStatic *caf)
{
    StgInd *bh;

    bh = lockCAF(reg, caf);
    if (!bh) return NULL;

    // Put this CAF on the mutable list for the old generation.
    // N.B. the nonmoving collector works a bit differently:
    // see Note [Static objects under the nonmoving collector].
    if (oldest_gen->no != 0 && !RtsFlags.GcFlags.useNonmoving) {
        recordMutableCap((StgClosure*)caf,
                         regTableToCapability(reg), oldest_gen->no);
    }

    return bh;
}

/* -----------------------------------------------------------------------------
   Nursery management.
   -------------------------------------------------------------------------- */

static bdescr *
allocNursery (uint32_t node, bdescr *tail, W_ blocks)
{
    bdescr *bd = NULL;
    W_ i, n;

    // We allocate the nursery as a single contiguous block and then
    // divide it into single blocks manually.  This way we guarantee
    // that the nursery blocks are adjacent, so that the processor's
    // automatic prefetching works across nursery blocks.  This is a
    // tiny optimisation (~0.5%), but it's free.

    while (blocks > 0) {
        n = stg_min(BLOCKS_PER_MBLOCK, blocks);
        // allocLargeChunk will prefer large chunks, but will pick up
        // small chunks if there are any available.  We must allow
        // single blocks here to avoid fragmentation (#7257)
        bd = allocLargeChunkOnNode(node, 1, n);
        n = bd->blocks;
        blocks -= n;

        for (i = 0; i < n; i++) {
            initBdescr(&bd[i], g0, g0);

            bd[i].blocks = 1;
            bd[i].flags = 0;

            if (i > 0) {
                bd[i].u.back = &bd[i-1];
            } else {
                bd[i].u.back = NULL;
            }

            if (i+1 < n) {
                bd[i].link = &bd[i+1];
            } else {
                bd[i].link = tail;
                if (tail != NULL) {
                    tail->u.back = &bd[i];
                }
            }

            bd[i].free = bd[i].start;
        }

        tail = &bd[0];
    }

    return &bd[0];
}

STATIC_INLINE void
assignNurseryToCapability (Capability *cap, uint32_t n)
{
    ASSERT(n < n_nurseries);
    cap->r.rNursery = &nurseries[n];
    cap->r.rCurrentNursery = nurseries[n].blocks;
    newNurseryBlock(nurseries[n].blocks);
    cap->r.rCurrentAlloc   = NULL;
    ASSERT(cap->r.rCurrentNursery->node == cap->node);
}

/*
 * Give each Capability a nursery from the pool. No need to do atomic increments
 * here, everything must be stopped to call this function.
 */
static void
assignNurseriesToCapabilities (uint32_t from, uint32_t to)
{
    uint32_t i, node;

    for (i = from; i < to; i++) {
        node = capabilities[i]->node;
        assignNurseryToCapability(capabilities[i], next_nursery[node]);
        next_nursery[node] += n_numa_nodes;
    }
}

static void
allocNurseries (uint32_t from, uint32_t to)
{
    uint32_t i;
    memcount n_blocks;

    if (RtsFlags.GcFlags.nurseryChunkSize) {
        n_blocks = RtsFlags.GcFlags.nurseryChunkSize;
    } else {
        n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
    }

    for (i = from; i < to; i++) {
        nurseries[i].blocks = allocNursery(capNoToNumaNode(i), NULL, n_blocks);
        nurseries[i].n_blocks = n_blocks;
    }
}

void
resetNurseries (void)
{
    uint32_t n;

    for (n = 0; n < n_numa_nodes; n++) {
        next_nursery[n] = n;
    }
    assignNurseriesToCapabilities(0, n_capabilities);

#if defined(DEBUG)
    bdescr *bd;
    for (n = 0; n < n_nurseries; n++) {
        for (bd = nurseries[n].blocks; bd; bd = bd->link) {
            ASSERT(bd->gen_no == 0);
            ASSERT(bd->gen == g0);
            ASSERT(bd->node == capNoToNumaNode(n));
            IF_DEBUG(zero_on_gc, memset(bd->start, 0xaa, BLOCK_SIZE));
        }
    }
#endif
}

W_
countNurseryBlocks (void)
{
    uint32_t i;
    W_ blocks = 0;

    for (i = 0; i < n_nurseries; i++) {
        blocks += nurseries[i].n_blocks;
    }
    return blocks;
}

//
// Resize each of the nurseries to the specified size.
//
static void
resizeNurseriesEach (W_ blocks)
{
    uint32_t i, node;
    bdescr *bd;
    W_ nursery_blocks;
    nursery *nursery;

    for (i = 0; i < n_nurseries; i++) {
        nursery = &nurseries[i];
        nursery_blocks = nursery->n_blocks;
        if (nursery_blocks == blocks) continue;

        node = capNoToNumaNode(i);
        if (nursery_blocks < blocks) {
            debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
                       blocks);
            nursery->blocks = allocNursery(node, nursery->blocks,
                                           blocks-nursery_blocks);
        }
        else
        {
            bdescr *next_bd;

            debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
                       blocks);

            bd = nursery->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;
            }
            nursery->blocks = bd;
            // might have gone just under, by freeing a large block, so make
            // up the difference.
            if (nursery_blocks < blocks) {
                nursery->blocks = allocNursery(node, nursery->blocks,
                                               blocks-nursery_blocks);
            }
        }
        nursery->n_blocks = blocks;
        ASSERT(countBlocks(nursery->blocks) == nursery->n_blocks);
    }
}

void
resizeNurseriesFixed (void)
{
    uint32_t blocks;

    if (RtsFlags.GcFlags.nurseryChunkSize) {
        blocks = RtsFlags.GcFlags.nurseryChunkSize;
    } else {
        blocks = RtsFlags.GcFlags.minAllocAreaSize;
    }

    resizeNurseriesEach(blocks);
}

//
// Resize the nurseries to the total specified size.
//
void
resizeNurseries (W_ blocks)
{
    // If there are multiple nurseries, then we just divide the number
    // of available blocks between them.
    resizeNurseriesEach(blocks / n_nurseries);
}

bool
getNewNursery (Capability *cap)
{
    StgWord i;
    uint32_t node = cap->node;
    uint32_t n;

    for(;;) {
        i = next_nursery[node];
        if (i < n_nurseries) {
            if (cas(&next_nursery[node], i, i+n_numa_nodes) == i) {
                assignNurseryToCapability(cap, i);
                return true;
            }
        } else if (n_numa_nodes > 1) {
            // Try to find an unused nursery chunk on other nodes.  We'll get
            // remote memory, but the rationale is that avoiding GC is better
            // than avoiding remote memory access.
            bool lost = false;
            for (n = 0; n < n_numa_nodes; n++) {
                if (n == node) continue;
                i = next_nursery[n];
                if (i < n_nurseries) {
                    if (cas(&next_nursery[n], i, i+n_numa_nodes) == i) {
                        assignNurseryToCapability(cap, i);
                        return true;
                    } else {
                        lost = true; /* lost a race */
                    }
                }
            }
            if (!lost) return false;
        } else {
            return false;
        }
    }
}

/* -----------------------------------------------------------------------------
   move_STACK is called to update the TSO structure after it has been
   moved from one place to another.
   -------------------------------------------------------------------------- */

void
move_STACK (StgStack *src, StgStack *dest)
{
    ptrdiff_t diff;

    // relocate the stack pointer...
    diff = (StgPtr)dest - (StgPtr)src; // In *words*
    dest->sp = (StgPtr)dest->sp + diff;
}

STATIC_INLINE void
accountAllocation(Capability *cap, W_ n)
{
    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_)));
    }

}

/* -----------------------------------------------------------------------------
   StgPtr allocate (Capability *cap, W_ n)

   Allocates an area of memory n *words* large, from the nursery of
   the supplied Capability, or from the global block pool if the area
   requested is larger than LARGE_OBJECT_THRESHOLD.  Memory is not
   allocated from the current nursery block, so as not to interfere
   with Hp/HpLim.

   The address of the allocated memory is returned. allocate() never
   fails; if it returns, the returned value is a valid address.  If
   the nursery is already full, then another block is allocated from
   the global block pool.  If we need to get memory from the OS and
   that operation fails, then the whole process will be killed.
   -------------------------------------------------------------------------- */

/*
 * Allocate some n words of heap memory; terminating
 * on heap overflow
 */
StgPtr
allocate (Capability *cap, W_ n)
{
    StgPtr p = allocateMightFail(cap, n);
    if (p == NULL) {
        reportHeapOverflow();
        // 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);
    }
    return p;
}

/*
 * Allocate some n words of heap memory; returning NULL
 * on heap overflow
 */
StgPtr
allocateMightFail (Capability *cap, W_ n)
{
    bdescr *bd;
    StgPtr p;

    if (RTS_UNLIKELY(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
        {
            return NULL;
        }

        // Only credit allocation after we've passed the size check above
        accountAllocation(cap, n);

        ACQUIRE_SM_LOCK
        bd = allocGroupOnNode(cap->node,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) */

    accountAllocation(cap, n);
    bd = cap->r.rCurrentAlloc;
    if (RTS_UNLIKELY(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 = allocBlockOnNode(cap->node);
            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.