#include <stdlib.h>
#include <string.h>
+#include "ffi.h"
+
/*
* All these globals require sm_mutex to access in THREADED_RTS mode.
*/
nat alloc_blocks; /* number of allocate()d blocks since GC */
nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
+static bdescr *exec_block;
+
generation *generations = NULL; /* all the generations */
generation *g0 = NULL; /* generation 0, for convenience */
generation *oldest_gen = NULL; /* oldest generation, for convenience */
* 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(&stg_BLACKHOLE_info));
+ ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLACKHOLE_info));
ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
alloc_blocks = 0;
alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
+ exec_block = NULL;
+
/* Tell GNU multi-precision pkg about our custom alloc functions */
mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
#ifdef THREADED_RTS
initSpinLock(&gc_alloc_block_sync);
- initSpinLock(&recordMutableGen_sync);
whitehole_spin = 0;
#endif
+ N = 0;
+
+ initGcThreads();
+
IF_DEBUG(gc, statDescribeGens());
RELEASE_SM_LOCK;
* any more and can use it as a STATIC_LINK.
*/
((StgIndStatic *)caf)->saved_info = NULL;
- recordMutableGen(caf, oldest_gen);
+ recordMutableGen(caf, oldest_gen->no);
}
RELEASE_SM_LOCK;
if (RtsFlags.GcFlags.maxHeapSize > 0 &&
req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
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);
}
bd = allocGroup(req_blocks);
dbl_link_onto(bd, &stp->large_objects);
stp->n_large_blocks += bd->blocks; // might be larger than req_blocks
+ alloc_blocks += bd->blocks;
bd->gen_no = g->no;
bd->step = stp;
bd->flags = BF_LARGE;
return allocated;
}
-// split N blocks off the start of the given bdescr, returning the
-// remainder as a new block group. We treat the remainder as if it
+// split N blocks off the front of the given bdescr, returning the
+// new block group. We treat the remainder as if it
// had been freshly allocated in generation 0.
bdescr *
splitLargeBlock (bdescr *bd, nat blocks)
new_bd->step = g0s0;
new_bd->flags = BF_LARGE;
new_bd->free = bd->free;
+ ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
// add the new number of blocks to the counter. Due to the gaps
// for block descriptor, new_bd->blocks + bd->blocks might not be
bd->step->n_large_blocks += bd->blocks;
return new_bd;
-}
+}
/* -----------------------------------------------------------------------------
allocateLocal()
bd->flags = 0;
// NO: alloc_blocks++;
// calcAllocated() uses the size of the nursery, and we've
- // already bumpted nursery->n_blocks above.
+ // already bumpted nursery->n_blocks above. We'll GC
+ // pretty quickly now anyway, because MAYBE_GC() will
+ // notice that CurrentNursery->link is NULL.
} else {
// we have a block in the nursery: take it and put
// it at the *front* of the nursery list, and use it
static void *
stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
{
+ size_t min_size;
void *new_stuff_ptr = stgAllocForGMP(new_size);
nat i = 0;
char *p = (char *) ptr;
char *q = (char *) new_stuff_ptr;
- for (; i < old_size; i++, p++, q++) {
+ min_size = old_size < new_size ? old_size : new_size;
+ for (; i < min_size; i++, p++, q++) {
*q = *p;
}
should be modified to use allocateExec instead of VirtualAlloc.
------------------------------------------------------------------------- */
-static bdescr *exec_block;
+#if defined(linux_HOST_OS)
+
+// On Linux we need to use libffi for allocating executable memory,
+// because it knows how to work around the restrictions put in place
+// by SELinux.
+
+void *allocateExec (nat bytes, void **exec_ret)
+{
+ void **ret, **exec;
+ ACQUIRE_SM_LOCK;
+ ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
+ RELEASE_SM_LOCK;
+ if (ret == NULL) return ret;
+ *ret = ret; // save the address of the writable mapping, for freeExec().
+ *exec_ret = exec + 1;
+ return (ret + 1);
+}
+
+// freeExec gets passed the executable address, not the writable address.
+void freeExec (void *addr)
+{
+ void *writable;
+ writable = *((void**)addr - 1);
+ ACQUIRE_SM_LOCK;
+ ffi_closure_free (writable);
+ RELEASE_SM_LOCK
+}
-void *allocateExec (nat bytes)
+#else
+
+void *allocateExec (nat bytes, void **exec_ret)
{
void *ret;
nat n;
exec_block->free += n + 1;
RELEASE_SM_LOCK
+ *exec_ret = ret;
return ret;
}
RELEASE_SM_LOCK
}
+#endif /* mingw32_HOST_OS */
+
/* -----------------------------------------------------------------------------
Debugging
countAllocdBlocks(stp->large_objects);
}
+// If memInventory() calculates that we have a memory leak, this
+// function will try to find the block(s) that are leaking by marking
+// all the ones that we know about, and search through memory to find
+// blocks that are not marked. In the debugger this can help to give
+// us a clue about what kind of block leaked. In the future we might
+// annotate blocks with their allocation site to give more helpful
+// info.
+static void
+findMemoryLeak (void)
+{
+ nat g, s, i;
+ for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
+ for (i = 0; i < n_capabilities; i++) {
+ markBlocks(capabilities[i].mut_lists[g]);
+ }
+ markBlocks(generations[g].mut_list);
+ for (s = 0; s < generations[g].n_steps; s++) {
+ markBlocks(generations[g].steps[s].blocks);
+ markBlocks(generations[g].steps[s].large_objects);
+ }
+ }
+
+ for (i = 0; i < n_nurseries; i++) {
+ markBlocks(nurseries[i].blocks);
+ markBlocks(nurseries[i].large_objects);
+ }
+
+#ifdef PROFILING
+ // TODO:
+ // if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
+ // markRetainerBlocks();
+ // }
+#endif
+
+ // count the blocks allocated by the arena allocator
+ // TODO:
+ // markArenaBlocks();
+
+ // count the blocks containing executable memory
+ markBlocks(exec_block);
+
+ reportUnmarkedBlocks();
+}
+
+
void
memInventory (rtsBool show)
{
leak = live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK;
- ASSERT(n_alloc_blocks == live_blocks);
-
if (show || leak)
{
if (leak) {
mblocks_allocated * BLOCKS_PER_MBLOCK, mblocks_allocated);
}
}
+
+ if (leak) {
+ debugBelch("\n");
+ findMemoryLeak();
+ }
+ ASSERT(n_alloc_blocks == live_blocks);
+ ASSERT(!leak);
}
== 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);
- }
}
}
checkFreeListSanity();
}
+
+#if defined(THREADED_RTS)
+ // check the stacks too in threaded mode, because we don't do a
+ // full heap sanity check in this case (see checkHeap())
+ checkMutableLists(rtsTrue);
+#else
+ checkMutableLists(rtsFalse);
+#endif
}
/* Nursery sanity check */