X-Git-Url: https://git.llucax.com/software/dgc/cdgc.git/blobdiff_plain/b28fd72842fc9ce935bed74f7b2ba79f9cc59711..3389bc359d65b04c74bb41ef4b640367cb9aa617:/rt/gc/cdgc/gc.d?ds=sidebyside

diff --git a/rt/gc/cdgc/gc.d b/rt/gc/cdgc/gc.d
index 78f673c..11944d8 100644
--- a/rt/gc/cdgc/gc.d
+++ b/rt/gc/cdgc/gc.d
@@ -50,6 +50,8 @@ import opts = rt.gc.cdgc.opts;
 
 import cstdlib = tango.stdc.stdlib;
 import cstring = tango.stdc.string;
+import cstdio = tango.stdc.stdio;
+debug(COLLECT_PRINTF) alias cstdio.printf printf;
 
 /*
  * This is a small optimization that proved it's usefulness. For small chunks
@@ -97,6 +99,11 @@ package bool has_pointermap(uint attrs)
     return !opts.options.conservative && !(attrs & BlkAttr.NO_SCAN);
 }
 
+private size_t round_up(size_t n, size_t to)
+{
+    return (n + to - 1) / to;
+}
+
 private
 {
     alias void delegate(Object) DEvent;
@@ -155,7 +162,8 @@ alias ubyte Bins;
 
 struct List
 {
-    List *next;
+    List* next;
+    Pool* pool;
 }
 
 
@@ -200,6 +208,9 @@ struct GC
     /// Turn off collections if > 0
     int disabled;
 
+    // PID of the fork()ed process doing the mark() (0 if is not running)
+    int mark_proc_pid;
+
     /// min(pool.baseAddr)
     byte *min_addr;
     /// max(pool.topAddr)
@@ -210,7 +221,7 @@ struct GC
 
     dynarray.DynArray!(void*) roots;
     dynarray.DynArray!(Range) ranges;
-    dynarray.DynArray!(Pool) pools;
+    dynarray.DynArray!(Pool*) pools;
 
     Stats stats;
 }
@@ -236,7 +247,7 @@ bool Invariant()
             if (i == 0)
                 assert(gc.min_addr == pool.baseAddr);
             if (i + 1 < gc.pools.length)
-                assert(*pool < gc.pools[i + 1]);
+                assert(*pool < *gc.pools[i + 1]);
             else if (i + 1 == gc.pools.length)
                 assert(gc.max_addr == pool.topAddr);
         }
@@ -250,10 +261,16 @@ bool Invariant()
             assert(gc.ranges[i].pbot <= gc.ranges[i].ptop);
         }
 
-        for (size_t i = 0; i < B_PAGE; i++)
-            for (List *list = gc.free_list[i]; list; list = list.next)
-            {
+        for (size_t i = 0; i < B_PAGE; i++) {
+            for (List *list = gc.free_list[i]; list; list = list.next) {
+                auto pool = list.pool;
+                assert (pool !is null);
+                auto p = cast(byte*) list;
+                assert (p >= pool.baseAddr);
+                assert (p < pool.topAddr);
+                assert (pool.freebits.test((p - pool.baseAddr) / 16));
             }
+        }
     }
     return true;
 }
@@ -302,8 +319,11 @@ BlkInfo getInfo(void* p)
         return BlkInfo.init;
     BlkInfo info;
     info.base = pool.findBase(p);
+    if (info.base is null)
+        return BlkInfo.init;
     info.size = pool.findSize(info.base);
-    info.attr = getAttr(pool, cast(size_t)(info.base - pool.baseAddr) / 16u);
+    size_t bit_i = (info.base - pool.baseAddr) / 16;
+    info.attr = getAttr(pool, bit_i);
     if (has_pointermap(info.attr)) {
         info.size -= size_t.sizeof; // PointerMap bitmask
         // Points to the PointerMap bitmask pointer, not user data
@@ -377,7 +397,7 @@ Bins findBin(size_t size)
 size_t reserve(size_t size)
 {
     assert(size != 0);
-    size_t npages = (size + PAGESIZE - 1) / PAGESIZE;
+    size_t npages = round_up(size, PAGESIZE);
     Pool*  pool = newPool(npages);
 
     if (!pool)
@@ -391,9 +411,14 @@ size_t reserve(size_t size)
  */
 void minimize()
 {
+    // Disabled if a parallel collection is in progress because the shared mark
+    // bits of the freed pool might be used by the mark process
+    if (gc.mark_proc_pid != 0)
+        return;
+
     size_t n;
     size_t pn;
-    Pool*  pool;
+    Pool* pool;
 
     for (n = 0; n < gc.pools.length; n++)
     {
@@ -406,6 +431,7 @@ void minimize()
         if (pn < pool.npages)
             continue;
         pool.Dtor();
+        cstdlib.free(pool);
         gc.pools.remove_at(n);
         n--;
     }
@@ -418,88 +444,50 @@ void minimize()
  * Allocate a chunk of memory that is larger than a page.
  * Return null if out of memory.
  */
-void *bigAlloc(size_t size)
+void* bigAlloc(size_t npages, out Pool* pool, size_t* pn)
 {
-    Pool*  pool;
-    size_t npages;
-    size_t n;
-    size_t pn;
-    size_t freedpages;
-    void*  p;
-    int    state;
+    // This code could use some refinement when repeatedly
+    // allocating very large arrays.
 
-    npages = (size + PAGESIZE - 1) / PAGESIZE;
-
-    for (state = 0; ; )
+    void* find_block()
     {
-        // This code could use some refinement when repeatedly
-        // allocating very large arrays.
-
-        for (n = 0; n < gc.pools.length; n++)
+        for (size_t n = 0; n < gc.pools.length; n++)
         {
             pool = gc.pools[n];
-            pn = pool.allocPages(npages);
-            if (pn != OPFAIL)
-                goto L1;
+            *pn = pool.allocPages(npages);
+            if (*pn != OPFAIL)
+                return pool.baseAddr + *pn * PAGESIZE;
         }
+        return null;
+    }
 
-        // Failed
-        switch (state)
-        {
-        case 0:
-            if (gc.disabled)
-            {
-                state = 1;
-                continue;
-            }
-            // Try collecting
-            freedpages = fullcollectshell();
-            if (freedpages >= gc.pools.length * ((POOLSIZE / PAGESIZE) / 4))
-            {
-                state = 1;
-                continue;
-            }
-            // Release empty pools to prevent bloat
-            minimize();
-            // Allocate new pool
-            pool = newPool(npages);
-            if (!pool)
-            {
-                state = 2;
-                continue;
-            }
-            pn = pool.allocPages(npages);
-            assert(pn != OPFAIL);
-            goto L1;
-        case 1:
-            // Release empty pools to prevent bloat
-            minimize();
-            // Allocate new pool
-            pool = newPool(npages);
-            if (!pool)
-                goto Lnomemory;
-            pn = pool.allocPages(npages);
-            assert(pn != OPFAIL);
-            goto L1;
-        case 2:
-            goto Lnomemory;
-        default:
-            assert(false);
-        }
+    void* alloc_more()
+    {
+        // Release empty pools to prevent bloat
+        minimize();
+        // Allocate new pool
+        pool = newPool(npages);
+        if (!pool)
+            return null; // let malloc handle the error
+        *pn = pool.allocPages(npages);
+        assert(*pn != OPFAIL);
+        return pool.baseAddr + *pn * PAGESIZE;
     }
 
-  L1:
-    pool.pagetable[pn] = B_PAGE;
-    if (npages > 1)
-        memset(&pool.pagetable[pn + 1], B_PAGEPLUS, npages - 1);
-    p = pool.baseAddr + pn * PAGESIZE;
-    memset(cast(char *)p + size, 0, npages * PAGESIZE - size);
-    if (opts.options.mem_stomp)
-        memset(p, 0xF1, size);
-    return p;
+    if (void* p = find_block())
+        return p;
+
+    if (gc.disabled)
+        return alloc_more();
+
+    // Try collecting
+    size_t freedpages = fullcollectshell();
+    if (freedpages >= gc.pools.length * ((POOLSIZE / PAGESIZE) / 4)) {
+        if (void* p = find_block())
+            return p;
+    }
 
-  Lnomemory:
-    return null; // let mallocNoSync handle the error
+    return alloc_more();
 }
 
 
@@ -532,20 +520,24 @@ Pool *newPool(size_t npages)
             npages = n;
     }
 
-    Pool p;
-    p.initialize(npages);
-    if (!p.baseAddr)
+    auto pool = cast(Pool*) cstdlib.calloc(1, Pool.sizeof);
+    if (pool is null)
+        return null;
+    pool.initialize(npages);
+    if (!pool.baseAddr)
     {
-        p.Dtor();
+        pool.Dtor();
         return null;
     }
 
-    Pool* pool = gc.pools.insert_sorted(p);
-    if (pool)
-    {
-        gc.min_addr = gc.pools[0].baseAddr;
-        gc.max_addr = gc.pools[gc.pools.length - 1].topAddr;
+    auto inserted_pool = *gc.pools.insert_sorted!("*a < *b")(pool);
+    if (inserted_pool is null) {
+        pool.Dtor();
+        return null;
     }
+    assert (inserted_pool is pool);
+    gc.min_addr = gc.pools[0].baseAddr;
+    gc.max_addr = gc.pools[gc.pools.length - 1].topAddr;
     return pool;
 }
 
@@ -558,12 +550,9 @@ Pool *newPool(size_t npages)
 int allocPage(Bins bin)
 {
     Pool*  pool;
-    size_t n;
     size_t pn;
-    byte*  p;
-    byte*  ptop;
 
-    for (n = 0; n < gc.pools.length; n++)
+    for (size_t n = 0; n < gc.pools.length; n++)
     {
         pool = gc.pools[n];
         pn = pool.allocPages(1);
@@ -577,14 +566,18 @@ int allocPage(Bins bin)
 
     // Convert page to free list
     size_t size = binsize[bin];
-    List **b = &gc.free_list[bin];
+    auto list_head = &gc.free_list[bin];
 
-    p = pool.baseAddr + pn * PAGESIZE;
-    ptop = p + PAGESIZE;
+    byte* p = pool.baseAddr + pn * PAGESIZE;
+    byte*  ptop = p + PAGESIZE;
+    size_t bit_i = pn * (PAGESIZE / 16);
+    pool.freebits.set_group(bit_i, PAGESIZE / 16);
     for (; p < ptop; p += size)
     {
-        (cast(List *)p).next = *b;
-        *b = cast(List *)p;
+        List* l = cast(List *) p;
+        l.next = *list_head;
+        l.pool = pool;
+        *list_head = l;
     }
     return 1;
 }
@@ -604,7 +597,7 @@ void mark_range(void *pbot, void *ptop, size_t* pm_bitmask)
     void **p1 = cast(void **)pbot;
     void **p2 = cast(void **)ptop;
     size_t pcache = 0;
-    uint changes = 0;
+    bool changes = false;
 
     size_t type_size = pm_bitmask[0];
     size_t* pm_bits = pm_bitmask + 1;
@@ -630,13 +623,17 @@ void mark_range(void *pbot, void *ptop, size_t* pm_bitmask)
             if (pool)
             {
                 size_t offset = cast(size_t)(p - pool.baseAddr);
-                size_t bit_i;
+                size_t bit_i = void;
                 size_t pn = offset / PAGESIZE;
                 Bins   bin = cast(Bins)pool.pagetable[pn];
 
+                // Cache B_PAGE, B_PAGEPLUS and B_FREE lookups
+                if (bin >= B_PAGE)
+                    pcache = cast(size_t)p & ~(PAGESIZE-1);
+
                 // Adjust bit to be at start of allocated memory block
                 if (bin <= B_PAGE)
-                    bit_i = (offset & notbinsize[bin]) >> 4;
+                    bit_i = (offset & notbinsize[bin]) / 16;
                 else if (bin == B_PAGEPLUS)
                 {
                     do
@@ -646,14 +643,8 @@ void mark_range(void *pbot, void *ptop, size_t* pm_bitmask)
                     while (cast(Bins)pool.pagetable[pn] == B_PAGEPLUS);
                     bit_i = pn * (PAGESIZE / 16);
                 }
-                else
-                {
-                    // Don't mark bits in B_FREE pages
+                else // Don't mark bits in B_FREE pages
                     continue;
-                }
-
-                if (bin >= B_PAGE) // Cache B_PAGE and B_PAGEPLUS lookups
-                    pcache = cast(size_t)p & ~(PAGESIZE-1);
 
                 if (!pool.mark.test(bit_i))
                 {
@@ -661,7 +652,7 @@ void mark_range(void *pbot, void *ptop, size_t* pm_bitmask)
                     if (!pool.noscan.test(bit_i))
                     {
                         pool.scan.set(bit_i);
-                        changes = 1;
+                        changes = true;
                     }
                 }
             }
@@ -769,17 +760,56 @@ size_t fullcollect(void *stackTop)
 {
     debug(COLLECT_PRINTF) printf("Gcx.fullcollect()\n");
 
-    // we always need to stop the world to make threads save the CPU registers
+    // If eager allocation is used, we need to check first if there is a mark
+    // process running. If there isn't, we start a new one (see the next code
+    // block). If there is, we check if it's still running or already finished.
+    // If it's still running, we tell the caller process no memory has been
+    // recovered (it will allocated more to fulfill the current request).  If
+    // the mark process is done, we lunch the sweep phase and hope enough
+    // memory is freed (if that not the case, the caller will allocate more
+    // memory and the next time it's exhausted it will run a new collection).
+    if (opts.options.eager_alloc) {
+        if (gc.mark_proc_pid != 0) { // there is a mark process in progress
+            os.WRes r = os.wait_pid(gc.mark_proc_pid, false); // don't block
+            assert (r != os.WRes.ERROR);
+            switch (r) {
+            case os.WRes.DONE:
+                debug(COLLECT_PRINTF) printf("\t\tmark proc DONE\n");
+                gc.mark_proc_pid = 0;
+                return sweep();
+            case os.WRes.RUNNING:
+                debug(COLLECT_PRINTF) printf("\t\tmark proc RUNNING\n");
+                return 0;
+            case os.WRes.ERROR:
+                debug(COLLECT_PRINTF) printf("\t\tmark proc ERROR\n");
+                disable_fork(); // Try to keep going without forking
+                break;
+            }
+        }
+    }
+
+    // We always need to stop the world to make threads save the CPU registers
     // in the stack and prepare themselves for thread_scanAll()
     thread_suspendAll();
     gc.stats.world_stopped();
 
+    // If forking is enabled, we fork() and start a new mark phase in the
+    // child. The parent process will tell the caller that no memory could be
+    // recycled if eager allocation is used, allowing the mutator to keep going
+    // almost instantly (at the expense of more memory consumption because
+    // a new allocation will be triggered to fulfill the current request). If
+    // no eager allocation is used, the parent will wait for the mark phase to
+    // finish before returning control to the mutator, but other threads are
+    // restarted and may run in parallel with the mark phase (unless they
+    // allocate or use the GC themselves, in which case the global GC lock will
+    // stop them).
     if (opts.options.fork) {
+        cstdio.fflush(null); // avoid duplicated FILE* output
         os.pid_t child_pid = os.fork();
         assert (child_pid != -1); // don't accept errors in non-release mode
         switch (child_pid) {
-        case -1: // if fork() fails, fallback to stop-the-world
-            opts.options.fork = false;
+        case -1: // if fork() fails, fall-back to stop-the-world
+            disable_fork();
             break;
         case 0: // child process (i.e. the collectors mark phase)
             mark(stackTop);
@@ -789,15 +819,28 @@ size_t fullcollect(void *stackTop)
             // start the world again and wait for the mark phase to finish
             thread_resumeAll();
             gc.stats.world_started();
-            int status = void;
-            os.pid_t wait_pid = os.waitpid(child_pid, &status, 0);
-            assert (wait_pid == child_pid);
-            return sweep();
+            if (opts.options.eager_alloc) {
+                gc.mark_proc_pid = child_pid;
+                return 0;
+            }
+            os.WRes r = os.wait_pid(child_pid); // block until it finishes
+            assert (r == os.WRes.DONE);
+            debug(COLLECT_PRINTF) printf("\t\tmark proc DONE (block)\n");
+            if (r == os.WRes.DONE)
+                return sweep();
+            debug(COLLECT_PRINTF) printf("\tmark() proc ERROR\n");
+            // If there was some error, try to keep going without forking
+            disable_fork();
+            // Re-suspend the threads to do the marking in this process
+            thread_suspendAll();
+            gc.stats.world_stopped();
         }
 
     }
 
-    // if we reach here, we are using the standard stop-the-world collection
+    // If we reach here, we are using the standard stop-the-world collection,
+    // either because fork was disabled in the first place, or because it was
+    // disabled because of some error.
     mark(stackTop);
     thread_resumeAll();
     gc.stats.world_started();
@@ -813,33 +856,13 @@ void mark(void *stackTop)
 {
     debug(COLLECT_PRINTF) printf("\tmark()\n");
 
-    gc.p_cache = null;
-    gc.size_cache = 0;
-
     gc.any_changes = false;
-    for (size_t n = 0; n < gc.pools.length; n++)
-    {
-        Pool* pool = gc.pools[n];
-        pool.mark.zero();
-        pool.scan.zero();
-        pool.freebits.zero();
-    }
-
-    // Mark each free entry, so it doesn't get scanned
-    for (size_t n = 0; n < B_PAGE; n++)
-    {
-        for (List *list = gc.free_list[n]; list; list = list.next)
-        {
-            Pool* pool = findPool(list);
-            assert(pool);
-            pool.freebits.set(cast(size_t)(cast(byte*)list - pool.baseAddr) / 16);
-        }
-    }
 
     for (size_t n = 0; n < gc.pools.length; n++)
     {
         Pool* pool = gc.pools[n];
         pool.mark.copy(&pool.freebits);
+        pool.scan.zero();
     }
 
     /// Marks a range of memory in conservative mode.
@@ -957,6 +980,8 @@ size_t sweep()
 {
     // Free up everything not marked
     debug(COLLECT_PRINTF) printf("\tsweep\n");
+    gc.p_cache = null;
+    gc.size_cache = 0;
     size_t freedpages = 0;
     size_t freed = 0;
     for (size_t n = 0; n < gc.pools.length; n++)
@@ -986,16 +1011,14 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
                 {
                     for (; p < ptop; p += size, bit_i += bit_stride)
                     {
-                        if (pool.finals.nbits && pool.finals.testClear(bit_i)) {
+                        if (pool.finals.testClear(bit_i)) {
                             if (opts.options.sentinel)
-                                rt_finalize(cast(List *)sentinel_add(p), false/*gc.no_stack > 0*/);
+                                rt_finalize(sentinel_add(p), false/*gc.no_stack > 0*/);
                             else
-                                rt_finalize(cast(List *)p, false/*gc.no_stack > 0*/);
+                                rt_finalize(p, false/*gc.no_stack > 0*/);
                         }
                         clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
 
-                        List *list = cast(List *)p;
-
                         if (opts.options.mem_stomp)
                             memset(p, 0xF3, size);
                     }
@@ -1012,16 +1035,14 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
                             sentinel_Invariant(sentinel_add(p));
 
                         pool.freebits.set(bit_i);
-                        if (pool.finals.nbits && pool.finals.testClear(bit_i)) {
+                        if (pool.finals.testClear(bit_i)) {
                             if (opts.options.sentinel)
-                                rt_finalize(cast(List *)sentinel_add(p), false/*gc.no_stack > 0*/);
+                                rt_finalize(sentinel_add(p), false/*gc.no_stack > 0*/);
                             else
-                                rt_finalize(cast(List *)p, false/*gc.no_stack > 0*/);
+                                rt_finalize(p, false/*gc.no_stack > 0*/);
                         }
                         clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
 
-                        List *list = cast(List *)p;
-
                         if (opts.options.mem_stomp)
                             memset(p, 0xF3, size);
 
@@ -1031,13 +1052,14 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
             }
             else if (bin == B_PAGE)
             {
-                size_t bit_i = pn * (PAGESIZE / 16);
+                size_t bit_stride = PAGESIZE / 16;
+                size_t bit_i = pn * bit_stride;
                 if (!pool.mark.test(bit_i))
                 {
                     byte *p = pool.baseAddr + pn * PAGESIZE;
                     if (opts.options.sentinel)
                         sentinel_Invariant(sentinel_add(p));
-                    if (pool.finals.nbits && pool.finals.testClear(bit_i)) {
+                    if (pool.finals.testClear(bit_i)) {
                         if (opts.options.sentinel)
                             rt_finalize(sentinel_add(p), false/*gc.no_stack > 0*/);
                         else
@@ -1045,8 +1067,9 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
                     }
                     clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
 
-                    debug(COLLECT_PRINTF) printf("\tcollecting big %x\n", p);
+                    debug(COLLECT_PRINTF) printf("\tcollecting big %p\n", p);
                     pool.pagetable[pn] = B_FREE;
+                    pool.freebits.set_group(bit_i, PAGESIZE / 16);
                     freedpages++;
                     if (opts.options.mem_stomp)
                         memset(p, 0xF3, PAGESIZE);
@@ -1054,6 +1077,8 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
                     {
                         pn++;
                         pool.pagetable[pn] = B_FREE;
+                        bit_i += bit_stride;
+                        pool.freebits.set_group(bit_i, PAGESIZE / 16);
                         freedpages++;
 
                         if (opts.options.mem_stomp)
@@ -1097,6 +1122,7 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
                         goto Lnotfree;
                 }
                 pool.pagetable[pn] = B_FREE;
+                pool.freebits.set_group(bit_base, PAGESIZE / 16);
                 recoveredpages++;
                 continue;
 
@@ -1107,10 +1133,14 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
                     bit_i = bit_base + u / 16;
                     if (pool.freebits.test(bit_i))
                     {
-                        List *list = cast(List *)(p + u);
-                        // avoid unnecessary writes
+                        assert ((p+u) >= pool.baseAddr);
+                        assert ((p+u) < pool.topAddr);
+                        List* list = cast(List*) (p + u);
+                        // avoid unnecesary writes (it really saves time)
                         if (list.next != gc.free_list[bin])
                             list.next = gc.free_list[bin];
+                        if (list.pool != pool)
+                            list.pool = pool;
                         gc.free_list[bin] = list;
                     }
                 }
@@ -1136,14 +1166,11 @@ in
 body
 {
     uint attrs;
-
-    if (pool.finals.nbits &&
-        pool.finals.test(bit_i))
+    if (pool.finals.test(bit_i))
         attrs |= BlkAttr.FINALIZE;
     if (pool.noscan.test(bit_i))
         attrs |= BlkAttr.NO_SCAN;
-//        if (pool.nomove.nbits &&
-//            pool.nomove.test(bit_i))
+//        if (pool.nomove.test(bit_i))
 //            attrs |= BlkAttr.NO_MOVE;
     return attrs;
 }
@@ -1161,8 +1188,6 @@ body
 {
     if (mask & BlkAttr.FINALIZE)
     {
-        if (!pool.finals.nbits)
-            pool.finals.alloc(pool.mark.nbits);
         pool.finals.set(bit_i);
     }
     if (mask & BlkAttr.NO_SCAN)
@@ -1188,7 +1213,7 @@ in
 }
 body
 {
-    if (mask & BlkAttr.FINALIZE && pool.finals.nbits)
+    if (mask & BlkAttr.FINALIZE)
         pool.finals.clear(bit_i);
     if (mask & BlkAttr.NO_SCAN)
         pool.noscan.clear(bit_i);
@@ -1197,6 +1222,13 @@ body
 }
 
 
+void disable_fork()
+{
+    // we have to disable both options, as eager_alloc assumes fork is enabled
+    opts.options.fork = false;
+    opts.options.eager_alloc = false;
+}
+
 
 void initialize()
 {
@@ -1206,10 +1238,18 @@ void initialize()
     // If we are going to fork, make sure we have the needed OS support
     if (opts.options.fork)
         opts.options.fork = os.HAVE_SHARED && os.HAVE_FORK;
+    // Eager allocation is only possible when forking
+    if (!opts.options.fork)
+        opts.options.eager_alloc = false;
     gc.lock = GCLock.classinfo;
     gc.inited = 1;
     setStackBottom(rt_stackBottom());
     gc.stats = Stats(gc);
+    if (opts.options.prealloc_npools) {
+        size_t pages = round_up(opts.options.prealloc_psize, PAGESIZE);
+        for (size_t i = 0; i < opts.options.prealloc_npools; ++i)
+            newPool(pages);
+    }
 }
 
 
@@ -1247,7 +1287,9 @@ private void *malloc(size_t size, uint attrs, size_t* pm_bitmask)
         lastbin = bin;
     }
 
-    size_t capacity; // to figure out where to store the bitmask
+    Pool* pool = void;
+    size_t bit_i = void;
+    size_t capacity = void; // to figure out where to store the bitmask
     if (bin < B_PAGE)
     {
         p = gc.free_list[bin];
@@ -1274,6 +1316,7 @@ private void *malloc(size_t size, uint attrs, size_t* pm_bitmask)
             if (!gc.free_list[bin] && !allocPage(bin))
             {
                 newPool(1);         // allocate new pool to find a new page
+                // TODO: hint allocPage() to use the pool we just created
                 int result = allocPage(bin);
                 if (!result)
                     onOutOfMemoryError();
@@ -1283,7 +1326,14 @@ private void *malloc(size_t size, uint attrs, size_t* pm_bitmask)
         capacity = binsize[bin];
 
         // Return next item from free list
-        gc.free_list[bin] = (cast(List*)p).next;
+        List* list = cast(List*) p;
+        assert ((cast(byte*)list) >= list.pool.baseAddr);
+        assert ((cast(byte*)list) < list.pool.topAddr);
+        gc.free_list[bin] = list.next;
+        pool = list.pool;
+        bit_i = (p - pool.baseAddr) / 16;
+        assert (pool.freebits.test(bit_i));
+        pool.freebits.clear(bit_i);
         if (!(attrs & BlkAttr.NO_SCAN))
             memset(p + size, 0, capacity - size);
         if (opts.options.mem_stomp)
@@ -1291,12 +1341,24 @@ private void *malloc(size_t size, uint attrs, size_t* pm_bitmask)
     }
     else
     {
-        p = bigAlloc(size);
+        size_t pn;
+        size_t npages = round_up(size, PAGESIZE);
+        p = bigAlloc(npages, pool, &pn);
         if (!p)
             onOutOfMemoryError();
-        // Round the size up to the number of pages needed to store it
-        size_t npages = (size + PAGESIZE - 1) / PAGESIZE;
+        assert (pool !is null);
+
         capacity = npages * PAGESIZE;
+        bit_i = pn * (PAGESIZE / 16);
+        pool.freebits.clear(bit_i);
+        pool.pagetable[pn] = B_PAGE;
+        if (npages > 1)
+            memset(&pool.pagetable[pn + 1], B_PAGEPLUS, npages - 1);
+        p = pool.baseAddr + pn * PAGESIZE;
+        memset(cast(char *)p + size, 0, npages * PAGESIZE - size);
+        if (opts.options.mem_stomp)
+            memset(p, 0xF1, size);
+
     }
 
     // Store the bit mask AFTER SENTINEL_POST
@@ -1313,13 +1375,11 @@ private void *malloc(size_t size, uint attrs, size_t* pm_bitmask)
         sentinel_init(p, size);
     }
 
-    if (attrs)
-    {
-        Pool *pool = findPool(p);
-        assert(pool);
-
-        setAttr(pool, cast(size_t)(p - pool.baseAddr) / 16, attrs);
+    if (attrs) {
+        setAttr(pool, bit_i, attrs);
+        assert (bin >= B_PAGE || !pool.freebits.test(bit_i));
     }
+
     return p;
 }
 
@@ -1403,7 +1463,7 @@ private void *realloc(void *p, size_t size, uint attrs,
             if (blk_size >= PAGESIZE && size >= PAGESIZE)
             {
                 auto psz = blk_size / PAGESIZE;
-                auto newsz = (size + PAGESIZE - 1) / PAGESIZE;
+                auto newsz = round_up(size, PAGESIZE);
                 if (newsz == psz)
                     return p;
 
@@ -1416,10 +1476,13 @@ private void *realloc(void *p, size_t size, uint attrs,
                         memset(p + size - pm_bitmask_size, 0xF2,
                                 blk_size - size - pm_bitmask_size);
                     pool.freePages(pagenum + newsz, psz - newsz);
+                    auto new_blk_size = (PAGESIZE * newsz);
+                    // update the size cache, assuming that is very likely the
+                    // size of this block will be queried in the near future
+                    pool.update_cache(p, new_blk_size);
                     if (has_pm) {
-                        auto end_of_blk = cast(size_t**)(
-                                blk_base_addr + (PAGESIZE * newsz) -
-                                pm_bitmask_size);
+                        auto end_of_blk = cast(size_t**)(blk_base_addr +
+                                new_blk_size - pm_bitmask_size);
                         *end_of_blk = pm_bitmask;
                     }
                     return p;
@@ -1437,10 +1500,14 @@ private void *realloc(void *p, size_t size, uint attrs,
                                         - pm_bitmask_size);
                             memset(pool.pagetable + pagenum +
                                     psz, B_PAGEPLUS, newsz - psz);
+                            auto new_blk_size = (PAGESIZE * newsz);
+                            // update the size cache, assuming that is very
+                            // likely the size of this block will be queried in
+                            // the near future
+                            pool.update_cache(p, new_blk_size);
                             if (has_pm) {
                                 auto end_of_blk = cast(size_t**)(
-                                        blk_base_addr +
-                                        (PAGESIZE * newsz) -
+                                        blk_base_addr + new_blk_size -
                                         pm_bitmask_size);
                                 *end_of_blk = pm_bitmask;
                             }
@@ -1520,8 +1587,8 @@ body
         return 0; // cannot extend buckets
 
     auto psz = blk_size / PAGESIZE;
-    auto minsz = (minsize + PAGESIZE - 1) / PAGESIZE;
-    auto maxsz = (maxsize + PAGESIZE - 1) / PAGESIZE;
+    auto minsz = round_up(minsize, PAGESIZE);
+    auto maxsz = round_up(maxsize, PAGESIZE);
 
     auto pagenum = (p - pool.baseAddr) / PAGESIZE;
 
@@ -1549,6 +1616,9 @@ body
     memset(pool.pagetable + pagenum + psz, B_PAGEPLUS, sz);
     gc.p_cache = null;
     gc.size_cache = 0;
+    // update the size cache, assuming that is very likely the size of this
+    // block will be queried in the near future
+    pool.update_cache(p, new_size);
 
     if (has_pm) {
         new_size -= size_t.sizeof;
@@ -1589,22 +1659,27 @@ private void free(void *p)
         // Free pages
         size_t npages = 1;
         size_t n = pagenum;
+        pool.freebits.set_group(bit_i, PAGESIZE / 16);
         while (++n < pool.npages && pool.pagetable[n] == B_PAGEPLUS)
             npages++;
         if (opts.options.mem_stomp)
             memset(p, 0xF2, npages * PAGESIZE);
         pool.freePages(pagenum, npages);
+        // just in case we were caching this pointer
+        pool.clear_cache(p);
     }
     else
     {
         // Add to free list
-        List *list = cast(List*)p;
+        List* list = cast(List*) p;
 
         if (opts.options.mem_stomp)
             memset(p, 0xF2, binsize[bin]);
 
         list.next = gc.free_list[bin];
+        list.pool = pool;
         gc.free_list[bin] = list;
+        pool.freebits.set(bit_i);
     }
 }
 
@@ -1694,9 +1769,7 @@ private void checkNoSync(void *p)
             if (bin < B_PAGE)
             {
                 // Check that p is not on a free list
-                List *list;
-
-                for (list = gc.free_list[bin]; list; list = list.next)
+                for (List* list = gc.free_list[bin]; list; list = list.next)
                 {
                     assert(cast(void*)list != p);
                 }
@@ -1762,7 +1835,7 @@ private GCStats getStats()
 
     for (n = 0; n < B_PAGE; n++)
     {
-        for (List *list = gc.free_list[n]; list; list = list.next)
+        for (List* list = gc.free_list[n]; list; list = list.next)
             flsize += binsize[n];
     }
 
@@ -1873,10 +1946,18 @@ struct Pool
     size_t cached_size;
     void* cached_ptr;
 
-    void clear_cache()
+    void clear_cache(void* ptr = null)
     {
-        this.cached_ptr = null;
-        this.cached_size = 0;
+        if (ptr is null || ptr is this.cached_ptr) {
+            this.cached_ptr = null;
+            this.cached_size = 0;
+        }
+    }
+
+    void update_cache(void* ptr, size_t size)
+    {
+        this.cached_ptr = ptr;
+        this.cached_size = size;
     }
 
     void initialize(size_t npages)
@@ -1893,7 +1974,6 @@ struct Pool
             npages = 0;
             poolsize = 0;
         }
-        //assert(baseAddr);
         topAddr = baseAddr + poolsize;
 
         size_t nbits = cast(size_t)poolsize / 16;
@@ -1904,10 +1984,17 @@ struct Pool
         if (opts.options.fork)
             vis = os.Vis.SHARED;
         mark.alloc(nbits, vis); // shared between mark and sweep
-        freebits.alloc(nbits, vis); // ditto
+        freebits.alloc(nbits); // not used by the mark phase
         scan.alloc(nbits); // only used in the mark phase
-        finals.alloc(nbits); // mark phase *MUST* have a snapshot
-        noscan.alloc(nbits); // ditto
+        finals.alloc(nbits); // not used by the mark phase
+        noscan.alloc(nbits); // mark phase *MUST* have a snapshot
+
+        // all is free when we start
+        freebits.set_all();
+
+        // avoid accidental sweeping of new pools while using eager allocation
+        if (gc.mark_proc_pid)
+            mark.set_all();
 
         pagetable = cast(ubyte*) cstdlib.malloc(npages);
         if (!pagetable)
@@ -1942,7 +2029,7 @@ struct Pool
         if (opts.options.fork)
             vis = os.Vis.SHARED;
         mark.Dtor(vis);
-        freebits.Dtor(vis);
+        freebits.Dtor();
         scan.Dtor();
         finals.Dtor();
         noscan.Dtor();
@@ -2093,8 +2180,9 @@ void sentinel_init(void *p, size_t size)
 
 void sentinel_Invariant(void *p)
 {
-    assert(*sentinel_pre(p) == SENTINEL_PRE);
-    assert(*sentinel_post(p) == SENTINEL_POST);
+    if (*sentinel_pre(p) != SENTINEL_PRE ||
+            *sentinel_post(p) != SENTINEL_POST)
+        cstdlib.abort();
 }