Tweak GC scheduler.

2 files changed, 258 insertions, 125 deletions

diff --git a/benchmarks/boehm-gc/patches/01-gc-include-private-gcconfig-h.patch b/benchmarks/boehm-gc/patches/01-gc-include-private-gcconfig-h.patch
index f8a0a1e..00b8c22 100644
--- a/benchmarks/boehm-gc/patches/01-gc-include-private-gcconfig-h.patch
+++ b/benchmarks/boehm-gc/patches/01-gc-include-private-gcconfig-h.patch
@@ -6,7 +6,7 @@ Tweak the configuration.
 
 diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.in gc/include/private/gcconfig.h gc.new/include/private/gcconfig.h
 --- gc/include/private/gcconfig.h	2007-06-29 02:00:09.000000000 +0200
-+++ gc.new/include/private/gcconfig.h	2009-02-23 22:59:56.000000000 +0100
++++ gc.new/include/private/gcconfig.h	2009-03-05 23:12:04.000000000 +0100
 @@ -451,7 +451,28 @@
  #     define  HURD
  #     define  I386
@@ -63,3 +63,12 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
  #   endif
  #   ifdef DARWIN
  #     define OS_TYPE "DARWIN"
+@@ -1842,6 +1855,8 @@
+ #   define CACHE_LINE_SIZE 64
+ #   ifdef LINUX
+ #	define OS_TYPE "LINUX"
++#       define USE_MMAP
++#       define USE_MUNMAP
+ #       define LINUX_STACKBOTTOM
+ #       if !defined(GC_LINUX_THREADS) || !defined(REDIRECT_MALLOC)
+ #	    define MPROTECT_VDB
diff --git a/benchmarks/boehm-gc/patches/05-viengoos-scheduler.patch b/benchmarks/boehm-gc/patches/05-viengoos-scheduler.patch
index d9cfb56..3ee49bd 100644
--- a/benchmarks/boehm-gc/patches/05-viengoos-scheduler.patch
+++ b/benchmarks/boehm-gc/patches/05-viengoos-scheduler.patch
@@ -6,10 +6,23 @@ Patch to support the Viengoos specific scheduler.  To disable it, set
 GC_viengoos_scheduler to 0.
 
 
-diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.in gc.orig/allchblk.c gc/allchblk.c
---- gc.orig/allchblk.c	2007-06-07 02:40:07.000000000 +0200
-+++ gc/allchblk.c	2009-02-24 03:07:28.000000000 +0100
-@@ -117,7 +117,9 @@ void GC_print_hblkfreelist()
+diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.in gc/allchblk.c gc.new/allchblk.c
+--- gc/allchblk.c	2007-06-07 02:40:07.000000000 +0200
++++ gc.new/allchblk.c	2009-03-07 21:53:56.000000000 +0100
+@@ -17,6 +17,12 @@
+ /* #define DEBUG */
+ #include <stdio.h>
+ #include "private/gc_priv.h"
++#include <stdint.h>
++#include <inttypes.h>
++
++#ifndef debug
++#define debug(lvl, fmt, ...) do {} while (0)
++#endif
+ 
+ GC_bool GC_use_entire_heap = 0;
+ 
+@@ -117,7 +123,9 @@ void GC_print_hblkfreelist()
        while (h != 0) {
          hhdr = HDR(h);
          sz = hhdr -> hb_sz;
@@ -20,7 +33,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
      	total_free += sz;
          if (GC_is_black_listed(h, HBLKSIZE) != 0) {
               GC_printf("start black listed\n");
-@@ -381,31 +383,142 @@ void GC_add_to_fl(struct hblk *h, hdr *h
+@@ -381,31 +389,160 @@ void GC_add_to_fl(struct hblk *h, hdr *h
  
  #ifdef USE_MUNMAP
  
@@ -33,7 +46,8 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +
  /* Unmap blocks that haven't been recently touched.  This is the only way */
  /* way blocks are ever unmapped.					  */
- void GC_unmap_old(void)
+-void GC_unmap_old(void)
++void GC_try_unmap(bool all)
  {
      struct hblk * h;
      hdr * hhdr;
@@ -41,28 +55,37 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
      unsigned short last_rec, threshold;
      int i;
  #   define UNMAP_THRESHOLD 6
--    
--    for (i = 0; i <= N_HBLK_FLS; ++i) {
 +
-+    uint64_t mapped_start = GC_get_heap_size() - GC_unmapped_bytes;
++    if (all)
++      GC_gcollect ();
 +
++    void unmap(ptr_t start, size_t bytes)
++    {
++      if (! all)
++	GC_unmap (start, bytes);
++      else
++	munmap (start, bytes);
++    }
+     
+-    for (i = 0; i <= N_HBLK_FLS; ++i) {
++    if (! all && GC_viengoos_scheduler && GC_mapped_bytes < MAPPED_THRESHOLD)
++      /* The amount of mapped memory does not exceed the low-water
++	 threshold.  */
++      return;
++
++    uint64_t mapped_at_start = GC_mapped_bytes;
 +   start:
-+    if (GC_viengoos_scheduler
-+	&& (GC_get_heap_size() - GC_unmapped_bytes
-+	    < 7 * (GC_available_bytes / 8))) {
-+      /* The number of mapped bytes is at most 7/8s the available
-+	 memory.  That's good enough for now.  (Recall: the high-water
-+	 mark is 15/16s the available memory.)  */
-+      uint64_t mapped_now = GC_get_heap_size() - GC_unmapped_bytes;
-+
-+      extern char *program_name;
-+      if (0)
-+	printf ("%s("VG_THREAD_ID_FMT"): After unmapping %"PRId64" mb, "
-+		"%"PRId64"mb used (%"PRId64"mb available)\n",
-+		program_name, hurd_myself (),
-+		(mapped_start - mapped_now) / 1024 / 1024,
-+		mapped_now / 1024 / 1024,
-+		(uint64_t) GC_available_bytes / 1024 / 1024);
++    if (! all
++	&& GC_viengoos_scheduler
++	&& GC_mapped_bytes < 7 * (MAPPED_THRESHOLD / 8)) {
++      /* Unmap at least enough memory such that we are below 7/8s
++	 threshold.  */
++    out:
++      debug (1, "After unmapping %"PRId64" mb, "
++	     "%"PRId64"mb mapped (threshold: %"PRId64" mb)",
++	     (mapped_at_start - GC_mapped_bytes) / 1024 / 1024,
++	     GC_mapped_bytes / 1024 / 1024,
++	     (uint64_t) MAPPED_THRESHOLD / 1024 / 1024);
 +      return;
 +    }
 +
@@ -80,10 +103,11 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 -	  hhdr -> hb_flags |= WAS_UNMAPPED;
 -    	}
 +
-+	if (GC_viengoos_scheduler
++	if (all
++	    || GC_viengoos_scheduler
 +	    || ((last_rec > GC_gc_no || last_rec < threshold)
 +		&& threshold < GC_gc_no /* not recently wrapped */)) {
-+	    GC_unmap((ptr_t)h, hhdr -> hb_sz);
++	    unmap((ptr_t)h, hhdr -> hb_sz);
 +	    hhdr -> hb_flags |= WAS_UNMAPPED;
 +
 +	    if (0)
@@ -133,7 +157,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +
 +		      /* Unmap if not already unmapped.  */
 +		      if (IS_MAPPED (nexthdr)) {
-+			  GC_unmap((ptr_t)next, nexthdr -> hb_sz);
++			  unmap((ptr_t)next, nexthdr -> hb_sz);
 +			  nexthdr -> hb_flags |= WAS_UNMAPPED;
 +		      }
 +
@@ -170,10 +194,17 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
        }
 -    }  
 +    }
++
++    goto out;
++}
++
++void GC_unmap_old(void)
++{
++  GC_try_unmap (false);
  }
  
  /* Merge all unmapped blocks that are adjacent to other free		*/
-@@ -413,6 +526,12 @@ void GC_unmap_old(void)
+@@ -413,6 +550,12 @@ void GC_unmap_old(void)
  /* fully mapped or fully unmapped.					*/
  void GC_merge_unmapped(void)
  {
@@ -186,7 +217,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
      struct hblk * h, *next;
      hdr * hhdr, *nexthdr;
      word size, nextsize;
-@@ -600,12 +719,26 @@ GC_allochblk_nth(size_t sz, int kind, un
+@@ -600,12 +743,26 @@ GC_allochblk_nth(size_t sz, int kind, un
  
      size_needed = HBLKSIZE * OBJ_SZ_TO_BLOCKS(sz);
  
@@ -213,7 +244,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
  	    if (size_avail != size_needed
  		&& !GC_use_entire_heap
  		&& !GC_dont_gc
-@@ -642,7 +775,7 @@ GC_allochblk_nth(size_t sz, int kind, un
+@@ -642,7 +799,7 @@ GC_allochblk_nth(size_t sz, int kind, un
  	        next_size = (signed_word)(thishdr -> hb_sz);
  	        if (next_size < size_avail
  	          && next_size >= size_needed
@@ -222,7 +253,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
  	          continue;
  	        }
  	      }
-@@ -748,10 +881,12 @@ GC_allochblk_nth(size_t sz, int kind, un
+@@ -748,10 +905,12 @@ GC_allochblk_nth(size_t sz, int kind, un
  		/* hbp may be on the wrong freelist; the parameter n	*/
  		/* is important.					*/
  		hbp = GC_get_first_part(hbp, hhdr, size_needed, n);
@@ -238,22 +269,26 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
      /* Add it to map of valid blocks */
 diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.in gc/alloc.c gc.new/alloc.c
 --- gc/alloc.c	2007-06-22 04:40:30.000000000 +0200
-+++ gc.new/alloc.c	2009-02-24 03:07:16.000000000 +0100
-@@ -15,6 +15,13 @@
++++ gc.new/alloc.c	2009-03-06 20:42:23.000000000 +0100
+@@ -15,6 +15,17 @@
   *
   */
  
 +#ifdef __gnu_hurd_viengoos__
 +#include <viengoos/activity.h>
 +#include <hurd/as.h>
++#else
++#define debug(lvl, fmt, ...) do {} while (0)
 +#endif
 +#include <stdint.h>
 +#include <sys/time.h>
 +#include <assert.h>
++#include <inttypes.h>
++#include <stdbool.h>
  
  # include "private/gc_priv.h"
  
-@@ -24,6 +31,8 @@
+@@ -24,6 +35,8 @@
  #   include <sys/types.h>
  # endif
  
@@ -262,16 +297,14 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
  /*
   * Separate free lists are maintained for different sized objects
   * up to MAXOBJBYTES.
-@@ -223,9 +232,160 @@ void GC_clear_a_few_frames()
+@@ -223,9 +236,188 @@ void GC_clear_a_few_frames()
  /* limits used by blacklisting.						*/
  static word GC_collect_at_heapsize = (word)(-1);
  
 +int GC_viengoos_scheduler = 1;
-+uintptr_t GC_available_bytes = (uintptr_t) ((256 + 128) * 1024 * 1024);
++uintptr_t GC_available_bytes = (uintptr_t) ((1024 + 717) * 1024 * 1024);
 +int GC_could_unmap;
 +
-+#define THRESHOLD ((uintptr_t) (15 * (GC_available_bytes / 16)))
-+
 +#ifdef __gnu_hurd_viengoos__
 +static void *
 +gather_stats (void *arg)
@@ -322,29 +355,27 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +	  panic ("Unknown event and unrequested event: %"PRIdPTR, info.event);
 +	}
 +
-+      extern char *program_name;
-+      if (0)
-+	printf ("%s("VG_THREAD_ID_FMT"): %s: "
-+		"%"PRId64" alloced: %"PRId64"mb, heap: %"PRId64"mb, "
-+		"mapped:%"PRId64"mb, unmapped: %"PRId64"mb, "
-+		"available: %"PRIdPTR"mb => %"PRIdPTR"mb, "
-+		"low-water: %"PRIdPTR"mb\n",
-+		program_name, hurd_myself (),
-+		info.event == vg_activity_info_stats
-+		? "Period" : DEBUG_BOLD ("PRESSURE"),
-+		(uint64_t) (info.event == vg_activity_info_stats
-+			    ? period : info.pressure.amount),
-+		(uint64_t) GC_adj_bytes_allocd() / 1024 / 1024,
-+		(uint64_t) GC_get_heap_size () / 1024 / 1024,
-+		(uint64_t) (GC_get_heap_size () - GC_unmapped_bytes)
-+		/ 1024 / 1024,
-+		(uint64_t) GC_unmapped_bytes / 1024 / 1024,
-+		(uint64_t) old_avail / 1024 / 1024,
-+		(uint64_t) GC_available_bytes / 1024 / 1024,
-+		(uint64_t) THRESHOLD / 1024 / 1024);
++      debug (1, "%s %"PRId64": "
++	     "available: %"PRIdPTR"mb => %"PRIdPTR"mb; "
++	     "%"PRId64"mb, heap: %"PRId64"mb - "
++	     "mapped: %"PRId64"mb = unmapped: %"PRId64"mb, "
++	     "(threshold: %"PRIdPTR"mb); %"PRId64" free",
++	     info.event == vg_activity_info_stats
++	     ? "Period" : DEBUG_BOLD ("PRESSURE"),
++	     (uint64_t) (info.event == vg_activity_info_stats
++			 ? period
++			 : (((uint64_t) info.pressure.amount
++			     * PAGESIZE) / 1024 / 1024)),
++	     (uint64_t) old_avail / 1024 / 1024,
++	     (uint64_t) GC_available_bytes / 1024 / 1024,
++	     (uint64_t) GC_get_heap_size () / 1024 / 1024,
++	     (uint64_t) GC_mapped_bytes / 1024 / 1024,
++	     (uint64_t) GC_unmapped_bytes / 1024 / 1024,
++	     (uint64_t) MAPPED_THRESHOLD / 1024 / 1024,
++	     (uint64_t) GC_free_bytes / 1024 / 1024);
 +      if (0 && info.event == vg_activity_info_pressure)
-+	printf ("pages to free: %d mb\n",
-+		info.pressure.amount * (PAGESIZE / 1024) / 1024);
++	s_printf ("pages to free: %d mb\n",
++		  info.pressure.amount * (PAGESIZE / 1024) / 1024);
 +    }
 +}
 +#endif
@@ -352,16 +383,11 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
  /* Have we allocated enough to amortize a collection? */
  GC_bool GC_should_collect(void)
  {
-+#ifndef USE_MUNMAP
-+# define GC_unmapped_bytes 0
-+#endif
-+
 +  if (GC_viengoos_scheduler)
 +    {
-+      int alloced = GC_adj_bytes_allocd();
 +#ifdef __gnu_hurd_viengoos__
 +      static int init;
-+      if (! init)
++      if (unlikely (! init))
 +	{
 +	  pthread_t tid;
 +	  error_t err = pthread_create (&tid, NULL, gather_stats, NULL);
@@ -371,59 +397,96 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +	}
 +#endif
 +
-+      /* Do some clean up if the mapped memory is more than 15/16s the
-+	 available memory.  We choose 15/16s as we need to consider
-+	 meta-data overhead and as we really want to avoid the system
-+	 pager kicking in.  */
-+      GC_bool r = GC_get_heap_size() - GC_unmapped_bytes > THRESHOLD;
++      /* GC_allochblk_nth calls this function repeatedly without
++	 immediately following up with a GC.  Once we decide to GC,
++	 just return true until it happens.  */
++      static int scheduled_gc_no;
++      if (scheduled_gc_no == GC_gc_no + 1)
++	return true;
++
++      bool r = false;
++
++      /* Unmap some memory if the number of mapped bytes exceeds
++	 MAPPED_THRESHOLD.  */
++      if (GC_mapped_bytes > MAPPED_THRESHOLD
++	  && GC_free_bytes > MAPPED_THRESHOLD / 3) {
++	debug (1, "Unmapping: "
++	       "heap: %"PRId64" - unmapped: %"PRId64" = "
++	       "mapped: %"PRId64" (threshold: %"PRId64")\n"
++	       "free: %"PRId64" (threshold: %"PRId64"); avail: %"PRId64,
++ 	       (uint64_t) GC_get_heap_size() / 1024 / 1024,
++	       (uint64_t) GC_unmapped_bytes / 1024 / 1024,
++	       (uint64_t) GC_mapped_bytes / 1024 / 1024,
++	       (uint64_t) MAPPED_THRESHOLD / 1024 / 1024,
++	       (uint64_t) GC_free_bytes / 1024 / 1024,
++	       MAPPED_THRESHOLD / 1024 / 1024 / 3,
++	       GC_available_bytes / 1024 / 1024);
 +
-+      if (r && (GC_get_heap_size () - GC_unmapped_bytes - alloced
-+		> GC_available_bytes / 3)){
 +	/* The number of unused mapped bytes is greater than a third
 +	   of the total available memory.  Before doing a GC, try to
 +	   unmap some free mapped pages.  */
++	  uint64_t mapped = GC_mapped_bytes;
 +	  GC_unmap_old ();
 +
-+	  r = GC_get_heap_size() - GC_unmapped_bytes > THRESHOLD;
++	  debug (1, DEBUG_BOLD ("After unmap: mapped %"PRId64" => %"PRId64" (threshold: %"PRId64"); delta: %"PRId64),
++		 (uint64_t) mapped / 1024 / 1024,
++		 (uint64_t) GC_mapped_bytes / 1024 / 1024,
++		 MAPPED_THRESHOLD / 1024 / 1024,
++		 (uint64_t) (mapped - GC_mapped_bytes) / 1024 / 1024);
++
++	  if (GC_mapped_bytes > 7 * (MAPPED_THRESHOLD / 8))
++	    /* Force a GC.  We didn't manage to unmap enough.  */
++	    r = true;
 +      }
 +
-+      if (r) {
-+	  if (alloced < GC_available_bytes / 100)
-+	      /* If we have not allocated anything since the last
-+		 collection, don't do a collection.  */
-+	      return FALSE;
-+
-+	  static int warning;
-+	  /* Only print this once per-gc.  GC_allochblk_nth calls this
-+	     function repeatedly without immediately following up with
-+	     a GC.  */
-+	  if (warning != GC_gc_no) {
-+	      warning = GC_gc_no;
-+	      extern char *program_name;
-+	      if (0)
-+		printf ("%s("VG_THREAD_ID_FMT"): Scheduling GC #%u: "
-+			"alloced: %"PRId64"mb, heap: %"PRId64"mb, "
-+			"mapped: %"PRId64"mb, unmapped: %"PRId64"mb, "
-+			"available: %"PRIdPTR"mb, low-water: %"PRIdPTR"mb\n",
-+			program_name, hurd_myself (),
-+			(uint64_t) GC_gc_no,
-+			(uint64_t) alloced / 1024 / 1024,
-+			(uint64_t) GC_get_heap_size () / 1024 / 1024,
-+			(uint64_t) (GC_get_heap_size () - GC_unmapped_bytes)
-+			/ 1024 / 1024,
-+			(uint64_t) GC_unmapped_bytes / 1024 / 1024,
-+			(uint64_t) GC_available_bytes / 1024 / 1024,
-+			(uint64_t) THRESHOLD / 1024 / 1024);
++      /* If 
++
++	   - the number of mapped bytes is approaching the mapped
++	     threshold and the number of unallocated heap bytes is low
++
++	 or
++
++	   - the number of mapped bytes exceeded the mapped threshold
++             and we couldn't unmap enough,
++
++	 then do a GC.  */
++      if ((GC_mapped_bytes > 15 * MAPPED_THRESHOLD / 16
++	   && GC_free_bytes < MAPPED_THRESHOLD / 16)
++	  || r) {
++	int alloced = GC_adj_bytes_allocd();
++
++	if (alloced < MAPPED_THRESHOLD / 100)
++	  /* But not if we've only allocated a few bytes since the
++	     last GC.  */
++	  {
++	    debug (1, "Delaying GC: alloced %d", alloced / 1024 / 1024);
++	    return FALSE;
 +	  }
++
++	debug (1, "Scheduling GC #%u: alloced since last gc: %"PRId64"\n"
++	       "heap: %"PRId64" mb - unmapped: %"PRId64" mb"
++	       " = mapped: %"PRId64" (threshold: %"PRId64"); "
++	       "free: %"PRId64" mb, available: %"PRId64" mb",
++	       (uint64_t) GC_gc_no,
++	       (uint64_t) alloced / 1024 / 1024,
++	       (uint64_t) GC_get_heap_size () / 1024 / 1024,
++	       (uint64_t) GC_unmapped_bytes / 1024 / 1024,
++	       (uint64_t) GC_mapped_bytes / 1024 / 1024,
++	       (uint64_t) MAPPED_THRESHOLD / 1024 / 1024,
++	       (uint64_t) GC_free_bytes / 1024 / 1024,
++	       (uint64_t) GC_available_bytes / 1024 / 1024);
++
++	  scheduled_gc_no = GC_gc_no + 1;
++	  return TRUE;
 +      }
 +
-+      return r;
++      return FALSE;
 +    }
 +  else
      return(GC_adj_bytes_allocd() >= min_bytes_allocd()
  	   || GC_heapsize >= GC_collect_at_heapsize);
  }
-@@ -256,6 +416,8 @@ void GC_maybe_gc(void)
+@@ -256,6 +448,8 @@ void GC_maybe_gc(void)
              n_partial_gcs = 0;
              return;
          } else {
@@ -432,7 +495,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
  #   	  ifdef PARALLEL_MARK
  	    GC_wait_for_reclaim();
  #   	  endif
-@@ -266,6 +428,7 @@ void GC_maybe_gc(void)
+@@ -266,6 +460,7 @@ void GC_maybe_gc(void)
       		  (unsigned long)GC_gc_no+1,
  		  (long)GC_bytes_allocd);
  	    }
@@ -440,7 +503,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
  	    GC_promote_black_lists();
  	    (void)GC_reclaim_all((GC_stop_func)0, TRUE);
  	    GC_clear_marks();
-@@ -275,7 +438,11 @@ void GC_maybe_gc(void)
+@@ -275,7 +470,11 @@ void GC_maybe_gc(void)
            } else {
              n_partial_gcs++;
            }
@@ -452,7 +515,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
          /* We try to mark with the world stopped.	*/
          /* If we run out of time, this turns into	*/
          /* incremental marking.			*/
-@@ -294,6 +461,8 @@ void GC_maybe_gc(void)
+@@ -294,6 +493,8 @@ void GC_maybe_gc(void)
  	        GC_n_attempts++;
  	    }
  	}
@@ -461,7 +524,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
      }
  }
  
-@@ -325,6 +494,10 @@ GC_bool GC_try_to_collect_inner(GC_stop_
+@@ -325,6 +526,10 @@ GC_bool GC_try_to_collect_inner(GC_stop_
  	   "Initiating full world-stop collection %lu after %ld allocd bytes\n",
  	   (unsigned long)GC_gc_no+1, (long)GC_bytes_allocd);
      }
@@ -472,7 +535,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
      GC_promote_black_lists();
      /* Make sure all blocks have been reclaimed, so sweep routines	*/
      /* don't see cleared mark bits.					*/
-@@ -337,7 +510,8 @@ GC_bool GC_try_to_collect_inner(GC_stop_
+@@ -337,7 +542,8 @@ GC_bool GC_try_to_collect_inner(GC_stop_
   	if ((GC_find_leak || stop_func != GC_never_stop_func)
  	    && !GC_reclaim_all(stop_func, FALSE)) {
  	    /* Aborted.  So far everything is still consistent.	*/
@@ -482,7 +545,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
  	}
      GC_invalidate_mark_state();  /* Flush mark stack.	*/
      GC_clear_marks();
-@@ -354,7 +528,8 @@ GC_bool GC_try_to_collect_inner(GC_stop_
+@@ -354,7 +560,8 @@ GC_bool GC_try_to_collect_inner(GC_stop_
  	GC_unpromote_black_lists();
        } /* else we claim the world is already still consistent.  We'll 	*/
          /* finish incrementally.					*/
@@ -492,7 +555,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
      }
      GC_finish_collection();
      if (GC_print_stats) {
-@@ -362,7 +537,10 @@ GC_bool GC_try_to_collect_inner(GC_stop_
+@@ -362,7 +569,10 @@ GC_bool GC_try_to_collect_inner(GC_stop_
          GC_log_printf("Complete collection took %lu msecs\n",
                    MS_TIME_DIFF(current_time,start_time));
      }
@@ -504,7 +567,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
  }
  
  
-@@ -390,6 +568,8 @@ void GC_collect_a_little_inner(int n)
+@@ -390,6 +600,8 @@ void GC_collect_a_little_inner(int n)
      
      if (GC_dont_gc) return;
      if (GC_incremental && GC_collection_in_progress()) {
@@ -513,7 +576,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
      	for (i = GC_deficit; i < GC_RATE*n; i++) {
      	    if (GC_mark_some((ptr_t)0)) {
      	        /* Need to finish a collection */
-@@ -415,6 +595,8 @@ void GC_collect_a_little_inner(int n)
+@@ -415,6 +627,8 @@ void GC_collect_a_little_inner(int n)
      	}
      	if (GC_deficit > 0) GC_deficit -= GC_RATE*n;
  	if (GC_deficit < 0) GC_deficit = 0;
@@ -522,7 +585,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
      } else {
          GC_maybe_gc();
      }
-@@ -904,6 +1086,8 @@ GC_bool GC_expand_hp_inner(word n)
+@@ -904,6 +1118,8 @@ GC_bool GC_expand_hp_inner(word n)
      /* Force GC before we are likely to allocate past expansion_slop */
        GC_collect_at_heapsize =
           GC_heapsize + expansion_slop - 2*MAXHINCR*HBLKSIZE;
@@ -531,7 +594,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
  #     if defined(LARGE_CONFIG)
          if (GC_collect_at_heapsize < GC_heapsize /* wrapped */)
           GC_collect_at_heapsize = (word)(-1);
-@@ -1021,3 +1205,410 @@ ptr_t GC_allocobj(size_t gran, int kind)
+@@ -1021,3 +1237,452 @@ ptr_t GC_allocobj(size_t gran, int kind)
      
      return(*flh);
  }
@@ -588,6 +651,7 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +
 +#ifdef RM_INTERN
 +# include "../viengoos/zalloc.h"
++# include "../viengoos/timer.h"
 +#else
 +# include <stdlib.h>
 +#endif
@@ -596,12 +660,33 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +#include <hurd/ihash.h>
 +#include <stddef.h>
 +#include <string.h>
++#include <stdbool.h>
 +
 +#ifdef __gnu_hurd_viengoos__
 +# include <viengoos/bits/sys.h>
 +# include <s-printf.h>
 +# include <hurd/thread.h>
 +# include <viengoos/misc.h>
++
++# ifndef NPROFILE
++#  include "../viengoos/debug.h"
++
++static void
++profile_command (int argc, char *argv[])
++{
++  profile_stats_dump ();
++}
++
++static struct debug_command profile_debug =
++  {
++    .name = "profile",
++    .short_name = 0,
++    .help = "Display profile data.",
++    .function = profile_command
++  };
++
++# endif
++
 +#else
 +# include <pthread.h>
 +# include <sys/time.h>
@@ -632,7 +717,11 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +now (void)
 +{
 +#ifdef __gnu_hurd_viengoos__
-+  return VG_TIME->ns_since_boot * 1000;
++# ifdef RM_INTERN
++  return time_data.ns_since_boot / 1000;
++# else
++  return VG_TIME->ns_since_boot / 1000;
++# endif
 +#else
 +  struct timeval t;
 +  struct timezone tz;
@@ -656,16 +745,26 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +};
 +
 +/* The number of threads we support.  */
-+#define THREADS 4
++#ifdef RM_INTERN
++# define THREADS 1
++#else
++# define THREADS 4
++#endif
 +static int thread_count;
 +
 +struct profile_block
 +{
 +#ifdef __gnu_hurd_viengoos__
-+# define MYSELF() hurd_myself ()
++# ifdef RM_INTERN
++#  define MYSELF() 0
++# else
++#  define MYSELF() hurd_myself ()
++# endif
++# define THREAD_FMT VG_THREAD_ID_FMT
 +  vg_thread_id_t tid;
 +#else
 +# define MYSELF() pthread_self ()
++# define THREAD_FMT "%d"
 +  pthread_t tid;
 +#endif
 +
@@ -705,6 +804,11 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +
 +  /* New thread.  */
 +
++#ifdef RM_INTERN
++  if (thread_count == 0)
++    debug_register (&profile_debug);
++#endif
++
 +  i = __sync_fetch_and_add (&thread_count, 1);
 +  if (i >= THREADS)
 +    panic ("More threads than profile space available!");
@@ -826,7 +930,8 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +	      order[j + 1] = t;
 +	    }
 +
-+      s_printf ("Thread: "VG_THREAD_ID_FMT"\n", pb->tid);
++#define MAX(a, b) ((a) > (b) ? (a) : (b))
++      s_printf ("Thread: "THREAD_FMT"\n", pb->tid);
 +      for (j = 0; j < count; j ++)
 +	{
 +	  i = order[j];
@@ -850,9 +955,10 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +		      pb->call_sites[i].time / pb->call_sites[i].calls,
 +		      (int) ((10 * pb->call_sites[i].time)
 +			     / pb->call_sites[i].calls) % 10,
-+		      (int) ((100 * pb->call_sites[i].time) / (n - pb->epoch)),
 +		      (int) ((100 * pb->call_sites[i].time)
-+			     / pb->profiled_time));
++			     / MAX (n - pb->epoch, 1)),
++		      (int) ((100 * pb->call_sites[i].time)
++			     / MAX (pb->profiled_time, 1)));
 +	}
 +
 +      s_printf ("profiled time: %"PRId64" ms (%d%%), calls: %"PRId64"\n",
@@ -940,11 +1046,29 @@ diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.i
 +  else
 +    site->nested_calls ++;
 +}
++#endif
+diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.in gc/include/private/gc_priv.h gc.new/include/private/gc_priv.h
+--- gc/include/private/gc_priv.h	2007-06-24 04:01:15.000000000 +0200
++++ gc.new/include/private/gc_priv.h	2009-03-06 16:53:23.000000000 +0100
+@@ -2037,4 +2037,15 @@ extern void GC_reset_fault_handler(void)
+ 
+ # endif /* Need to handle address faults.	*/
+ 
++#include <stdint.h>
++extern uintptr_t GC_available_bytes;
 +
++#ifndef USE_MUNMAP
++# define GC_unmapped_bytes 0
 +#endif
-diff -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.in -rup gc.orig/os_dep.c gc/os_dep.c
---- gc.orig/os_dep.c	2007-06-29 21:17:44.000000000 +0200
-+++ gc/os_dep.c	2009-02-21 02:33:55.000000000 +0100
++
++#define MAPPED_THRESHOLD (15 * (GC_available_bytes / 16))
++#define GC_mapped_bytes (GC_heapsize - GC_unmapped_bytes)
++#define GC_free_bytes (GC_large_free_bytes - GC_unmapped_bytes)
++
+ # endif /* GC_PRIVATE_H */
+diff -urp -x config.guess -x config.sub -x 'aclocal*' -x configure -x Makefile.in gc/os_dep.c gc.new/os_dep.c
+--- gc/os_dep.c	2007-06-29 21:17:44.000000000 +0200
++++ gc.new/os_dep.c	2009-03-01 21:06:48.000000000 +0100
 @@ -2064,6 +2064,17 @@ ptr_t GC_unmap_end(ptr_t start, size_t b
      return end_addr;
  }