Merge remote-tracking branch 'origin/stable-2.0'
[bpt/guile.git] / libguile / gc.c
index 40b158a..71efd03 100644 (file)
@@ -1,4 +1,4 @@
-/* Copyright (C) 1995,1996,1997,1998,1999,2000,2001, 2002, 2003, 2006, 2008, 2009, 2010, 2011 Free Software Foundation, Inc.
+/* Copyright (C) 1995,1996,1997,1998,1999,2000,2001, 2002, 2003, 2006, 2008, 2009, 2010, 2011, 2012 Free Software Foundation, Inc.
  *
  * This library is free software; you can redistribute it and/or
  * modify it under the terms of the GNU Lesser General Public License
@@ -27,6 +27,8 @@
 #include <stdio.h>
 #include <errno.h>
 #include <string.h>
+#include <stdlib.h>
+#include <math.h>
 
 #ifdef __ia64__
 #include <ucontext.h>
@@ -45,7 +47,6 @@ extern unsigned long * __libc_ia64_register_backing_store_base;
 #include "libguile/root.h"
 #include "libguile/strings.h"
 #include "libguile/vectors.h"
-#include "libguile/weaks.h"
 #include "libguile/hashtab.h"
 #include "libguile/tags.h"
 
@@ -64,10 +65,6 @@ extern unsigned long * __libc_ia64_register_backing_store_base;
 #include "libguile/debug-malloc.h"
 #endif
 
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
-
 #ifdef HAVE_UNISTD_H
 #include <unistd.h>
 #endif
@@ -83,9 +80,8 @@ int scm_expensive_debug_cell_accesses_p = 0;
 int scm_debug_cells_gc_interval = 0;
 
 /* Hash table that keeps a reference to objects the user wants to protect from
-   garbage collection.  It could arguably be private but applications have come
-   to rely on it (e.g., Lilypond 2.13.9).  */
-SCM scm_protects;
+   garbage collection.  */
+static SCM scm_protects;
 
 
 #if (SCM_DEBUG_CELL_ACCESSES == 1)
@@ -190,6 +186,36 @@ SCM_DEFINE (scm_set_debug_cell_accesses_x, "set-debug-cell-accesses!", 1, 0, 0,
 #endif  /* SCM_DEBUG_CELL_ACCESSES == 1 */
 
 \f
+
+/* Compatibility.  */
+
+#ifndef HAVE_GC_GET_HEAP_USAGE_SAFE
+static void
+GC_get_heap_usage_safe (GC_word *pheap_size, GC_word *pfree_bytes,
+                        GC_word *punmapped_bytes, GC_word *pbytes_since_gc,
+                        GC_word *ptotal_bytes)
+{
+  *pheap_size = GC_get_heap_size ();
+  *pfree_bytes = GC_get_free_bytes ();
+#ifdef HAVE_GC_GET_UNMAPPED_BYTES
+  *punmapped_bytes = GC_get_unmapped_bytes ();
+#else
+  *punmapped_bytes = 0;
+#endif
+  *pbytes_since_gc = GC_get_bytes_since_gc ();
+  *ptotal_bytes = GC_get_total_bytes ();
+}
+#endif
+
+#ifndef HAVE_GC_GET_FREE_SPACE_DIVISOR
+static GC_word
+GC_get_free_space_divisor (void)
+{
+  return GC_free_space_divisor;
+}
+#endif
+
+\f
 /* Hooks.  */
 scm_t_c_hook scm_before_gc_c_hook;
 scm_t_c_hook scm_before_mark_c_hook;
@@ -201,6 +227,10 @@ scm_t_c_hook scm_after_gc_c_hook;
 static void
 run_before_gc_c_hook (void)
 {
+  if (!SCM_I_CURRENT_THREAD)
+    /* GC while a thread is spinning up; punt.  */
+    return;
+
   scm_c_hook_run (&scm_before_gc_c_hook, NULL);
 }
 
@@ -211,6 +241,9 @@ unsigned long scm_gc_ports_collected = 0;
 static long gc_time_taken = 0;
 static long gc_start_time = 0;
 
+static unsigned long free_space_divisor;
+static unsigned long minimum_free_space_divisor;
+static double target_free_space_divisor;
 
 static unsigned long protected_obj_count = 0;
 
@@ -272,14 +305,12 @@ SCM_DEFINE (scm_gc_stats, "gc-stats", 0, 0, 0,
 #define FUNC_NAME s_scm_gc_stats
 {
   SCM answer;
-  size_t heap_size, free_bytes, bytes_since_gc, total_bytes;
+  GC_word heap_size, free_bytes, unmapped_bytes, bytes_since_gc, total_bytes;
   size_t gc_times;
 
-  heap_size      = GC_get_heap_size ();
-  free_bytes     = GC_get_free_bytes ();
-  bytes_since_gc = GC_get_bytes_since_gc ();
-  total_bytes    = GC_get_total_bytes ();
-  gc_times       = GC_gc_no;
+  GC_get_heap_usage_safe (&heap_size, &free_bytes, &unmapped_bytes,
+                          &bytes_since_gc, &total_bytes);
+  gc_times = GC_gc_no;
 
   answer =
     scm_list_n (scm_cons (sym_gc_time_taken, scm_from_long (gc_time_taken)),
@@ -353,6 +384,9 @@ SCM_DEFINE (scm_gc, "gc", 0, 0, 0,
 #define FUNC_NAME s_scm_gc
 {
   scm_i_gc ("call");
+  /* If you're calling scm_gc(), you probably want synchronous
+     finalization.  */
+  GC_invoke_finalizers ();
   return SCM_UNSPECIFIED;
 }
 #undef FUNC_NAME
@@ -577,11 +611,23 @@ scm_getenv_int (const char *var, int def)
   return res;
 }
 
+#ifndef HAVE_GC_SET_FINALIZE_ON_DEMAND
+static void
+GC_set_finalize_on_demand (int foo)
+{
+  GC_finalize_on_demand = foo;
+}
+#endif
+
 void
 scm_storage_prehistory ()
 {
   GC_all_interior_pointers = 0;
-  GC_set_free_space_divisor (scm_getenv_int ("GC_FREE_SPACE_DIVISOR", 3));
+  free_space_divisor = scm_getenv_int ("GC_FREE_SPACE_DIVISOR", 3);
+  minimum_free_space_divisor = free_space_divisor;
+  target_free_space_divisor = free_space_divisor;
+  GC_set_free_space_divisor (free_space_divisor);
+  GC_set_finalize_on_demand (1);
 
   GC_INIT ();
 
@@ -725,7 +771,8 @@ accumulate_gc_timer (void * hook_data SCM_UNUSED,
                 void *data SCM_UNUSED)
 {
   if (gc_start_time)
-    { long now = scm_c_get_internal_run_time ();
+    {
+      long now = scm_c_get_internal_run_time ();
       gc_time_taken += now - gc_start_time;
       gc_start_time = 0;
     }
@@ -733,6 +780,168 @@ accumulate_gc_timer (void * hook_data SCM_UNUSED,
   return NULL;
 }
 
+/* Return some idea of the memory footprint of a process, in bytes.
+   Currently only works on Linux systems.  */
+static size_t
+get_image_size (void)
+{
+  unsigned long size, resident, share;
+  size_t ret = 0;
+
+  FILE *fp = fopen ("/proc/self/statm", "r");
+
+  if (fp && fscanf (fp, "%lu %lu %lu", &size, &resident, &share) == 3)
+    ret = resident * 4096;
+
+  if (fp)
+    fclose (fp);
+
+  return ret;
+}
+
+/* These are discussed later.  */
+static size_t bytes_until_gc;
+static scm_i_pthread_mutex_t bytes_until_gc_lock = SCM_I_PTHREAD_MUTEX_INITIALIZER;
+
+/* Make GC run more frequently when the process image size is growing,
+   measured against the number of bytes allocated through the GC.
+
+   If Guile is allocating at a GC-managed heap size H, libgc will tend
+   to limit the process image size to H*N.  But if at the same time the
+   user program is mallocating at a rate M bytes per GC-allocated byte,
+   then the process stabilizes at H*N*M -- assuming that collecting data
+   will result in malloc'd data being freed.  It doesn't take a very
+   large M for this to be a bad situation.  To limit the image size,
+   Guile should GC more often -- the bigger the M, the more often.
+
+   Numeric functions that produce bigger and bigger integers are
+   pessimal, because M is an increasing function of time.  Here is an
+   example of such a function:
+
+      (define (factorial n)
+        (define (fac n acc)
+          (if (<= n 1)
+            acc
+            (fac (1- n) (* n acc))))
+        (fac n 1))
+
+   It is possible for a process to grow for reasons that will not be
+   solved by faster GC.  In that case M will be estimated as
+   artificially high for a while, and so GC will happen more often on
+   the Guile side.  But when it stabilizes, Guile can ease back the GC
+   frequency.
+
+   The key is to measure process image growth, not mallocation rate.
+   For maximum effectiveness, Guile reacts quickly to process growth,
+   and exponentially backs down when the process stops growing.
+
+   See http://thread.gmane.org/gmane.lisp.guile.devel/12552/focus=12936
+   for further discussion.
+ */
+static void *
+adjust_gc_frequency (void * hook_data SCM_UNUSED,
+                     void *fn_data SCM_UNUSED,
+                     void *data SCM_UNUSED)
+{
+  static size_t prev_image_size = 0;
+  static size_t prev_bytes_alloced = 0;
+  size_t image_size;
+  size_t bytes_alloced;
+  
+  scm_i_pthread_mutex_lock (&bytes_until_gc_lock);
+  bytes_until_gc = GC_get_heap_size ();
+  scm_i_pthread_mutex_unlock (&bytes_until_gc_lock);
+
+  image_size = get_image_size ();
+  bytes_alloced = GC_get_total_bytes ();
+
+#define HEURISTICS_DEBUG 0
+
+#if HEURISTICS_DEBUG
+  fprintf (stderr, "prev image / alloced: %lu / %lu\n", prev_image_size, prev_bytes_alloced);
+  fprintf (stderr, "     image / alloced: %lu / %lu\n", image_size, bytes_alloced);
+  fprintf (stderr, "divisor %lu / %f\n", free_space_divisor, target_free_space_divisor);
+#endif
+
+  if (prev_image_size && bytes_alloced != prev_bytes_alloced)
+    {
+      double growth_rate, new_target_free_space_divisor;
+      double decay_factor = 0.5;
+      double hysteresis = 0.1;
+
+      growth_rate = ((double) image_size - prev_image_size)
+        / ((double)bytes_alloced - prev_bytes_alloced);
+      
+#if HEURISTICS_DEBUG
+      fprintf (stderr, "growth rate %f\n", growth_rate);
+#endif
+
+      new_target_free_space_divisor = minimum_free_space_divisor;
+
+      if (growth_rate > 0)
+        new_target_free_space_divisor *= 1.0 + growth_rate;
+
+#if HEURISTICS_DEBUG
+      fprintf (stderr, "new divisor %f\n", new_target_free_space_divisor);
+#endif
+
+      if (new_target_free_space_divisor < target_free_space_divisor)
+        /* Decay down.  */
+        target_free_space_divisor =
+          (decay_factor * target_free_space_divisor
+           + (1.0 - decay_factor) * new_target_free_space_divisor);
+      else
+        /* Jump up.  */
+        target_free_space_divisor = new_target_free_space_divisor;
+
+#if HEURISTICS_DEBUG
+      fprintf (stderr, "new target divisor %f\n", target_free_space_divisor);
+#endif
+
+      if (free_space_divisor + 0.5 + hysteresis < target_free_space_divisor
+          || free_space_divisor - 0.5 - hysteresis > target_free_space_divisor)
+        {
+          free_space_divisor = lround (target_free_space_divisor);
+#if HEURISTICS_DEBUG
+          fprintf (stderr, "new divisor %lu\n", free_space_divisor);
+#endif
+          GC_set_free_space_divisor (free_space_divisor);
+        }
+    }
+
+  prev_image_size = image_size;
+  prev_bytes_alloced = bytes_alloced;
+
+  return NULL;
+}
+
+/* The adjust_gc_frequency routine handles transients in the process
+   image size.  It can't handle instense non-GC-managed steady-state
+   allocation though, as it decays the FSD at steady-state down to its
+   minimum value.
+
+   The only real way to handle continuous, high non-GC allocation is to
+   let the GC know about it.  This routine can handle non-GC allocation
+   rates that are similar in size to the GC-managed heap size.
+ */
+
+void
+scm_gc_register_allocation (size_t size)
+{
+  scm_i_pthread_mutex_lock (&bytes_until_gc_lock);
+  if (bytes_until_gc - size > bytes_until_gc)
+    {
+      bytes_until_gc = GC_get_heap_size ();
+      scm_i_pthread_mutex_unlock (&bytes_until_gc_lock);
+      GC_gcollect ();
+    }
+  else
+    {
+      bytes_until_gc -= size;
+      scm_i_pthread_mutex_unlock (&bytes_until_gc_lock);
+    }
+}
+
 
 \f
 
@@ -835,6 +1044,16 @@ scm_init_gc ()
   scm_c_hook_add (&scm_before_gc_c_hook, start_gc_timer, NULL, 0);
   scm_c_hook_add (&scm_after_gc_c_hook, accumulate_gc_timer, NULL, 0);
 
+#if HAVE_GC_GET_HEAP_USAGE_SAFE
+  /* GC_get_heap_usage does not take a lock, and so can run in the GC
+     start hook.  */
+  scm_c_hook_add (&scm_before_gc_c_hook, adjust_gc_frequency, NULL, 0);
+#else
+  /* GC_get_heap_usage might take a lock (and did from 7.2alpha1 to
+     7.2alpha7), so call it in the after_gc_hook.  */
+  scm_c_hook_add (&scm_after_gc_c_hook, adjust_gc_frequency, NULL, 0);
+#endif
+
 #ifdef HAVE_GC_SET_START_CALLBACK
   GC_set_start_callback (run_before_gc_c_hook);
 #endif