Better resolution for issue #11849: Ensure that free()d memory arenas are really released

on POSIX systems supporting anonymous memory mappings.  Patch by Charles-François Natali.

Misc/NEWS

@@ -10,6 +10,10 @@ What's New in Python 3.3 Alpha 1?
 Core and Builtins
 -----------------
 
+- Issue #11849: Ensure that free()d memory arenas are really released
+  on POSIX systems supporting anonymous memory mappings.  Patch by
+  Charles-François Natali.
+
 - Issue #13436: Fix a bogus error message when an AST object was passed
   an invalid integer value.
 
@@ -260,10 +264,6 @@ Core and Builtins
   interpreter is shutting down; instead, these threads are now killed when
   they try to take the GIL.
 
-- Issue #11849: Make it more likely for the system allocator to release
-  free()d memory arenas on glibc-based systems.  Patch by Charles-François
-  Natali.
-
 - Issue #9756: When calling a method descriptor or a slot wrapper descriptor,
   the check of the object type doesn't read the __class__ attribute anymore.
   Fix a crash if a class override its __class__ attribute (e.g. a proxy of the

Objects/obmalloc.c

@@ -2,8 +2,11 @@
 
 #ifdef WITH_PYMALLOC
 
-#ifdef HAVE_MALLOPT_MMAP_THRESHOLD
+#ifdef HAVE_MMAP
-  #include <malloc.h>
+ #include <sys/mman.h>
+ #ifdef MAP_ANONYMOUS
+  #define ARENAS_USE_MMAP
+ #endif
 #endif
 
 #ifdef WITH_VALGRIND
@@ -183,15 +186,15 @@ static int running_on_valgrind = -1;
 /*
  * The allocator sub-allocates <Big> blocks of memory (called arenas) aligned
  * on a page boundary. This is a reserved virtual address space for the
- * current process (obtained through a malloc call). In no way this means
+ * current process (obtained through a malloc()/mmap() call). In no way this
- * that the memory arenas will be used entirely. A malloc(<Big>) is usually
+ * means that the memory arenas will be used entirely. A malloc(<Big>) is
- * an address range reservation for <Big> bytes, unless all pages within this
+ * usually an address range reservation for <Big> bytes, unless all pages within
- * space are referenced subsequently. So malloc'ing big blocks and not using
+ * this space are referenced subsequently. So malloc'ing big blocks and not
- * them does not mean "wasting memory". It's an addressable range wastage...
+ * using them does not mean "wasting memory". It's an addressable range
+ * wastage...
  *
- * Therefore, allocating arenas with malloc is not optimal, because there is
+ * Arenas are allocated with mmap() on systems supporting anonymous memory
- * some address space wastage, but this is the most portable way to request
+ * mappings to reduce heap fragmentation.
- * memory from the system across various platforms.
  */
 #define ARENA_SIZE              (256 << 10)     /* 256KB */
 
@@ -556,11 +559,6 @@ new_arena(void)
 #if SIZEOF_SIZE_T <= SIZEOF_INT
         if (numarenas > PY_SIZE_MAX / sizeof(*arenas))
             return NULL;                /* overflow */
-#endif
-#ifdef HAVE_MALLOPT_MMAP_THRESHOLD
-        /* Ensure arenas are allocated by mmap to avoid heap fragmentation. */
-        if (numarenas == INITIAL_ARENA_OBJECTS)
-            mallopt(M_MMAP_THRESHOLD, ARENA_SIZE);
 #endif
         nbytes = numarenas * sizeof(*arenas);
         arenaobj = (struct arena_object *)realloc(arenas, nbytes);
@@ -594,7 +592,12 @@ new_arena(void)
     arenaobj = unused_arena_objects;
     unused_arena_objects = arenaobj->nextarena;
     assert(arenaobj->address == 0);
+#ifdef ARENAS_USE_MMAP
+    arenaobj->address = (uptr)mmap(NULL, ARENA_SIZE, PROT_READ|PROT_WRITE,
+                                   MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
+#else
     arenaobj->address = (uptr)malloc(ARENA_SIZE);
+#endif
     if (arenaobj->address == 0) {
         /* The allocation failed: return NULL after putting the
          * arenaobj back.
@@ -1071,7 +1074,11 @@ PyObject_Free(void *p)
                 unused_arena_objects = ao;
 
                 /* Free the entire arena. */
+#ifdef ARENAS_USE_MMAP
+                munmap((void *)ao->address, ARENA_SIZE);
+#else
                 free((void *)ao->address);
+#endif
                 ao->address = 0;                        /* mark unassociated */
                 --narenas_currently_allocated;
 

configure

@@ -9401,8 +9401,8 @@ for ac_func in alarm accept4 setitimer getitimer bind_textdomain_codeset chown \
  getgrouplist getgroups getlogin getloadavg getpeername getpgid getpid \
  getpriority getresuid getresgid getpwent getspnam getspent getsid getwd \
  if_nameindex \
- initgroups kill killpg lchmod lchown lockf linkat lstat lutimes memrchr \
+ initgroups kill killpg lchmod lchown lockf linkat lstat lutimes mmap \
- mbrtowc mkdirat mkfifo \
+ memrchr mbrtowc mkdirat mkfifo \
  mkfifoat mknod mknodat mktime mremap nice openat pathconf pause pipe2 plock poll \
  posix_fallocate posix_fadvise pread \
  pthread_init pthread_kill putenv pwrite readlink readlinkat readv realpath renameat \
@@ -9784,34 +9784,6 @@ $as_echo "no" >&6; }
 fi
 rm -f core conftest.err conftest.$ac_objext conftest.$ac_ext
 
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking whether mallopt can set malloc mmap threshold" >&5
-$as_echo_n "checking whether mallopt can set malloc mmap threshold... " >&6; }
-cat confdefs.h - <<_ACEOF >conftest.$ac_ext
-/* end confdefs.h.  */
-
-#include <malloc.h>
-
-int
-main ()
-{
-mallopt(M_MMAP_THRESHOLD, 256 * 1024)
-  ;
-  return 0;
-}
-_ACEOF
-if ac_fn_c_try_compile "$LINENO"; then :
-
-$as_echo "#define HAVE_MALLOPT_MMAP_THRESHOLD 1" >>confdefs.h
-
-   { $as_echo "$as_me:${as_lineno-$LINENO}: result: yes" >&5
-$as_echo "yes" >&6; }
-else
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
-
-fi
-rm -f core conftest.err conftest.$ac_objext conftest.$ac_ext
-
 { $as_echo "$as_me:${as_lineno-$LINENO}: checking for broken unsetenv" >&5
 $as_echo_n "checking for broken unsetenv... " >&6; }
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
@@ -14591,8 +14563,8 @@ esac
 
 cat >>$CONFIG_STATUS <<_ACEOF || ac_write_fail=1
 # Files that config.status was made for.
-config_files="$ac_config_files"
+config_files="`echo $ac_config_files`"
-config_headers="$ac_config_headers"
+config_headers="`echo $ac_config_headers`"
 
 _ACEOF
 

configure.in

@@ -2567,8 +2567,8 @@ AC_CHECK_FUNCS(alarm accept4 setitimer getitimer bind_textdomain_codeset chown \
  getgrouplist getgroups getlogin getloadavg getpeername getpgid getpid \
  getpriority getresuid getresgid getpwent getspnam getspent getsid getwd \
  if_nameindex \
- initgroups kill killpg lchmod lchown lockf linkat lstat lutimes memrchr \
+ initgroups kill killpg lchmod lchown lockf linkat lstat lutimes mmap \
- mbrtowc mkdirat mkfifo \
+ memrchr mbrtowc mkdirat mkfifo \
  mkfifoat mknod mknodat mktime mremap nice openat pathconf pause pipe2 plock poll \
  posix_fallocate posix_fadvise pread \
  pthread_init pthread_kill putenv pwrite readlink readlinkat readv realpath renameat \
@@ -2679,15 +2679,6 @@ AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
   [AC_MSG_RESULT(no)
 ])
 
-AC_MSG_CHECKING(whether mallopt can set malloc mmap threshold)
-AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
-#include <malloc.h>
-]], [[mallopt(M_MMAP_THRESHOLD, 256 * 1024)]])],
-  [AC_DEFINE(HAVE_MALLOPT_MMAP_THRESHOLD, 1, Define if mallopt can set malloc mmap threshold.)
-   AC_MSG_RESULT(yes)],
-  [AC_MSG_RESULT(no)
-])
-
 AC_MSG_CHECKING(for broken unsetenv)
 AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
 #include <stdlib.h>

pyconfig.h.in

@@ -512,9 +512,6 @@
 /* Define this if you have the makedev macro. */
 #undef HAVE_MAKEDEV
 
-/* Define if mallopt can set malloc mmap threshold. */
-#undef HAVE_MALLOPT_MMAP_THRESHOLD
-
 /* Define to 1 if you have the `mbrtowc' function. */
 #undef HAVE_MBRTOWC
 
@@ -545,6 +542,9 @@
 /* Define to 1 if you have the `mktime' function. */
 #undef HAVE_MKTIME
 
+/* Define to 1 if you have the `mmap' function. */
+#undef HAVE_MMAP
+
 /* Define to 1 if you have the `mremap' function. */
 #undef HAVE_MREMAP