On Sat, Feb 23, 2008 at 05:27:15PM +0530, Arun Balakrishnan (WT01 - Computing, 
Storage & Software Products) wrote:
> 
>    Hi,
>    We are currently working on a project wherein we are porting a library
>    from GNU/Linux to FreeBSD 6.0 - RELEASE 32-bit and 64-bit. As part of
>    the standard memory leak tests, we noticed that the ported library is
>    leaking memory. After lots of analysis we found something very
>    strange. Just repeatedly loading and unloading our library was itself
>    throwing up a leak. We are able to reproduce a similar leak using the
>    following steps:
>    1. SimpleLib.cpp - Simple dummy library
>    2. LibLoader.cpp - Utility to repeatedly load the library
>    3. Compile as mentioned
>    4. Run under Valgrind for multiple times (31 times in our example.
>    Hard coded for simpilicity)
>    =================SimpleLib.cpp===================
>    #include <stdio.h>
>    #include <stdlib.h>
>    class CLeaker
>    {
>    public:
>      CLeaker() { };
>      virtual ~CLeaker() { };
>    };
>    CLeaker obj;
>    ================LibLoader.cpp======================
>    #include "stdio.h"
>    #include "dlfcn.h"
>    #include <stdlib.h>
>    #include <unistd.h>
>    #include <sys/time.h>
>    int main()
>    {
>      int i = 0;
>      int loop = 31;
>      while (i<loop)
>        {
>          i++;
>          void *handle = dlopen(argv[1], RTLD_LAZY);
>          if ( !handle )
>            exit(1);
>          dlclose(handle);
>        }
>      return 0;
>    }
>    ======================================================================
>    ==
>    Compilation:
>    g++ -shared -Wl,-soname,SimpleLib.so -o SimpleLib.so SimpleLib.cpp -g
>    g++ -o LibLoader_FreeBSD LibLoader.cpp -g
>    ======================================================================
>    ===
>    Execution:
>    valgrind --trace-pthread=all --show-below-main=yes
>    --show-reachable=yes --leak-check=yes ./LibLoader_FreeBSD
>    ./SimpleLib.so
>    ======================================================================
>    ===
>    Output: (snipped off irrelevant portions)
>    ==1155== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from
>    0)
>    ==1155== malloc/free: in use at exit: 520 bytes in 1 blocks.
>    ==1155== malloc/free: 1 allocs, 0 frees, 520 bytes allocated.
>    ==1155== For counts of detected errors, rerun with: -v
>    ==1155== searching for pointers to 1 not-freed blocks.
>    ==1155== checked 2140912 bytes.
>    ==1155==
>    ==1155== 520 bytes in 1 blocks are still reachable in loss record 1 of
>    1
>    ==1155==    at 0x3C032183: malloc (in
>    /usr/local/lib/valgrind/vgpreload_memcheck.so)
>    ==1155==    by 0x3C1CB018: (within /lib/libc.so.6)
>    ==1155==    by 0x3C1CB206: __cxa_atexit (in /lib/libc.so.6)
>    ==1155==    by 0x3C1F0898: ???
>    ==1155==
>    ==1155== LEAK SUMMARY:
>    ==1155==    definitely lost: 0 bytes in 0 blocks.
>    ==1155==    possibly lost:   0 bytes in 0 blocks.
>    ==1155==    still reachable: 520 bytes in 1 blocks.
>    ==1155==         suppressed: 0 bytes in 0 blocks.
>    ======================================================================
>    ===
>    Queries:
>    1. As seen in the Valgrind output, there is a 520bytes leak. This
>    happens only after around 31 loops and keeps increasing. By 100 loops,
>    the leak goes up to 1560 bytes. In our situation with our library, the
>    520bytes leak starts by the third iteration itself and by around 23
>    iterations it reaches 5KB. We are really stumped as to what could be
>    the possible reason for this leak? Where is the malloc called from?
>    Why only after executing 31 times? Executing the same code under
>    GNU/Linux does not show any leak even for over 1000 iterations.
>    2. While executing this without Valgrind, in another terminal we did a
>    "ps -Aopid,rss | grep LibLoader_" continuously in a loop and saw that
>    the RSS (resident set size) field value keeps increasing by 4KB every
>    now and then. The same experiment on GNU/Linux shows that RSS remains
>    at the same value. What could be the cause for the ever rising RSS
>    value?
>    Any help in this regard would be really helpful. Thanks in advance.
>    Rgds,
>    ~Arun

The valgrind report points to memory used by the atexit_register()
for keeping the information on the functions registered by means of
atexit(3) and __cxa_atexit(). See the lib/libc/stdlib/atexit.c. In your
(non-compilable) example, __cxa_atexit() is used by shared objects to
register the destructor for global objects to be called at the dso
unload.

The handling of the memory is complicated because atexit() specification
states that:
- functions shall be called in the reverse order of their registration;
- at least 32 functions can be registered with atexit().

The current implementation never frees the struct atexit to try to
conform to the requirement of order.

The static __atexit0, intended to guarantee success of the first 32
atexit() calls, may not guarantee it, because the space can be consumed
by the interleaved __cxa_atexit() instead.

Patch below may help with the libc leak.

diff --git a/lib/libc/stdlib/atexit.c b/lib/libc/stdlib/atexit.c
index 05dad84..8389637 100644
--- a/lib/libc/stdlib/atexit.c
+++ b/lib/libc/stdlib/atexit.c
@@ -41,6 +41,7 @@ __FBSDID("$FreeBSD: src/lib/libc/stdlib/atexit.c,v 1.8 
2007/01/09 00:28:09 imp E
 #include <stdlib.h>
 #include <unistd.h>
 #include <pthread.h>
+#include <sys/queue.h>
 #include "atexit.h"
 #include "un-namespace.h"
 
@@ -56,7 +57,7 @@ static pthread_mutex_t atexit_mutex = 
PTHREAD_MUTEX_INITIALIZER;
 #define _MUTEX_UNLOCK(x)       if (__isthreaded) _pthread_mutex_unlock(x)
 
 struct atexit {
-       struct atexit *next;                    /* next in list */
+       LIST_ENTRY(atexit) link;
        int ind;                                /* next index in this table */
        struct atexit_fn {
                int fn_type;                    /* ATEXIT_? from above */
@@ -69,7 +70,10 @@ struct atexit {
        } fns[ATEXIT_SIZE];                     /* the table itself */
 };
 
-static struct atexit *__atexit;                /* points to head of LIFO stack 
*/
+/* Head of LIFO stack */
+LIST_HEAD(, atexit) __atexit = LIST_HEAD_INITIALIZER(__atexit);
+static struct atexit __atexit0;                        /* one guaranteed table 
*/
+static unsigned long __atexit_gen;
 
 /*
  * Register the function described by 'fptr' to be called at application
@@ -79,30 +83,33 @@ static struct atexit *__atexit;             /* points to 
head of LIFO stack */
 static int
 atexit_register(struct atexit_fn *fptr)
 {
-       static struct atexit __atexit0; /* one guaranteed table */
        struct atexit *p;
+       unsigned long old__atexit_gen;
 
        _MUTEX_LOCK(&atexit_mutex);
-       if ((p = __atexit) == NULL)
-               __atexit = p = &__atexit0;
-       else while (p->ind >= ATEXIT_SIZE) {
-               struct atexit *old__atexit;
-               old__atexit = __atexit;
-               _MUTEX_UNLOCK(&atexit_mutex);
-               if ((p = (struct atexit *)malloc(sizeof(*p))) == NULL)
-                       return (-1);
-               _MUTEX_LOCK(&atexit_mutex);
-               if (old__atexit != __atexit) {
-                       /* Lost race, retry operation */
+       if (LIST_EMPTY(&__atexit)) {
+               p = &__atexit0;
+               LIST_INSERT_HEAD(&__atexit, p, link);
+       } else {
+       retry:
+               p = LIST_FIRST(&__atexit);
+               if (p->ind >= ATEXIT_SIZE) {
+                       old__atexit_gen = __atexit_gen;
                        _MUTEX_UNLOCK(&atexit_mutex);
-                       free(p);
+                       if ((p = (struct atexit *)malloc(sizeof(*p))) == NULL)
+                               return (-1);
                        _MUTEX_LOCK(&atexit_mutex);
-                       p = __atexit;
-                       continue;
+                       if (old__atexit_gen != __atexit_gen) {
+                               /* Lost race, retry operation */
+                               _MUTEX_UNLOCK(&atexit_mutex);
+                               free(p);
+                               _MUTEX_LOCK(&atexit_mutex);
+                               goto retry;
+                       }
+                       p->ind = 0;
+                       LIST_INSERT_HEAD(&__atexit, p, link);
+                       __atexit_gen++;
                }
-               p->ind = 0;
-               p->next = __atexit;
-               __atexit = p;
        }
        p->fns[p->ind++] = *fptr;
        _MUTEX_UNLOCK(&atexit_mutex);
@@ -119,7 +126,7 @@ atexit(void (*func)(void))
        int error;
 
        fn.fn_type = ATEXIT_FN_STD;
-       fn.fn_ptr.std_func = func;;
+       fn.fn_ptr.std_func = func;
        fn.fn_arg = NULL;
        fn.fn_dso = NULL;
 
@@ -138,7 +145,7 @@ __cxa_atexit(void (*func)(void *), void *arg, void *dso)
        int error;
 
        fn.fn_type = ATEXIT_FN_CXA;
-       fn.fn_ptr.cxa_func = func;;
+       fn.fn_ptr.cxa_func = func;
        fn.fn_arg = arg;
        fn.fn_dso = dso;
 
@@ -154,32 +161,55 @@ __cxa_atexit(void (*func)(void *), void *arg, void *dso)
 void
 __cxa_finalize(void *dso)
 {
-       struct atexit *p;
-       struct atexit_fn fn;
-       int n;
+       struct atexit *p, *p1, cp;
+       struct atexit_fn *fn;
+       int i, n, inuse;
+       unsigned long orig__atexit_gen;
 
        _MUTEX_LOCK(&atexit_mutex);
-       for (p = __atexit; p; p = p->next) {
+ restart:
+       inuse = 0;
+       LIST_FOREACH_SAFE(p, &__atexit, link, p1) {
+               cp.ind = 0;
                for (n = p->ind; --n >= 0;) {
                        if (p->fns[n].fn_type == ATEXIT_FN_EMPTY)
                                continue; /* already been called */
-                       if (dso != NULL && dso != p->fns[n].fn_dso)
+                       if (dso != NULL && dso != p->fns[n].fn_dso) {
+                               inuse = 1;
                                continue; /* wrong DSO */
-                       fn = p->fns[n];
+                       }
+                       cp.fns[cp.ind++] = p->fns[n];
                        /*
                          Mark entry to indicate that this particular handler
                          has already been called.
                        */
                        p->fns[n].fn_type = ATEXIT_FN_EMPTY;
-                       _MUTEX_UNLOCK(&atexit_mutex);
-               
+               }
+               if (!inuse && p != &__atexit0) {
+                       LIST_REMOVE(p, link);
+                       __atexit_gen++;
+               } else {
+                       /*
+                        * The current entry cannot be removed, and so
+                        * any consequent entries.
+                        */
+                       inuse = 1;
+                       p = NULL;
+               }
+               orig__atexit_gen = __atexit_gen;
+               _MUTEX_UNLOCK(&atexit_mutex);
+               free(p);
+               for (i = 0; i < cp.ind; i++) {
+                       fn = &cp.fns[i];
                        /* Call the function of correct type. */
-                       if (fn.fn_type == ATEXIT_FN_CXA)
-                               fn.fn_ptr.cxa_func(fn.fn_arg);
-                       else if (fn.fn_type == ATEXIT_FN_STD)
-                               fn.fn_ptr.std_func();
-                       _MUTEX_LOCK(&atexit_mutex);
+                       if (fn->fn_type == ATEXIT_FN_CXA)
+                               fn->fn_ptr.cxa_func(fn->fn_arg);
+                       else if (fn->fn_type == ATEXIT_FN_STD)
+                               fn->fn_ptr.std_func();
                }
+               _MUTEX_LOCK(&atexit_mutex);
+               if (orig__atexit_gen != __atexit_gen)
+                       goto restart;
        }
        _MUTEX_UNLOCK(&atexit_mutex);
 }

Attachment: pgpPgt81rlHoK.pgp
Description: PGP signature

  • Memo... Arun Balakrishnan (WT01 - Computing, Storage & Software Products)
    • ... Kostik Belousov
      • ... Arun Balakrishnan
        • ... Kostik Belousov
          • ... Arun Balakrishnan
            • ... Kostik Belousov

Reply via email to