Re: FW: [Myricom help #22513] MPICH-GM 1.2.5..10 on Intell C++ 7.1 for Itanium2 realloc problem]

> -----Original Message-----
> From: Alexey V. Medvedev [mailto:[email protected]] Sent: 
> Friday, March 05, 2004 12:30 AM
> To: [email protected]
> Cc: [email protected]; [email protected]
> Subject: Re: [Myricom help #22513] MPICH-GM 1.2.5..10 on Intell C++ 
> 7.1 for Itanium2 realloc problem]
>
> [...]
>
> I think I have found bug in ptmalloc2 code used in MPICH-GM
> which causes described failures.  
>

Dear Alexey,

We would like to thank you for tracking this bug, and for the excellent 
work you did on it.

We are sorry that we were not able to help you, but we were originally 
not able to reproduce the problem,
even taking time to take your malloc traces and generating artifical 
programs doing the same
allocation patterns. 

As you point out, our modification of the MMAP related constants in the 
ptmalloc code create a bug,
 the symptoms will be:
 - realloc() might fail (i.e. return null prematurely) for multithreaded 
programs when extending
 a block of less than 1Mbyte to something between 1Mb and 4Mb.
 - Multithreaded programs with the possibility of concurrent malloc or 
reallocs in several threads are affected,
  mono-threaded programs are immune to this bug due to the way ptmalloc 
works.

The Intel C++ runtime had an influence on malloc/realloc usage() (and 
that sent us on the wrong path).

> I think you must fix the problem in the nearest mpich-gm release ASAP.
>  
>

We have a mpich-1.2.5..12pre5.tar.gz tar-ball on the ftp/web page, which 
contains a fix for that among other
changes, and which will become mpich-1.2.5..12 if no problem is found 
during final testing.

It contains the patch attached to this mail  as a fix to the realloc() 
problem. Note that we made a little test case to reproduce the bug which 
is in the second attachment.

Regards,

Loic

_______________________________________________
Myrinet mailing list
[email protected]
http://email.osc.edu/mailman/listinfo/myrinet

diff -u ../mpich-1.2.5..12pre5/mpid/ch_gm/ptmalloc2/arena.c mpid/ch_gm/ptmalloc2/arena.c
--- ../mpich-1.2.5..12pre5/mpid/ch_gm/ptmalloc2/arena.c	Wed Mar 12 20:13:46 2003
+++ mpid/ch_gm/ptmalloc2/arena.c	Fri Mar  5 12:22:02 2004
@@ -24,7 +24,7 @@
 
 #define HEAP_MIN_SIZE (32*1024)
 #ifndef HEAP_MAX_SIZE
-#define HEAP_MAX_SIZE (1024*1024) /* must be a power of two */
+#define HEAP_MAX_SIZE (4*1024*1024) /* must be a power of two */
 #endif
 
 /* HEAP_MIN_SIZE and HEAP_MAX_SIZE limit the size of mmap()ed heaps
diff -u ../mpich-1.2.5..12pre5/mpid/ch_gm/ptmalloc2/malloc.c mpid/ch_gm/ptmalloc2/malloc.c
--- ../mpich-1.2.5..12pre5/mpid/ch_gm/ptmalloc2/malloc.c	Wed Apr 16 23:57:31 2003
+++ mpid/ch_gm/ptmalloc2/malloc.c	Fri Mar  5 12:22:31 2004
@@ -15,7 +15,7 @@
 
 #define MORECORE gmpi_sbrk
 #define munmap(a,b) gmpi_munmap(a,b)
-#define DEFAULT_MMAP_THRESHOLD 4*1024*1024
+#define DEFAULT_MMAP_THRESHOLD (2*1024*1024)
 #define HAVE_MREMAP 0

#include <stdio.h>
#include <stdlib.h>
#include <malloc.h>
#include <pthread.h>
#include <assert.h>


pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;

#define DO_LOCK 0

void *bug(void *a)
{
  void *ptr;
  while (1) {
    if (DO_LOCK) pthread_mutex_lock(&mutex);
    ptr = malloc(300);
    ptr = realloc(ptr,1500000);
    assert(ptr);
    free(ptr);
    if (DO_LOCK) pthread_mutex_unlock(&mutex);
  }
  return 0;
}

int main()
{
  pthread_t ph;
  pthread_create(&ph,0, bug,0);
  bug(0);
  return 0;
}