Ticket History #6308: Alignment problem in resolver test

Hello!
In compiling krb5, version 1.6.3, I ran into a core dump (bus error) during 'gmake check' on a Solaris 2.10/Sun Fire T200 build system, when it tried to run tests/resolve/resolve.

It seems there is an alignment problem with the automatic 'addrcopy' variable in src/tests/resolve/resolve.c. I forced it to align (and execute successfully) after the change below. I figured this was an easier test, than changing the code to use an in_addr struct -- which is perhaps a more "correct" fix.

Thanks
/Jorgen Wahlsten


--- resolve.c~  2003-07-22 14:02:34.000000000 -0400
+++ resolve.c   2008-12-23 10:09:31.475658000 -0500
@@ -77,7 +77,11 @@
 {
        char myname[MAXHOSTNAMELEN+1];
        char *ptr;
-       char addrcopy[4];
+       /* char addrcopy[4]; */
+       union {
+           char addrcopy[4];
+           int make_me_aligned;
+       } u;
        struct hostent *host;
        int quiet = 0;
 
@@ -123,10 +127,10 @@
            printf("Host address: %d.%d.%d.%d\n",
                   UC(ptr[0]), UC(ptr[1]), UC(ptr[2]), UC(ptr[3]));
 
-       memcpy(addrcopy, ptr, 4);
+       memcpy(u.addrcopy, ptr, 4);
 
        /* Convert back to full name */
-       if((host = gethostbyaddr(addrcopy, 4, AF_INET)) == NULL) {
+       if((host = gethostbyaddr(u.addrcopy, 4, AF_INET)) == NULL) {
                fprintf(stderr, "Error looking up IP address - fatal\n");
                exit(2);
        }

Hello Ezra.

Yes, I used a 32-bit compiled gcc 4.1.1.

I do not have Sun's C compiler and can not verify that it is working with it, but if it's important for you to verify that it is working with Sun's C compiler, I can try to get it. Let me know.

If I turn on optimization (-O2 or -O3) the program works, while using -O0/-O1 the program still core dumps.

Compiling resolve.c as a 32-bit binary or a 64-bit binary does not seem to matter. It core dumps no matter what.

I was thinking this had something to do with automatic variable alignment, and therefore changed:

char myname[MAXHOSTNAMELEN+1]

to

char myname[MAXHOSTNAMELEN+8]

(MAXHOSTNAMELEN is defined as 256 in netdb.h),

which makes the program work again (not core dump).

For what it's worth, I think Sun's C compiler *always* aligns the automatic variables on 8-byte boundaries, while GCC tries to fit the addrcopy[4] into the "padding" of myname[256+1] (7 bytes padding?), as an optimization. It does seem strange that this is done *without* optimization though.

Regardless, of the differences between Sun's C compiler and GCC, I would be inclined to agree with you that this is a solaris C library bug, that is being triggered by using GCC in this particular case.

As for my comment about using an in_addr pointer, I think I merely looked at the EXAMPLE section in the solaris 2.10 gethostbyaddr man page. There is also a NOTE in that man page about INET_ADDR.

Let me know if you want to any additional information.

Thanks,
Jorgen







--
_______________________________________________________________________
Jorgen Wahlsten -- http://www.wahlsten.com/
AIM: JorgenWahlsten
YIM: jorgenwahlsten
ICQ: 171198501