Milter-greylist

On 02/18/2010 11:40 AM, manu@... wrote:
>
>
> Dietmar Rieder <adrieder@...
> <mailto:adrieder%40sbox.tugraz.at>> wrote:
>
>  > libumem.so.1`malloc+0x23
>  > libresolv.so.2`__res_vinit+0x118
>  > libresolv.so.2`res_ninit+0x1a
>  > dnsrbl_check_source+0x14a
>
> A quick look at NetBSD's res_ninit() sources (it's from BIND, and I
> beleive most system have the same), indeed shows a malloc():
> statp->_u._ext.ext = malloc(sizeof(*statp->_u._ext.ext));
>
> This means that any successful call to res_ninit() must have a
> corresponding res_ndestroy() call.
>
> In dnsrbl_check_source, once res_ninit() is called, there is no way out
> of the function without executing res_ndestroy(). But the trick is that
> res_ninit() and res_ndestroy() may be macros in milter-greylist, so that
> we can cope with older BIND resolvers without having #ifdef's everywhere
> in our code.
>
> In some situation, milter-greylist can define res_ndestroy() as
> res_nclose(), and this is a problem, because at least in NetBSD's
> resolver, res_ndestroy() does free statp->_u._ext.ext whereas
> res_nclose() does not.
>
> Hence, here is the offending code in dnsrbl.c
>
> #if (defined(res_ninit) || (__RES >= 19991006) )
> #define HAVE_RESN 1
> #ifndef res_ndestroy
> #define res_ndestroy(res) res_nclose(res) /* <--- XXXhereXXX */
> #endif
> #else
> #define res_ninit(res) \
> ((_res.options & RES_INIT) == 0 && res_init())
> #define res_nquery(res, req, class, type, ans, anslen) \
> res_query(req, class, type, ans, anslen)
> #define res_ndestroy(res)
> #endif
>
> The ifdef on res_ninit and res_ndestroy is an error, it will never be
> evaluated as true. Please try this instead, and tell us if you get the
> warning at compile time.
>
> #if (__RES >= 19991006) )
> #define HAVE_RESN 1
> #else
> #warn "no res_ninit!"
> #define res_ninit(res) \
> ((_res.options & RES_INIT) == 0 && res_init())
> #define res_nquery(res, req, class, type, ans, anslen) \
> res_query(req, class, type, ans, anslen)
> #define res_ndestroy(res)
> #endif
>
> Handling the BIND versions where res_nclose() exists but not
> res_ndestroy() must be done by an autoconf test.


Hi thanks for the quick response and analysis. I changed the dnsrbl.c 
file as indicated and recompiled the milter. I didn't get a warning at 
compile time.

I genereated a core dump after some minute, and now it looks really 
promising!!!!

gcore 6891
mdb core.6891
 > ::findleaks -dvf
findleaks:                maximum buffers => 873580
findleaks:                 actual buffers => 872852
findleaks:
findleaks:             potential pointers => 8311361
findleaks:                     dismissals => 2091319       (25.1%)
findleaks:                         misses => 4770578       (57.3%)
findleaks:                           dups => 576612        ( 6.9%)
findleaks:                        follows => 872852        (10.5%)
findleaks:
findleaks:              elapsed wall time => 5 seconds
findleaks:
CACHE     LEAKED   BUFCTL CALLER
----------------------------------------------------------------------
    Total       0 buffers, 0 bytes

I'll observe it for a while and I'll let you know if there is any 
leaking left.

THANK you so much!

Didi