[openib-general] OpenSM realloc error

[Hal Rosenstock]

Hi again Owen,

On Thu, 2006-02-16 at 20:01, Owen Stampflee wrote:
> http://cvs.terraplex.com/~owen/osm.log
> 
> I'm currently using the Fedora FC5 packages that have been rebuilt,

I'm not sure what svn the FC5 package corresponds to. Can you check the
following in osm/libvendor/osm_vendor_ibumad.c:

Index: libvendor/osm_vendor_ibumad.c
===================================================================
--- libvendor/osm_vendor_ibumad.c       (revision 5029)
+++ libvendor/osm_vendor_ibumad.c       (revision 5030)
@@ -137,7 +137,7 @@
 get_madw(osm_vendor_t *p_vend, ib_net64_t *tid)
 {
        umad_match_t *m, *e;
-       ib_net64_t mtid = (*tid & 0xffffffff00000000llu);
+       ib_net64_t mtid = (*tid & cl_ntoh64(0x00000000ffffffffllu));
 
        cl_spinlock_acquire( &p_vend->match_tbl_lock );
        for (m = p_vend->mtbl.tbl, e = m + p_vend->mtbl.max; m < e; m++) {

This was fixed in svn 5030:

r5030 | halr | 2006-01-16 13:05:22 -0500 (Mon, 16 Jan 2006) | 8 lines

In osm_vendor_ibumad.c::get_madw, fix endian of mask for 
transaction ID used for comparison

The lack of this caused issued on running OpenSM on PPC
where get_madw would fail finding the matching transaction

Signed-off-by: Hal Rosenstock <halr at voltaire.com>

I am guessing that the FC5 OpenSM package is earlier than 5030. Can you
check get_madw (and make the one line change if it does not handle the
tid properly.

> I'm going to try the OpenIB 5411 source.

That will be another way to tell this.

-- Hal