[ewg] Issues with ping-pong tests

[Sharma, Karun]

The segmentation fault which is coming with srq_pingpong, is intermittent. I have seen it on one server while on other server it fails with a different error message. However, I am attaching the gdb trace. Let me know if you need any other info.
 
You may also want to look into bug # 525, I have opened for an issue with rc_pingpong.
 
Thanks
Karun
 
##########################################################
(gdb) r -m 500
Starting program: /usr/local/bin/ibv_srq_pingpong -m 500
[Thread debugging using libthread_db enabled]
[New Thread 47443289843408 (LWP 26790)]
Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 47443289843408 (LWP 26790)]
0x0000000000000000 in ?? ()
(gdb) bt
#0  0x0000000000000000 in ?? ()
#1  0x00002b26401b5062 in ibv_create_qp (pd=0x50cd30, 
    qp_init_attr=0x7fff6ab3fab0) at src/verbs.c:342
#2  0x0000000000401eb2 in main (argc=<value optimized out>, 
    argv=<value optimized out>) at examples/srq_pingpong.c:399
(gdb) fr 1
#1  0x00002b26401b5062 in ibv_create_qp (pd=0x50cd30, 
    qp_init_attr=0x7fff6ab3fab0) at src/verbs.c:342
342             struct ibv_qp *qp = pd->context->ops.create_qp(pd, qp_init_attr);
(gdb) fr 2
#2  0x0000000000401eb2 in main (argc=<value optimized out>, 
    argv=<value optimized out>) at examples/srq_pingpong.c:399
399                     ctx->qp[i] = ibv_create_qp(ctx->pd, &attr);
(gdb) 
##############################################################

________________________________

From: Roland Dreier [mailto:rdreier at cisco.com]
Sent: Fri 4/6/2007 10:44 PM
To: Sharma, Karun
Cc: ewg at lists.openfabrics.org
Subject: Re: [ewg] Issues with ping-pong tests



 > 1. ibv_ud_pingpong :  doesn't work if size > 2048. I am getting error saying " Failed status 1 for wr_id 2".
 > For sizes <= 2048, it is working fine.

Not surprising, since the max UD message size on your system is almost
certainly 2048.  Perhaps the diagnostic could be improved, but I don't
think this is a big deal.

 > 2. -m option of ibv_srq_pingpong doesn't work. The error which we
 > are getting is also not related to MTU. In the man page, there is no
 > mention about this option, but if we do "ibv_srq_pingpong -h", we can
 > see -m option.

There was a missing "break" in the switch statement in srq_pingpong.c
between the 'm' case and the 'q' case, so setting the MTU also
affected the number of QPs.  I checked in a fix.

 > ss23:~ # ibv_srq_pingpong -m 500
 > Segmentation fault

I don't see this failure mode here.  Can you run this under gdb and do
a bt to get a traceback?

 - R.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/ewg/attachments/20070409/70826677/attachment.html>