Zafar Gilani wrote:
Thanks for the reply. I read your name under the Author for may be all
the IBV structs/operations at linux.die.net <http://linux.die.net>. So
I am highly impressed by the work (can only dream of it myself). :)
thanks ...
1. I don't know why the original author (Roland Dreir) has used
ntohl() for rkey and remote_addr. Though to use it on buffer (buf) is
essential in order to transfer the byte order from network to host.
I need to check the test code to check the reason....
2. I am using QP type RC for reliable connection.
This means that if there is an error you will get a bad completion
(it you would have use UC QP, in case of an error in the receiver side,
the packet would have dropped).
3. Yes I am checking for that but the code gets stuck before that,
around when I call ibv_get_cq_event to wait for next completion event
in the event channel. I think the second write (second iteration of
for loop) is not working properly since when I try to send buf[0] via
RDMA_WRITE and buf[1] via SEND then it works fine.
Using ibv_get_cq_event can be tricky, you must arm the CQ (call
ibv_req_notify_cq) BEFORE a completion can enter to that CQ.
To make things more clear, why won't you just poll the CQ for completion
directly? (without using the CQ events)
I believe that you will get a completion with error...
Dotan
The code I am walking about:
while(1) {
if(ibv_get_cq_event(comp_chan, &evt_cq, &cq_context)) // here it
gets stuck
return 1;
// does not print this
printf("after get_cq_event\n"); fflush(stdout);
if(ibv_req_notify_cq(cq, 0))
return 1;
printf("after req_notify_cq\n"); fflush(stdout);
if(ibv_poll_cq(cq, 1, &wc) != 1)
return 1;
printf("after poll_cq\n"); fflush(stdout);
if(wc.status != IBV_WC_SUCCESS)
return 1;
printf("after wc.status\n"); fflush(stdout);
if(wc.wr_id == 0) {
printf("%d\n", ntohl(buf[0])); fflush(stdout);
return 0;
}
}
Your thoughts on this?
Thanks,
zafar
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
