[openib-general] 'Couldn't post send' error?

[Troy Benjegerdes]

On Fri, Aug 12, 2005 at 09:42:58PM -0500, Troy Benjegerdes wrote:
> What's this mean?
> 
> da4:~/NetPIPE_3.6.2# ibv_rc_pingpong 10.1.5.218
>   local address:  LID 0x0002, QPN 0x0d0404, PSN 0x599dea
>   remote address: LID 0x0001, QPN 0x090404, PSN 0x93b0c8
> Couldn't post send
> 
> I'm trying to get NetPIPE to work again, and it gets a similiar error.
> 
> It looks like 'uc' and 'ud' versions work just fine.

Interesting.. it looks like it's size dependent..

da5:~# ibv_rc_pingpong -s 65536
  local address:  LID 0x0001, QPN 0x250404, PSN 0xabcbf7
  remote address: LID 0x0002, QPN 0x2a0404, PSN 0xa2f603
131072000 bytes in 0.15 seconds = 7225.03 Mbit/sec
1000 iters in 0.15 seconds = 145.13 usec/iter
da5:~# ibv_rc_pingpong -s 512
  local address:  LID 0x0001, QPN 0x260404, PSN 0xa7be00
  remote address: LID 0x0002, QPN 0x2b0404, PSN 0x534952
Couldn't post send


I'm running the kernel modules from 2.6.13-rc6, and userspace libs
checked out from subversion sometime last week.

Do I need to have matching kernel and userspace versions? I can't find
anything in the kernel log messages. This is also with a new single-port
memfree card in a dual opteron iwill DK8NES motherboard, with the cards
connected back-to-back with no switch.