[ofa-general] [PATCH] SRPT for current SCST

[Vu Pham]

You can get hardware max rdma atomic by doing ib_query_qp

CM save the initiator's req->init_depth to its 
cm_id_priv->responder_resources in cm_req_handler (Win's srp 
  initiator set it at 16)

srpt call ib_cm_init_qp_attr() which will initialize 
qp_attr->max_dest_rd_atomic = 
cm_id_priv->responder_resources = 16;

then srpt call ib_modify_qp() without checking/reseting the 
hw max rdma atomic in qp_attr; therefore, ib_modify_qp() failed

You can try a quick dirty fix by applying this patch to 
ib_srpt.c

-vu




> So the client is requesting a QP with a max rdma_atomic of 15, but the
> target initializes the QP with a max of 4. Is this max rdma_atomic of 4
> based on hardware capabilities (Memory vs MemFree adapters), or a target
> QP initialization preset value? I'll swap adapters if needed.
> 
> -----Original Message-----
> From: chas williams - CONTRACTOR [mailto:chas at cmf.nrl.navy.mil] 
> Sent: Wednesday, August 22, 2007 4:28 AM
> To: Sufficool, Stanley
> Cc: Vu Pham; Stanley Sufficool; general
> Subject: Re: [ofa-general] [PATCH] SRPT for current SCST 
> 
> In message
> <C2F174F99918D54CA2A96E57C5079B6F35504C at sbc-exmsg2.sbcounty.gov>,"Su
> fficool, Stanley" writes:
>> I'm still researching where the rdma_atomic is getting the invalid 
>> attribute. I will give an update if I find anything else.
> 
> srp clients set this value as the responder_resources during login i
> believe.  this seems to be done in srp_get_responder_resources() on the
> windows srp client.  its taken from the qp attributes instead of picking
> a fixed number, like 4 in the linux version of the srp client.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: qp_max_atomic_res.patch
Type: text/x-diff
Size: 917 bytes
Desc: not available
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20070822/34ef90b8/attachment.patch>