Hello Stan, First of all, thank you for the report. This value was changed due to continuous process of IPoIB performance improvement (Csum and LSO patches). Of course, we checked it on both x86 and x64 platforms for connectx, but we had no chance to check all possible platforms x all possible hca's.
MAX_SEND_SGE will be set to appropriate value for mthca in the next commit. XaleX -----Original Message----- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Smith, Stan Sent: Saturday, August 30, 2008 1:16 AM To: [email protected] Subject: [ofw] trunk IPoIB for x86 & ia64 is broken - fails QP create. Along about svn.1499 ipoib_port.h was committed with a modified MAX_SEND_SGE. MAX_SEND_SGE was boosted from 8 to 30 which worked fine for x64 but broke ipoib for x86 and ia64 using mthca hardware. An SGE value of 30 causes create_qp() for x86 & ia64 to fail with IB_INVALID_MAX_SGE. Next time, make sure your test coverage is complete!!! I can see some slack for ia64 but there is NO EXCUSE for not testing x86!! In the WOF2-0 branch, in order to make a RC0 release, I set MAX_SEND_SGE to 16 which works fine for all supported arches; the trunk remains broken. Will the person who created the ipoib_port.h MAX_SEND_SGE change, investigate a more appropriate MAX_SEND_SGE value, test on x86 and fix the trunk! Grumpy Stan. PS: Curious how the trunk can remain broken for 2 of 3 supported architectures for numerous days. PSS: USA on holiday until Tuesday. _______________________________________________ ofw mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw _______________________________________________ ofw mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw
