> libmlx4 has this comments: > > /* FIXME flush wc buffers */ > > and since it does *not* currently actually flush the buffers, if we > enable WC for blueflame, WRs gets mixed in the WC buffer, and QP gets > corrupted/stuck. > > It seems we should we have arch.h under mthca and stick > some macro like wc_wmb() in there. > > Or, would infiniband/arch.h under libibverbs be a better place?
I think we should add it to infiniband/arch.h but then also have an #ifndef wc_wmb in libmlx4 until libibverbs with the define is ubiquitous. > If WC is not enabled, userspace can avoid the flush - so, should we > return such a bit as part of kernel abi? Maybe, although I'm not sure it's worth it. - R. _______________________________________________ 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
