> > Error handling looks a bit bogus here - we'll never give the task > > it's rlimit back. Wouldn't it be a bit cleaner to allocate > > the work object together with umem? > > Yes, that's a good point. Mainline already has this issue, so I'll > make that a separate patch before this one.
Actually, it's a pain to fix in mainline, because the lifetime of struct ib_umem is not controlled by ib_umem_release (it's embedded in other structures instead). So I'll fix this as a patch on top of this one (and it becomes yet another virtue of this approach -- it makes this problem easily fixable). By the way, what's your overall opinion of this patch? Do you like this approach for mlx4 queues (and in general)? - 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
