> First I should add the case that triggered his patch: the userspace code > calculated a smaller buffer size than the kernel code, which caused > get_user_pages() to fail since part of the buffer did not belong to the > process's address space.
OK, in this case it seems the bug is in the kernel -- since it is overestimating the size of the WQEs needed. So we might as well fix it in the kernel. > As Mihcael said in a subsequent post, we still need this code both in > user and in kernel. Yes, but I think this issue really convinces me that we should decouple the two calculations, so the kernel code is only used for kernel QPs. And then change the mlx4 ABI so that userspace tells the kernel the wqe buffer size and rq/sq wqe shift/offset. That will allow for different SQ BB sizes and also make things more robust against bugs like this. - 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
