Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Tue, Mar 19, 2013 at 05:11:05PM +0100, Stephan Budach wrote: > Am 18.03.13 20:05, schrieb Stephan Budach: > >Am 18.03.13 10:07, schrieb Lars Ellenberg: > >>On Sun, Mar 17, 2013 at 08:19:26AM +0100, Stephan Budach wrote: > >>>Hi all, > >>> > >>>I am running DRBD 8.4.3 in a dual-pri setup. DRBD interlink is > >>>through a 10 GbE Intel Cx520 link, which is piggy-backed between the > >>>two Supermicro boxes. I have configured two SCST iscsi targets from > >>>two DRBD volumes, which configured as multipathed targets on three > >>Don't. > >> > >>Simply: Don't. > >> > >> > >>Multipath expects to talk via more than one path to *the same target*. > >> > >>What your setup does it have multipath talk (via one path each) to > >>*two different targets*, that don't know of each other, > >>and happen to be able to present the same data most of the time. > >> > >>As iSCSI is not just about READs and WRITEs, but, especially in a > >>clustered setup, about reservations and other management commands, > >>this CAN NOT WORK, and that is NOT drbd's fault. > >> > >>So just don't. > >> > >>Thank you. > >> > >> > >> Lars > >> > >Hi Lars, > > > > > >okay, point taken, but what about the issue with drbd > >disconnecting when both primary get under high load? Even without > >iSCSI access this shouldn't happen, should it? > > > > > >Thanks, > >Stephan > > > Btw, LIO does support SCSI3 persistent reservations, so one could > swap scst with LIO and build multipathed iSCSI targets with it, > unless I am mistaken and have misread anything. You are absolutely mistaken. First, SCST does support SCSI-3 persistent reservations as well. But the point is not about wether or not the target supports some SCSI command subset. But wether or not it is cluster aware. The initiators are assuming to talk to *ONE* and the same target, whereas in this setup, it would talk to *TWO* independent targets (which happen to be able to present the same data most of the time). So, staying in the "reservation" example, one initiator would succeed in checking out reservations on the pink target, while an other initiator succeeds in checking out the same reservation on the purple target. As the targets are not cluster aware, all management commands (even "just" LUN resets) can (and thus will!) cause serious harm. -- : Lars Ellenberg : LINBIT | Your Way to High Availability : DRBD/HA support and consulting http://www.linbit.com