scsi_target with multiple luns
lydianconcepts at gmail.com
Fri Jun 1 04:47:08 UTC 2007
Remember that scsi_target is an example. A fine example, but an
example. Having a single user process that can be shot and killed, no
matter how multithreaded or AIO'd, is not necessarily the wisest
choice for building a target device.
On 5/31/07, Scott Long <scottl at samsco.org> wrote:
> Hidetoshi Shimokawa wrote:
> > I have not fully analyzed the problem but I'll describe it just for a note.
> > I'd like to ask maintainers of scsi_target for further analysis.
> > I experiance a problem with small number(1) of simq and multiple
> > scsi_target(8) instances.
> > As far as I understand, the following situation could occur under a fairly
> > heavy load.
> > 1. process A send a request -> cam send to sim
> > 2. process B send a request -> blocked because the simq is full
> > 3. the request of process A is finished (in the context of process A)
> > 4. cam/scsi_target tries to send the request of process B.
> > But the mapped memory is of process A, and scsi_target send wrong
> > ccb to sim.
> > Maybe, we should rewrite scsi_target in kernel space with GEOM support..
> I'm unclear on how GEOM would fix this. Also, scsi targets aren't
> always DA devices. I dedicated scsi_da_target device that is backed
> by GEOM might be interesting, though. Even more interesting would be
> a direct DMA method that required no KVA mappings for the data.
> freebsd-scsi at freebsd.org mailing list
> To unsubscribe, send any mail to "freebsd-scsi-unsubscribe at freebsd.org"
More information about the freebsd-scsi