[ofa-general] Re: sg_reset can trigger a NULL pointer dereference in the SRP initiator

[James Bottomley]

On Thu, 2009-08-06 at 17:38 +0200, Bart Van Assche wrote:
> On Thu, Aug 6, 2009 at 10:30 AM, Boaz Harrosh <bharrosh at panasas.com> wrote:
> > [Just out of memory, I've not inspected the code for a long time]
> >
> > It looks like an srp_request was never allocated for the reset
> > command. (since it never went through .queuecommand)
> >
> > static int srp_find_req(struct srp_target_port *target,
> >                        struct scsi_cmnd *scmnd,
> >                        struct srp_request **req)
> > {
> >        if (scmnd->host_scribble == (void *) -1L)
> >                return -1;
> >
> >        *req = &target->req_ring[(long) scmnd->host_scribble];
> >
> >        return 0;
> > }
> >
> > Specifically scmnd->host_scribble can just be Zero.
> > When queues are active that does not matter and a device is found
> > since the reset does not really need the scsi_cmnd. But in above
> > scenario the queues were never used and the array entry is empty.
> 
> Hello Boaz,
> 
> Thanks for the info. Do you know by heart which SCSI drivers process
> the SG_SCSI_RESET ioctl correctly and that could be used as an example
> for fixing the SRP initiator ?

Basically all of them which are in regular use for clustering; so SAN:
qla2xxx; lpfc.  And for legacy SPI clusters: aic7xxx;mptspi

James