On Thu, 2026-07-23 at 15:30 +0000, [email protected] wrote:
> Thank you for your contribution! Sashiko AI review found 2 potential
> issue(s) to consider:
> 
> New issues:
> - [Low] The patch calls the unexported internal midlayer function
> `scsi_scan_host_selected()` and includes the private header
> `"scsi_priv.h"` from a hardware driver, which will break module
> builds.
> 
> Pre-existing issues:
> - [High] Timeouts on targets scanned without an `fc_rport` parent
> will cause a NULL pointer dereference in the FC transport error
> handler.
> --
> 
> --- Patch [1]: [PATCH] scsi: storvsc: Support manual scans for all
> Hyper-V targets ---
> commit ce267d7e85b7a51757a8afe1188939d3c5146d27
> Author: Laurence Oberman <[email protected]>
> 
> scsi: storvsc: Support manual scans for all Hyper-V targets
> 
> This patch provides a storvsc-specific user_scan callback that
> directly
> invokes scsi_scan_host_selected(). This bypasses the single-rport
> lookup
> in fc_user_scan() to allow userspace to explicitly scan any target
> supported by storvsc. The goal is to avoid requiring storvsc to
> create
> and manage synthetic fc_rports for every Hyper-V target.
> 
> > diff --git a/drivers/scsi/storvsc_drv.c
> > b/drivers/scsi/storvsc_drv.c
> > index f03dc03a42c35..eb40473a8aad6 100644
> > --- a/drivers/scsi/storvsc_drv.c
> > +++ b/drivers/scsi/storvsc_drv.c
> > @@ -43,6 +43,7 @@
> >  #include <scsi/scsi_dbg.h>
> >  #include <scsi/scsi_transport_fc.h>
> >  #include <scsi/scsi_transport.h>
> > +#include "scsi_priv.h"
> 
> [Severity: Low]
> Will this cause an architectural layering violation? Including a
> private
> midlayer header like scsi_priv.h from a low-level hardware driver
> normally
> breaks subsystem encapsulation.
> 
> [ ... ]
> 
> > @@ -550,6 +551,17 @@ static void storvsc_host_scan(struct
> > work_struct *work)
> >     scsi_scan_host(host);
> >  }
> >  
> > +#if IS_ENABLED(CONFIG_SCSI_FC_ATTRS)
> > +static int storvsc_user_scan(struct Scsi_Host *host,
> > +                        unsigned int channel,
> > +                        unsigned int id,
> > +                        u64 lun)
> > +{
> > +   return scsi_scan_host_selected(host, channel, id, lun,
> > +                                  SCSI_SCAN_MANUAL);
> > +}
> 
> [Severity: Low]
> Does this cause a module build failure during the modpost link stage?
> 
> The function scsi_scan_host_selected() in drivers/scsi/scsi_scan.c
> lacks an
> EXPORT_SYMBOL directive. If the kernel is configured to build storvsc
> as a
> module (CONFIG_HYPERV_STORAGE=m), it seems it will fail to link.
> 
> [Severity: High]
> This is a pre-existing issue, but does creating a topology without an
> fc_rport parent lead to a NULL pointer dereference during error
> handling?
> 
> If a SCSI command times out on a target scanned directly under the
> SCSI host
> (which this manual scan facilitates), storvsc_eh_timed_out()
> delegates to
> the FC transport's fc_eh_timed_out(). Because the target was scanned
> under
> the host device instead of an fc_rport, starget_to_rport() evaluates
> to NULL.
> 
> This NULL pointer is then unconditionally dereferenced in
> drivers/scsi/scsi_transport_fc.c:fc_eh_timed_out():
> 
> fc_eh_timed_out()
> {
>     ...
>     struct fc_rport *rport = starget_to_rport(scsi_target(scmd-
> >device));
> 
>     if (rport->port_state == FC_PORTSTATE_BLOCKED)
>     ...
> }
> 
> Can timeouts under hardware or network load on these targets trigger
> a
> kernel panic?
> 
> > +#endif
> > +
> 
> [ ... ]
Thanks for the review.

The module-build observation is correct. scsi_scan_host_selected() is
an internal SCSI midlayer helper and is not exported for use by a
modular storvsc driver. The successful built-in kernel build did not
expose that problem so I never caught it.

The callback should instead use the exported scsi_scan_target()
interface and should not include scsi_priv.h.

I am also reviewing the FC error-handler concern before posting a
revised patch. Devices discovered directly below the SCSI host do not
have an fc_rport parent, while the storvsc timeout path delegates to
fc_eh_timed_out(). A revised patch must ensure that this topology
cannot lead to a NULL rport dereference.

I will address both points in v2.

Thanks,
Laurence


Reply via email to