Re: [PATCH] scsi: core: Don't free dev_name() manually

[James Bottomley]

On Tue, 2026-01-20 at 21:11 +0800, Tzung-Bi Shih wrote:
> On Mon, Jan 19, 2026 at 11:02:55AM -0500, James Bottomley wrote:
> > On Mon, 2026-01-19 at 22:28 +0800, Tzung-Bi Shih wrote:
> > > On Sun, Jan 18, 2026 at 09:45:26AM -0500, James Bottomley wrote:
> > > > On Sun, 2026-01-18 at 03:32 +0800, Tzung-Bi Shih wrote:
> > [...]
> > > > > >  
> > > > > >  static struct class shost_class = {
> > > > > > @@ -279,11 +278,9 @@ int scsi_add_host_with_dma(struct
> > > > > > Scsi_Host
> > > > > > *shost, struct device *dev,
> > > > > >   goto out_disable_runtime_pm;
> > > > > >  
> > > > > >   scsi_host_set_state(shost, SHOST_RUNNING);
> > > > > > - get_device(shost->shost_gendev.parent);
> > > > 
> > > > We need a reference to the parent to prevent surprise removal
> > > > ...
> > > > where else is the reference held?
> > > 
> > > It looks to me the same question as above.  IIUC, device_add()
> > > holds
> > > a reference count to its parent[3].  Drivers don't need to do it
> > > explicitly.
> > 
> > That's not good enough for SCSI: we have a rather complicated state
> > model for hosts.  device_add() doesn't occur until the host moves
> > out of the SHOST_CREATED state, which can be quite a time after
> > device _initialize() so something has to pin the resources until
> > then, which is why these references are taken.   You're certainly
> > free to suggest a different way of doing this, but you can't just
> > get rid of the existing mechanism without replacing it with
> > something else.
> 
> I may misunderstand: isn't the initial reference count from
> device_initialize() held for the purpose (i.e., pin the resource)? 
> The driver calls scsi_host_put() to drop the reference count when the
> underlying chip is removing.
> 
> The proposed code to remove the get_device() just right before
> device_add() in scsi_add_host_with_dma().  I don't see what else
> resources it can pin.

And the reverse, when the host is going away and device_del gets called
but something has the sysfs node open?

Regards,

James