Re: Synchronizing scsi_remove_host and the error handler

[James Bottomley <James.Bottomley@SteelEye.com>]

On Mon, 2005-08-08 at 16:41 -0400, Alan Stern wrote:
> > > Would you agree to a patch adding such a kref?
> > 
> > Well, not really.  The host template usually exists as a variable in the
> > module, so its lifetime is tied to the lifetime of the module.  Adding a
> > kref wouldn't help because it will still be freed when the module is
> > removed.  If there's a case where the template is being freed
> > prematurely because the module is being removed, then we have the module
> > refcounting wrong somewhere.  Have you run across such a case?
> 
> I tried to provoke such a case, but failed.  Evidently I've been barking
> up the wrong tree.  Normally the SCSI core doesn't call the LLD unless
> some process has opened a device file for something on that host.  This
> will automatically do try_module_get on the LLD, making it impossible for
> the LLD to be removed from memory.
> 
> I'm not certain this reasoning is 100% reliable, though -- does it cover
> _every_ case where the core calls the LLD?  Maybe something like this 
> little patch would be a good idea:

Yes, I was considering something similar, since all the last put of a
driver does is send the completion that driver_unregister() should be
waiting for

> +	struct device_driver	*parent_driver;

I don't think we need this.  The underlying device has to be the parent
of shost_gendev, so you can get the parent driver as

shost_gendev.parent->driver

> +	if (shost->parent_driver)
> +		put_driver(shost->parent_driver);

And just before this would be the place to do the final release of all
the resources the HBA is holding.

James