Re: [Patch] Fix oops on rmmod usb-storage

[Mike Anderson <andmike@us.ibm.com>]

Alan Stern [stern@rowland.harvard.edu] wrote:
> Mike, thanks for adding me to this thread...
> 
> On Wed, 29 Sep 2004, Christoph Hellwig wrote:
> 
> > > In some cases the knowledge would be in the driver would it not. When the
> > > drivers remove function is called the driver knows the state of the
> > > transport or in the case of the usb storage device it knows it will
> > > not process anymore commands. 
> > > 
> > > Alan / Matthew / others can correct if I have mis-read something in
> > > usb_stor_release_resources but it knows that it will not process anymore
> > > commands prior to calling scsi_remove_host so if we had a method to flag
> > > this it would help in our shutdown if we want to support both devices
> > > that can and cannot process commands.
> 
> That's right; when usb-storage calls scsi_remove_host it will not process 
> any more commands.
> 
> > Well, we're certainly not going to change the scsi_remove_host signature,
> > but we could add a separate one.
> > 
> > But I don't like that change at all as we would still have that problem
> > with all driver that don't have a way to magically find out.  We should
> > really try to make scsi_remove_host foolprof.
> 
> Please take a look at my questions in
> 
> http://marc.theaimsgroup.com/?l=linux-scsi&m=109647697621208&w=2
> 
> Part of the problem here may be that ownership of SCSI commands isn't 
> clear during the host removal process.  As things stand, usb-storage has 
> no choice but to ignore commands queued after scsi_remove_host is called.  
> This certainly will cause problems if it means we have to wait around for 
> the commands to time out and be aborted.
> 

Is it possible usb storage could do something else beside just ignore
the commands like return them with DID_NO_CONNECT.

In storage_disconnect it looks like you set some state, wait for the
current command, but do not set anything that queuecommand can look at
to start rejecting commands.

-andmike
--
Michael Anderson
andmike@us.ibm.com