Re: IDE locking problem

[Bartlomiej Zolnierkiewicz <B.Zolnierkiewicz@elka.pw.edu.pl>]

On 5 Aug 2003, Benjamin Herrenschmidt wrote:

> > It only works with default/legacy io bases.
>
> Which is all I care about in this specific situation. The "mediabay"
> hotswap controller is driven by ide-pmac, which provides it's own
> ide_init_hwif_ports() and ide_default_io_base() arch hooks. (In the
> PPC arch, those arch functions can be themselves hooked in by the
> actual platform code, on pmac, this goes to ide-pmac, though I also
> deal with added PCI cards).
>
> > > Also, look at my patch, we also _NEED_ to add some proper
> > > device_unregister calls to ide_unregister() or this function will
> > > leave dangling entries in the device list, and since those have the
> > > same restrictions as the new blk_cleanup_queue(), we really need to
> > > do that without the lock held.
> >
> > Yes.
>
> That reminds me that without this fix, even ide-cs will break things
> when ejecting a CF card for example.

I am aware of that.

> > > I'd suggest merging my patch for now, it won't make things much
> > > worse than what they are today regarding racyness of IDE registration
> > > and unregistration, we an look into sanitizing this as a 2.7 goal.
> >
> > Okay :\.
>
> Heh, thanks... Well... who else uses ide_unregister except ide-pmac and
> ide-cs anyway ? If it's really that little used, it's quite under control
> and we can try to clean it up a bit more then during early 2.6.0...

HDIO_UNREGISTER_HWIF ioctl.  Together with HDIO_SCAN_HWIF ioctl it
provides dirty hotswap functionality.  Thats why I mentioned about problem
with ide_unregister() and non default io bases.

> Right now, I'm waiting to see what we can come up to with Jens Axboe
> regarding race-free teardown of the blk queue. Once that's agreed, we
> can think about just removing the call to init_hwif_data and just reinit
> the drive array ourselves. That would be a good thing anyway as the
> current ide_unregister allocation of a full hwif_t on stack is causing
> other problems (stack bloat typically, an issue with people trying to
> shrink the kernel process stack to 1 page).

Decreasing stack usage is an additional motivation to simplify
ide_unregister().  We really need to remove ide_unregister() from Jörn's
top stack (l)users list ;-).

Thanks,
--
Bartlomiej