[PATCH] Change the PCI bus driver to free resources leaked by drivers

[Warner Losh]

On Jun 24, 2013, at 2:59 PM, John Baldwin wrote:

> Currently our driver model trusts drivers to DTRT and properly release any 
> resources they allocated during probe() and attach().  I've added a new
> resource_list() helper method to release active resources on a resource
> list and used this to write a pci_child_detached() which cleans up any
> active resources when a device fails to probe or a driver finishes
> detach.  It also fixes an issue where we did not power down devices when
> the driver was detached (e.g. via kldunload).  I've tested the resource
> bits by writing a dummy driver that intentionally attached to an unattached
> device and leaked a memory BAR and verified that the bus warned about the
> leak and cleaned it up.
> 
> http://www.FreeBSD.org/~jhb/patches/pci_clean_detach.patch

I think most of pci_child_detached() could be a generic thing (except for the weird interaction with the msi-wart).  This is likely fixable.

We don't tear down any interrupt handlers that the device established. This is fixable, but the PCI bus would need to start tracking interrupts that are established...

We don't tear down any timers the device may have setup. This likely isn't fixable.

We likely don't want to tear down interrupts in bus_deactivate_resource, since I think it is theoretically legal to deactivate an interrupt resource without tearing down the interrupt. But maybe it shouldn't be...

Warner