On 02/12/2012 02:15 PM, Michael S. Tsirkin wrote:
On Sun, Feb 12, 2012 at 02:04:29PM -0600, Anthony Liguori wrote:
On 02/12/2012 11:57 AM, Michael S. Tsirkin wrote:
On Sun, Feb 12, 2012 at 11:38:24AM -0600, Anthony Liguori wrote:
From: Anthony Liguori<aligu...@us.ibm.com>
Date: Sun, 12 Feb 2012 11:36:24 -0600
Subject: [PATCH] device_add: don't add a /peripheral link until init is complete
Otherwise we end up with a dangling reference which causes qdev_free() to fail.
Reported-by: Michael Tsirkin<m...@redhat.com>
Signed-off-by: Anthony Liguori<aligu...@us.ibm.com>
This handles the option parsing but what about hotplug
failures (when bus->hotplug returns an error)?
Sorry, I don't follow.
The assert you reported was that object_free() noted a reference
count of !0 which indicates something else was holding the reference
to the object. In this case, it was the child link in /peripheral.
By delaying creating the link in /peripheral, we eliminate the problem
completely.
Th other problem was internal in pci which calls ->hostplug
during initialization. This doesn't seem affected?
But I didn't try, maybe I misundertand.
Yeah, from qdev's perspective it's all just init failing. hotplug is entirely a
PCI concept.
BTW, the explicit calls to do_pci_unregister are redundant.
finalize() will be called during cleanup which means exit() will be
invoked (which already calls do_pci_unregister). I'm not sure why
this isn't failing more aggressively but it looks clearly wrong to
me.
Regards,
Anthony Liguori
Me too. Want to try to drop them?
Yeah, I'll make this a two patch series.
Regards,
Anthony Liguori
