[Xen-devel] race with restart and device removal

[John Levon <levon@xxxxxxxxxxxxxxxxx>]

We've observed that sometimes during restart some of a domain's vbds
failed to be created. From some debugging it appears to be this code:

        # Marshall devices (running or from configuration)
        if not ignore_devices:
            for cls in XendDevices.valid_devices():
                found = False

                # figure if there is a device that is running
                if domain:
                    try:
                        controller = domain.getDeviceController(cls)
                        configs = controller.configurations()
                        for config in configs:
                            sxpr.append(['device', config])
                            found = True
                    except:
                        log.exception("dumping sxp from device controllers")
                        pass

                # if we didn't find that device, check the existing config
                # for a device in the same class
                if not found:
                    for dev_type, dev_info in self.all_devices_sxpr():
                        if dev_type == cls:
                            sxpr.append(['device', dev_info])

'configs' comes from the xen store:

    def deviceIDs(self, transaction = None):
        """@return The IDs of each of the devices currently configured for
        this instance's deviceClass.
        """
        fe = self.backendRoot()
        if transaction:
            return map(lambda x: int(x.split('/')[-1]), transaction.list(fe))
        else:
            return map(int, xstransact.List(fe))


But this can race against the disk backend removing the xenstore entries
itself: so we may have removed one device but not the other.

I wonder why we're looking at the running devices at all here? What's the best 
fix?

thanks
john

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel