On Mon, Apr 21, 2014 at 01:35:34PM +1000, Benjamin Herrenschmidt wrote: >On Mon, 2014-04-21 at 10:25 +0800, Wei Yang wrote: >> When pcibios_remove_pci_devices() is removing pci devices, it will release >> pci device respectively. When the refcount of the device is 0, the pci_dev >> structure will be destroyed. >> >> On PowerNV platform, the pci_dev will not be destroyed since the refcount is >> not 0. >> >> After applying the patch, this warning is cleared during the EEH hotplug >> event. > >You have to be careful here. We take a reference to the device in the >structure eeh_dev, that means we might access it after it's freed if >we don't increase the refcount.
Ben, Thanks for reminding. Hmm, I checked the eeh_dev structure, there is a field pdev in eeh_dev, which is not a ref to the device. Then I did a check in eeh hotplug code patch. The pci_dev remove/install happens in eeh_reset_device(), between pcibios_remove_pci_devices() and pcibios_add_pci_devices(). During this period, we will try to clear/set the PE state and restore the BAR of a pci device. But the BAR is restored through the device node instead of pci_dev structure. As my understanding the eeh code here has the assumption that these pci devices in the same PE are removed. Maybe I missed something, needs confirmation from Gavin. The code I removed in this patch is introduced in commit 184cd4a3(powerpc/powernv: PCI support for p7IOC under OPAL v2). Sounds the original purpose is to make sure a pci device is covered by a Bus PE, so don't remove it. After booting up, Bus PE will always be there, which means we will always hold a refcount to a pci device. I didn't come up with a better idea, or leaving the warning is fine. Any comments are welcome :) > >Cheers, >Ben. > _______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@lists.ozlabs.org https://lists.ozlabs.org/listinfo/linuxppc-dev
