Messages in this thread | | | Date | Thu, 20 Sep 2012 18:51:04 -0700 | Subject | Re: [PATCH 4/5] PCI/IOV: simplify code by hotplug safe pci_get_domain_bus_and_slot() | From | Yinghai Lu <> |
| |
On Thu, Sep 20, 2012 at 4:59 PM, Bjorn Helgaas <bhelgaas@google.com> wrote: > On Thu, Sep 20, 2012 at 2:38 PM, Yinghai Lu <yinghai@kernel.org> wrote: >> in that case, VFs are stopped before PF, so they are not in device >> tree anymore. >> so pci_get_domain_bus_and_slot will not find those VFs. >> >> So the reference to PF is not released. Also vit_bus may not be released too. >> >> So you have to rework >> pci_get_domain_bus_and_slot to make it work on pci devices get stopped only. >> >> or just drop this from the tree. > > pci_find_bus() is a broken interface (because there's no reference > counting or safety with respect to hot-plug), and if the design > depends on it, that means the design is broken, too. I don't think > reworking pci_get_domain_bus_and_slot() is the right answer. > > It's not clear to me why we need the split between stopping and > removing devices. That split leads to these zombie devices that have > been stopped and are no longer findable by bus_find_device() (which is > used by pci_get_domain_bus_and_slot()), but still "exist" in some > fashion until they're removed. It's unreasonable for random PCI and > driver code to have to worry about that zombie state.
That is not zombie state. that is driver unloaded, and not in /sys, /proc. that pci device only can be found under bus->devices.
just like we have pci_device_add and pci_bus_add_device or acpi add and acpi start.
Splitting stop and remove should be more clean than mixing stop and remove.
> > I'll revert this patch for now to fix the regression. Of course, that > means we will still have the old problem of using the unsafe > pci_find_bus(). > >> BTW, Bjorn, >> for the similar reason, you need to apply >> http://git.kernel.org/?p=linux/kernel/git/yinghai/linux-yinghai.git;a=commitdiff;h=e5a50aa3dfca1331c7c783412b1524bea06d2752 >> ... >> the reason is: >> >> We stop all VFs at first , stop PF >> before we stop PF, we can not remove VFs, >> >> otherwise virtfn_remove does not work properly to remove reference and >> virtfn bus while can not find vf. > > I'll apply this one, too, for now. I put them both on the > pci/yinghai-revert-pci_find_bus-and-remove-cleanup branch. Let me > know if that's not what you expected.
Yes, they are good.
> > I'm not happy about either reverting Jiang's patch or splitting > stop/remove again. It complicates the design and the code. I'll > apply them because they're regressions, and we don't have time for > redesign before 3.7. But I encourage you to think about how to do > this more cleanly.
That will need to redesign sriov implementation.
Also that pci root bus add/start, stop/remove will need special sequence to make ioapic and dmar to be started early before normal pci device drivers and stopped after normal pci device drivers.
-Yinghai
| |