On Fri, Aug 02, 2019 at 08:31:26PM +0000, Dexuan Cui wrote:

quoted

From: Bjorn Helgaas <helgaas@kernel.org>
Sent: Friday, August 2, 2019 12:41 PM
The subject line only describes the mechanical code change, which is
obvious from the patch.  It would be better if we could say something
about *why* we need this.

Hi Bjorn,
Sorry. I'll try to write a better changelog in v2. :-)
 

quoted

On Fri, Aug 02, 2019 at 01:32:28AM +0000, Dexuan Cui wrote:

quoted

When a slot is removed, the pci_dev must still exist.

pci_remove_root_bus() removes and free all the pci_devs, so
hv_pci_remove_slots() must be called before pci_remove_root_bus(),
otherwise a general protection fault can happen, if the kernel is built

"general protection fault" is an x86 term that doesn't really say what
the issue is.  I suspect this would be a "use-after-free" problem.

Yes, it's use-after-free. I'll fix the the wording.
 

quoted

quoted

--- a/drivers/pci/controller/pci-hyperv.c
+++ b/drivers/pci/controller/pci-hyperv.c

@@ -2757,8 +2757,8 @@ static int hv_pci_remove(struct hv_device *hdev)
 		/* Remove the bus from PCI's point of view. */
 		pci_lock_rescan_remove();
 		pci_stop_root_bus(hbus->pci_bus);
-		pci_remove_root_bus(hbus->pci_bus);
 		hv_pci_remove_slots(hbus);
+		pci_remove_root_bus(hbus->pci_bus);

I'm curious about why we need hv_pci_remove_slots() at all.  None of
the other callers of pci_stop_root_bus() and pci_remove_root_bus() do
anything similar to hv_pci_remove_slots().

Surely some of those callers also support slots, so there must be some
other path that calls pci_destroy_slot() in those cases.  Can we use a
similar strategy here?

Originally Stephen Heminger added the slot code for pci-hyperv.c:
a15f2c08c708 ("PCI: hv: support reporting serial number as slot information")
So he may know this better. My understanding is: we can not use the similar
stragegy used in the 2 other users of pci_create_slot():

drivers/pci/hotplug/pci_hotplug_core.c calls pci_create_slot().
It looks drivers/pci/hotplug/ is quite different from pci-hyperv.c because
pci-hyper-v uses a simple *private* hot-plug protocol, making it impossible
to use the API pci_hp_register() and pci_hp_destroy() -> pci_destroy_slot().

drivers/acpi/pci_slot.c calls pci_create_slot(), and saves the created slots in
the static "slot_list" list in the same file. Again, since pci-hyper-v uses a private
PCI-device-discovery protocol (which is based on VMBus rather the emulated
ACPI and PCI), acpi_pci_slot_enumerate() can not find the PCI devices that are
discovered by pci-hyperv, so we can not use the standard register_slot() ->
pci_create_slot() to create the slots and hence acpi_pci_slot_remove() -> 
pci_destroy_slot() can not work for pci-hyperv.

Hmm, ok.  This still doesn't seem right to me, but I think the bottom
line will be that the current slot registration interfaces just don't
work quite right for all the cases we want them to.

Maybe it would be a good project for somebody to rethink them, but it
doesn't seem practical for *this* patch.  Thanks for looking into it
this far!

I think I can use this as the v2 changelog:

The slot must be removed before the pci_dev is removed, otherwise a panic
can happen due to use-after-free.

Sounds good.

Bjorn