Re: [PATCH] x86, reboot: skip reboot_fixups in early boot phase

[Bjorn Helgaas <bhelgaas@google.com>]

[+cc Greg for driver core]

On Fri, Jan 25, 2013 at 10:13:03AM +0900, Joonsoo Kim wrote:
> Hello, Bjorn.
> 
> On Thu, Jan 24, 2013 at 10:45:13AM -0700, Bjorn Helgaas wrote:
> > On Fri, Dec 28, 2012 at 6:50 AM, Joonsoo Kim <js1304@gmail.com> wrote:
> > > During early boot phase, PCI bus subsystem is not yet initialized.
> > > If panic is occured in early boot phase and panic_timeout is set,
> > > code flow go into emergency_restart() and hit mach_reboot_fixups(), then
> > > encounter another panic. When second panic, we can't hold a panic_lock, so
> > > code flow go into panic_smp_self_stop() which prevent system to restart.
> > >
> > > For avoid second panic, skip reboot_fixups in early boot phase.
> > > It makes panic_timeout works in early boot phase.
> > >
> > > Cc: Thomas Gleixner <tglx@linutronix.de>
> > > Cc: Ingo Molnar <mingo@redhat.com>
> > > Cc: "H. Peter Anvin" <hpa@zytor.com>
> > > Signed-off-by: Joonsoo Kim <js1304@gmail.com>
> > >
> > > diff --git a/arch/x86/kernel/reboot_fixups_32.c b/arch/x86/kernel/reboot_fixups_32.c
> > > index c8e41e9..b9b8ec9 100644
> > > --- a/arch/x86/kernel/reboot_fixups_32.c
> > > +++ b/arch/x86/kernel/reboot_fixups_32.c
> > > @@ -89,6 +89,10 @@ void mach_reboot_fixups(void)
> > >         if (in_interrupt())
> > >                 return;
> > >
> > > +       /* During early boot phase, PCI is not yet initialized */
> > > +       if (system_state == SYSTEM_BOOTING)
> > > +               return;
> > > +
> > >         for (i=0; i < ARRAY_SIZE(fixups_table); i++) {
> > >                 cur = &(fixups_table[i]);
> > >                 dev = pci_get_device(cur->vendor, cur->device, NULL);
> > 
> > I guess you're saying that if we call pci_get_device() too early, it
> > panics?  Did you figure out why that happens?
> > 
> > If we call pci_get_device() before PCI has been initialized, it would
> > be good if it just returned NULL, indicating that we didn't find any
> > matching device.  I looked briefly, and I thought that's what would
> > happen, but apparently I'm missing something.
> 
> In bus_find_device(), klist_iter_init_node() is called with 
> @bus->p->klist_device. Before initialization, bus->p is NULL,
> so panic is occured.

I see.  pci_bus_type.p is initialized by __bus_register() in this path:

  pci_driver_init        # postcore_initcall
    bus_register(&pci_bus_type)
      __bus_register
        priv = kzalloc(sizeof(struct subsys_private))
	bus->p = priv
	klist_init(&priv->klist_devices, klist_devices_get, klist_devices_put)

I was hoping we could statically initialize the klist, but that doesn't
seem likely.

But I wonder if we could do something like the following.  If we could,
then callers wouldn't have to worry about whether or not the bus has been
initialized.

diff --git a/drivers/base/bus.c b/drivers/base/bus.c
index 24eb078..ede19b8 100644
--- a/drivers/base/bus.c
+++ b/drivers/base/bus.c
@@ -290,7 +290,7 @@ int bus_for_each_dev(struct bus_type *bus, struct device *start,
 	struct device *dev;
 	int error = 0;
 
-	if (!bus)
+	if (!bus || !bus->p)
 		return -EINVAL;
 
 	klist_iter_init_node(&bus->p->klist_devices, &i,
@@ -324,7 +324,7 @@ struct device *bus_find_device(struct bus_type *bus,
 	struct klist_iter i;
 	struct device *dev;
 
-	if (!bus)
+	if (!bus || !bus->p)
 		return NULL;
 
 	klist_iter_init_node(&bus->p->klist_devices, &i,
@@ -440,7 +440,7 @@ int bus_for_each_drv(struct bus_type *bus, struct device_driver *start,
 	struct device_driver *drv;
 	int error = 0;
 
-	if (!bus)
+	if (!bus || !bus->p)
 		return -EINVAL;
 
 	klist_iter_init_node(&bus->p->klist_drivers, &i,