Re: [PATCH] PCI: hv: Fix a bug on removing child devices on the bus

[Rob Herring]

On Wed, Aug 25, 2021 at 2:11 PM Michael Kelley <mikelley@xxxxxxxxxxxxx> wrote:
>
> From: Long Li <longli@xxxxxxxxxxxxx> Sent: Tuesday, August 24, 2021 10:28 AM
> >
> > > Subject: Re: [PATCH] PCI: hv: Fix a bug on removing child devices on the bus
> > >
> > > On Tue, Aug 24, 2021 at 12:20:20AM -0700, longli@xxxxxxxxxxxxxxxxx wrote:
> > > > From: Long Li <longli@xxxxxxxxxxxxx>
> > > >
> > > > In hv_pci_bus_exit, the code is holding a spinlock while calling
> > > > pci_destroy_slot(), which takes a mutex.
> > > >
> > > > This is not safe for spinlock. Fix this by moving the children to be
> > > > deleted to a list on the stack, and removing them after spinlock is
> > > > released.
> > > >
> > > > Fixes: 94d22763207a ("PCI: hv: Fix a race condition when removing the
> > > > device")
> > > >
> > > > Cc: "K. Y. Srinivasan" <kys@xxxxxxxxxxxxx>
> > > > Cc: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>
> > > > Cc: Stephen Hemminger <sthemmin@xxxxxxxxxxxxx>
> > > > Cc: Wei Liu <wei.liu@xxxxxxxxxx>
> > > > Cc: Dexuan Cui <decui@xxxxxxxxxxxxx>
> > > > Cc: Lorenzo Pieralisi <lorenzo.pieralisi@xxxxxxx>
> > > > Cc: Rob Herring <robh@xxxxxxxxxx>
> > > > Cc: "Krzysztof Wilczyński" <kw@xxxxxxxxx>
> > > > Cc: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
> > > > Cc: Michael Kelley <mikelley@xxxxxxxxxxxxx>
> > > > Cc: Dan Carpenter <dan.carpenter@xxxxxxxxxx>
> > > > Reported-by: Dan Carpenter <dan.carpenter@xxxxxxxxxx>
> > > > Signed-off-by: Long Li <longli@xxxxxxxxxxxxx>
> > > > ---
> > > >  drivers/pci/controller/pci-hyperv.c | 15 ++++++++++++---
> > > >  1 file changed, 12 insertions(+), 3 deletions(-)
> > > >
> > > > diff --git a/drivers/pci/controller/pci-hyperv.c
> > > > b/drivers/pci/controller/pci-hyperv.c
> > > > index a53bd8728d0d..d4f3cce18957 100644
> > > > --- a/drivers/pci/controller/pci-hyperv.c
> > > > +++ b/drivers/pci/controller/pci-hyperv.c
> > > > @@ -3220,6 +3220,7 @@ static int hv_pci_bus_exit(struct hv_device *hdev,
> > > bool keep_devs)
> > > >   struct hv_pci_dev *hpdev, *tmp;
> > > >   unsigned long flags;
> > > >   int ret;
> > > > + struct list_head removed;
> > >
> > > This can be moved to where it is needed -- the if(!keep_dev) branch -- to limit its
> > > scope.
> > >
> > > >
> > > >   /*
> > > >    * After the host sends the RESCIND_CHANNEL message, it doesn't @@
> > > > -3229,9 +3230,18 @@ static int hv_pci_bus_exit(struct hv_device *hdev, bool
> > > keep_devs)
> > > >           return 0;
> > > >
> > > >   if (!keep_devs) {
> > > > -         /* Delete any children which might still exist. */
> > > > +         INIT_LIST_HEAD(&removed);
> > > > +
> > > > +         /* Move all present children to the list on stack */
> > > >           spin_lock_irqsave(&hbus->device_list_lock, flags);
> > > > -         list_for_each_entry_safe(hpdev, tmp, &hbus->children,
> > > list_entry) {
> > > > +         list_for_each_entry_safe(hpdev, tmp, &hbus->children,
> > > list_entry)
> > > > +                 list_move_tail(&hpdev->list_entry, &removed);
> > > > +         spin_unlock_irqrestore(&hbus->device_list_lock, flags);
> > > > +
> > > > +         /* Remove all children in the list */
> > > > +         while (!list_empty(&removed)) {
> > > > +                 hpdev = list_first_entry(&removed, struct hv_pci_dev,
> > > > +                                          list_entry);
> > >
> > > list_for_each_entry_safe can also be used here, right?
> > >
> > > Wei.
> >
> > I will address your comments.
> >
> > Long
>
> I thought list_for_each_entry_safe() is for use when list manipulation
> is *not* protected by a lock and you want to safely walk the list
> even if an entry gets removed.  If the list is protected by a lock or
> not subject to contention (as is the case here), then
> list_for_each_entry() is the simpler implementation.  The original
> implementation didn't need to use the _safe version because of
> the spin lock.
>
> Or do I have it backwards?

"_safe" only means "safe against removal of list entry" as the
kerneldoc says. But that means removal within the loop iteration, not
any writer. A lock is needed in either case if there's another writer.

Don't ask me about the RCU variant though...

Rob