> Subject: RE: [PATCH] PCI: hv: Fix a race condition when removing the device
>
> From: longli@xxxxxxxxxxxxxxxxx <longli@xxxxxxxxxxxxxxxxx> Sent:
> Monday, April 19, 2021 12:21 PM
> >
> > On removing the device, any work item (hv_pci_devices_present() or
> > hv_pci_eject_device()) scheduled on workqueue hbus->wq may still be
> > running and race with hv_pci_remove().
> >
> > This can happen because the host may send PCI_EJECT or
> > PCI_BUS_RELATIONS(2) and decide to rescind the channel immediately
> after that.
> >
> > Fix this by flushing/stopping the workqueue of hbus before doing hbus
> remove.
>
> I can see that this change follows the same pattern as in hv_pci_suspend().
> The
> comments there give a full explanation of the issue and the solution. But
> interestingly, the current code also has a reference count mechanism on the
> hbus. And code near the end of hv_pci_remove() decrements the reference
> count and then waits for all users to finish before destroying the workqueue.
> With this change, is this reference counting mechanism still needed? If the
> workqueue has already been emptied, it seems like the
> wait_for_completion() near the end of hv_pci_remove() would never be
> waiting for anything. It makes me wonder if moving the reference count
> checking code from near the end of
> hv_pci_remove() up to near the beginning would solve the problem as well
> (and maybe in hv_pci_suspend also?).
Yes I think put_hvpcibus() and get_hvpcibus() can be removed, as we have changed to use a dedicated workqueue for hbus since they were introduced.
But we still need to call tasklet_disable/enable() the same way hv_pci_suspend() does, the reason is that we need to protect hbus->state. This value needs to be consistent for the driver. For example, a CPU may decide to schedule a work on a work queue that we just flushed or destroyed, by reading the wrong hbus->state.
>
> Michael
>
> >
> > Signed-off-by: Long Li <longli@xxxxxxxxxxxxx>
> > ---
> > drivers/pci/controller/pci-hyperv.c | 11 +++++++++++
> > 1 file changed, 11 insertions(+)
> >
> > diff --git a/drivers/pci/controller/pci-hyperv.c
> > b/drivers/pci/controller/pci-hyperv.c
> > index 27a17a1e4a7c..116815404313 100644
> > --- a/drivers/pci/controller/pci-hyperv.c
> > +++ b/drivers/pci/controller/pci-hyperv.c
> > @@ -3305,6 +3305,17 @@ static int hv_pci_remove(struct hv_device
> > *hdev)
> >
> > hbus = hv_get_drvdata(hdev);
> > if (hbus->state == hv_pcibus_installed) {
> > + tasklet_disable(&hdev->channel->callback_event);
> > + hbus->state = hv_pcibus_removing;
> > + tasklet_enable(&hdev->channel->callback_event);
> > +
> > + flush_workqueue(hbus->wq);
> > + /*
> > + * At this point, no work is running or can be scheduled
> > + * on hbus-wq. We can't race with hv_pci_devices_present()
> > + * or hv_pci_eject_device(), it's safe to proceed.
> > + */
> > +
> > /* Remove the bus from PCI's point of view. */
> > pci_lock_rescan_remove();
> > pci_stop_root_bus(hbus->pci_bus);
> > --
> > 2.27.0
