[2/3] PCI: hv: serialize the present/eject work items

When we hot-remove the device, we first receive a PCI_EJECT message and
then receive a PCI_BUS_RELATIONS message with bus_rel->device_count == 0.

The first message is offloaded to hv_eject_device_work(), and the second
is offloaded to pci_devices_present_work(). Both the paths can be running
list_del(&hpdev->list_entry), causing general protection fault, because
system_wq can run them concurrently.

The patch eliminates the race condition.

Signed-off-by: Dexuan Cui <decui@microsoft.com>
Tested-by: Adrian Suhov <v-adsuho@microsoft.com>
Tested-by: Chris Valean <v-chvale@microsoft.com>
Cc: Vitaly Kuznetsov <vkuznets@redhat.com>
Cc: Jack Morgenstein <jackm@mellanox.com>
Cc: stable@vger.kernel.org
Cc: Stephen Hemminger <sthemmin@microsoft.com>
Cc: K. Y. Srinivasan <kys@microsoft.com>
---
 drivers/pci/host/pci-hyperv.c | 19 ++++++++++++++++---
 1 file changed, 16 insertions(+), 3 deletions(-)

> -----Original Message-----
> From: linux-kernel-owner@vger.kernel.org <linux-kernel-owner@vger.kernel.org> On Behalf
> Of Dexuan Cui
> Sent: Friday, March 2, 2018 4:21 PM
> To: bhelgaas@google.com; linux-pci@vger.kernel.org; KY Srinivasan <kys@microsoft.com>;
> Stephen Hemminger <sthemmin@microsoft.com>
> Cc: linux-kernel@vger.kernel.org; driverdev-devel@linuxdriverproject.org; Haiyang Zhang
> <haiyangz@microsoft.com>; olaf@aepfle.de; apw@canonical.com; jasowang@redhat.com;
> vkuznets@redhat.com; marcelo.cerri@canonical.com; Dexuan Cui <decui@microsoft.com>;
> Jack Morgenstein <jackm@mellanox.com>; stable@vger.kernel.org
> Subject: [PATCH 2/3] PCI: hv: serialize the present/eject work items
> 
> When we hot-remove the device, we first receive a PCI_EJECT message and
> then receive a PCI_BUS_RELATIONS message with bus_rel->device_count == 0.
> 
> The first message is offloaded to hv_eject_device_work(), and the second
> is offloaded to pci_devices_present_work(). Both the paths can be running
> list_del(&hpdev->list_entry), causing general protection fault, because
> system_wq can run them concurrently.
> 
> The patch eliminates the race condition.

With this patch, the enum_sem field in struct hv_pcibus_device
is no longer needed.  The semaphore serializes execution in
hv_pci_devices_present_work(), and that serialization is now done
with the ordered workqueue.  Also, the last paragraph of the top level
comment for hv_pci_devices_present_work() should be updated to
reflect the new ordering assumptions.

Separately, an unrelated bug:  At the top of hv_eject_device_work(),
the first test may do a put_pcichild() and return.  This exit path also
needs to do put_hvpcibus() to balance the ref counts, or do a goto
the last two lines at the bottom of the function.

> 
> Signed-off-by: Dexuan Cui <decui@microsoft.com>
> Tested-by: Adrian Suhov <v-adsuho@microsoft.com>
> Tested-by: Chris Valean <v-chvale@microsoft.com>
> Cc: Vitaly Kuznetsov <vkuznets@redhat.com>
> Cc: Jack Morgenstein <jackm@mellanox.com>
> Cc: stable@vger.kernel.org
> Cc: Stephen Hemminger <sthemmin@microsoft.com>
> Cc: K. Y. Srinivasan <kys@microsoft.com>
> ---
>  drivers/pci/host/pci-hyperv.c | 19 ++++++++++++++++---
>  1 file changed, 16 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/pci/host/pci-hyperv.c b/drivers/pci/host/pci-hyperv.c
> index 1233300f41c6..57b1fb3ebdb9 100644
> --- a/drivers/pci/host/pci-hyperv.c
> +++ b/drivers/pci/host/pci-hyperv.c
> @@ -461,6 +461,8 @@ struct hv_pcibus_device {
>  	struct retarget_msi_interrupt retarget_msi_interrupt_params;
> 
>  	spinlock_t retarget_msi_interrupt_lock;
> +
> +	struct workqueue_struct *wq;
>  };
> 
>  /*
> @@ -1770,7 +1772,7 @@ static void hv_pci_devices_present(struct hv_pcibus_device *hbus,
>  	spin_unlock_irqrestore(&hbus->device_list_lock, flags);
> 
>  	get_hvpcibus(hbus);
> -	schedule_work(&dr_wrk->wrk);
> +	queue_work(hbus->wq, &dr_wrk->wrk);

This invocation of get_hvpcibus() and queue_work() could be made
conditional on whether the preceding list_add_tail() transitioned
the list from empty to non-empty.  If the list was already non-empty,
a previously queued invocation of hv_pci_devices_present_work()
will find the new entry and process it.   But this is an
optimization in a non-perf sensitive code path, so may not be
worth it.

>  }
> 
>  /**
> @@ -1848,7 +1850,7 @@ static void hv_pci_eject_device(struct hv_pci_dev *hpdev)
>  	get_pcichild(hpdev, hv_pcidev_ref_pnp);
>  	INIT_WORK(&hpdev->wrk, hv_eject_device_work);
>  	get_hvpcibus(hpdev->hbus);
> -	schedule_work(&hpdev->wrk);
> +	queue_work(hpdev->hbus->wq, &hpdev->wrk);
>  }
> 
>  /**
> @@ -2463,11 +2465,17 @@ static int hv_pci_probe(struct hv_device *hdev,
>  	spin_lock_init(&hbus->retarget_msi_interrupt_lock);
>  	sema_init(&hbus->enum_sem, 1);
>  	init_completion(&hbus->remove_event);
> +	hbus->wq = alloc_ordered_workqueue("hv_pci_%x", 0,
> +					   hbus->sysdata.domain);
> +	if (!hbus->wq) {
> +		ret = -ENOMEM;
> +		goto free_bus;
> +	}
> 
>  	ret = vmbus_open(hdev->channel, pci_ring_size, pci_ring_size, NULL, 0,
>  			 hv_pci_onchannelcallback, hbus);
>  	if (ret)
> -		goto free_bus;
> +		goto destroy_wq;
> 
>  	hv_set_drvdata(hdev, hbus);
> 
> @@ -2536,6 +2544,9 @@ static int hv_pci_probe(struct hv_device *hdev,
>  	hv_free_config_window(hbus);
>  close:
>  	vmbus_close(hdev->channel);
> +destroy_wq:
> +	drain_workqueue(hbus->wq);

The drain_workqueue() call isn't necessary.  destroy_workqueue() calls
drain_workqueue() and there better not be anything in the workqueue
anyway since all the ref counts are zero.

> +	destroy_workqueue(hbus->wq);
>  free_bus:
>  	free_page((unsigned long)hbus);
>  	return ret;
> @@ -2615,6 +2626,8 @@ static int hv_pci_remove(struct hv_device *hdev)
>  	irq_domain_free_fwnode(hbus->sysdata.fwnode);
>  	put_hvpcibus(hbus);
>  	wait_for_completion(&hbus->remove_event);
> +	drain_workqueue(hbus->wq);

Same here -- drain_workqueue() isn't needed.  The workqueue
must be empty anyway since the remove_event has completed
and the ref counts will all be zero.

> +	destroy_workqueue(hbus->wq);
>  	free_page((unsigned long)hbus);
>  	return 0;
>  }
> --
> 2.7.4

> From: Michael Kelley (EOSG)
> Sent: Saturday, March 3, 2018 08:10
> > From: linux-kernel-owner@vger.kernel.org <linux-kernel-
> owner@vger.kernel.org> On Behalf Of Dexuan Cui
> > Sent: Friday, March 2, 2018 4:21 PM
> > When we hot-remove the device, we first receive a PCI_EJECT message and
> > then receive a PCI_BUS_RELATIONS message with bus_rel->device_count ==
> 0.
> >
> > The first message is offloaded to hv_eject_device_work(), and the second
> > is offloaded to pci_devices_present_work(). Both the paths can be running
> > list_del(&hpdev->list_entry), causing general protection fault, because
> > system_wq can run them concurrently.
> >
> > The patch eliminates the race condition.
> 
> With this patch, the enum_sem field in struct hv_pcibus_device
> is no longer needed.  The semaphore serializes execution in
> hv_pci_devices_present_work(), and that serialization is now done
> with the ordered workqueue.  Also, the last paragraph of the top level
> comment for hv_pci_devices_present_work() should be updated to
> reflect the new ordering assumptions.

Thanks! I'll make an extra patch for this.

> Separately, an unrelated bug:  At the top of hv_eject_device_work(),
> the first test may do a put_pcichild() and return.  This exit path also
> needs to do put_hvpcibus() to balance the ref counts, or do a goto
> the last two lines at the bottom of the function.

When we're in hv_eject_device_work(), IMO hpdev->state must be 
hv_pcichild_ejecting, so I'm going to make a patch like this:

@@ -1867,10 +1867,7 @@ static void hv_eject_device_work(struct work_struct *work)

        hpdev = container_of(work, struct hv_pci_dev, wrk);

-       if (hpdev->state != hv_pcichild_ejecting) {
-               put_pcichild(hpdev, hv_pcidev_ref_pnp);
-               return;
-       }
+       WARN_ON(hpdev->state != hv_pcichild_ejecting);

        /*
         * Ejection can come before or after the PCI bus has been set up, so
 
> > @@ -1770,7 +1772,7 @@ static void hv_pci_devices_present(struct
> hv_pcibus_device *hbus,
> >  	spin_unlock_irqrestore(&hbus->device_list_lock, flags);
> >
> >  	get_hvpcibus(hbus);
> > -	schedule_work(&dr_wrk->wrk);
> > +	queue_work(hbus->wq, &dr_wrk->wrk);
> 
> This invocation of get_hvpcibus() and queue_work() could be made
> conditional on whether the preceding list_add_tail() transitioned
> the list from empty to non-empty.  If the list was already non-empty,
> a previously queued invocation of hv_pci_devices_present_work()
> will find the new entry and process it.   But this is an
> optimization in a non-perf sensitive code path, so may not be
> worth it.

Exactly. I'll add the the optimization.

> > @@ -1848,7 +1850,7 @@ static void hv_pci_eject_device(struct hv_pci_dev
> *hpdev)
> >  	get_pcichild(hpdev, hv_pcidev_ref_pnp);
> >  	INIT_WORK(&hpdev->wrk, hv_eject_device_work);
> >  	get_hvpcibus(hpdev->hbus);
> > -	schedule_work(&hpdev->wrk);
> > +	queue_work(hpdev->hbus->wq, &hpdev->wrk);
> >  }
> >
> >  /**
> > @@ -2463,11 +2465,17 @@ static int hv_pci_probe(struct hv_device *hdev,
> >  	spin_lock_init(&hbus->retarget_msi_interrupt_lock);
> >  	sema_init(&hbus->enum_sem, 1);
> >  	init_completion(&hbus->remove_event);
> > +	hbus->wq = alloc_ordered_workqueue("hv_pci_%x", 0,
> > +					   hbus->sysdata.domain);
> > +	if (!hbus->wq) {
> > +		ret = -ENOMEM;
> > +		goto free_bus;
> > +	}
> >
> >  	ret = vmbus_open(hdev->channel, pci_ring_size, pci_ring_size, NULL, 0,
> >  			 hv_pci_onchannelcallback, hbus);
> >  	if (ret)
> > -		goto free_bus;
> > +		goto destroy_wq;
> >
> >  	hv_set_drvdata(hdev, hbus);
> >
> > @@ -2536,6 +2544,9 @@ static int hv_pci_probe(struct hv_device *hdev,
> >  	hv_free_config_window(hbus);
> >  close:
> >  	vmbus_close(hdev->channel);
> > +destroy_wq:
> > +	drain_workqueue(hbus->wq);
> 
> The drain_workqueue() call isn't necessary.  destroy_workqueue() calls
> drain_workqueue() and there better not be anything in the workqueue
> anyway since all the ref counts are zero.

OK. Will remove it.

> > +	destroy_workqueue(hbus->wq);
> >  free_bus:
> >  	free_page((unsigned long)hbus);
> >  	return ret;
> > @@ -2615,6 +2626,8 @@ static int hv_pci_remove(struct hv_device *hdev)
> >  	irq_domain_free_fwnode(hbus->sysdata.fwnode);
> >  	put_hvpcibus(hbus);
> >  	wait_for_completion(&hbus->remove_event);
> > +	drain_workqueue(hbus->wq);
> 
> Same here -- drain_workqueue() isn't needed.  The workqueue
> must be empty anyway since the remove_event has completed
> and the ref counts will all be zero.

Will remove it.

> > +	destroy_workqueue(hbus->wq);

I'm going to post a v2 patchset tomorrow.

Thanks,
-- Dexuan