Message-ID: <20170119151056.44b559d1@t450s.home>
Date:   Thu, 19 Jan 2017 15:10:56 -0700
From:   Alex Williamson <alex.williamson@...hat.com>
To:     "Michael S. Tsirkin" <mst@...hat.com>
Cc:     Cao jin <caoj.fnst@...fujitsu.com>, linux-kernel@...r.kernel.org,
        kvm@...r.kernel.org, qemu-devel@...gnu.org,
        izumi.taku@...fujitsu.com
Subject: Re: [PATCH RFC] vfio error recovery: kernel support

On Thu, 19 Jan 2017 22:16:03 +0200
"Michael S. Tsirkin" <mst@...hat.com> wrote:

> This is a design and an initial patch for kernel side for AER
> support in VFIO.
> 
> 0. What happens now (PCIE AER only)
>    Fatal errors cause a link reset.
>    Non fatal errors don't.
>    All errors stop the VM eventually, but not immediately
>    because it's detected and reported asynchronously.
>    Interrupts are forwarded as usual.
>    Correctable errors are not reported to guest at all.
>    Note: PPC EEH is different. This focuses on AER.
> 
> 1. Correctable errors
>    I don't see a need to report these to guest. So let's not.
> 
> 2. Fatal errors
>    It's not easy to handle them gracefully since link reset
>    is needed. As a first step, let's use the existing mechanism
>    in that case.
>    
> 2. Non-fatal errors
>    Here we could make progress by reporting them to guest
>    and have guest handle them.
>    Issues:
>     a. this behaviour should only be enabled with new userspace
>        old userspace should work without changes
>     Suggestion: One way to address this would be to add a new eventfd
>     non_fatal_err_trigger. If not set, invoke err_trigger.
> 
>     b. drivers are supposed to stop MMIO when error is reported
>     if vm keeps going, we will keep doing MMIO/config
>     Suggestion 1: ignore this. vm stop happens much later when userspace runs anyway,
>     so we are not making things much worse
>     Suggestion 2: try to stop MMIO/config, resume on resume call
> 
>     Patch below implements Suggestion 1.

Although this is really against the documentation, which states
error_detected() is the point at which the driver should quiesce the
device and not touch it further (until diagnostic poking at
mmio_enabled or full access at resume callback).
 
>     c. PF driver might detect that function is completely broken,
>     if vm keeps going, we will keep doing MMIO/config
>     Suggestion 1: ignore this. vm stop happens much later when userspace runs anyway,
>     so we are not making things much worse
>     Suggestion 2: detect this and invoke err_trigger to stop VM
> 
>     Patch below implements Suggestion 2.
> 
> Aside: we currently return PCI_ERS_RESULT_DISCONNECT when device
> is not attached. This seems bogus, likely based on the confusing name.
> We probably should return PCI_ERS_RESULT_CAN_RECOVER.

Not sure I agree here, if we get called for the error_detected callback
and we can't find a handle for the device, we certainly don't want to
see any of the other callbacks for this device and we can't do anything
about recovering it.  What's wrong with putting the device into a
failed state in that case?

I actually question whether CAN_RECOVER is really the right return for
the existing path.  If we consider this to be a fatal error, should we
be voting NEED_RESET?  We're certainly not doing anything to return the
device to a working state.  Should we be more harsh if err_trigger is
not registered, putting the device into DISCONNECT?  Should only the new
path you've added below for non-fatal errors return CAN_RECOVER?

> The following patch does not change that.
> 
> Signed-off-by: Michael S. Tsirkin <mst@...hat.com>
> 
> ---
> 
> The patch is completely untested. Let's discuss the design first.
> Cao jin, if this is deemed acceptable please take it from here.
> 
> diff --git a/drivers/vfio/pci/vfio_pci.c b/drivers/vfio/pci/vfio_pci.c
> index dce511f..fdca683 100644
> --- a/drivers/vfio/pci/vfio_pci.c
> +++ b/drivers/vfio/pci/vfio_pci.c
> @@ -1292,7 +1292,9 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>  
>  	mutex_lock(&vdev->igate);
>  
> -	if (vdev->err_trigger)
> +	if (state == pci_channel_io_normal && vdev->non_fatal_err_trigger)
> +		eventfd_signal(vdev->err_trigger, 1);

s/err_trigger/non_fatal_err_trigger/

> +	else if (vdev->err_trigger)
>  		eventfd_signal(vdev->err_trigger, 1);
>  
>  	mutex_unlock(&vdev->igate);
> @@ -1302,8 +1304,38 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>  	return PCI_ERS_RESULT_CAN_RECOVER;
>  }
>  
> +static pci_ers_result_t vfio_pci_aer_slot_reset(struct pci_dev *pdev,
> +						pci_channel_state_t state)
> +{
> +	struct vfio_pci_device *vdev;
> +	struct vfio_device *device;
> +
> +	device = vfio_device_get_from_dev(&pdev->dev);
> +	if (!device)
> +		goto err_dev;
> +
> +	vdev = vfio_device_data(device);
> +	if (!vdev)
> +		goto err_dev;

s/err_dev/err_data/

> +
> +	mutex_lock(&vdev->igate);
> +
> +	if (vdev->err_trigger)
> +		eventfd_signal(vdev->err_trigger, 1);
> +
> +	mutex_unlock(&vdev->igate);
> +
> +	vfio_device_put(device);
> +
> +err_data:
> +	vfio_device_put(device);
> +err_dev:
> +	return PCI_ERS_RESULT_RECOVERED;
> +}
> +
>  static const struct pci_error_handlers vfio_err_handlers = {
>  	.error_detected = vfio_pci_aer_err_detected,
> +	.slot_reset = vfio_pci_aer_slot_reset,
>  };
>  
>  static struct pci_driver vfio_pci_driver = {
> diff --git a/drivers/vfio/pci/vfio_pci_intrs.c b/drivers/vfio/pci/vfio_pci_intrs.c
> index 1c46045..e883db5 100644
> --- a/drivers/vfio/pci/vfio_pci_intrs.c
> +++ b/drivers/vfio/pci/vfio_pci_intrs.c
> @@ -611,6 +611,17 @@ static int vfio_pci_set_err_trigger(struct vfio_pci_device *vdev,
>  					       count, flags, data);
>  }
>  
> +static int vfio_pci_set_non_fatal_err_trigger(struct vfio_pci_device *vdev,
> +				    unsigned index, unsigned start,
> +				    unsigned count, uint32_t flags, void *data)
> +{
> +	if (index != VFIO_PCI_NON_FATAL_ERR_IRQ_INDEX || start != 0 || count > 1)
> +		return -EINVAL;
> +
> +	return vfio_pci_set_ctx_trigger_single(&vdev->non_fatal_err_trigger,
> +					       count, flags, data);
> +}
> +
>  static int vfio_pci_set_req_trigger(struct vfio_pci_device *vdev,
>  				    unsigned index, unsigned start,
>  				    unsigned count, uint32_t flags, void *data)
> @@ -664,6 +675,14 @@ int vfio_pci_set_irqs_ioctl(struct vfio_pci_device *vdev, uint32_t flags,
>  			break;
>  		}
>  		break;
> +	case VFIO_PCI_NON_FATAL_ERR_IRQ_INDEX:
> +		switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
> +		case VFIO_IRQ_SET_ACTION_TRIGGER:
> +			if (pci_is_pcie(vdev->pdev))
> +				func = vfio_pci_set_err_trigger;
> +			break;
> +		}
> +		break;
>  	case VFIO_PCI_REQ_IRQ_INDEX:
>  		switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
>  		case VFIO_IRQ_SET_ACTION_TRIGGER:
> diff --git a/drivers/vfio/pci/vfio_pci_private.h b/drivers/vfio/pci/vfio_pci_private.h
> index f37c73b..c27a507 100644
> --- a/drivers/vfio/pci/vfio_pci_private.h
> +++ b/drivers/vfio/pci/vfio_pci_private.h
> @@ -93,6 +93,7 @@ struct vfio_pci_device {
>  	struct pci_saved_state	*pci_saved_state;
>  	int			refcnt;
>  	struct eventfd_ctx	*err_trigger;
> +	struct eventfd_ctx	*non_fatal_err_trigger;
>  	struct eventfd_ctx	*req_trigger;
>  	struct list_head	dummy_resources_list;
>  };
> 

VFIO_PCI_NON_FATAL_ERR_IRQ_INDEX never got defined.

So if we think the link is ok, we notify a non-fatal event to the user,
but we don't do anything about preventing access to the device between
error_detected and resume as the documentation indicates we should.  If
the system does a slot reset anyway, perhaps as a response to another
driver on the same bus, we promote to fatal error signaling.  If we
have no user signaling mechanism, shouldn't that also mark the device
failed via returning DISCONNECT?  On the QEMU side, we'd still need to
try to guess whether the VM is attempting a link reset is in response to
the AER event and QEMU would need to vm_stop() in that case, right?
Thanks,

Alex

Powered by blists - more mailing lists