On Tue, Oct 21, 2025 at 08:03:28PM -0300, Jason Gunthorpe wrote:
> On Sat, Oct 11, 2025 at 09:38:47PM +0200, Michał Winiarski wrote:
> > + /*
> > + * "STOP" handling is reused for "RUNNING_P2P", as the device doesn't
> > have the capability to
> > + * selectively block p2p DMA transfers.
> > + * The device is not processing new workload requests when the VF is
> > stopped, and both
> > + * memory and MMIO communication channels are transferred to
> > destination (where processing
> > + * will be resumed).
> > + */
> > + if ((cur == VFIO_DEVICE_STATE_RUNNING && new == VFIO_DEVICE_STATE_STOP)
> > ||
> > + (cur == VFIO_DEVICE_STATE_RUNNING && new ==
> > VFIO_DEVICE_STATE_RUNNING_P2P)) {
> > + ret = xe_sriov_vfio_stop(xe_vdev->pf, xe_vdev->vfid);
>
> This comment is not right, RUNNING_P2P means the device can still
> receive P2P activity on it's BAR. Eg a GPU will still allow read/write
> to its framebuffer.
>
> But it is not initiating any new transactions.
/*
* "STOP" handling is reused for "RUNNING_P2P", as the device doesn't
* have the capability to selectively block outgoing p2p DMA transfers.
* While the device is allowing BAR accesses when the VF is stopped, it
* is not processing any new workload requests, effectively stopping
* any outgoing DMA transfers (not just p2p).
* Both memory and MMIO communication channels with the workload
* scheduling firmware are transferred to destination (where processing
* will be resumed).
*/
Does this work better?
>
> > +static void xe_vfio_pci_migration_init(struct vfio_device *core_vdev)
> > +{
> > + struct xe_vfio_pci_core_device *xe_vdev =
> > + container_of(core_vdev, struct xe_vfio_pci_core_device,
> > core_device.vdev);
> > + struct pci_dev *pdev = to_pci_dev(core_vdev->dev);
> > +
> > + if (!xe_sriov_vfio_migration_supported(pdev->physfn))
> > + return;
> > +
> > + /* vfid starts from 1 for xe */
> > + xe_vdev->vfid = pci_iov_vf_id(pdev) + 1;
> > + xe_vdev->pf = pdev->physfn;
>
> No, this has to use pci_iov_get_pf_drvdata, and this driver should
> never have a naked pf pointer flowing around.
>
> The entire exported interface is wrongly formed:
>
> +bool xe_sriov_vfio_migration_supported(struct pci_dev *pdev);
> +int xe_sriov_vfio_wait_flr_done(struct pci_dev *pdev, unsigned int vfid);
> +int xe_sriov_vfio_stop(struct pci_dev *pdev, unsigned int vfid);
> +int xe_sriov_vfio_run(struct pci_dev *pdev, unsigned int vfid);
> +int xe_sriov_vfio_stop_copy_enter(struct pci_dev *pdev, unsigned int vfid);
>
> None of these should be taking in a naked pci_dev, it should all work
> on whatever type the drvdata is.
I'll change it to:
struct xe_device *xe_sriov_vfio_get_xe_device(struct pci_dev *pdev);
bool xe_sriov_vfio_migration_supported(struct xe_device *xe);
int xe_sriov_vfio_wait_flr_done(struct xe_device *xe, unsigned int vfid);
int xe_sriov_vfio_stop(struct xe_device *xe, unsigned int vfid);
int xe_sriov_vfio_run(struct xe_device *xe, unsigned int vfid);
int xe_sriov_vfio_stop_copy_enter(struct xe_device *xe, unsigned int vfid);
(...)
>
> And this gross thing needs to go away too:
>
> > + if (pdev->is_virtfn && strcmp(pdev->physfn->dev.driver->name, "xe")
> > == 0)
> > + xe_vfio_pci_migration_init(core_vdev);
Right. With using pci_iov_get_pf_drvdata() it just goes away
automatically.
Thanks,
-Michał
>
> Jason
