Re: [PATCH v4 02/16] vfio/pci: Preserve vfio-pci device files across Live Update

[Samiullah Khawaja <skhawaja@google.com>]

On Tue, May 12, 2026 at 02:29:19PM -0700, Vipin Sharma wrote:
>On Tue, May 12, 2026 at 01:59:51PM -0700, David Matlack wrote:
>> On Mon, May 11, 2026 at 4:48 PM Vipin Sharma <vipinsh@google.com> wrote:
>>
>> > diff --git a/drivers/vfio/pci/Kconfig b/drivers/vfio/pci/Kconfig
>> > index c12d614fc6c4..019de053f116 100644
>> > --- a/drivers/vfio/pci/Kconfig
>> > +++ b/drivers/vfio/pci/Kconfig
>> > @@ -45,13 +45,15 @@ config VFIO_PCI_IGD
>> >
>> >  config VFIO_PCI_LIVEUPDATE
>> >         bool "VFIO PCI support for Live Update (EXPERIMENTAL)"
>> > -       depends on PCI_LIVEUPDATE
>> > +       depends on PCI_LIVEUPDATE && VFIO_DEVICE_CDEV
>> >         help
>> >           Support for preserving devices bound to vfio-pci across a Live
>> >           Update. This option should only be enabled by developers working on
>> >           implementing this support. Once enough support has landed in the
>> >           kernel, this option will no longer be marked EXPERIMENTAL.
>> >
>> > +         Enabling this will disable support for VFIO PCI DMA buffer.
>> > +
>> >           If you don't know what to do here, say N.
>> >
>> >  endif
>> > @@ -68,7 +70,7 @@ config VFIO_PCI_ZDEV_KVM
>> >           To enable s390x KVM vfio-pci extensions, say Y.
>> >
>> >  config VFIO_PCI_DMABUF
>> > -       def_bool y if VFIO_PCI_CORE && PCI_P2PDMA && DMA_SHARED_BUFFER
>> > +       def_bool y if VFIO_PCI_CORE && PCI_P2PDMA && DMA_SHARED_BUFFER && !VFIO_PCI_LIVEUPDATE
>>
>> Why does enabling VFIO_PCI_LIVEUPDATE require disabling
>> VFIO_PCI_DMABUF? I saw the cover letter says "to keep things simple",
>> but what specific problem does this solve or simplify?
>
>I should have provided more details there.
>
>When device is getting reset in vfio_pci_liveupdate_freeze(), we are
>zapping userspace mapped bars, we also need to use
>vfio_pci_dma_buf_move() to revoke dma buffer access or
>vfio_pci_dma_buf_cleanup() combination. Cleanup takes the memory lock
>which freeze already takes, and there are some refcounts which are
>managed in both of these APIs. This was causing complexities with code
>flow based on result of pci_load_saved_state(). All this was adding more
>refactoring than I wanted in the series.

Maybe we can return -EOPNOTSUPP if any dmabufs for this vfio cdev are
exported during preserve?
>
>I decided to just drop the change and disable the support of DMA Buffer for
>now to keep number of patches less in the series. This will go away once
>we remove reset condition in freeze.