[Intel-gfx] [PATCH v6 09/10] vfio/pci: Extend VFIO_DEVICE_GET_PCI_HOT_RESET_INFO for vfio device cdev

[Alex Williamson]

On Thu, 1 Jun 2023 06:06:17 +0000
"Liu, Yi L" <yi.l.liu at intel.com> wrote:

> > From: Jason Gunthorpe <jgg at nvidia.com>
> > Sent: Thursday, June 1, 2023 3:00 AM
> > 
> > On Fri, May 26, 2023 at 10:04:27AM +0800, Baolu Lu wrote:  
> > > On 5/25/23 9:02 PM, Liu, Yi L wrote:  
> > > > >   It's possible that requirement
> > > > > might be relaxed in the new DMA ownership model, but as it is right
> > > > > now, the code enforces that requirement and any new discussion about
> > > > > what makes hot-reset available should note both the ownership and
> > > > > dev_set requirement.  Thanks,  
> > > > I think your point is that if an iommufd_ctx has acquired DMA ownerhisp
> > > > of an iommu_group, it means the device is owned. And it should not
> > > > matter whether all the devices in the iommu_group is present in the
> > > > dev_set. It is allowed that some devices are bound to pci-stub or
> > > > pcieport driver. Is it?
> > > >
> > > > Actually I have a doubt on it. IIUC, the above requirement on dev_set
> > > > is to ensure the reset to the devices are protected by the dev_set->lock.
> > > > So that either the reset issued by driver itself or a hot reset request
> > > > from user, there is no race. But if a device is not in the dev_set, then
> > > > hot reset request from user might race with the bound driver. DMA ownership
> > > > only guarantees the drivers won't handle DMA via DMA API which would have
> > > > conflict with DMA mappings from user. I'm not sure if it is able to
> > > > guarantee reset is exclusive as well. I see pci-stub and pcieport driver
> > > > are the only two drivers that set the driver_managed_dma flag besides the
> > > > vfio drivers. pci-stub may be fine. not sure about pcieport driver.  
> > >
> > > commit c7d469849747 ("PCI: portdrv: Set driver_managed_dma") described
> > > the criteria of adding driver_managed_dma to the pcieport driver.
> > >
> > > "
> > > We achieve this by setting ".driver_managed_dma = true" in pci_driver
> > > structure. It is safe because the portdrv driver meets below criteria:
> > >
> > > - This driver doesn't use DMA, as you can't find any related calls like
> > >   pci_set_master() or any kernel DMA API (dma_map_*() and etc.).
> > > - It doesn't use MMIO as you can't find ioremap() or similar calls. It's
> > >   tolerant to userspace possibly also touching the same MMIO registers
> > >   via P2P DMA access.
> > > "
> > >
> > > pci_rest_device() definitely shouldn't be done by the kernel drivers
> > > that have driver_managed_dma set.  
> > 
> > Right
> > 
> > The only time it is safe to reset is if you know there is no attached
> > driver or you know VFIO is the attached driver and the caller owns the
> > VFIO too.
> > 
> > We haven't done a no attached driver test due to races.  
> 
> Ok. @Alex, should we relax the above dev_set requirement now or should
> be in a separate series?


Sounds like no, you should be rejecting enhancements that increase
scope at this point and I don't see consensus here.  My concern was
that we're not correctly describing the dev_set restriction which is
already in place but needs to be more explicitly described in an
implied ownership model vs proof of ownership model.  Thanks,

Alex