Re: [PATCH v2 2/2] pci: Support "removable" attribute for PCI devices

[Rajat Jain <rajatja@google.com>]

On Tue, May 11, 2021 at 4:02 PM Bjorn Helgaas <helgaas@kernel.org> wrote:
>
> On Tue, May 11, 2021 at 03:15:11PM -0700, Rajat Jain wrote:
> > On Tue, May 11, 2021 at 2:30 PM Bjorn Helgaas <helgaas@kernel.org> wrote:
> > > On Fri, Apr 23, 2021 at 07:16:31PM -0700, Rajat Jain wrote:
> > > ...
> > > This looks like a good start.  I think it would be useful to have a
> > > more concrete example of how this information will be used.  I know
> > > that use would be in userspace, so an example probably would not be a
> > > kernel patch.  If you have user code published anywhere, that would
> > > help.  Or even a patch to an existing daemon.  Or pointers to how
> > > "removable" is used for USB devices.
> >
> > Sure, I'll point to some existing user space code (which will be using
> > a similar attribute we are carrying internally).
>
> Great, thanks!
>
> > > > +     set_pci_dev_removable(dev);
> > >
> > > So this *only* sets the "removable" attribute based on the
> > > ExternalFacingPort or external-facing properties.  I think Oliver and
> > > David were hinting that maybe we should also set it for devices in
> > > hotpluggable slots.  What do you think?
> >
> > I did think about it. So I have a mixed feeling about this. Primarily
> > because I have seen the use of hotpluggable slots in situations where
> > we wouldn't want to classify the device as removable:
> >
> > - Using link-state based hotplug as a way to work around unstable PCIe
> > links. I have seen PCIe devices marked as hot-pluggable only to ensure
> > that if the PCIe device falls off PCI bus due to some reason (e.g. due
> > to SI issues or device firmware bugs), the kernel should be able to
> > detect it if it does come back up (remember quick "Link-Down" /
> > "Link-Up" events in succession?).
> >
> > - Internal hot-pluggable PCI devices. In my past life, I was working
> > on a large system that would have hot-pluggable daughter cards, but
> > those wouldn't be user removable. Also, it is conceivable to have
> > hot-pluggable M.2 slots for PCIe devices such as NVMEs etc, but they
> > may still not be removable by user. I don't think these should be
> > treated as "removable". I was also looking at USB as an example where
> > this originally came from, USB does ensure that only devices that are
> > "user visible" devices are marked as "removable":
> >
> > 54d3f8c63d69 ("usb: Set device removable state based on ACPI USB data")
> > d35e70d50a06 ("usb: Use hub port data to determine whether a port is removable")
>
> IIUC your main concern is consumer platforms where PCI devices would
> be hotplugged via a Thunderbolt or similar cable, and that port
> would be marked as an "ExternalFacingPort" so we'd mark them as
> "removable".

Yes.

>
> A device in a server hotplug slot would probably *not* be marked as
> "removable".  The same device in an external chassis connected via an
> iPass or similar cable *might* be "removable" depending on whether the
> firmware calls the iPass port an "ExternalFacingPort".

Yes.

>
> Does the following capture some of what you're thinking?  Maybe some
> wordsmithed version of it would be useful in a comment and/or commit
> log?

Yes, you captured my thoughts perfectly. I shall update the commit log
and / or provide comments to reflect this.

>
>   We're mainly concerned with consumer platforms with accessible
>   Thunderbolt ports that are vulnerable to DMA attacks, and we expect
>   those ports to be identified as "ExternalFacingPort".
>
>   Devices in traditional hotplug slots are also "removable," but not
>   as vulnerable because these slots are less accessible to users.
>
> > > I wonder if this (and similar hooks like set_pcie_port_type(),
> > > set_pcie_untrusted(), set_pcie_thunderbolt(), etc) should go *after*
> > > the early fixups so we could use fixups to work around issues?
> >
> > I agree. We can do that if none of the early fixups actually use the
> > fields set by these functions. I think it should be ok to move
> > set_pcie_untrusted(), set_pcie_thunderbolt(), but I wonder if any
> > early fixups already use the pcie_cap or any other fields set by
> > set_pcie_port_type().
>
> I think you should move the one you're adding
> (set_pci_dev_removable()) and leave the others where they are for now.

Ack, will do.

Thanks,

Rajat

>
> No need to expand the scope of your patch; I was just thinking they're
> all basically similar and should ideally be done at similar times.
>
> > > >       /* Early fixups, before probing the BARs */
> > > >       pci_fixup_device(pci_fixup_early, dev);
> > > >
> > > > --
> > > > 2.31.1.498.g6c1eba8ee3d-goog
> > > >