On Fri, Apr 29, 2022 at 12:20 PM Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
>
> On 4/29/22 11:47, Dan Williams wrote:
> > On Fri, Apr 29, 2022 at 11:34 AM Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
> >>
> >> On 4/29/22 10:48, Dan Williams wrote:
> >>>> But, neither of those really help with, say, a device-DAX mapping of
> >>>> TDX-*IN*capable memory handed to KVM. The "new syscall" would just
> >>>> throw up its hands and leave users with the same result: TDX can't be
> >>>> used. The new sysfs ABI for NUMA nodes wouldn't clearly apply to
> >>>> device-DAX because they don't respect the NUMA policy ABI.
> >>> They do have "target_node" attributes to associate node specific
> >>> metadata, and could certainly express target_node capabilities in its
> >>> own ABI. Then it's just a matter of making pfn_to_nid() do the right
> >>> thing so KVM kernel side can validate the capabilities of all inbound
> >>> pfns.
> >>
> >> Let's walk through how this would work with today's kernel on tomorrow's
> >> hardware, without KVM validating PFNs:
> >>
> >> 1. daxaddr mmap("/dev/dax1234")
> >> 2. kvmfd = open("/dev/kvm")
> >> 3. ioctl(KVM_SET_USER_MEMORY_REGION, { daxaddr };
> >
> > At least for a file backed mapping the capability lookup could be done
> > here, no need to wait for the fault.
>
> For DAX mappings, sure. But, anything that's backed by page cache, you
> can't know until the RAM is allocated.
>
> ...
> >> Those pledges are hard for anonymous memory though. To fulfill the
> >> pledge, we not only have to validate that the NUMA policy is compatible
> >> at KVM_SET_USER_MEMORY_REGION, we also need to decline changes to the
> >> policy that might undermine the pledge.
> >
> > I think it's less that the kernel needs to enforce a pledge and more
> > that an interface is needed to communicate the guest death reason.
> > I.e. "here is the impossible thing you asked for, next time set this
> > policy to avoid this problem".
>
> IF this code is booted on a system where non-TDX-capable memory is
> discovered, do we:
> 1. Disable TDX, printk() some nasty message, then boot as normal
> or,
> 2a. Boot normally with TDX enabled
> 2b. Add enhanced error messages in case of TDH.MEM.PAGE.AUG/ADD failure
> (the "SEAMCALLs" which are the last line of defense and will reject
> the request to add non-TDX-capable memory to a guest). Or maybe
> an even earlier message.
>
> For #1, if TDX is on, we are quite sure it will work. But, it will
> probably throw up its hands on tomorrow's hardware. (This patch set).
>
> For #2, TDX might break (guests get killed) at runtime on tomorrow's
> hardware, but it also might be just fine. Users might be able to work
> around things by, for instance, figuring out a NUMA policy which
> excludes TDX-incapable memory. (I think what Dan is looking for)
>
> Is that a fair summary?
Yes, just the option for TDX and non-TDX to live alongside each
other... although in the past I have argued to do option-1 and enforce
it at the lowest level [1]. Like platform BIOS is responsible to
disable CXL if CXL support for a given CPU security feature is
missing. However, I think end users will want to have their
confidential computing and capacity too. As long as that is not
precluded to be added after the fact, option-1 can be a way forward
until a concrete user for mixed mode shows up.
Is there something already like this today for people that, for
example, attempt to use PCI BAR mappings as memory? Or does KVM simply
allow for garbage-in garbage-out?
In the end the patches shouldn't talk about whether or not PMEM is
supported on a platform or not, that's irrelevant. What matters is
that misconfigurations can happen, should be rare to non-existent on
current platforms, and if it becomes a problem the kernel can grow ABI
to let userspace enumerate the conflicts.
[1]: https://lore.kernel.org/linux-cxl/CAPcyv4jMQbHYQssaDDDQFEbOR1v14VUnejcSwOP9VGUnZSsCKw@xxxxxxxxxxxxxx/
