[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Understanding osdep_xenforeignmemory_map mmap behaviour



On 31.08.22 18:02, Alex Bennée wrote:

Juergen Gross <jgross@xxxxxxxx> writes:

[[PGP Signed Part:Undecided]]
On 24.08.22 17:58, Alex Bennée wrote:
Juergen Gross <jgross@xxxxxxxx> writes:

[[PGP Signed Part:Undecided]]
On 24.08.22 13:22, Alex Bennée wrote:
Andrew Cooper <Andrew.Cooper3@xxxxxxxxxx> writes:

On 24/08/2022 10:19, Viresh Kumar wrote:
On 24-03-22, 06:12, Juergen Gross wrote:
For a rather long time we were using "normal" user pages for this purpose,
which were just locked into memory for doing the hypercall.

Unfortunately there have been very rare problems with that approach, as
the Linux kernel can set a user page related PTE to invalid for short
periods of time, which led to EFAULT in the hypervisor when trying to
access the hypercall data.

In Linux this can avoided only by using kernel memory, which is the
reason why the hypercall buffers are allocated and mmap()-ed through the
privcmd driver.
Hi Juergen,

I understand why we moved from user pages to kernel pages, but I don't
fully understand why we need to make two separate calls to map the
guest memory, i.e. mmap() followed by ioctl(IOCTL_PRIVCMD_MMAPBATCH).

Why aren't we doing all of it from mmap() itself ? I hacked it up to
check on it and it works fine if we do it all from mmap() itself.
As I understand it the MMAPBATCH ioctl is being treated like every
other
hypercall proxy through the ioctl interface. Which makes sense from the
point of view of having a consistent interface to the hypervisor but not
from point of view of providing a consistent userspace interface for
mapping memory which doesn't care about the hypervisor details.
The privcmd_mmapbatch_v2 interface is slightly richer than what you
could expose via mmap() because it allows the handling of partial
mappings with what I presume is a per-page *err array. If you issued the
hypercall directly from the mmap() and one of the pages wasn't mapped by
the hypervisor you would have to unwind everything before returning
EFAULT to the user.

Aren't we abusing the Linux userspace ABI here ? As standard userspace
code would expect just mmap() to be enough to map the memory. Yes, the
current user, Xen itself, is adapted to make two calls, but it breaks
as soon as we want to use something that relies on Linux userspace
ABI.

For instance, in our case, where we are looking to create
hypervisor-agnostic virtio backends, the rust-vmm library [1] issues
mmap() only and expects it to work. It doesn't know it is running on a
Xen system, and it shouldn't know that as well.

Use /dev/xen/hypercall which has a sane ABI for getting "safe" memory.
privcmd is very much not sane.

In practice you'll need to use both.  /dev/xen/hypercall for getting
"safe" memory, and /dev/xen/privcmd for issuing hypercalls for now.
I'm unsure what is meant by safe memory here. privcmd_buf_mmap()
looks
like it just allocates a bunch of GFP_KERNEL pages rather than
interacting with the hypervisor directly. Are these the same pages that
get used when you eventually call privcmd_ioctl_mmap_batch()?

privcmd_buf_mmap() is allocating kernel pages which are used for data being
accessed by the hypervisor when doing the hypercall later. This is a generic
interface being used for all hypercalls, not only for
privcmd_ioctl_mmap_batch().

The fact that /dev/xen/hypercall is specified by xen_privcmdbuf_dev is a
little confusing TBH.
Anyway the goal here is to provide a non-xen aware userspace with
standard userspace API to access the guests memory. Perhaps messing

This is what the Xen related libraries are meant for. Your decision to
ignore those is firing back now.
We didn't ignore them - the initial version of the xen-vhost-master
binary was built with the rust and linking to the Xen libraries. We are
however in the process of moving to more pure rust (with the xen-sys
crate being a pure rust ioctl/hypercall wrapper).

Ah, okay, I wasn't aware of this.

However I was under the impression there where two classes of
hypercalls. ABI stable ones which won't change (which is all we are
planning to implement for xen-sys) and non-stable ABIs which would need
mediating by the xen libs. We are hoping we can do all of VirtIO with
just the stable ABI.

Okay.


around with the semantics of the /dev/xen/[hypercall|privcmd] devices
nodes is too confusing.
Maybe we could instead:
    1. Have the Xen aware VMM ask to make the guests memory visible to
the
       host kernels address space.

Urgh. This would be a major breach of the Xen security concept.

    2. When this is done explicitly create a device node to represent it 
(/dev/xen/dom-%d-mem?)
    3. Pass this new device to the non-Xen aware userspace which uses the
       standard mmap() call to make the kernel pages visible to userspace
Does that make sense?

Maybe from your point of view, but not from the Xen architectural point
of view IMHO. You are removing basically the main security advantages of
Xen by generating a kernel interface for mapping arbitrary guest memory
easily.
We are not talking about doing an end-run around the Xen
architecture.
The guest still has to instruct the hypervisor to grant access to its
memory. Currently this is a global thing (i.e. whole address space or
nothing) but obviously more fine grained grants can be done on a
transaction by transaction basis although we are exploring more
efficient mechanisms for this (shared pools and carve outs).

Happy to hear that.

This does raise questions for the mmap interface though - each
individually granted region would need to be mapped into the dom0
userspace virtual address space or perhaps a new flag for mmap() so we
can map the whole address space but expect SIGBUS faults if we access
something that hasn't been granted.

Do I understand that correctly? You want the guest to grant a memory
region to the backend, and the backend should be able to map this region
not using grants, but the guest physical addresses?

Yes - although it doesn't have to be the whole GPA range. The vhost-user
protocol communicates what offset into the GPA space the various memory
regions exist at.

How would the interface with the hypervisor look like then?

In order to make this secure, the hypervisor would need to scan the
grant table of the guest to look for a physical address the backend
wants to map. I don't think this is an acceptable interface.


Juergen

Attachment: OpenPGP_0xB0DE9DD628BF132F.asc
Description: OpenPGP public key

Attachment: OpenPGP_signature
Description: OpenPGP digital signature


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.