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

Re: [PATCH for-4.14] mm: fix public declaration of struct xen_mem_acquire_resource



Hi Jan,

On 24/06/2020 11:12, Jan Beulich wrote:
On 23.06.2020 19:26, Roger Pau Monné wrote:
On Tue, Jun 23, 2020 at 06:18:52PM +0200, Jan Beulich wrote:
On 23.06.2020 17:56, Roger Pau Monné wrote:
On Tue, Jun 23, 2020 at 05:02:04PM +0200, Jan Beulich wrote:
On 23.06.2020 15:52, Roger Pau Monne wrote:
XENMEM_acquire_resource and it's related structure is currently inside
a __XEN__ or __XEN_TOOLS__ guarded section to limit it's scope to the
hypervisor or the toolstack only. This is wrong as the hypercall is
already being used by the Linux kernel at least, and as such needs to
be public.

Also switch the usage of uint64_aligned_t to plain uint64_t, as
uint64_aligned_t is only to be used by the toolstack. Note that the
layout of the structure will be the same, as the field is already
naturally aligned to a 8 byte boundary.

I'm afraid it's more complicated, and hence ...

No functional change expected.

... this doesn't hold: As I notice only now the struct also wrongly
(if it was meant to be a tools-only interface) failed to use
XEN_GUEST_HANDLE_64(), which would in principle have been a problem
for 32-bit callers (passing garbage in the upper half of a handle).
It's not an actual problem because, unlike would typically be the
case for tools-only interfaces, there is compat translation for this
struct.

Yes, there's already compat translation for the frame_list array.

With this, however, the problem of your change becomes noticeable:
The structure's alignment for 32-bit x86, and with it the structure's
sizeof(), both change. You should be able to see this by comparing
xen/common/compat/memory.c's disassembly before and after your
change - the number of bytes to copy by the respective
copy_from_guest() ought to have changed.

Right, this is the layout with frame aligned to 8 bytes:

struct xen_mem_acquire_resource {
        uint16_t                   domid;                /*     0     2 */
        uint16_t                   type;                 /*     2     2 */
        uint32_t                   id;                   /*     4     4 */
        uint32_t                   nr_frames;            /*     8     4 */
        uint32_t                   pad;                  /*    12     4 */
        uint64_t                   frame;                /*    16     8 */
        long unsigned int *        frame_list;           /*    24     4 */

        /* size: 32, cachelines: 1, members: 7 */
        /* padding: 4 */
        /* last cacheline: 32 bytes */
};

And this is without:

struct xen_mem_acquire_resource {
        uint16_t                   domid;                /*     0     2 */
        uint16_t                   type;                 /*     2     2 */
        uint32_t                   id;                   /*     4     4 */
        uint32_t                   nr_frames;            /*     8     4 */
        uint32_t                   pad;                  /*    12     4 */
        uint64_t                   frame;                /*    16     8 */
        long unsigned int *        frame_list;           /*    24     4 */

        /* size: 28, cachelines: 1, members: 7 */
        /* last cacheline: 28 bytes */
};

Note Xen has already been copying those extra 4 padding bytes from
32bit Linux kernels AFAICT, as the struct declaration in Linux has
dropped the aligned attribute.

The question now is whether we're willing to accept such a (marginal)
incompatibility. The chance of a 32-bit guest actually running into
it shouldn't be very high, but with the right placement of an
instance of the struct at the end of a page it would be possible to
observe afaict.

Or whether we go the alternative route and pad the struct suitably
for 32-bit.

I guess we are trapped with what Linux has on it's headers now, so we
should handle the struct size difference in Xen?

How do you mean to notice the difference in Xen? You don't know
what the caller's environment's sizeof() would yield.

I'm confused. Couldn't we switch from uint64_aligned_t to plain
uint64_t (like it's currently on the Linux headers), and then use the
compat layer in Xen to handle the size difference when called from
32bit environments?

And which size would we use there? The old or the new one (breaking
future or existing callers respectively)? Meanwhile I think that if
this indeed needs to not be tools-only (which I still question),

I think we now agreed on a subthread that the kernel needs to know the layout of the hypercall.

then our only possible route is to add explicit padding for the
32-bit case alongside the change you're already making.

AFAICT Linux 32-bit doesn't have this padding. So wouldn't it make incompatible the two incompatible?

Cheers,

--
Julien Grall



 


Rackspace

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