On Wed, Jan 17, 2024 at 08:07:49AM +0800, Kevin Loughlin wrote:

On Mon, Jan 15, 2024 at 12:47 PM Borislav Petkov [off-list ref] wrote:

quoted

On Thu, Jan 11, 2024 at 10:36:50PM +0000, Kevin Loughlin wrote:

quoted

SEV/SME code can execute prior to page table fixups for kernel
relocation. However, as with global variables accessed in
__startup_64(), the compiler is not required to generate RIP-relative
accesses for SEV/SME global variables, causing certain flavors of SEV
hosts and guests built with clang to crash during boot.

So, about that. If I understand my gcc toolchain folks correctly:

mcmodel=kernel - everything fits into the high 31 bit of the address
space

-fPIE/PIC - position independent

And supplied both don't make a whole lotta of sense: if you're building
position-independent, then mcmodel=kernel would be overridden by the
first.

I have no clue why clang enabled it...

So, *actually* the proper fix here should be not to add this "fixed_up"
gunk everywhere but remove mcmodel=kernel from the build and simply do
-fPIE/PIC.

I believe that the key distinction is that using mcmodel=kernel (upper
2 GB of address space) or the similar mcmodel=small (lower 2 GB) means
the compiler *can* use RIP-relative addressing for globals (because
everything is within +/- 2GB of RIP) but is not *required* to do so.
In contrast, fPIE/fPIC *requires* relative addressing but does not
necessarily require a specific 2 GB placement range. Altogether, I do
think there are use cases for both options individually. I can't think
of a reason why gcc wouldn't be able to support mcmodel=kernel in
conjunction with fPIE off the top of my head, but I admittedly haven't
looked into it; I simply observed that the combination is not
currently supported.

RE: compiling the whole x86-64 kernel with fPIE/fPIC, I believe the
required changes would be very extensive (see "[PATCH RFC 00/43]
x86/pie: Make kernel image's virtual address flexible" at
https://lore.kernel.org/lkml/cover.1682673542.git.houwenlong.hwl@antgroup.com/ (local)).

FYI I have sent a small patchset[0] that attempts to build the head code
(head64.c and mem_encrypt_identity.c) as PIE. However, there are still a
few functions that are called by the head code but are not in these two
files.

[0] https://lore.kernel.org/lkml/cover.1689130310.git.houwenlong.hwl@antgroup.com (local)