Re: [PATCH v12 19/46] x86/kernel: Make the .bss..decrypted section shared in RMP table

[Sean Christopherson <seanjc@google.com>]

On Tue, Jun 14, 2022, Sean Christopherson wrote:
> s/Brijesh/Michael
> 
> On Mon, Mar 07, 2022, Brijesh Singh wrote:
> > The encryption attribute for the .bss..decrypted section is cleared in the
> > initial page table build. This is because the section contains the data
> > that need to be shared between the guest and the hypervisor.
> > 
> > When SEV-SNP is active, just clearing the encryption attribute in the
> > page table is not enough. The page state need to be updated in the RMP
> > table.
> > 
> > Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
> > ---
> >  arch/x86/kernel/head64.c | 13 +++++++++++++
> >  1 file changed, 13 insertions(+)
> > 
> > diff --git a/arch/x86/kernel/head64.c b/arch/x86/kernel/head64.c
> > index 83514b9827e6..656d2f3e2cf0 100644
> > --- a/arch/x86/kernel/head64.c
> > +++ b/arch/x86/kernel/head64.c
> > @@ -143,7 +143,20 @@ static unsigned long __head sme_postprocess_startup(struct boot_params *bp, pmdv
> >  	if (sme_get_me_mask()) {
> >  		vaddr = (unsigned long)__start_bss_decrypted;
> >  		vaddr_end = (unsigned long)__end_bss_decrypted;
> > +
> >  		for (; vaddr < vaddr_end; vaddr += PMD_SIZE) {
> > +			/*
> > +			 * On SNP, transition the page to shared in the RMP table so that
> > +			 * it is consistent with the page table attribute change.
> > +			 *
> > +			 * __start_bss_decrypted has a virtual address in the high range
> > +			 * mapping (kernel .text). PVALIDATE, by way of
> > +			 * early_snp_set_memory_shared(), requires a valid virtual
> > +			 * address but the kernel is currently running off of the identity
> > +			 * mapping so use __pa() to get a *currently* valid virtual address.
> > +			 */
> > +			early_snp_set_memory_shared(__pa(vaddr), __pa(vaddr), PTRS_PER_PMD);
> 
> This breaks SME on Rome and Milan when compiling with clang-13.  I haven't been
> able to figure out exactly what goes wrong.  printk isn't functional at this point,
> and interactive debug during boot on our test systems is beyond me.  I can't even
> verify that the bug is specific to clang because the draconian build system for our
> test systems apparently is stuck pointing at gcc-4.9.
> 
> I suspect the issue is related to relocation and/or encrypting memory, as skipping
> the call to early_snp_set_memory_shared() if SNP isn't active masks the issue.
> I've dug through the assembly and haven't spotted a smoking gun, e.g. no obvious
> use of absolute addresses.
> 
> Forcing a VM through the same path doesn't fail.  I can't test an SEV guest at the
> moment because INIT_EX is also broken.

The SEV INIT_EX was a PEBKAC issue.  An SEV guest boots just fine with a clang-built
kernel, so either it's a finnicky relocation issue or something specific to SME.

> The crash incurs a very, very slow reboot, and I was out of cycles to work on this
> about three hours ago.  If someone on the AMD side can repro, it would be much
> appreciated.