Re: [RFC PATCH] x86/sev: Disallow userspace access to BIOS region for SEV-SNP guests

[Dan Williams <dan.j.williams@intel.com>]

Nikolay Borisov wrote:
> 
> 
> On 9.04.25 г. 2:55 ч., Dan Williams wrote:
> > Dave Hansen wrote:
> >> On 4/8/25 06:43, Tom Lendacky wrote:
> >>>> Tom/Boris, do you see a problem blocking access to /dev/mem for SEV
> >>>> guests?
> >>> Not sure why we would suddenly not allow that.
> >>
> >> Both TDX and SEV-SNP have issues with allowing access to /dev/mem.
> >> Disallowing access to the individually troublesome regions can fix
> >> _part_ of the problem. But suddenly blocking access is guaranteed to fix
> >> *ALL* the problems forever.
> > 
> > ...or at least solicits practical use cases for why the kernel needs to
> > poke holes in the policy.
> > 
> >> Or, maybe we just start returning 0's for all reads and throw away all
> >> writes. That is probably less likely to break userspace that doesn't
> >> know what it's doing in the first place.
> > 
> > Yes, and a bulk of the regression risk has already been pipe-cleaned by
> > KERNEL_LOCKDOWN that shuts down /dev/mem and PCI resource file mmap in
> > many scenarios.
> > 
> > Here is an updated patch that includes some consideration for mapping
> > zeros for known legacy compatibility use cases.
> > 
> > -- 8< --
> > From: Dan Williams <dan.j.williams@intel.com>
> > Subject: [PATCH] x86: Restrict /dev/mem access for potentially unaccepted
> >   memory by default
> > 
> > Nikolay reports [1] that accessing BIOS data (first 1MB of the physical
> > address space) via /dev/mem results in an SEPT violation.
> > 
> > The cause is ioremap() (via xlate_dev_mem_ptr()) establishes an
> > unencrypted mapping where the kernel had established an encrypted
> > mapping previously.
> > 
> > An initial attempt to fix this revealed that TDX and SEV-SNP have
> > different expectations about which and when address ranges can be mapped
> > via /dev/mem.
> > 
> > Rather than develop a precise set of allowed /dev/mem capable TVM
> > address ranges, lean on the observation that KERNEL_LOCKDOWN is already
> > blocking /dev/mem access in many cases to do the same by default for x86
> > TVMs. This can still be later relaxed as specific needs arise, but in
> > the meantime close off this source of mismatched IORES_MAP_ENCRYPTED
> > expectations.
> > 
> > Note that this is careful to map zeroes rather than reject mappings of
> > the BIOS data space.
> > 
> > Cc: <x86@kernel.org>
> > Cc: Vishal Annapurve <vannapurve@google.com>
> > Cc: Kirill Shutemov <kirill.shutemov@linux.intel.com>
> > Reported-by: Nikolay Borisov <nik.borisov@suse.com>
> > Closes: http://lore.kernel.org/20250318113604.297726-1-nik.borisov@suse.com [1]
> > Fixes: 9aa6ea69852c ("x86/tdx: Make pages shared in ioremap()")
> > Cc: <stable@vger.kernel.org>
> > Signed-off-by: Dan Williams <dan.j.williams@intel.com>
> > ---
> >   arch/x86/Kconfig                |  2 ++
> >   arch/x86/include/asm/x86_init.h |  2 ++
> >   arch/x86/kernel/x86_init.c      |  6 ++++++
> >   arch/x86/mm/init.c              | 14 +++++++++++---
> >   4 files changed, 21 insertions(+), 3 deletions(-)
> > 
> > diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> > index 15f346f02af0..6d4f94a79314 100644
> > --- a/arch/x86/Kconfig
> > +++ b/arch/x86/Kconfig
> > @@ -888,6 +888,7 @@ config INTEL_TDX_GUEST
> >   	depends on X86_64 && CPU_SUP_INTEL
> >   	depends on X86_X2APIC
> >   	depends on EFI_STUB
> > +	depends on STRICT_DEVMEM
> >   	select ARCH_HAS_CC_PLATFORM
> >   	select X86_MEM_ENCRYPT
> >   	select X86_MCE
> > @@ -1507,6 +1508,7 @@ config AMD_MEM_ENCRYPT
> >   	bool "AMD Secure Memory Encryption (SME) support"
> >   	depends on X86_64 && CPU_SUP_AMD
> >   	depends on EFI_STUB
> > +	depends on STRICT_DEVMEM
> >   	select DMA_COHERENT_POOL
> >   	select ARCH_USE_MEMREMAP_PROT
> >   	select INSTRUCTION_DECODER
> > diff --git a/arch/x86/include/asm/x86_init.h b/arch/x86/include/asm/x86_init.h
> > index 213cf5379a5a..0ae436b34b88 100644
> > --- a/arch/x86/include/asm/x86_init.h
> > +++ b/arch/x86/include/asm/x86_init.h
> > @@ -305,6 +305,7 @@ struct x86_hyper_runtime {
> >    * 				semantics.
> >    * @realmode_reserve:		reserve memory for realmode trampoline
> >    * @realmode_init:		initialize realmode trampoline
> > + * @devmem_is_allowed		restrict /dev/mem and PCI sysfs resource access
> >    * @hyper:			x86 hypervisor specific runtime callbacks
> >    */
> >   struct x86_platform_ops {
> > @@ -323,6 +324,7 @@ struct x86_platform_ops {
> >   	void (*set_legacy_features)(void);
> >   	void (*realmode_reserve)(void);
> >   	void (*realmode_init)(void);
> > +	bool (*devmem_is_allowed)(unsigned long pfn);
> >   	struct x86_hyper_runtime hyper;
> >   	struct x86_guest guest;
> >   };
> > diff --git a/arch/x86/kernel/x86_init.c b/arch/x86/kernel/x86_init.c
> > index 0a2bbd674a6d..346301375bd4 100644
> > --- a/arch/x86/kernel/x86_init.c
> > +++ b/arch/x86/kernel/x86_init.c
> > @@ -143,6 +143,11 @@ static void enc_kexec_begin_noop(void) {}
> >   static void enc_kexec_finish_noop(void) {}
> >   static bool is_private_mmio_noop(u64 addr) {return false; }
> >   
> > +static bool platform_devmem_is_allowed(unsigned long pfn)
> > +{
> > +	return !cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT);
> > +}
> > +
> >   struct x86_platform_ops x86_platform __ro_after_init = {
> >   	.calibrate_cpu			= native_calibrate_cpu_early,
> >   	.calibrate_tsc			= native_calibrate_tsc,
> > @@ -156,6 +161,7 @@ struct x86_platform_ops x86_platform __ro_after_init = {
> >   	.restore_sched_clock_state	= tsc_restore_sched_clock_state,
> >   	.realmode_reserve		= reserve_real_mode,
> >   	.realmode_init			= init_real_mode,
> > +	.devmem_is_allowed		= platform_devmem_is_allowed,
> >   	.hyper.pin_vcpu			= x86_op_int_noop,
> >   	.hyper.is_private_mmio		= is_private_mmio_noop,
> >   
> > diff --git a/arch/x86/mm/init.c b/arch/x86/mm/init.c
> > index bfa444a7dbb0..c8679ae1bc8b 100644
> > --- a/arch/x86/mm/init.c
> > +++ b/arch/x86/mm/init.c
> > @@ -867,6 +867,8 @@ void __init poking_init(void)
> >    */
> >   int devmem_is_allowed(unsigned long pagenr)
> >   {
> > +	bool platform_allowed = x86_platform.devmem_is_allowed(pagenr);
> > +
> >   	if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE,
> >   				IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE)
> >   			!= REGION_DISJOINT) {
> > @@ -885,14 +887,20 @@ int devmem_is_allowed(unsigned long pagenr)
> >   	 * restricted resource under CONFIG_STRICT_DEVMEM.
> >   	 */
> >   	if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) {
> > -		/* Low 1MB bypasses iomem restrictions. */
> > -		if (pagenr < 256)
> > +		/*
> > +		 * Low 1MB bypasses iomem restrictions unless the
> > +		 * platform says "no", in which case map zeroes
> > +		 */
> > +		if (pagenr < 256) {
> > +			if (!platform_allowed)
> > +				return 2;
> 
> That'll work but I hate the way this interface works. The sole user of 
> this 0/1/2 convention is page_is_allowed() and the check for 1  inside 
> write_mem(). The proper patch will need to document this...

That's good feedback. I will introduce some defines for those magic
values: DEVMEM_{ALLOW,DENY,ZEROES}.

> Reviewed-by: Nikolay Borisov <nik.borisov@suse.com>

Thanks for taking a look.