Index | Thread | Search

From:
Mark Kettenis <mark.kettenis@xs4all.nl>
Subject:
Re: SEV-SNP: Prepare for SNP page validation 1/4
To:
Hans-Jörg Höxer <hshoexer@genua.de>
Cc:
tech@openbsd.org
Date:
Tue, 11 Aug 2026 15:19:18 +0200

Download raw body.

Thread
> Date: Tue, 11 Aug 2026 11:33:30 +0200
> From: Hans-Jörg Höxer <hshoexer@genua.de>

This includes the SEV-SNP code unconditionally in all amd64 kernels.
We probably want it on RAMDISK_CD and VMBOOT.  But I'm not sure about
the (small) RAMDISK.

> this is patch 1/4 that implemnts the missing pieces for SEV-SNP guest
> support.
> 
> One essential functionality of SEV-SNP is page validation: An SNP
> enabled guest claims pages using the pvalidate instruction.  After
> successfully claiming a page, the CPU ensures that only that guest
> accesses that page.  If the hypervisor or another guest accesses
> that page, they will receive a protection fault trap.
> 
> For bounce buffers, the guest explicitly rescinds the page to the
> hypervisor.
> 
> This change implements claiming and rescinding pages for SEV-SNP.
> It also provides a new GHCB request, "Page Stage Change" (PSC),
> that is used alongside with pvalidate.
> 
> Handling of an AMD EPYC cache coherency CPU bug by Stefan Fritsch
> and Sebastian Sturm.
> 
> ok? comments?
> 
> Take care,
> HJ.
> 
> ---
>  sys/arch/amd64/amd64/ghcb.c         |  58 +++++++
>  sys/arch/amd64/amd64/snp.c          | 234 ++++++++++++++++++++++++++++
>  sys/arch/amd64/conf/files.amd64     |   1 +
>  sys/arch/amd64/include/ghcb.h       |  17 ++
>  sys/arch/amd64/include/snp.h        |  28 ++++
>  sys/arch/amd64/include/specialreg.h |   8 +
>  sys/arch/amd64/include/vmmvar.h     |   3 +-
>  7 files changed, 348 insertions(+), 1 deletion(-)
>  create mode 100644 sys/arch/amd64/amd64/snp.c
>  create mode 100644 sys/arch/amd64/include/snp.h
> 
> diff --git a/sys/arch/amd64/amd64/ghcb.c b/sys/arch/amd64/amd64/ghcb.c
> index a0539c76eb4..0083444de72 100644
> --- a/sys/arch/amd64/amd64/ghcb.c
> +++ b/sys/arch/amd64/amd64/ghcb.c
> @@ -434,3 +434,61 @@ _ghcb_io_rw(uint16_t port, int valsz, uint32_t *val, bool read)
>  	if (read)
>  		*val = frame.tf_rax;
>  }
> +
> +/*
> + * ghcb_psc_vmgexit
> + *
> + * Request a page state change from the hypervisor.
> + */
> +int
> +ghcb_psc_vmgexit(struct ghcb_psc *psc, size_t sz)
> +{
> +	struct ghcb_sa	*ghcb;
> +	uint64_t		 s, sw_exitinfo1, sw_exitinfo2;
> +	uint8_t		 valid_bm[GHCB_VB_SZ], expected_bm[GHCB_VB_SZ];
> +	int		 error = 0;
> +
> +	if (sz > sizeof(ghcb->v_sharedbuf))
> +		return (EINVAL);
> +
> +	memset(valid_bm, 0, sizeof(valid_bm));
> +	memset(expected_bm, 0, sizeof(expected_bm));
> +
> +	s = intr_disable();
> +
> +	ghcb = (struct ghcb_sa *)ghcb_vaddr;
> +	ghcb_clear(ghcb);
> +	memcpy(ghcb->v_sharedbuf, psc, sz);
> +
> +	ghcb->v_sw_exitcode = SEV_VMGEXIT_PAGE_STATE_CHANGE;
> +	ghcb->v_sw_exitinfo1 = 0;
> +	ghcb->v_sw_exitinfo2 = 0;
> +	ghcb->v_sw_scratch = ghcb_paddr + offsetof(struct ghcb_sa,
> +	    v_sharedbuf);
> +	ghcb_valbm_set(valid_bm, GHCB_SW_EXITCODE);
> +	ghcb_valbm_set(valid_bm, GHCB_SW_EXITINFO1);
> +	ghcb_valbm_set(valid_bm, GHCB_SW_EXITINFO2);
> +	ghcb_valbm_set(valid_bm, GHCB_SW_SCRATCH);
> +	ghcb_valbm_set(expected_bm, GHCB_SW_EXITINFO1);
> +	ghcb_valbm_set(expected_bm, GHCB_SW_EXITINFO2);
> +
> +	memcpy(ghcb->valid_bitmap, valid_bm, sizeof(ghcb->valid_bitmap));
> +
> +	vmgexit();
> +
> +	memcpy(valid_bm, ghcb->valid_bitmap, sizeof(valid_bm));
> +	sw_exitinfo1 = ghcb->v_sw_exitinfo1;
> +	sw_exitinfo2 = ghcb->v_sw_exitinfo2;
> +	ghcb_clear(ghcb);
> +
> +	if (ghcb_verify_bm(valid_bm, expected_bm))
> +		panic("invalid hypervisor response");
> +
> +	if (sw_exitinfo1 != 0)
> +		panic("page state change failed: 0x%llx 0x%llx", sw_exitinfo1,
> +		    sw_exitinfo2);
> +
> +	intr_restore(s);
> +
> +	return (error);
> +}
> diff --git a/sys/arch/amd64/amd64/snp.c b/sys/arch/amd64/amd64/snp.c
> new file mode 100644
> index 00000000000..64a277ba19a
> --- /dev/null
> +++ b/sys/arch/amd64/amd64/snp.c
> @@ -0,0 +1,234 @@
> +/*	$OpenBSD$	*/
> +
> +/*
> + * Copyright (c) 2025 Hans-Joerg Hoexer <hshoexer@genua.de>
> + *
> + * Permission to use, copy, modify, and distribute this software for any
> + * purpose with or without fee is hereby granted, provided that the above
> + * copyright notice and this permission notice appear in all copies.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
> + * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
> + * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
> + * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
> + * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
> + * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
> + * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
> + */
> +
> +#include <sys/param.h>
> +#include <sys/types.h>
> +#include <sys/systm.h>
> +
> +#include <machine/ghcb.h>
> +#include <machine/snp.h>
> +
> +#define SNP_PSC_PRIVATE		0x0001
> +#define SNP_PSC_SHARED		0x0002
> +
> +/*
> + * Workaround for cache coherency CPU bug named CVE-2025-38560 in linux and
> + * CVE-2024-36331 in OVMF.  It seems this can cause memory corruption even
> + * without any attacker, corrupting page tables and causing panics while
> + * booting.
> + */
> +static void
> +snp_evict_cache(vaddr_t va, unsigned int npages)
> +{
> +	volatile uint8_t dummy;
> +	uint8_t *ptr = (uint8_t *)va;
> +	unsigned int i;
> +
> +	if (curcpu()->ci_feature_amdsev_ebx & CPUIDEBX_COHERENCY_SFW_NO)
> +		return;
> +
> +	for (i = 0; i < npages; i++) {
> +		/*
> +		 * Reading from the first and last cache line of the page from
> +		 * the guest causes a flush of all cache-lines of that page
> +		 * and is cheaper than a complete clflush.
> +		 */
> +		dummy = ptr[i * PAGE_SIZE];
> +		dummy = ptr[i * PAGE_SIZE + PAGE_SIZE - 1];
> +	 }
> +}
> +
> +/*
> + * _pvalidate
> + *
> + * pvalidate single 4K page.
> + */
> +static __inline int
> +_pvalidate(vaddr_t va, uint32_t state)
> +{
> +	uint64_t	rax, rv;
> +	uint32_t	ecx, edx;
> +	uint8_t		unchanged = 0;
> +
> +	rax = va;
> +	ecx = 0;	/* 4K page */
> +	if (state == SNP_PSC_PRIVATE)
> +		edx = 1;	/* set "validated" in RMP */
> +	else
> +		edx = 0;	/* clear "validated" RMP */
> +
> +	__asm volatile(
> +		"	pvalidate				;"
> +		"	setc	%b0				;"
> +	    : "=qm" (unchanged), "=a" (rv)
> +	    : "a" (rax), "c" (ecx), "d" (edx)
> +	    : "memory", "cc");
> +
> +	/*
> +	 * pvalidate was successful, however, the "validated" entry
> +	 * was already set to "state" (ie. no change).
> +	 */
> +	if (rv == 0 && unchanged)
> +		return (-1);
> +
> +	if (state == SNP_PSC_PRIVATE)
> +		snp_evict_cache(va, 1);
> +
> +	return (rv);
> +}
> +
> +/*
> + * pvalidate
> + *
> + * pvalidate range in 4K chunks.
> + */
> +static int
> +pvalidate(paddr_t start, paddr_t end, uint32_t state, int early)
> +{
> +	paddr_t		pa;
> +	size_t		left;
> +	int		error;
> +
> +	left = end - start;
> +	for (pa = start; pa < end; pa += PAGE_SIZE, left -= PAGE_SIZE) {
> +		if ((error = _pvalidate(PMAP_DIRECT_MAP(pa), state)) != 0) {
> +			/*
> +                         * During early bootstrap it is acceptable,
> +                         * when we initially claim a page, that is
> +                         * already private.  Therefore the valid
> +                         * flags was not changed in the RMP.
> +			 *
> +			 * However, after bootstrap, this may not happen.
> +			 */
> +			if (early && state == SNP_PSC_PRIVATE && error == -1)
> +				continue;
> +			return (error);
> +		}
> +	}
> +
> +	return (0);
> +}
> +
> +/*
> + * _snp_page_state_change
> + *
> + * Set the host physical page, that is assigned to our VM as guest
> + * physical page to either "private" (ie. encrypted) or "shared" (ie.
> + * unencrypted).
> + *
> + * When sharing a page with hypervisor, ie. a page shall be used
> + * as bounce buffer, we require the following order of events:
> + *
> + * 1) Set "validate" in the RMP to 0 using pvalidate.
> + * 2) Request the hypervisor to set "assigend" in the RMP to
> + *    "hypervisor owned".
> + * 3) Establish a mapping with the C-bit cleared; with this mapping
> + *    the page can be accessed unencrypted.
> + *
> + * When claiming a page, ie. when freeing a bounce buffer page, we
> + * require the following order:
> + *
> + * 1) Remove the mapping for unencryped access.
> + * 2) Request the hypervisor to set "assigend" to "guest owned"
> + * 3) Set "validated" in the RMP to 1; this requires a mapping with
> + *    C-bit set; we use the PMAP DIRECT mapping of that page, as
> + *    these are always encrypted.
> + */
> +static void
> +_snp_page_state_change(paddr_t start, paddr_t end, int state, int early)
> +{
> +	struct ghcb_psc	psc;
> +	paddr_t		pa;
> +	int		error;
> +	size_t		sz, left;
> +
> +	left = end - start;
> +	for (pa = start; pa < end; pa += sz, left -= sz) {
> +		if ((pa & ~L2_FRAME) == 0 && left >= NBPD_L2)
> +			sz = NBPD_L2;
> +		else
> +			sz = NBPD_L1;
> +		memset(&psc, 0, sizeof(psc));
> +		psc.psc_hdr.end_entry = 0;
> +		psc.psc_entry.cur_page = 0;
> +		psc.psc_entry.gfn = pa >> PAGE_SHIFT;
> +		psc.psc_entry.operation = state;
> +		psc.psc_entry.pagesize = (sz == NBPD_L2);
> +
> +		if (state == SNP_PSC_SHARED &&
> +		    (error = pvalidate(pa, pa + sz, state, early)) != 0) {
> +			panic("pvalidate failed: %d", error);
> +		}
> +
> +		if ((error = ghcb_psc_vmgexit(&psc, sizeof(psc))) != 0)
> +			panic("ghcb_psc_vmgexit() failed: %d", error);
> +
> +		if (state == SNP_PSC_PRIVATE &&
> +		    (error =  pvalidate(pa, pa + sz, state, early)) != 0) {
> +			panic("pvalidate failed: %d", error);
> +		}
> +	}
> +}
> +
> +static void
> +_snp_claim(paddr_t start, paddr_t end, int early)
> +{
> +	_snp_page_state_change(start, end, SNP_PSC_PRIVATE, early);
> +}
> +
> +static void
> +_snp_rescind(paddr_t start, paddr_t end, int early)
> +{
> +	_snp_page_state_change(start, end, SNP_PSC_SHARED, early);
> +}
> +
> +
> +void
> +snp_claim_early(paddr_t start, paddr_t end)
> +{
> +	if (!ISSET(cpu_sev_guestmode, SEV_STAT_SNP_ACTIVE))
> +		return;
> +
> +	return _snp_claim(start, end, 1);
> +}
> +
> +void
> +snp_claim_pages(struct pglist *mlist)
> +{
> +	struct vm_page *pg;
> +
> +	if (!ISSET(cpu_sev_guestmode, SEV_STAT_SNP_ACTIVE))
> +		return;
> +
> +	TAILQ_FOREACH(pg, mlist, pageq)
> +		_snp_claim(VM_PAGE_TO_PHYS(pg),
> +		    VM_PAGE_TO_PHYS(pg) + PAGE_SIZE, 0);
> +}
> +
> +void
> +snp_rescind_pages(struct pglist *mlist)
> +{
> +	struct vm_page *pg;
> +
> +	if (!ISSET(cpu_sev_guestmode, SEV_STAT_SNP_ACTIVE))
> +		return;
> +
> +	TAILQ_FOREACH(pg, mlist, pageq)
> +		_snp_rescind(VM_PAGE_TO_PHYS(pg),
> +		    VM_PAGE_TO_PHYS(pg) + PAGE_SIZE, 0);
> +}
> diff --git a/sys/arch/amd64/conf/files.amd64 b/sys/arch/amd64/conf/files.amd64
> index 0c1e224f5c2..5474475bb41 100644
> --- a/sys/arch/amd64/conf/files.amd64
> +++ b/sys/arch/amd64/conf/files.amd64
> @@ -30,6 +30,7 @@ file	arch/amd64/amd64/i8259.c
>  file	arch/amd64/amd64/cacheinfo.c
>  file	arch/amd64/amd64/ghcb.c
>  file	arch/amd64/amd64/sev_bus_space.c
> +file	arch/amd64/amd64/snp.c
>  file	arch/amd64/amd64/vector.S
>  file	arch/amd64/amd64/copy.S
>  file	arch/amd64/amd64/spl.S
> diff --git a/sys/arch/amd64/include/ghcb.h b/sys/arch/amd64/include/ghcb.h
> index 55e184186ab..a821abdfb23 100644
> --- a/sys/arch/amd64/include/ghcb.h
> +++ b/sys/arch/amd64/include/ghcb.h
> @@ -111,6 +111,22 @@ struct ghcb_sync {
>  
>  #ifndef _LOCORE
>  
> +struct ghcb_psc {
> +	struct {
> +		uint16_t	cur_entry;
> +		uint16_t	end_entry;
> +		uint32_t	reserved;
> +	} psc_hdr;
> +
> +	struct {
> +		uint64_t	cur_page:12;
> +		uint64_t	gfn:40;
> +		uint64_t	operation:4;
> +		uint64_t	pagesize:1;
> +		uint64_t	reserved:7;
> +	} psc_entry;
> +};
> +
>  extern vaddr_t ghcb_vaddr;
>  extern paddr_t ghcb_paddr;
>  
> @@ -129,6 +145,7 @@ int	ghcb_valbm_isset(uint8_t *, int);
>  int	ghcb_verify_bm(uint8_t *, uint8_t *);
>  int	ghcb_valid(struct ghcb_sa *);
>  int	ghcb_empty(struct ghcb_sa *);
> +int	ghcb_psc_vmgexit(struct ghcb_psc *, size_t);
>  
>  void	ghcb_sync_val(int, int, struct ghcb_sync *);
>  void	ghcb_sync_out(struct trapframe *, const struct ghcb_extra_regs *,
> diff --git a/sys/arch/amd64/include/snp.h b/sys/arch/amd64/include/snp.h
> new file mode 100644
> index 00000000000..07dcea3a813
> --- /dev/null
> +++ b/sys/arch/amd64/include/snp.h
> @@ -0,0 +1,28 @@
> +/*	$OpenBSD$	*/
> +
> +/*
> + * Copyright (c) 2025 Hans-Joerg Hoexer <hshoexer@genua.de>
> + *
> + * Permission to use, copy, modify, and distribute this software for any
> + * purpose with or without fee is hereby granted, provided that the above
> + * copyright notice and this permission notice appear in all copies.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
> + * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
> + * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
> + * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
> + * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
> + * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
> + * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
> + */
> +
> +#ifndef _MACHINE_SNP_H_
> +#define _MACHINE_SNP_H_
> +
> +#include <uvm/uvm.h>
> +
> +void	snp_claim_early(paddr_t, paddr_t);
> +void	snp_claim_pages(struct pglist *);
> +void	snp_rescind_pages(struct pglist *);
> +
> +#endif /* !_MACHINE_SNP_H_ */
> diff --git a/sys/arch/amd64/include/specialreg.h b/sys/arch/amd64/include/specialreg.h
> index 17bcc8cb29f..1a32da87ed7 100644
> --- a/sys/arch/amd64/include/specialreg.h
> +++ b/sys/arch/amd64/include/specialreg.h
> @@ -431,6 +431,14 @@
>       "\023VTOMMSR" "\024IBSVIRT" "\031VMSARPROT" "\032SMTPROT" \
>       "\035SVSMPAGEMSR" "\036NVSMSR" )
>  
> +/*
> + * AMD CPUID function 0x8000001F EBX bits
> + */
> +#define CPUIDEBX_COHERENCY_SFW_NO	(1ULL << 31) /* no coherency cpu bug */
> +
> + /* Number of encrypted guests */
> + #define CPUID_AMDSEV_ECX_BITS ("\20")
> +
>  /* Number of encrypted guests */
>  #define CPUID_AMDSEV_ECX_BITS ("\20")
>  
> diff --git a/sys/arch/amd64/include/vmmvar.h b/sys/arch/amd64/include/vmmvar.h
> index 25b1618ad1f..215ee20988f 100644
> --- a/sys/arch/amd64/include/vmmvar.h
> +++ b/sys/arch/amd64/include/vmmvar.h
> @@ -267,10 +267,11 @@
>  #define SVM_VMEXIT_INVALID			-1
>  
>  /*
> - *  Additional VMEXIT codes used in SEV-ES/SNP in the GHCB
> + * Additional VMEXIT codes used in SEV-ES/SNP in the GHCB
>   */
>  #define SEV_VMGEXIT_MMIO_READ			0x80000001
>  #define SEV_VMGEXIT_MMIO_WRITE			0x80000002
> +#define SEV_VMGEXIT_PAGE_STATE_CHANGE		0x80000010
>  
>  #ifndef _LOCORE
>  
> -- 
> 2.53.0
> 
> 
> [2:application/pkcs7-signature Show Save:smime.p7s (6kB)]
>