[PATCH] x86/sev: Skip DR7 write during kexec when it would trigger an unserviceable #VC

From: Ashish Kalra <[email protected]>

machine_kexec() calls hw_breakpoint_disable(), which writes DR7. On
SEV-ES/SNP guests that do not have DebugSwap enabled, DR7 accesses are
intercepted and delivered to the guest as a #VC exception, which must be
serviced through the GHCB.

By the time machine_kexec() runs, snp_kexec_finish() (called from
native_machine_shutdown()) has already converted the per-CPU GHCBs and
the boot GHCB back to private and set boot_ghcb to NULL, tearing down the
GHCB infrastructure. The DR7 write therefore raises a #VC that
dereferences a NULL GHCB pointer, causing a page fault and the kexec to
fail.

The NULL boot_ghcb pointer is a result of commit 3645eb7e3915 ("x86/fred:
Fix early boot failures on SEV-ES/SNP guests"): that change makes
__sev_get_ghcb() return the now-NULL boot_ghcb in this window, turning
the previously-benign teardown-window #VC (a bogus VMGEXIT the host
rejected but survived) into a NULL dereference.

Skip the DR7 write in this case. Its value is not needed across kexec
because the new kernel re-initializes DR7. Guests with DebugSwap enabled
context-switch DR7 in hardware and do not intercept the write, and
non-encrypted guests are unaffected, so both continue to call
hw_breakpoint_disable() as before.

Add a CC_ATTR_GUEST_DEBUG_VIRT confidential-computing attribute, so the
debug-register-virtualization state (AMD DebugSwap) is queried through
cc_platform_has().

Reported-by: Srikanth Aithal <[email protected]>
Suggested-by: Tom Lendacky <[email protected]>
Fixes: 3074152e56c9 ("x86/sev: Convert shared memory back to private on kexec")
Cc: [email protected]
Signed-off-by: Ashish Kalra <[email protected]>
Reviewed-by: Nikunj A Dadhania <[email protected]>
Reviewed-by: Tom Lendacky <[email protected]>
Tested-by: Srikanth Aithal <[email protected]>
---
 arch/x86/coco/core.c               |  3 +++
 arch/x86/kernel/machine_kexec_64.c | 11 ++++++++++-
 include/linux/cc_platform.h        |  8 ++++++++
 3 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/arch/x86/coco/core.c b/arch/x86/coco/core.c
index 989ca9f72ba3..d4c90a2da2bf 100644
--- a/arch/x86/coco/core.c
+++ b/arch/x86/coco/core.c
@@ -87,6 +87,9 @@ static bool noinstr amd_cc_platform_has(enum cc_attr attr)
 	case CC_ATTR_GUEST_STATE_ENCRYPT:
 		return sev_status & MSR_AMD64_SEV_ES_ENABLED;
 
+	case CC_ATTR_GUEST_DEBUG_VIRT:
+		return sev_status & MSR_AMD64_SNP_DEBUG_SWAP;
+
 	/*
 	 * With SEV, the rep string I/O instructions need to be unrolled
 	 * but SEV-ES supports them through the #VC handler.
diff --git a/arch/x86/kernel/machine_kexec_64.c b/arch/x86/kernel/machine_kexec_64.c
index c3f4a389992d..60d25ceada03 100644
--- a/arch/x86/kernel/machine_kexec_64.c
+++ b/arch/x86/kernel/machine_kexec_64.c
@@ -398,7 +398,16 @@ void __nocfi machine_kexec(struct kimage *image)
 
 	/* Interrupts aren't acceptable while we reboot */
 	local_irq_disable();
-	hw_breakpoint_disable();
+
+	/*
+	 * On SEV-ES/SEV-SNP guests without debug virtualization enabled, DR7
+	 * writes are intercepted and generate a #VC. The GHCBs have already
+	 * been torn down at this point so the #VC cannot be handled. Skip the
+	 * DR7 write as the new kernel re-initializes DR7 during boot.
+	 */
+	if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT) ||
+	    cc_platform_has(CC_ATTR_GUEST_DEBUG_VIRT))
+		hw_breakpoint_disable();
 	cet_disable();
 
 	if (image->preserve_context) {
diff --git a/include/linux/cc_platform.h b/include/linux/cc_platform.h
index 559353ad64ac..24ba1000ac94 100644
--- a/include/linux/cc_platform.h
+++ b/include/linux/cc_platform.h
@@ -62,6 +62,14 @@ enum cc_attr {
 	 */
 	CC_ATTR_GUEST_STATE_ENCRYPT,
 
+	/**
+	 * @CC_ATTR_GUEST_DEBUG_VIRT: Guest debug register virtualization is active.
+	 *
+	 * The platform/OS is running as a guest/virtual machine and is
+	 * virtualizing the debug register state.
+	 */
+	CC_ATTR_GUEST_DEBUG_VIRT,
+
 	/**
 	 * @CC_ATTR_GUEST_UNROLL_STRING_IO: String I/O is implemented with
 	 *                                  IN/OUT instructions
-- 
2.43.0