[RFC,v2.1] x86/sev-es: Handle NMI State

Commit Message

Joerg Roedel March 20, 2020, 1:17 p.m. UTC

On Thu, Mar 19, 2020 at 08:35:59AM -0700, Andy Lutomirski wrote:
> 1. Just put the NMI unmask in do_nmi().  The kernel *already* knows
> how to handle running do_nmi() with NMIs unmasked.  This is much, much
> simpler than your code.

Okay, attached is the updated patch which implements this approach. I
tested it in an SEV-ES guest with 'perf top' running for a little more
than 30 minutes and all looked good. I also removed the dead code from
the patch.


From ec3b021c5d9130fd66e00d823c4fabc675c4b49e Mon Sep 17 00:00:00 2001
From: Joerg Roedel <[e-mail]>
Date: Tue, 28 Jan 2020 17:31:05 +0100
Subject: [PATCH] x86/sev-es: Handle NMI State

When running under SEV-ES the kernel has to tell the hypervisor when to
open the NMI window again after an NMI was injected. This is done with
an NMI-complete message to the hypervisor.

Add code to the kernels NMI handler to send this message right at the
beginning of do_nmi(). This always allows nesting NMIs.

Signed-off-by: Joerg Roedel <[e-mail]>
---
 arch/x86/include/asm/sev-es.h   |  2 ++
 arch/x86/include/uapi/asm/svm.h |  1 +
 arch/x86/kernel/nmi.c           |  8 ++++++++
 arch/x86/kernel/sev-es.c        | 18 ++++++++++++++++++
 4 files changed, 29 insertions(+)

Comments

Dave Hansen March 20, 2020, 2:42 p.m. UTC | #1

On 3/20/20 6:17 AM, Joerg Roedel wrote:
> On Thu, Mar 19, 2020 at 08:35:59AM -0700, Andy Lutomirski wrote:
>> 1. Just put the NMI unmask in do_nmi().  The kernel *already* knows
>> how to handle running do_nmi() with NMIs unmasked.  This is much, much
>> simpler than your code.
> Okay, attached is the updated patch which implements this approach. I
> tested it in an SEV-ES guest with 'perf top' running for a little more
> than 30 minutes and all looked good. I also removed the dead code from
> the patch.

FWIW, perf plus the x86 selftests run in a big loop was my best way of
stressing the NMI path when we mucked with it for PTI.  The selftests
make sure to hit some of the more rare entry/exit paths.

Joerg Roedel March 20, 2020, 7:42 p.m. UTC | #2

On Fri, Mar 20, 2020 at 07:42:09AM -0700, Dave Hansen wrote:
> FWIW, perf plus the x86 selftests run in a big loop was my best way of
> stressing the NMI path when we mucked with it for PTI.  The selftests
> make sure to hit some of the more rare entry/exit paths.

Yeah, I ran the x86 selftests in an SEV-ES guest on-top of these
patches, that works. But doing this together with 'perf top' is also on
the list of tests to do.

Regards,

	Joerg

Patch

diff --git a/arch/x86/include/asm/sev-es.h b/arch/x86/include/asm/sev-es.h
index 63acf50e6280..441ec1ba2cc7 100644
--- a/arch/x86/include/asm/sev-es.h
+++ b/arch/x86/include/asm/sev-es.h
@@ -82,11 +82,13 @@  struct real_mode_header;
 
 #ifdef CONFIG_AMD_MEM_ENCRYPT
 int sev_es_setup_ap_jump_table(struct real_mode_header *rmh);
+void sev_es_nmi_complete(void);
 #else /* CONFIG_AMD_MEM_ENCRYPT */
 static inline int sev_es_setup_ap_jump_table(struct real_mode_header *rmh)
 {
 	return 0;
 }
+static inline void sev_es_nmi_complete(void) { }
 #endif /* CONFIG_AMD_MEM_ENCRYPT*/
 
 #endif
diff --git a/arch/x86/include/uapi/asm/svm.h b/arch/x86/include/uapi/asm/svm.h
index 20a05839dd9a..0f837339db66 100644
--- a/arch/x86/include/uapi/asm/svm.h
+++ b/arch/x86/include/uapi/asm/svm.h
@@ -84,6 +84,7 @@ 
 /* SEV-ES software-defined VMGEXIT events */
 #define SVM_VMGEXIT_MMIO_READ			0x80000001
 #define SVM_VMGEXIT_MMIO_WRITE			0x80000002
+#define SVM_VMGEXIT_NMI_COMPLETE		0x80000003
 #define SVM_VMGEXIT_AP_HLT_LOOP			0x80000004
 #define SVM_VMGEXIT_AP_JUMP_TABLE		0x80000005
 #define		SVM_VMGEXIT_SET_AP_JUMP_TABLE			0
diff --git a/arch/x86/kernel/nmi.c b/arch/x86/kernel/nmi.c
index 54c21d6abd5a..fc872a7e0ed1 100644
--- a/arch/x86/kernel/nmi.c
+++ b/arch/x86/kernel/nmi.c
@@ -37,6 +37,7 @@ 
 #include <asm/reboot.h>
 #include <asm/cache.h>
 #include <asm/nospec-branch.h>
+#include <asm/sev-es.h>
 
 #define CREATE_TRACE_POINTS
 #include <trace/events/nmi.h>
@@ -510,6 +511,13 @@  NOKPROBE_SYMBOL(is_debug_stack);
 dotraplinkage notrace void
 do_nmi(struct pt_regs *regs, long error_code)
 {
+	/*
+	 * Re-enable NMIs right here when running as an SEV-ES guest. This might
+	 * cause nested NMIs, but those can be handled safely.
+	 */
+	if (sev_es_active())
+		sev_es_nmi_complete();
+
 	if (IS_ENABLED(CONFIG_SMP) && cpu_is_offline(smp_processor_id()))
 		return;
 
diff --git a/arch/x86/kernel/sev-es.c b/arch/x86/kernel/sev-es.c
index 3c22f256645e..a7e2739771e7 100644
--- a/arch/x86/kernel/sev-es.c
+++ b/arch/x86/kernel/sev-es.c
@@ -270,6 +270,24 @@  static phys_addr_t vc_slow_virt_to_phys(struct ghcb *ghcb, long vaddr)
 /* Include code shared with pre-decompression boot stage */
 #include "sev-es-shared.c"
 
+void sev_es_nmi_complete(void)
+{
+	struct ghcb_state state;
+	struct ghcb *ghcb;
+
+	ghcb = sev_es_get_ghcb(&state);
+
+	vc_ghcb_invalidate(ghcb);
+	ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_NMI_COMPLETE);
+	ghcb_set_sw_exit_info_1(ghcb, 0);
+	ghcb_set_sw_exit_info_2(ghcb, 0);
+
+	sev_es_wr_ghcb_msr(__pa(ghcb));
+	VMGEXIT();
+
+	sev_es_put_ghcb(&state);
+}
+
 static u64 sev_es_get_jump_table_addr(void)
 {
 	struct ghcb_state state;