Skip to contentSkip to conversation

x86/sev: KEXEC/KDUMP support for SEV-ES guests

PATCH In review Joerg Roedel 5 participants 17 messages Jan 27, 2022 – Jun 4, 2023

Subsystems (6)
Stable branch x86 KVM virtio core +2 more subsystems
lore
Share & tools
lore
#1 lore raw

PATCH v3 1/10 x86/kexec/64: Disable kexec when SEV-ES is active

In reply to Joerg Roedel

From: Joerg Roedel <[email protected]>

SEV-ES needs special handling to support kexec. Disable it when SEV-ES is active until support is implemented.

---
 arch/x86/kernel/machine_kexec_64.c | 11 +++++++++++
 1 file changed, 11 insertions(+)
arch/x86/kernel/machine_kexec_64.c +11 -0
index f5da4a18070a..5079a75f8944 100644
@@ -269,11 +269,22 @@ static void load_segments(void)
 		);
 }
 
+static bool machine_kexec_supported(void)
+{
+	if (cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
+		return false;
+
+	return true;
+}
+
 int machine_kexec_prepare(struct kimage *image)
 {
 	unsigned long start_pgtable;
 	int result;
 
+	if (!machine_kexec_supported())
+		return -ENOSYS;
+
 	/* Calculate the offsets */
 	start_pgtable = page_to_pfn(image->control_code_page) << PAGE_SHIFT;
 
Signature · 1 line
2.34.1
#0 lore raw

v3 cover x86/sev: KEXEC/KDUMP support for SEV-ES guests

From: Joerg Roedel <[email protected]>

Hi,

here are changes to enable kexec/kdump in SEV-ES guests. The biggest problem for supporting kexec/kdump under SEV-ES is to find a way to hand the non-boot CPUs (APs) from one kernel to another.

Without SEV-ES the first kernel parks the CPUs in a HLT loop until they get reset by the kexec'ed kernel via an INIT-SIPI-SIPI sequence. For virtual machines the CPU reset is emulated by the hypervisor, which sets the vCPU registers back to reset state.

This does not work under SEV-ES, because the hypervisor has no access to the vCPU registers and can't make modifications to them. So an SEV-ES guest needs to reset the vCPU itself and park it using the AP-reset-hold protocol. Upon wakeup the guest needs to jump to real-mode and to the reset-vector configured in the AP-Jump-Table.

The code to do this is the main part of this patch-set. It works by placing code on the AP Jump-Table page itself to park the vCPU and for jumping to the reset vector upon wakeup. The code on the AP Jump Table runs in 16-bit protected mode with segment base set to the beginning of the page. The AP Jump-Table is usually not within the first 1MB of memory, so the code can't run in real-mode.

The AP Jump-Table is the best place to put the parking code, because the memory is owned, but read-only by the firmware and writeable by the OS. Only the first 4 bytes are used for the reset-vector, leaving the rest of the page for code/data/stack to park a vCPU. The code can't be in kernel memory because by the time the vCPU wakes up the memory will be owned by the new kernel, which might have overwritten it already.

The other patches add initial GHCB Version 2 protocol support, because kexec/kdump need the MSR-based (without a GHCB) AP-reset-hold VMGEXIT, which is a GHCB protocol version 2 feature.

The kexec'ed kernel is also entered via the decompressor and needs MMIO support there, so this patch-set also adds MMIO #VC support to the decompressor and support for handling CLFLUSH instructions.

Finally there is also code to disable kexec/kdump support at runtime when the environment does not support it (e.g. no GHCB protocol version 2 support or AP Jump Table over 4GB).

The diffstat looks big, but most of it is moving code for MMIO #VC support around to make it available to the decompressor.

The previous version of this patch-set can be found here:

	https://lore.kernel.org/lkml/[email protected]/

Please review.

Thanks,

Joerg

Changes v2->v3:

- Rebased to v5.17-rc1
- Applied most review comments by Boris
- Use the name 'AP jump table' consistently
- Make kexec-disabling for unsupported guests x86-specific
- Cleanup and consolidate patches to detect GHCB v2 protocol support

Joerg Roedel (10):
x86/kexec/64: Disable kexec when SEV-ES is active x86/sev: Save and print negotiated GHCB protocol version x86/sev: Set GHCB data structure version
x86/sev: Cache AP Jump Table Address
x86/sev: Setup code to park APs in the AP Jump Table x86/sev: Park APs on AP Jump Table with GHCB protocol version 2 x86/sev: Use AP Jump Table blob to stop CPU
x86/sev: Add MMIO handling support to boot/compressed/ code x86/sev: Handle CLFLUSH MMIO events
x86/kexec/64: Support kexec under SEV-ES with AP Jump Table Blob

 arch/x86/boot/compressed/sev.c          |  45 +-
 arch/x86/include/asm/insn-eval.h        |   1 +
 arch/x86/include/asm/realmode.h         |   5 +
 arch/x86/include/asm/sev-ap-jumptable.h |  29 +
 arch/x86/include/asm/sev.h              |  11 +-
 arch/x86/kernel/machine_kexec_64.c      |  12 +
 arch/x86/kernel/process.c               |   8 +
 arch/x86/kernel/sev-shared.c            | 233 +++++-
 arch/x86/kernel/sev.c                   | 404 +++++------
 arch/x86/lib/insn-eval-shared.c         | 913 ++++++++++++++++++++++++
 arch/x86/lib/insn-eval.c                | 909 +----------------------
 arch/x86/realmode/Makefile              |   9 +-
 arch/x86/realmode/rm/Makefile           |  11 +-
 arch/x86/realmode/rm/header.S           |   3 +
 arch/x86/realmode/rm/sev.S              |  85 +++
 arch/x86/realmode/rmpiggy.S             |   6 +
 arch/x86/realmode/sev/Makefile          |  33 +
 arch/x86/realmode/sev/ap_jump_table.S   | 131 ++++
 arch/x86/realmode/sev/ap_jump_table.lds |  24 +
 19 files changed, 1730 insertions(+), 1142 deletions(-)
 create mode 100644 arch/x86/include/asm/sev-ap-jumptable.h
 create mode 100644 arch/x86/lib/insn-eval-shared.c
 create mode 100644 arch/x86/realmode/rm/sev.S
 create mode 100644 arch/x86/realmode/sev/Makefile
 create mode 100644 arch/x86/realmode/sev/ap_jump_table.S
 create mode 100644 arch/x86/realmode/sev/ap_jump_table.lds

base-commit: e783362eb54cd99b2cac8b3a9aeac942e6f6ac07

Signature · 1 line
2.34.1
#2 lore raw

PATCH v3 2/10 x86/sev: Save and print negotiated GHCB protocol version

From: Joerg Roedel <[email protected]>

Save the results of the GHCB protocol negotiation into a data structure and print information about versions supported and used to the kernel log.

This is useful for debugging kexec issues in SEV-ES guests down the road to quickly spot whether kexec is supported on the given host.

---
 arch/x86/include/asm/sev.h   |  4 ++--
 arch/x86/kernel/sev-shared.c | 36 ++++++++++++++++++++++++++++++++++--
 arch/x86/kernel/sev.c        |  8 ++++++++
 3 files changed, 44 insertions(+), 4 deletions(-)
arch/x86/include/asm/sev.h +2 -2
index ec060c433589..17b75f6ee11a 100644
@@ -12,8 +12,8 @@
 #include <asm/insn.h>
 #include <asm/sev-common.h>
 
-#define GHCB_PROTO_OUR		0x0001UL
-#define GHCB_PROTOCOL_MAX	1ULL
+#define GHCB_PROTOCOL_MIN	1ULL
+#define GHCB_PROTOCOL_MAX	2ULL
 #define GHCB_DEFAULT_USAGE	0ULL
 
 #define	VMGEXIT()			{ asm volatile("rep; vmmcall\n\r"); }
arch/x86/kernel/sev-shared.c +34 -2
index ce987688bbc0..60ca7dd64d64 100644
@@ -14,6 +14,23 @@
 #define has_cpuflag(f)	boot_cpu_has(f)
 #endif
 
+/*
+ * struct ghcb_info - Used to return GHCB protocol
+ *				   negotiation details.
+ *
+ * @hv_proto_min:	Minimum GHCB protocol version supported by Hypervisor
+ * @hv_proto_max:	Maximum GHCB protocol version supported by Hypervisor
+ * @vm_proto:		Protocol version the VM (this kernel) will use
+ */
+struct ghcb_info {
+	unsigned int hv_proto_min;
+	unsigned int hv_proto_max;
+	unsigned int vm_proto;
+};
+
+/* Negotiated GHCB protocol version */
+static struct ghcb_info ghcb_info __ro_after_init;
+
 static bool __init sev_es_check_cpu_features(void)
 {
 	if (!has_cpuflag(X86_FEATURE_RDRAND)) {
@@ -44,6 +61,7 @@ static void __noreturn sev_es_terminate(unsigned int reason)
 
 static bool sev_es_negotiate_protocol(void)
 {
+	unsigned int protocol;
 	u64 val;
 
 	/* Do the GHCB protocol version negotiation */
@@ -54,10 +72,24 @@ static bool sev_es_negotiate_protocol(void)
 	if (GHCB_MSR_INFO(val) != GHCB_MSR_SEV_INFO_RESP)
 		return false;
 
-	if (GHCB_MSR_PROTO_MAX(val) < GHCB_PROTO_OUR ||
-	    GHCB_MSR_PROTO_MIN(val) > GHCB_PROTO_OUR)
+	/* Sanity check untrusted input */
+	if (GHCB_MSR_PROTO_MIN(val) > GHCB_MSR_PROTO_MAX(val))
 		return false;
 
+	/* Use maximum supported protocol version */
+	protocol = min_t(unsigned int, GHCB_MSR_PROTO_MAX(val), GHCB_PROTOCOL_MAX);
+
+	/*
+	 * Hypervisor does not support any protocol version required for this
+	 * kernel.
+	 */
+	if (protocol < GHCB_MSR_PROTO_MIN(val))
+		return false;
+
+	ghcb_info.hv_proto_min = GHCB_MSR_PROTO_MIN(val);
+	ghcb_info.hv_proto_max = GHCB_MSR_PROTO_MAX(val);
+	ghcb_info.vm_proto     = protocol;
+
 	return true;
 }
 
arch/x86/kernel/sev.c +8 -0
index e6d316a01fdd..8a4317fa699a 100644
@@ -779,6 +779,14 @@ void __init sev_es_init_vc_handling(void)
 
 	/* Secondary CPUs use the runtime #VC handler */
 	initial_vc_handler = (unsigned long)kernel_exc_vmm_communication;
+
+	/*
+	 * Print information about supported and negotiated GHCB protocol
+	 * versions.
+	 */
+	pr_info("Hypervisor GHCB protocol version support: min=%u max=%u\n",
+		ghcb_info.hv_proto_min, ghcb_info.hv_proto_max);
+	pr_info("Using GHCB protocol version %u\n", ghcb_info.vm_proto);
 }
 
 static void __init vc_early_forward_exception(struct es_em_ctxt *ctxt)
Signature · 1 line
2.34.1
#3 lore raw

PATCH v3 3/10 x86/sev: Set GHCB data structure version

In reply to Joerg Roedel

From: Joerg Roedel <[email protected]>

It turned out that the GHCB->protocol field does not declare the version of the guest-hypervisor communication protocol, but rather the version of the GHCB data structure. Reflect that in the define used to set the protocol field.

---
 arch/x86/include/asm/sev.h   | 3 +++
 arch/x86/kernel/sev-shared.c | 2 +-
 2 files changed, 4 insertions(+), 1 deletion(-)
arch/x86/include/asm/sev.h +3 -0
index 17b75f6ee11a..09eb2cc6f54f 100644
@@ -16,6 +16,9 @@
 #define GHCB_PROTOCOL_MAX	2ULL
 #define GHCB_DEFAULT_USAGE	0ULL
 
+/* Version of the GHCB data structure */
+#define GHCB_VERSION		1
+
 #define	VMGEXIT()			{ asm volatile("rep; vmmcall\n\r"); }
 
 enum es_result {
arch/x86/kernel/sev-shared.c +1 -1
index 60ca7dd64d64..4468150a42bb 100644
@@ -162,7 +162,7 @@ enum es_result sev_es_ghcb_hv_call(struct ghcb *ghcb, bool set_ghcb_msr,
 				   u64 exit_info_1, u64 exit_info_2)
 {
 	/* Fill in protocol and format specifiers */
-	ghcb->protocol_version = GHCB_PROTOCOL_MAX;
+	ghcb->protocol_version = GHCB_VERSION;
 	ghcb->ghcb_usage       = GHCB_DEFAULT_USAGE;
 
 	ghcb_set_sw_exit_code(ghcb, exit_code);
Signature · 1 line
2.34.1
#4 lore raw

PATCH v3 4/10 x86/sev: Cache AP Jump Table Address

In reply to Joerg Roedel

From: Joerg Roedel <[email protected]>

Store the physical address of the AP jump table in kernel memory so that it does not need to be fetched from the Hypervisor again.

---
 arch/x86/kernel/sev.c | 28 +++++++++++++++-------------
 1 file changed, 15 insertions(+), 13 deletions(-)
arch/x86/kernel/sev.c +15 -13
index 8a4317fa699a..969ef9855bb5 100644
@@ -43,6 +43,9 @@ static struct ghcb boot_ghcb_page __bss_decrypted __aligned(PAGE_SIZE);
  */
 static struct ghcb __initdata *boot_ghcb;
 
+/* Cached AP jump table Address */
+static phys_addr_t jump_table_pa;
+
 /* #VC handler runtime per-CPU data */
 struct sev_es_runtime_data {
 	struct ghcb ghcb_page;
@@ -523,12 +526,14 @@ void noinstr __sev_es_nmi_complete(void)
 	__sev_put_ghcb(&state);
 }
 
-static u64 get_jump_table_addr(void)
+static phys_addr_t get_jump_table_addr(void)
 {
 	struct ghcb_state state;
 	unsigned long flags;
 	struct ghcb *ghcb;
-	u64 ret = 0;
+
+	if (jump_table_pa)
+		return jump_table_pa;
 
 	local_irq_save(flags);
 
@@ -544,39 +549,36 @@ static u64 get_jump_table_addr(void)
 
 	if (ghcb_sw_exit_info_1_is_valid(ghcb) &&
 	    ghcb_sw_exit_info_2_is_valid(ghcb))
-		ret = ghcb->save.sw_exit_info_2;
+		jump_table_pa = (phys_addr_t)ghcb->save.sw_exit_info_2;
 
 	__sev_put_ghcb(&state);
 
 	local_irq_restore(flags);
 
-	return ret;
+	return jump_table_pa;
 }
 
 int sev_es_setup_ap_jump_table(struct real_mode_header *rmh)
 {
 	u16 startup_cs, startup_ip;
-	phys_addr_t jump_table_pa;
-	u64 jump_table_addr;
 	u16 __iomem *jump_table;
+	phys_addr_t pa;
 
-	jump_table_addr = get_jump_table_addr();
+	pa = get_jump_table_addr();
 
 	/* On UP guests there is no jump table so this is not a failure */
-	if (!jump_table_addr)
+	if (!pa)
 		return 0;
 
-	/* Check if AP Jump Table is page-aligned */
-	if (jump_table_addr & ~PAGE_MASK)
+	/* Check if AP jump table is page-aligned */
+	if (pa & ~PAGE_MASK)
 		return -EINVAL;
 
-	jump_table_pa = jump_table_addr & PAGE_MASK;
-
 	startup_cs = (u16)(rmh->trampoline_start >> 4);
 	startup_ip = (u16)(rmh->sev_es_trampoline_start -
 			   rmh->trampoline_start);
 
-	jump_table = ioremap_encrypted(jump_table_pa, PAGE_SIZE);
+	jump_table = ioremap_encrypted(pa, PAGE_SIZE);
 	if (!jump_table)
 		return -EIO;
 
Signature · 1 line
2.34.1
#6 lore raw

PATCH v3 5/10 x86/sev: Setup code to park APs in the AP Jump Table

In reply to Joerg Roedel

From: Joerg Roedel <[email protected]>

The AP jump table under SEV-ES contains the reset vector where non-boot CPUs start executing when coming out of reset. This means that a CPU coming out of the AP-reset-hold VMGEXIT also needs to start executing at the reset vector stored in the AP jump table.

The problem is to find a safe place to put the real-mode code which executes the VMGEXIT and jumps to the reset vector. The code can not be in kernel memory, because after kexec that memory is owned by the new kernel and the code might have been overwritten.

Fortunately the AP jump table itself is a safe place, because the memory is not owned by the OS and will not be overwritten by a new kernel started through kexec. The table is 4k in size and only the first 4 bytes are used for the reset vector. This leaves enough space for some 16-bit code to do the job and even a small stack.

The AP jump table must be 4K in size, in encrypted memory and it must be 4K (page) aligned. There can only be one AP jump table and it should reside in memory that has been marked as reserved by UEFI.

Install 16-bit code into the AP jump table under SEV-ES after the APs have been brought up. The code will do an AP-reset-hold VMGEXIT and jump to the reset vector after being woken up.

---
 arch/x86/include/asm/realmode.h         |   2 +
 arch/x86/include/asm/sev-ap-jumptable.h |  29 ++++++
 arch/x86/kernel/sev.c                   | 104 +++++++++++++++++++
 arch/x86/realmode/Makefile              |   9 +-
 arch/x86/realmode/rmpiggy.S             |   6 ++
 arch/x86/realmode/sev/Makefile          |  33 ++++++
 arch/x86/realmode/sev/ap_jump_table.S   | 131 ++++++++++++++++++++++++
 arch/x86/realmode/sev/ap_jump_table.lds |  24 +++++
 8 files changed, 337 insertions(+), 1 deletion(-)
 create mode 100644 arch/x86/include/asm/sev-ap-jumptable.h
 create mode 100644 arch/x86/realmode/sev/Makefile
 create mode 100644 arch/x86/realmode/sev/ap_jump_table.S
 create mode 100644 arch/x86/realmode/sev/ap_jump_table.lds
arch/x86/include/asm/realmode.h +2 -0
index 331474b150f1..d17f495e86cd 100644
@@ -62,6 +62,8 @@ extern unsigned long initial_gs;
 extern unsigned long initial_stack;
 #ifdef CONFIG_AMD_MEM_ENCRYPT
 extern unsigned long initial_vc_handler;
+extern unsigned char rm_ap_jump_table_blob[];
+extern unsigned char rm_ap_jump_table_blob_end[];
 #endif
 
 extern unsigned char real_mode_blob[];
arch/x86/include/asm/sev-ap-jumptable.h new file +29 -0
new file mode 100644
index 000000000000..710547999dee
@@ -0,0 +1,29 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * AMD Encrypted Register State Support
+ *
+ * Author: Joerg Roedel <[email protected]>
+ */
+#ifndef __ASM_SEV_AP_JUMPTABLE_H
+#define __ASM_SEV_AP_JUMPTABLE_H
+
+#define	SEV_APJT_CS16	0x8
+#define	SEV_APJT_DS16	0x10
+
+#define SEV_APJT_ENTRY	0x10
+
+#ifndef __ASSEMBLY__
+
+/*
+ * The reset_ip and reset_cs members are fixed and defined through the GHCB
+ * specification. Do not change or move them around.
+ */
+struct sev_ap_jump_table_header {
+	u16	reset_ip;
+	u16	reset_cs;
+	u16	ap_jumptable_gdt;
+};
+
+#endif /* !__ASSEMBLY__ */
+
+#endif /* __ASM_SEV_AP_JUMPTABLE_H */
arch/x86/kernel/sev.c +104 -0

122 lines not on this page · raw message

arch/x86/realmode/Makefile +8 -1
index a0b491ae2de8..00f3cceb9580 100644
@@ -11,12 +11,19 @@
 KASAN_SANITIZE			:= n
 KCSAN_SANITIZE			:= n
 
+RMPIGGY-y				 = $(obj)/rm/realmode.bin
+RMPIGGY-$(CONFIG_AMD_MEM_ENCRYPT)	+= $(obj)/sev/ap_jump_table.bin
+
 subdir- := rm
+subdir- := sev
 
 obj-y += init.o
 obj-y += rmpiggy.o
 
-$(obj)/rmpiggy.o: $(obj)/rm/realmode.bin
+$(obj)/rmpiggy.o: $(RMPIGGY-y)
 
 $(obj)/rm/realmode.bin: FORCE
 	$(Q)$(MAKE) $(build)=$(obj)/rm $@
+
+$(obj)/sev/ap_jump_table.bin: FORCE
+	$(Q)$(MAKE) $(build)=$(obj)/sev $@
arch/x86/realmode/rmpiggy.S +6 -0
index c8fef76743f6..a659f98617ff 100644
@@ -17,3 +17,9 @@ SYM_DATA_END_LABEL(real_mode_blob, SYM_L_GLOBAL, real_mode_blob_end)
 SYM_DATA_START(real_mode_relocs)
 	.incbin	"arch/x86/realmode/rm/realmode.relocs"
 SYM_DATA_END(real_mode_relocs)
+
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+SYM_DATA_START(rm_ap_jump_table_blob)
+	.incbin "arch/x86/realmode/sev/ap_jump_table.bin"
+SYM_DATA_END_LABEL(rm_ap_jump_table_blob, SYM_L_GLOBAL, rm_ap_jump_table_blob_end)
+#endif
arch/x86/realmode/sev/Makefile new file +33 -0
new file mode 100644
index 000000000000..7cf5f31f6419
@@ -0,0 +1,33 @@
+# SPDX-License-Identifier: GPL-2.0
+
+# Sanitizer runtimes are unavailable and cannot be linked here.
+KASAN_SANITIZE			:= n
+KCSAN_SANITIZE			:= n
+OBJECT_FILES_NON_STANDARD	:= y
+
+# Prevents link failures: __sanitizer_cov_trace_pc() is not linked in.
+KCOV_INSTRUMENT		:= n
+
+always-y 	:= ap_jump_table.bin
+ap_jump_table-y	+= ap_jump_table.o
+targets		+= $(ap_jump_table-y)
+
+APJUMPTABLE_OBJS = $(addprefix $(obj)/,$(ap_jump_table-y))
+
+LDFLAGS_ap_jump_table.elf := -m elf_i386 -T
+
+targets 	+= ap_jump_table.elf
+$(obj)/ap_jump_table.elf: $(obj)/ap_jump_table.lds $(APJUMPTABLE_OBJS) FORCE
+	$(call if_changed,ld)
+
+OBJCOPYFLAGS_ap_jump_table.bin := -O binary
+
+targets 	+= ap_jump_table.bin
+$(obj)/ap_jump_table.bin: $(obj)/ap_jump_table.elf FORCE
+	$(call if_changed,objcopy)
+
+# ---------------------------------------------------------------------------
+
+KBUILD_AFLAGS	:= $(REALMODE_CFLAGS) -D__ASSEMBLY__
+GCOV_PROFILE := n
+UBSAN_SANITIZE := n
arch/x86/realmode/sev/ap_jump_table.S new file +131 -0

131 lines not on this page · raw message

arch/x86/realmode/sev/ap_jump_table.lds new file +24 -0
new file mode 100644
index 000000000000..4e47f1a6eb4e
@@ -0,0 +1,24 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * ap_jump_table.lds
+ *
+ * Linker script for the SEV-ES AP jump table code
+ */
+
+OUTPUT_FORMAT("elf32-i386")
+OUTPUT_ARCH(i386)
+ENTRY(ap_park)
+
+SECTIONS
+{
+	. = 0;
+	.text : {
+		*(.text)
+		*(.text.*)
+	}
+
+	/DISCARD/ : {
+		*(.note*)
+		*(.debug*)
+	}
+}
Signature · 1 line
2.34.1
#8 lore raw

PATCH v3 6/10 x86/sev: Park APs on AP Jump Table with GHCB protocol version 2

In reply to Joerg Roedel

From: Joerg Roedel <[email protected]>

GHCB protocol version 2 adds the MSR-based AP-reset-hold VMGEXIT which does not need a GHCB. Use that to park APs in 16-bit protected mode on the AP jump table.

---
 arch/x86/include/asm/realmode.h |  3 ++
 arch/x86/kernel/sev.c           | 51 ++++++++++++++++++--
 arch/x86/realmode/rm/Makefile   | 11 +++--
 arch/x86/realmode/rm/header.S   |  3 ++
 arch/x86/realmode/rm/sev.S      | 85 +++++++++++++++++++++++++++++++++
 5 files changed, 143 insertions(+), 10 deletions(-)
 create mode 100644 arch/x86/realmode/rm/sev.S
arch/x86/include/asm/realmode.h +3 -0
index d17f495e86cd..12f18782b0e0 100644
@@ -23,6 +23,9 @@ struct real_mode_header {
 	u32	trampoline_header;
 #ifdef CONFIG_AMD_MEM_ENCRYPT
 	u32	sev_es_trampoline_start;
+	u32	sev_ap_park;
+	u32	sev_ap_park_seg;
+	u32	sev_ap_park_gdt;
 #endif
 #ifdef CONFIG_X86_64
 	u32	trampoline_pgd;
arch/x86/kernel/sev.c +46 -5

82 lines not on this page · raw message

arch/x86/realmode/rm/Makefile +6 -5
index 83f1b6a56449..955610480ab8 100644
@@ -27,11 +27,12 @@ wakeup-objs	+= video-vga.o
 wakeup-objs	+= video-vesa.o
 wakeup-objs	+= video-bios.o
 
-realmode-y			+= header.o
-realmode-y			+= trampoline_$(BITS).o
-realmode-y			+= stack.o
-realmode-y			+= reboot.o
-realmode-$(CONFIG_ACPI_SLEEP)	+= $(wakeup-objs)
+realmode-y				+= header.o
+realmode-y				+= trampoline_$(BITS).o
+realmode-y				+= stack.o
+realmode-y				+= reboot.o
+realmode-$(CONFIG_ACPI_SLEEP)		+= $(wakeup-objs)
+realmode-$(CONFIG_AMD_MEM_ENCRYPT)	+= sev.o
 
 targets	+= $(realmode-y)
 
arch/x86/realmode/rm/header.S +3 -0
index 8c1db5bf5d78..6c17f8fd1eb4 100644
@@ -22,6 +22,9 @@ SYM_DATA_START(real_mode_header)
 	.long	pa_trampoline_header
 #ifdef CONFIG_AMD_MEM_ENCRYPT
 	.long	pa_sev_es_trampoline_start
+	.long	pa_sev_ap_park_asm
+	.long	__KERNEL32_CS
+	.long	pa_sev_ap_park_gdt;
 #endif
 #ifdef CONFIG_X86_64
 	.long	pa_trampoline_pgd;
arch/x86/realmode/rm/sev.S new file +85 -0

85 lines not on this page · raw message

Signature · 1 line
2.34.1
#9 lore raw

PATCH v3 7/10 x86/sev: Use AP Jump Table blob to stop CPU

In reply to Joerg Roedel

From: Joerg Roedel <[email protected]>

To support kexec under SEV-ES the APs can't be parked with HLT. Upon wakeup the AP needs to find its way to execute at the reset vector set by the new kernel and in real-mode.

This is what the AP jump table blob provides, so stop the APs the SEV-ES way by calling the AP-reset-hold VMGEXIT from the AP jump table.

---
 arch/x86/include/asm/sev.h |  2 ++
 arch/x86/kernel/process.c  |  8 ++++++++
 arch/x86/kernel/sev.c      | 14 +++++++++++++-
 3 files changed, 23 insertions(+), 1 deletion(-)
arch/x86/include/asm/sev.h +2 -0
index 09eb2cc6f54f..e342dce3e7a1 100644
@@ -90,12 +90,14 @@ extern enum es_result sev_es_ghcb_hv_call(struct ghcb *ghcb,
 					  struct es_em_ctxt *ctxt,
 					  u64 exit_code, u64 exit_info_1,
 					  u64 exit_info_2);
+void sev_es_stop_this_cpu(void);
 #else
 static inline void sev_es_ist_enter(struct pt_regs *regs) { }
 static inline void sev_es_ist_exit(void) { }
 static inline int sev_es_setup_ap_jump_table(struct real_mode_header *rmh) { return 0; }
 static inline void sev_es_nmi_complete(void) { }
 static inline int sev_es_efi_map_ghcbs(pgd_t *pgd) { return 0; }
+static inline void sev_es_stop_this_cpu(void) { }
 #endif
 
 #endif
arch/x86/kernel/process.c +8 -0
index 81d8ef036637..09982f0fbb87 100644
@@ -46,6 +46,7 @@
 #include <asm/proto.h>
 #include <asm/frame.h>
 #include <asm/unwind.h>
+#include <asm/sev.h>
 
 #include "process.h"
 
@@ -769,6 +770,13 @@ void stop_this_cpu(void *dummy)
 	if (boot_cpu_has(X86_FEATURE_SME))
 		native_wbinvd();
 	for (;;) {
+		/*
+		 * SEV-ES guests need a special stop routine to support
+		 * kexec. Try this first, if it fails the function will
+		 * return and native_halt() is used.
+		 */
+		sev_es_stop_this_cpu();
+
 		/*
 		 * Use native_halt() so that memory contents don't change
 		 * (stack usage and variables) after possibly issuing the
arch/x86/kernel/sev.c +13 -1
index fcff39475fbe..71301016c3ea 100644
@@ -673,7 +673,6 @@ static bool __init sev_es_setup_ghcb(void)
 	return true;
 }
 
-#ifdef CONFIG_HOTPLUG_CPU
 void __noreturn sev_jumptable_ap_park(void)
 {
 	local_irq_disable();
@@ -706,6 +705,19 @@ void __noreturn sev_jumptable_ap_park(void)
 }
 STACK_FRAME_NON_STANDARD(sev_jumptable_ap_park);
 
+void sev_es_stop_this_cpu(void)
+{
+	if (!static_branch_unlikely(&sev_es_enable_key))
+		return;
+
+	/* Only park in the AP jump table when the code has been installed */
+	if (!sev_ap_jumptable_blob_installed)
+		return;
+
+	sev_jumptable_ap_park();
+}
+
+#ifdef CONFIG_HOTPLUG_CPU
 static void sev_es_ap_hlt_loop(void)
 {
 	struct ghcb_state state;
Signature · 1 line
2.34.1
#10 lore raw

PATCH v3 8/10 x86/sev: Add MMIO handling support to boot/compressed/ code

In reply to Joerg Roedel

From: Joerg Roedel <[email protected]>

Move the code for MMIO handling in the #VC handler to sev-shared.c so that it can be used in the decompressor code. The decompressor needs to handle MMIO events for writing to the VGA framebuffer.

When the kernel is booted via UEFI the VGA console is not enabled that early, but a kexec boot will enable it and the decompressor needs MMIO support to write to the frame buffer.

This also requires to share some code from lib/insn-eval.c. Since insn-eval.c can't be included into the decompressor code directly, move the relevant parts into lib/insn-eval-shared.c and include that file.

---
 arch/x86/boot/compressed/sev.c  |  45 +-
 arch/x86/kernel/sev-shared.c    | 192 +++++++
 arch/x86/kernel/sev.c           | 192 -------
 arch/x86/lib/insn-eval-shared.c | 906 +++++++++++++++++++++++++++++++
 arch/x86/lib/insn-eval.c        | 909 +-------------------------------
 5 files changed, 1128 insertions(+), 1116 deletions(-)
 create mode 100644 arch/x86/lib/insn-eval-shared.c
arch/x86/boot/compressed/sev.c +26 -19
index 28bcf04c022e..f4c0af184b17 100644
@@ -26,25 +26,6 @@
 struct ghcb boot_ghcb_page __aligned(PAGE_SIZE);
 struct ghcb *boot_ghcb;
 
-/*
- * Copy a version of this function here - insn-eval.c can't be used in
- * pre-decompression code.
- */
-static bool insn_has_rep_prefix(struct insn *insn)
-{
-	insn_byte_t p;
-	int i;
-
-	insn_get_prefixes(insn);
-
-	for_each_insn_prefix(insn, i, p) {
-		if (p == 0xf2 || p == 0xf3)
-			return true;
-	}
-
-	return false;
-}
-
 /*
  * Only a dummy for insn_get_seg_base() - Early boot-code is 64bit only and
  * doesn't use segments.
@@ -54,6 +35,16 @@ static unsigned long insn_get_seg_base(struct pt_regs *regs, int seg_reg_idx)
 	return 0UL;
 }
 
+static int get_seg_base_limit(struct insn *insn, struct pt_regs *regs,
+			      int regoff, unsigned long *base,
+			      unsigned long *limit)
+{
+	if (base)
+		*base = 0ULL;
+	if (limit)
+		*limit = ~0ULL;
+}
+
 static inline u64 sev_es_rd_ghcb_msr(void)
 {
 	unsigned long low, high;
@@ -105,6 +96,14 @@ static enum es_result vc_read_mem(struct es_em_ctxt *ctxt,
 	return ES_OK;
 }
 
+static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt *ctxt,
+					   unsigned long vaddr, phys_addr_t *paddr)
+{
+	*paddr = (phys_addr_t)vaddr;
+
+	return ES_OK;
+}
+
 #undef __init
 #undef __pa
 #define __init
@@ -112,9 +111,14 @@ static enum es_result vc_read_mem(struct es_em_ctxt *ctxt,
 
 #define __BOOT_COMPRESSED
 
+#undef WARN_ONCE
+#define WARN_ONCE(condition, format...)
+
 /* Basic instruction decoding support needed */
+#include <asm/insn-eval.h>
 #include "../../lib/inat.c"
 #include "../../lib/insn.c"
+#include "../../lib/insn-eval-shared.c"
 
 /* Include code for early handlers */
 #include "../../kernel/sev-shared.c"
@@ -193,6 +197,9 @@ void do_boot_stage2_vc(struct pt_regs *regs, unsigned long exit_code)
 	case SVM_EXIT_CPUID:
 		result = vc_handle_cpuid(boot_ghcb, &ctxt);
 		break;
+	case SVM_EXIT_NPF:
+		result = vc_handle_mmio(boot_ghcb, &ctxt);
+		break;
 	default:
 		result = ES_UNSUPPORTED;
 		break;
arch/x86/kernel/sev-shared.c +192 -0

195 lines not on this page · raw message

arch/x86/kernel/sev.c +0 -192

198 lines not on this page · raw message

arch/x86/lib/insn-eval-shared.c new file +906 -0

906 lines not on this page · raw message

arch/x86/lib/insn-eval.c +4 -905

937 lines not on this page · raw message

Signature · 1 line
2.34.1
#11 lore raw

PATCH v3 9/10 x86/sev: Handle CLFLUSH MMIO events

In reply to Joerg Roedel

From: Joerg Roedel <[email protected]>

Handle CLFLUSH instruction to MMIO memory in the #VC handler. The instruction is ignored by the handler, as the Hypervisor is responsible for cache management of emulated MMIO memory.

---
 arch/x86/include/asm/insn-eval.h | 1 +
 arch/x86/kernel/sev-shared.c     | 3 +++
 arch/x86/lib/insn-eval-shared.c  | 7 +++++++
 3 files changed, 11 insertions(+)
arch/x86/include/asm/insn-eval.h +1 -0
index f07faa61c7f3..c3eb753a912b 100644
@@ -40,6 +40,7 @@ enum mmio_type {
 	MMIO_READ_ZERO_EXTEND,
 	MMIO_READ_SIGN_EXTEND,
 	MMIO_MOVS,
+	MMIO_IGNORE,
 };
 
 enum mmio_type insn_decode_mmio(struct insn *insn, int *bytes);
arch/x86/kernel/sev-shared.c +3 -0
index b12fb063a30e..1aa33509c7b5 100644
@@ -698,6 +698,9 @@ static enum es_result vc_handle_mmio(struct ghcb *ghcb, struct es_em_ctxt *ctxt)
 	if (mmio == MMIO_DECODE_FAILED)
 		return ES_DECODE_FAILED;
 
+	if (mmio == MMIO_IGNORE)
+		return ES_OK;
+
 	if (mmio != MMIO_WRITE_IMM && mmio != MMIO_MOVS) {
 		reg_data = insn_get_modrm_reg_ptr(insn, ctxt->regs);
 		if (!reg_data)
arch/x86/lib/insn-eval-shared.c +7 -0
index ec310b5e6cd5..ddec72fccdd2 100644
@@ -898,6 +898,13 @@ enum mmio_type insn_decode_mmio(struct insn *insn, int *bytes)
 				*bytes = 2;
 			type = MMIO_READ_SIGN_EXTEND;
 			break;
+		case 0xae: /* CLFLUSH */
+			/*
+			 * Ignore CLFLUSHes - those go to emulated MMIO anyway and the
+			 * hypervisor is responsible for cache management.
+			 */
+			type = MMIO_IGNORE;
+			break;
 		}
 		break;
 	}
Signature · 1 line
2.34.1
#12 lore raw

PATCH v3 10/10 x86/kexec/64: Support kexec under SEV-ES with AP Jump Table Blob

In reply to Joerg Roedel

From: Joerg Roedel <[email protected]>

When the AP jump table blob is installed the kernel can hand over the APs from the old to the new kernel. Enable kexec when the AP jump table blob has been installed.

---
 arch/x86/include/asm/sev.h         |  2 ++
 arch/x86/kernel/machine_kexec_64.c |  3 ++-
 arch/x86/kernel/sev.c              | 15 +++++++++++++++
 3 files changed, 19 insertions(+), 1 deletion(-)
arch/x86/include/asm/sev.h +2 -0
index e342dce3e7a1..41e07d037b6e 100644
@@ -91,6 +91,7 @@ extern enum es_result sev_es_ghcb_hv_call(struct ghcb *ghcb,
 					  u64 exit_code, u64 exit_info_1,
 					  u64 exit_info_2);
 void sev_es_stop_this_cpu(void);
+bool sev_kexec_supported(void);
 #else
 static inline void sev_es_ist_enter(struct pt_regs *regs) { }
 static inline void sev_es_ist_exit(void) { }
@@ -98,6 +99,7 @@ static inline int sev_es_setup_ap_jump_table(struct real_mode_header *rmh) { ret
 static inline void sev_es_nmi_complete(void) { }
 static inline int sev_es_efi_map_ghcbs(pgd_t *pgd) { return 0; }
 static inline void sev_es_stop_this_cpu(void) { }
+static inline bool sev_kexec_supported(void) { return true; }
 #endif
 
 #endif
arch/x86/kernel/machine_kexec_64.c +2 -1
index 5079a75f8944..c58808fe3fb5 100644
@@ -27,6 +27,7 @@
 #include <asm/kexec-bzimage64.h>
 #include <asm/setup.h>
 #include <asm/set_memory.h>
+#include <asm/sev.h>
 
 #ifdef CONFIG_ACPI
 /*
@@ -271,7 +272,7 @@ static void load_segments(void)
 
 static bool machine_kexec_supported(void)
 {
-	if (cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
+	if (!sev_kexec_supported())
 		return false;
 
 	return true;
arch/x86/kernel/sev.c +15 -0
index 1bced5b49150..17dcbcddd6ab 100644
@@ -884,6 +884,21 @@ static int __init sev_setup_ap_jump_table(void)
 }
 core_initcall(sev_setup_ap_jump_table);
 
+bool sev_kexec_supported(void)
+{
+	if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
+		return true;
+
+	/*
+	 * KEXEC with SEV-ES and more than one CPU is only supported
+	 * when the AP jump table is installed.
+	 */
+	if (num_possible_cpus() > 1)
+		return sev_ap_jumptable_blob_installed;
+	else
+		return true;
+}
+
 static void __init alloc_runtime_data(int cpu)
 {
 	struct sev_es_runtime_data *data;
Signature · 1 line
2.34.1
#5 lore raw

Re: [PATCH v3 04/10] x86/sev: Cache AP Jump Table Address

In reply to Joerg Roedel
Quoted 4 lines · Joerg Roedel

From: Joerg Roedel <[email protected]>

Store the physical address of the AP jump table in kernel memory so that it does not need to be fetched from the Hypervisor again.

This doesn't explain why the kernel would retrieve the jump table more than once, e.g. at this point in the series, this can only ever be called once.

Quoted 20 lines
---
 arch/x86/kernel/sev.c | 28 +++++++++++++++-------------
 1 file changed, 15 insertions(+), 13 deletions(-)
arch/x86/kernel/sev.c +3 -0
index 8a4317fa699a..969ef9855bb5 100644
@@ -43,6 +43,9 @@ static struct ghcb boot_ghcb_page __bss_decrypted __aligned(PAGE_SIZE);
  */
 static struct ghcb __initdata *boot_ghcb;
 
+/* Cached AP jump table Address */
+static phys_addr_t jump_table_pa;
+
 /* #VC handler runtime per-CPU data */
 struct sev_es_runtime_data {
 	struct ghcb ghcb_page;

arch/x86/kernel/sev.c void noinstr __sev_es_nmi_complete(void) · line 529

Quoted 5 lines
arch/x86/kernel/sev.c +1 -1
@@ -523,12 +526,14 @@ void noinstr __sev_es_nmi_complete(void)
 	__sev_put_ghcb(&state);
 }
 
-static u64 get_jump_table_addr(void)
+static phys_addr_t get_jump_table_addr(void)

Not new, but I believe this can be tagged __init.

Quoted 17 lines
arch/x86/kernel/sev.c +4 -2
 {
 	struct ghcb_state state;
 	unsigned long flags;
 	struct ghcb *ghcb;
-	u64 ret = 0;
+
+	if (jump_table_pa)
+		return jump_table_pa;
 
 	local_irq_save(flags);
 
@@ -544,39 +549,36 @@ static u64 get_jump_table_addr(void)
 
 	if (ghcb_sw_exit_info_1_is_valid(ghcb) &&
 	    ghcb_sw_exit_info_2_is_valid(ghcb))
-		ret = ghcb->save.sw_exit_info_2;
+		jump_table_pa = (phys_addr_t)ghcb->save.sw_exit_info_2;

arch/x86/kernel/sev.c static u64 get_jump_table_addr(void) · line 561

Quoted 10 lines
arch/x86/kernel/sev.c +1 -1
 
 	__sev_put_ghcb(&state);
 
 	local_irq_restore(flags);
 
-	return ret;
+	return jump_table_pa;
 }
 
 int sev_es_setup_ap_jump_table(struct real_mode_header *rmh)

__init here too.

arch/x86/kernel/sev.c static u64 get_jump_table_addr(void) · line 569

Quoted 11 lines
arch/x86/kernel/sev.c +2 -3
 {
 	u16 startup_cs, startup_ip;
-	phys_addr_t jump_table_pa;
-	u64 jump_table_addr;
 	u16 __iomem *jump_table;
+	phys_addr_t pa;
 
-	jump_table_addr = get_jump_table_addr();
+	pa = get_jump_table_addr();
 
 	/* On UP guests there is no jump table so this is not a failure */

Does anything actually check that the jump table is valid for SMP guests?

arch/x86/kernel/sev.c static u64 get_jump_table_addr(void) · line 570

Quoted 2 lines
arch/x86/kernel/sev.c +1 -1
-	if (!jump_table_addr)
+	if (!pa)

Using '0' for "not valid" is funky because '0' isn't technically an illegal GPA, and because it means the address (or lack thereof) isn't cached on a single-vCPU guest.

#7 lore raw

Re: [PATCH v3 05/10] x86/sev: Setup code to park APs in the AP Jump Table

In reply to Joerg Roedel
Quoted 1 line · Joerg Roedel

+static int __init sev_setup_ap_jump_table(void)

This name is really confusing. AFAICT, it's specific to SEV-ES, but used only "sev" for the namespace because sev_es_setup_ap_jump_table() already exists. I assume this variant is purely for parking/offlining vCPUs? Adding that in the name would be helpful.

The two flows are also very, very similar, but apparently do slightly different things. Even more odd is that this version applies different sanity checks on the address than the existing code. It should be fairly simple to extract a common helper. That would likely help with naming problem too.

Quoted 56 lines
arch/x86/kernel/sev.c +56 -0
+{
+	size_t blob_size = rm_ap_jump_table_blob_end - rm_ap_jump_table_blob;
+	u16 startup_cs, startup_ip;
+	u16 __iomem *jump_table;
+	phys_addr_t pa;
+
+	if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
+		return 0;
+
+	if (ghcb_info.vm_proto < 2) {
+		pr_warn("AP jump table parking requires at least GHCB protocol version 2\n");
+		return 0;
+	}
+
+	pa = get_jump_table_addr();
+
+	/* On UP guests there is no jump table so this is not a failure */
+	if (!pa)
+		return 0;
+
+	/* Check overflow and size for untrusted jump table address */
+	if (pa + PAGE_SIZE < pa || pa + PAGE_SIZE > SZ_4G) {
+		pr_info("AP jump table is above 4GB or address overflow - not enabling AP jump table parking\n");
+		return 0;
+	}
+
+	jump_table = ioremap_encrypted(pa, PAGE_SIZE);
+	if (WARN_ON(!jump_table))
+		return -EINVAL;
+
+	/*
+	 * Save reset vector to restore it later because the blob will
+	 * overwrite it.
+	 */
+	startup_ip = jump_table[0];
+	startup_cs = jump_table[1];
+
+	/* Install AP jump table Blob with real mode AP parking code */
+	memcpy_toio(jump_table, rm_ap_jump_table_blob, blob_size);
+
+	/* Setup AP jump table GDT */
+	sev_es_setup_ap_jump_table_data(jump_table, (u32)pa);
+
+	writew(startup_ip, &jump_table[0]);
+	writew(startup_cs, &jump_table[1]);
+
+	iounmap(jump_table);
+
+	pr_info("AP jump table Blob successfully set up\n");
+
+	/* Mark AP jump table blob as available */
+	sev_ap_jumptable_blob_installed = true;
+
+	return 0;
+}
+core_initcall(sev_setup_ap_jump_table);
#13 lore raw
In reply to Joerg Roedel
Quoted 108 lines · Joerg Roedel

From: Joerg Roedel <[email protected]>

Hi,

here are changes to enable kexec/kdump in SEV-ES guests. The biggest problem for supporting kexec/kdump under SEV-ES is to find a way to hand the non-boot CPUs (APs) from one kernel to another.

Without SEV-ES the first kernel parks the CPUs in a HLT loop until they get reset by the kexec'ed kernel via an INIT-SIPI-SIPI sequence. For virtual machines the CPU reset is emulated by the hypervisor, which sets the vCPU registers back to reset state.

This does not work under SEV-ES, because the hypervisor has no access to the vCPU registers and can't make modifications to them. So an SEV-ES guest needs to reset the vCPU itself and park it using the AP-reset-hold protocol. Upon wakeup the guest needs to jump to real-mode and to the reset-vector configured in the AP-Jump-Table.

The code to do this is the main part of this patch-set. It works by placing code on the AP Jump-Table page itself to park the vCPU and for jumping to the reset vector upon wakeup. The code on the AP Jump Table runs in 16-bit protected mode with segment base set to the beginning of the page. The AP Jump-Table is usually not within the first 1MB of memory, so the code can't run in real-mode.

The AP Jump-Table is the best place to put the parking code, because the memory is owned, but read-only by the firmware and writeable by the OS. Only the first 4 bytes are used for the reset-vector, leaving the rest of the page for code/data/stack to park a vCPU. The code can't be in kernel memory because by the time the vCPU wakes up the memory will be owned by the new kernel, which might have overwritten it already.

The other patches add initial GHCB Version 2 protocol support, because kexec/kdump need the MSR-based (without a GHCB) AP-reset-hold VMGEXIT, which is a GHCB protocol version 2 feature.

The kexec'ed kernel is also entered via the decompressor and needs MMIO support there, so this patch-set also adds MMIO #VC support to the decompressor and support for handling CLFLUSH instructions.

Finally there is also code to disable kexec/kdump support at runtime when the environment does not support it (e.g. no GHCB protocol version 2 support or AP Jump Table over 4GB).

The diffstat looks big, but most of it is moving code for MMIO #VC support around to make it available to the decompressor.

The previous version of this patch-set can be found here:

	https://lore.kernel.org/lkml/20210913155603.28383-1-joro at 8bytes.org/

Please review.

Thanks,

Joerg

Changes v2->v3:

- Rebased to v5.17-rc1
- Applied most review comments by Boris
- Use the name 'AP jump table' consistently
- Make kexec-disabling for unsupported guests x86-specific
- Cleanup and consolidate patches to detect GHCB v2 protocol support

Joerg Roedel (10):
x86/kexec/64: Disable kexec when SEV-ES is active x86/sev: Save and print negotiated GHCB protocol version x86/sev: Set GHCB data structure version
x86/sev: Cache AP Jump Table Address
x86/sev: Setup code to park APs in the AP Jump Table x86/sev: Park APs on AP Jump Table with GHCB protocol version 2 x86/sev: Use AP Jump Table blob to stop CPU
x86/sev: Add MMIO handling support to boot/compressed/ code x86/sev: Handle CLFLUSH MMIO events
x86/kexec/64: Support kexec under SEV-ES with AP Jump Table Blob

 arch/x86/boot/compressed/sev.c          |  45 +-
 arch/x86/include/asm/insn-eval.h        |   1 +
 arch/x86/include/asm/realmode.h         |   5 +
 arch/x86/include/asm/sev-ap-jumptable.h |  29 +
 arch/x86/include/asm/sev.h              |  11 +-
 arch/x86/kernel/machine_kexec_64.c      |  12 +
 arch/x86/kernel/process.c               |   8 +
 arch/x86/kernel/sev-shared.c            | 233 +++++-
 arch/x86/kernel/sev.c                   | 404 +++++------
 arch/x86/lib/insn-eval-shared.c         | 913 ++++++++++++++++++++++++
 arch/x86/lib/insn-eval.c                | 909 +----------------------
 arch/x86/realmode/Makefile              |   9 +-
 arch/x86/realmode/rm/Makefile           |  11 +-
 arch/x86/realmode/rm/header.S           |   3 +
 arch/x86/realmode/rm/sev.S              |  85 +++
 arch/x86/realmode/rmpiggy.S             |   6 +
 arch/x86/realmode/sev/Makefile          |  33 +
 arch/x86/realmode/sev/ap_jump_table.S   | 131 ++++
 arch/x86/realmode/sev/ap_jump_table.lds |  24 +
 19 files changed, 1730 insertions(+), 1142 deletions(-)
 create mode 100644 arch/x86/include/asm/sev-ap-jumptable.h
 create mode 100644 arch/x86/lib/insn-eval-shared.c
 create mode 100644 arch/x86/realmode/rm/sev.S
 create mode 100644 arch/x86/realmode/sev/Makefile
 create mode 100644 arch/x86/realmode/sev/ap_jump_table.S
 create mode 100644 arch/x86/realmode/sev/ap_jump_table.lds
Quoted 4 lines

base-commit: e783362eb54cd99b2cac8b3a9aeac942e6f6ac07

Signature · 1 line
2.34.1

Hi Joerg,

I tried the patch set with 5.17.0-rc1 kernel, and I have a few questions:

1) Is it a bug or should qemu-kvm 6.2.0 be patched with specific patch? Because I found it will exit with 0 when I tried to reboot the VM with sev-es enabled. However with only sev enabled, the VM can do reboot with no problem:

[root at dell-per7525-03 ~]# virsh start TW-SEV-ES --console
....
Fedora Linux 35 (Server Edition)
Kernel 5.17.0-rc1 on an x86_64 (ttyS0)
....
[root at fedora ~]# reboot
.....

[   48.077682] reboot: Restarting system
[   48.078109] reboot: machine restart
                       ^^^^^^^^^^^^^^^ guest vm reached restart

[root at dell-per7525-03 ~]# echo $?
0
^^^ qemu-kvm exit with 0, no reboot back to normal VM kernel [root at dell-per7525-03 ~]#

2) With sev-es enabled and the 2 patch sets applied: A) [PATCH v3 00/10] x86/sev:
KEXEC/KDUMP support for SEV-ES guests, and B) [PATCH v6 0/7] KVM: SVM: Add initial GHCB protocol version 2 support. I can enable kdump and have vmcore generated:

[root at fedora ~]# dmesg|grep -i sev

[    0.030600] SEV: Hypervisor GHCB protocol version support: min=1 max=2
[    0.030602] SEV: Using GHCB protocol version 2
[    0.296144] AMD Memory Encryption Features active: SEV SEV-ES
[    0.450991] SEV: AP jump table Blob successfully set up

[root at fedora ~]# kdumpctl status
kdump: Kdump is operational

However without the 2 patch sets, I can also enable kdump and have vmcore generated:

[root at fedora ~]# dmesg|grep -i sev

[    0.295754] AMD Memory Encryption Features active: SEV SEV-ES
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ patch set A & B 
	       not applied, so only have this string.
[root at fedora ~]# echo c > /proc/sysrq-trigger
...
[    2.759403] kdump[549]: saving vmcore-dmesg.txt to /sysroot/var/crash/127.0.0.1-2022-04-18-05:58:50/
[    2.804355] kdump[555]: saving vmcore-dmesg.txt complete
[    2.806915] kdump[557]: saving vmcore
                           ^^^^^^^^^^^^^ vmcore can still be generated
...
[    7.068981] reboot: Restarting system
[    7.069340] reboot: machine restart

[root at dell-per7525-03 ~]# echo $?
0
^^^ same exit issue as question 1.

I doesn't have a complete technical background of the patch set, but isn't it the issue which this patch set is trying to solve? Or I missed something?

Thanks,
Tao Liu

Quoted 4 lines

_______________________________________________
Virtualization mailing list
Virtualization at lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

#14 lore raw
Quoted 10 lines · Tao Liu

On Thu, Jan 27, 2022 at 11:10:34AM +0100, Joerg Roedel wrote:

Hi Joerg,

I tried the patch set with 5.17.0-rc1 kernel, and I have a few questions:

1) Is it a bug or should qemu-kvm 6.2.0 be patched with specific patch? Because I found it will exit with 0 when I tried to reboot the VM with sev-es enabled. However with only sev enabled, the VM can do reboot with no problem:

Qemu was specifically patched to exit on reboot with SEV-ES guests. Qemu performs a reboot by resetting the vCPU state, which can't be done with an SEV-ES guest because the vCPU state is encrypted.

Quoted 44 lines

[root at dell-per7525-03 ~]# virsh start TW-SEV-ES --console
....
Fedora Linux 35 (Server Edition)
Kernel 5.17.0-rc1 on an x86_64 (ttyS0)
....
[root at fedora ~]# reboot
.....

[   48.077682] reboot: Restarting system
[   48.078109] reboot: machine restart
                        ^^^^^^^^^^^^^^^ guest vm reached restart

[root at dell-per7525-03 ~]# echo $?
0
^^^ qemu-kvm exit with 0, no reboot back to normal VM kernel [root at dell-per7525-03 ~]#

2) With sev-es enabled and the 2 patch sets applied: A) [PATCH v3 00/10] x86/sev:
KEXEC/KDUMP support for SEV-ES guests, and B) [PATCH v6 0/7] KVM: SVM: Add initial GHCB protocol version 2 support. I can enable kdump and have vmcore generated:

[root at fedora ~]# dmesg|grep -i sev

[    0.030600] SEV: Hypervisor GHCB protocol version support: min=1 max=2
[    0.030602] SEV: Using GHCB protocol version 2
[    0.296144] AMD Memory Encryption Features active: SEV SEV-ES
[    0.450991] SEV: AP jump table Blob successfully set up

[root at fedora ~]# kdumpctl status
kdump: Kdump is operational

However without the 2 patch sets, I can also enable kdump and have vmcore generated:

[root at fedora ~]# dmesg|grep -i sev

[    0.295754] AMD Memory Encryption Features active: SEV SEV-ES
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ patch set A & B
	       not applied, so only have this string.
[root at fedora ~]# echo c > /proc/sysrq-trigger
...
[    2.759403] kdump[549]: saving vmcore-dmesg.txt to /sysroot/var/crash/127.0.0.1-2022-04-18-05:58:50/
[    2.804355] kdump[555]: saving vmcore-dmesg.txt complete
[    2.806915] kdump[557]: saving vmcore
                            ^^^^^^^^^^^^^ vmcore can still be generated
...
[    7.068981] reboot: Restarting system
[    7.069340] reboot: machine restart
Quoted 6 lines

[root at dell-per7525-03 ~]# echo $?
0
^^^ same exit issue as question 1.

I doesn't have a complete technical background of the patch set, but isn't it the issue which this patch set is trying to solve? Or I missed something?

The main goal of this patch set is to really to solve the ability to perform a kexec. I would expect kdump to work since kdump shuts down all but the executing vCPU and performs its operations before "rebooting" (which will exit Qemu as I mentioned above). But kexec requires the need to restart the APs from within the guest after they have been stopped. That requires specific support and actions on the part of the guest kernel in how the APs are stopped and restarted.

Thanks,
Tom

Quoted 9 lines

Thanks,
Tao Liu

Quoted 4 lines

_______________________________________________
Virtualization mailing list
Virtualization at lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

#15 lore raw

Hi Tom,

Quoted 8 lines · Tom Lendacky
Quoted 7 lines · Tao Liu

On Thu, Jan 27, 2022 at 11:10:34AM +0100, Joerg Roedel wrote:

Hi Joerg,

I tried the patch set with 5.17.0-rc1 kernel, and I have a few questions:

Quoted 8 lines
Quoted 3 lines

1) Is it a bug or should qemu-kvm 6.2.0 be patched with specific patch? Because I found it will exit with 0 when I tried to reboot the VM with sev-es enabled. However with only sev enabled, the VM can do reboot with no problem:

Qemu was specifically patched to exit on reboot with SEV-ES guests. Qemu performs a reboot by resetting the vCPU state, which can't be done with an SEV-ES guest because the vCPU state is encrypted.

Sorry for the late response, and thank you for the explanation!

Quoted 51 lines
Quoted 44 lines

[root@dell-per7525-03 ~]# virsh start TW-SEV-ES --console
....
Fedora Linux 35 (Server Edition)
Kernel 5.17.0-rc1 on an x86_64 (ttyS0)
....
[root@fedora ~]# reboot
.....

[   48.077682] reboot: Restarting system
[   48.078109] reboot: machine restart
                        ^^^^^^^^^^^^^^^ guest vm reached restart

[root@dell-per7525-03 ~]# echo $?
0
^^^ qemu-kvm exit with 0, no reboot back to normal VM kernel [root@dell-per7525-03 ~]#

2) With sev-es enabled and the 2 patch sets applied: A) [PATCH v3 00/10] x86/sev:
KEXEC/KDUMP support for SEV-ES guests, and B) [PATCH v6 0/7] KVM: SVM: Add initial GHCB protocol version 2 support. I can enable kdump and have vmcore generated:

[root@fedora ~]# dmesg|grep -i sev

[    0.030600] SEV: Hypervisor GHCB protocol version support: min=1 max=2
[    0.030602] SEV: Using GHCB protocol version 2
[    0.296144] AMD Memory Encryption Features active: SEV SEV-ES
[    0.450991] SEV: AP jump table Blob successfully set up

[root@fedora ~]# kdumpctl status
kdump: Kdump is operational

However without the 2 patch sets, I can also enable kdump and have vmcore generated:

[root@fedora ~]# dmesg|grep -i sev

[    0.295754] AMD Memory Encryption Features active: SEV SEV-ES
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ patch set A & B
	       not applied, so only have this string.
[root@fedora ~]# echo c > /proc/sysrq-trigger
...
[    2.759403] kdump[549]: saving vmcore-dmesg.txt to /sysroot/var/crash/127.0.0.1-2022-04-18-05:58:50/
[    2.804355] kdump[555]: saving vmcore-dmesg.txt complete
[    2.806915] kdump[557]: saving vmcore
                            ^^^^^^^^^^^^^ vmcore can still be generated
...
[    7.068981] reboot: Restarting system
[    7.069340] reboot: machine restart
Quoted 6 lines

[root@dell-per7525-03 ~]# echo $?
0
^^^ same exit issue as question 1.

I doesn't have a complete technical background of the patch set, but isn't it the issue which this patch set is trying to solve? Or I missed something?

Quoted 7 lines

The main goal of this patch set is to really to solve the ability to perform a kexec. I would expect kdump to work since kdump shuts down all but the executing vCPU and performs its operations before "rebooting" (which will exit Qemu as I mentioned above). But kexec requires the need to restart the APs from within the guest after they have been stopped. That requires specific support and actions on the part of the guest kernel in how the APs are stopped and restarted.

Recently I got one sev-es flaged machine borrowed and retested the patch, which worked fine for kexec when sev-es enabled. With the patchset applied in 5.17.0-rc1, kexec'ed kernel can bring up all APs with no problem.

However as for kdump, I find one issue. Although kdump kernel can work well on one cpu, but we can still enable multi-cpus by removing the "nr_cpus=1" kernel parameter in kdump sysconfig. I was expecting kdump kernel can bring up all APs as kexec did, however:

[    0.000000] Command line: elfcorehdr=0x5b000000 BOOT_IMAGE=(hd0,gpt2)/vmlinuz-5.17.0-rc1+ ro resume=/dev/mapper/rhel-swap biosdevname=0 net.ifnames=0 console=ttyS0 irqpoll reset_devices cgroup_disable=memory mce=off numa=off udev.children-max=2 panic=10 rootflags=nofail acpi_no_memhotplug transparent_hugepage=never nokaslr novmcoredd hest_disable disable_cpu_apicid=0 iTCO_wdt.pretimeout=0
...
[    0.376663] smp: Bringing up secondary CPUs ...
[    0.377599] x86: Booting SMP configuration:
[    0.378342] .... node  #0, CPUs:      #1
[   10.377698] smpboot: do_boot_cpu failed(-1) to wakeup CPU#1
[   10.379882]  #2
[   20.379645] smpboot: do_boot_cpu failed(-1) to wakeup CPU#2
[   20.380648] smp: Brought up 1 node, 1 CPU
[   20.381600] smpboot: Max logical packages: 4
[   20.382597] smpboot: Total of 1 processors activated (4192.00 BogoMIPS)

Turns out for kdump, the APs were not stopped properly, so I modified the following code:

arch/x86/kernel/reboot.c +2 -0
@@ -26,6 +26,7 @@
 #include <asm/cpu.h>
 #include <asm/nmi.h>
 #include <asm/smp.h>
+#include <asm/sev.h>
 
 #include <linux/ctype.h>
 #include <linux/mc146818rtc.h>
@@ -821,6 +822,7 @@ static int crash_nmi_callback(unsigned int val, struct pt_regs *regs)
 
        atomic_dec(&waiting_for_crash_ipi);
        /* Assume hlt works */
+       sev_es_stop_this_cpu();
        halt();
        for (;;)
                cpu_relax();
[    0.000000] Command line: elfcorehdr=0x5b000000 BOOT_IMAGE=(hd0,gpt2)/vmlinuz-5.17.0-rc1-hack+ ro resume=/dev/mapper/rhel-swap biosdevname=0 net.ifnames=0 console=ttyS0 irqpoll reset_devices cgroup_disable=memory mce=off numa=off udev.children-max=2 panic=10 rootflags=nofail acpi_no_memhotplug transparent_hugepage=never nokaslr novmcoredd hest_disable disable_cpu_apicid=0 iTCO_wdt.pretimeout=0
...
[    0.402618] smp: Bringing up secondary CPUs ...
[    0.403308] x86: Booting SMP configuration:
[    0.404171] .... node  #0, CPUs:      #1 #2 #3
[    0.407362] smp: Brought up 1 node, 4 CPUs
[    0.408907] smpboot: Max logical packages: 4
[    0.409172] smpboot: Total of 4 processors activated (16768.01 BogoMIPS)

Now all APs can work in kdump kernel.

Thanks,
Tao Liu

Quoted 13 lines

Thanks,
Tom

Quoted 8 lines

Thanks,
Tao Liu

Quoted 4 lines

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

#16 lore raw
In reply to Joerg Roedel

Hi Joerg and Tom,

Quoted 4 lines · Joerg Roedel

From: Joerg Roedel <[email protected]>

Hi,

Quoted 8 lines

here are changes to enable kexec/kdump in SEV-ES guests. The biggest problem for supporting kexec/kdump under SEV-ES is to find a way to hand the non-boot CPUs (APs) from one kernel to another.

Without SEV-ES the first kernel parks the CPUs in a HLT loop until they get reset by the kexec'ed kernel via an INIT-SIPI-SIPI sequence. For virtual machines the CPU reset is emulated by the hypervisor, which sets the vCPU registers back to reset state.

Is there any plan for this patchset to proceed? Without this patchset, it does fail kexec and kdump with multiple cpus.

Thanks
Baoquan

Quoted 100 lines

This does not work under SEV-ES, because the hypervisor has no access to the vCPU registers and can't make modifications to them. So an SEV-ES guest needs to reset the vCPU itself and park it using the AP-reset-hold protocol. Upon wakeup the guest needs to jump to real-mode and to the reset-vector configured in the AP-Jump-Table.

The code to do this is the main part of this patch-set. It works by placing code on the AP Jump-Table page itself to park the vCPU and for jumping to the reset vector upon wakeup. The code on the AP Jump Table runs in 16-bit protected mode with segment base set to the beginning of the page. The AP Jump-Table is usually not within the first 1MB of memory, so the code can't run in real-mode.

The AP Jump-Table is the best place to put the parking code, because the memory is owned, but read-only by the firmware and writeable by the OS. Only the first 4 bytes are used for the reset-vector, leaving the rest of the page for code/data/stack to park a vCPU. The code can't be in kernel memory because by the time the vCPU wakes up the memory will be owned by the new kernel, which might have overwritten it already.

The other patches add initial GHCB Version 2 protocol support, because kexec/kdump need the MSR-based (without a GHCB) AP-reset-hold VMGEXIT, which is a GHCB protocol version 2 feature.

The kexec'ed kernel is also entered via the decompressor and needs MMIO support there, so this patch-set also adds MMIO #VC support to the decompressor and support for handling CLFLUSH instructions.

Finally there is also code to disable kexec/kdump support at runtime when the environment does not support it (e.g. no GHCB protocol version 2 support or AP Jump Table over 4GB).

The diffstat looks big, but most of it is moving code for MMIO #VC support around to make it available to the decompressor.

The previous version of this patch-set can be found here:

	https://lore.kernel.org/lkml/[email protected]/

Please review.

Thanks,

Joerg

Changes v2->v3:

- Rebased to v5.17-rc1
- Applied most review comments by Boris
- Use the name 'AP jump table' consistently
- Make kexec-disabling for unsupported guests x86-specific
- Cleanup and consolidate patches to detect GHCB v2 protocol support

Joerg Roedel (10):
x86/kexec/64: Disable kexec when SEV-ES is active x86/sev: Save and print negotiated GHCB protocol version x86/sev: Set GHCB data structure version
x86/sev: Cache AP Jump Table Address
x86/sev: Setup code to park APs in the AP Jump Table x86/sev: Park APs on AP Jump Table with GHCB protocol version 2 x86/sev: Use AP Jump Table blob to stop CPU
x86/sev: Add MMIO handling support to boot/compressed/ code x86/sev: Handle CLFLUSH MMIO events
x86/kexec/64: Support kexec under SEV-ES with AP Jump Table Blob

 arch/x86/boot/compressed/sev.c          |  45 +-
 arch/x86/include/asm/insn-eval.h        |   1 +
 arch/x86/include/asm/realmode.h         |   5 +
 arch/x86/include/asm/sev-ap-jumptable.h |  29 +
 arch/x86/include/asm/sev.h              |  11 +-
 arch/x86/kernel/machine_kexec_64.c      |  12 +
 arch/x86/kernel/process.c               |   8 +
 arch/x86/kernel/sev-shared.c            | 233 +++++-
 arch/x86/kernel/sev.c                   | 404 +++++------
 arch/x86/lib/insn-eval-shared.c         | 913 ++++++++++++++++++++++++
 arch/x86/lib/insn-eval.c                | 909 +----------------------
 arch/x86/realmode/Makefile              |   9 +-
 arch/x86/realmode/rm/Makefile           |  11 +-
 arch/x86/realmode/rm/header.S           |   3 +
 arch/x86/realmode/rm/sev.S              |  85 +++
 arch/x86/realmode/rmpiggy.S             |   6 +
 arch/x86/realmode/sev/Makefile          |  33 +
 arch/x86/realmode/sev/ap_jump_table.S   | 131 ++++
 arch/x86/realmode/sev/ap_jump_table.lds |  24 +
 19 files changed, 1730 insertions(+), 1142 deletions(-)
 create mode 100644 arch/x86/include/asm/sev-ap-jumptable.h
 create mode 100644 arch/x86/lib/insn-eval-shared.c
 create mode 100644 arch/x86/realmode/rm/sev.S
 create mode 100644 arch/x86/realmode/sev/Makefile
 create mode 100644 arch/x86/realmode/sev/ap_jump_table.S
 create mode 100644 arch/x86/realmode/sev/ap_jump_table.lds

base-commit: e783362eb54cd99b2cac8b3a9aeac942e6f6ac07

Signature · 1 line
2.34.1