Skip to contentSkip to conversation

x86/hyperv/sev: Add AMD sev-snp enlightened guest support on hyperv

RFC PATCH Superseded Tianyu Lan 8 participants 63 messages Nov 19, 2022 – Jan 12, 2023

Subsystems (9)

v2 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Tianyu Lan

author
#0 lore raw

v2 cover x86/hyperv/sev: Add AMD sev-snp enlightened guest support on hyperv

From: Tianyu Lan <[email protected]>

This patchset is to add AMD sev-snp enlightened guest support on hyperv. Hyperv uses Linux direct boot mode to boot up Linux kernel and so it needs to pvalidate system memory by itself.

In hyperv case, there is no boot loader and so cc blob is prepared by hypervisor. In this series, hypervisor set the cc blob address directly into boot parameter of Linux kernel. If the magic number on cc blob address is valid, kernel will read cc blob.

Shared memory between guests and hypervisor should be decrypted and zero memory after decrypt memory. The data in the target address. It maybe smearedto avoid smearing data.

Introduce #HV exception support in AMD sev snp code and #HV handler.

Change since v1:
- Remove boot param changes for cc blob address and use setup head to pass cc blob info
- Remove unnessary WARN and BUG check
- Add system vector table map in the #HV exception
- Fix interrupt exit issue when use #HV exception

Ashish Kalra (2):
x86/sev: optimize system vector processing invoked from #HV exception x86/sev: Fix interrupt exit code paths from #HV exception

Tianyu Lan (16):
x86/sev: Pvalidate memory gab for decompressing kernel x86/hyperv: Add sev-snp enlightened guest specific config x86/hyperv: apic change for sev-snp enlightened guest x86/hyperv: Decrypt hv vp assist page in sev-snp enlightened guest x86/hyperv: Get Virtual Trust Level via hvcall x86/hyperv: Use vmmcall to implement hvcall in sev-snp enlightened
guest
clocksource: hyper-v: decrypt hyperv tsc page in sev-snp enlightened
guest
x86/hyperv: decrypt vmbus pages for sev-snp enlightened guest x86/hyperv: set target vtl in the vmbus init message
drivers: hv: Decrypt percpu hvcall input arg page in sev-snp
enlightened guest
Drivers: hv: vmbus: Decrypt vmbus ring buffer
x86/hyperv: Initialize cpu and memory for sev-snp enlightened guest x86/hyperv: Add smp support for sev-snp guest
x86/hyperv: Add hyperv-specific hadling for VMMCALL under SEV-ES x86/sev: Add a #HV exception handler
x86/sev: Initialize #HV doorbell and handle interrupt requests

 arch/x86/boot/compressed/head_64.S    |   8 +
 arch/x86/boot/compressed/sev.c        |  84 ++++++
 arch/x86/entry/entry_64.S             |  82 +++++
 arch/x86/hyperv/hv_apic.c             |  79 +++--
 arch/x86/hyperv/hv_init.c             |  47 +++
 arch/x86/hyperv/ivm.c                 |  12 +-
 arch/x86/include/asm/cpu_entry_area.h |   6 +
 arch/x86/include/asm/idtentry.h       | 107 ++++++-
 arch/x86/include/asm/irqflags.h       |  19 ++
 arch/x86/include/asm/mem_encrypt.h    |   2 +
 arch/x86/include/asm/mshyperv.h       |  68 +++--
 arch/x86/include/asm/msr-index.h      |   6 +
 arch/x86/include/asm/page_64_types.h  |   1 +
 arch/x86/include/asm/sev.h            |  13 +
 arch/x86/include/asm/svm.h            |  55 +++-
 arch/x86/include/asm/trapnr.h         |   1 +
 arch/x86/include/asm/traps.h          |   1 +
 arch/x86/include/uapi/asm/svm.h       |   4 +
 arch/x86/kernel/cpu/common.c          |   1 +
 arch/x86/kernel/cpu/mshyperv.c        | 267 ++++++++++++++++-
 arch/x86/kernel/dumpstack_64.c        |   9 +-
 arch/x86/kernel/idt.c                 |   1 +
 arch/x86/kernel/sev.c                 | 412 ++++++++++++++++++++++----
 arch/x86/kernel/traps.c               |  50 ++++
 arch/x86/kernel/vmlinux.lds.S         |   7 +
 arch/x86/mm/cpu_entry_area.c          |   2 +
 drivers/clocksource/hyperv_timer.c    |   2 +-
 drivers/hv/connection.c               |  14 +
 drivers/hv/hv.c                       |  32 +-
 drivers/hv/hv_common.c                |  22 ++
 drivers/hv/ring_buffer.c              |   7 +-
 include/asm-generic/hyperv-tlfs.h     |  19 ++
 include/asm-generic/mshyperv.h        |   2 +
 include/linux/hyperv.h                |   4 +-
 34 files changed, 1340 insertions(+), 106 deletions(-)
Signature · 1 line
2.25.1
author
#1 lore raw

RFC PATCH v2 1/18 x86/sev: Pvalidate memory gab for decompressing kernel

From: Tianyu Lan <[email protected]>

Pvalidate needed pages for decompressing kernel. The E820_TYPE_RAM entry includes only validated memory. The kernel expects that the RAM entry's addr is fixed while the entry size is to be extended to cover addresses to the start of next entry. This patch increases the RAM entry size to cover all possilble memory addresses until init_size.

---
 arch/x86/boot/compressed/head_64.S |  8 +++
 arch/x86/boot/compressed/sev.c     | 84 ++++++++++++++++++++++++++++++
 2 files changed, 92 insertions(+)
arch/x86/boot/compressed/head_64.S +8 -0
index d33f060900d2..818edaf5d0cf 100644
@@ -348,6 +348,14 @@ SYM_CODE_START(startup_64)
 	cld
 	cli
 
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/* pvalidate memory on demand if SNP is enabled. */
+	pushq	%rsi
+	movq    %rsi, %rdi
+	call 	pvalidate_for_startup_64
+	popq	%rsi
+#endif
+
 	/* Setup data segments. */
 	xorl	%eax, %eax
 	movl	%eax, %ds
arch/x86/boot/compressed/sev.c +84 -0

107 lines not on this page · raw message

Signature · 1 line
2.25.1
author
#8 lore raw

RFC PATCH v2 2/18 x86/hyperv: Add sev-snp enlightened guest specific config

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

Introduce static key isolation_type_en_snp for enlightened guest check and add some specific options in ms_hyperv_init_ platform().

---
 arch/x86/hyperv/ivm.c           | 12 +++++++++++-
 arch/x86/include/asm/mshyperv.h |  2 ++
 arch/x86/kernel/cpu/mshyperv.c  | 29 ++++++++++++++++++++++++-----
 drivers/hv/hv_common.c          |  7 +++++++
 4 files changed, 44 insertions(+), 6 deletions(-)
arch/x86/hyperv/ivm.c +11 -1
index 1dbcbd9da74d..e9c30dad3419 100644
@@ -259,10 +259,20 @@ bool hv_is_isolation_supported(void)
 }
 
 DEFINE_STATIC_KEY_FALSE(isolation_type_snp);
+DEFINE_STATIC_KEY_FALSE(isolation_type_en_snp);
+
+/*
+ * hv_isolation_type_en_snp - Check system runs in the AMD SEV-SNP based
+ * isolation enlightened VM.
+ */
+bool hv_isolation_type_en_snp(void)
+{
+	return static_branch_unlikely(&isolation_type_en_snp);
+}
 
 /*
  * hv_isolation_type_snp - Check system runs in the AMD SEV-SNP based
- * isolation VM.
+ * isolation VM with vTOM support.
  */
 bool hv_isolation_type_snp(void)
 {
arch/x86/include/asm/mshyperv.h +2 -0
index 61f0c206bff0..9b8c3f638845 100644
@@ -14,6 +14,7 @@
 union hv_ghcb;
 
 DECLARE_STATIC_KEY_FALSE(isolation_type_snp);
+DECLARE_STATIC_KEY_FALSE(isolation_type_en_snp);
 
 typedef int (*hyperv_fill_flush_list_func)(
 		struct hv_guest_mapping_flush_list *flush,
@@ -32,6 +33,7 @@ extern u64 hv_current_partition_id;
 
 extern union hv_ghcb * __percpu *hv_ghcb_pg;
 
+extern bool hv_isolation_type_en_snp(void);
 int hv_call_deposit_pages(int node, u64 partition_id, u32 num_pages);
 int hv_call_add_logical_proc(int node, u32 lp_index, u32 acpi_id);
 int hv_call_create_vp(int node, u64 partition_id, u32 vp_index, u32 flags);
arch/x86/kernel/cpu/mshyperv.c +24 -5
index 831613959a92..2ea4f21c6172 100644
@@ -273,6 +273,21 @@ static void __init ms_hyperv_init_platform(void)
 	ms_hyperv.misc_features = cpuid_edx(HYPERV_CPUID_FEATURES);
 	ms_hyperv.hints    = cpuid_eax(HYPERV_CPUID_ENLIGHTMENT_INFO);
 
+	/*
+	 * Add custom configuration for SEV-SNP Enlightened guest
+	 */
+	if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+		ms_hyperv.features |= HV_ACCESS_FREQUENCY_MSRS;
+		ms_hyperv.misc_features |= HV_FEATURE_FREQUENCY_MSRS_AVAILABLE;
+		ms_hyperv.misc_features &= ~HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE;
+		ms_hyperv.hints |= HV_DEPRECATING_AEOI_RECOMMENDED;
+		ms_hyperv.hints |= HV_X64_APIC_ACCESS_RECOMMENDED;
+		ms_hyperv.hints |= HV_X64_CLUSTER_IPI_RECOMMENDED;
+	}
+
+	pr_info("Hyper-V: enlightment features 0x%x, hints 0x%x, misc 0x%x\n",
+		ms_hyperv.features, ms_hyperv.hints, ms_hyperv.misc_features);
+
 	hv_max_functions_eax = cpuid_eax(HYPERV_CPUID_VENDOR_AND_MAX_FUNCTIONS);
 
 	pr_info("Hyper-V: privilege flags low 0x%x, high 0x%x, hints 0x%x, misc 0x%x\n",
@@ -328,18 +343,22 @@ static void __init ms_hyperv_init_platform(void)
 		ms_hyperv.shared_gpa_boundary =
 			BIT_ULL(ms_hyperv.shared_gpa_boundary_bits);
 
-		pr_info("Hyper-V: Isolation Config: Group A 0x%x, Group B 0x%x\n",
-			ms_hyperv.isolation_config_a, ms_hyperv.isolation_config_b);
-
-		if (hv_get_isolation_type() == HV_ISOLATION_TYPE_SNP) {
+		if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+			static_branch_enable(&isolation_type_en_snp);
+		} else if (hv_get_isolation_type() == HV_ISOLATION_TYPE_SNP) {
 			static_branch_enable(&isolation_type_snp);
 #ifdef CONFIG_SWIOTLB
 			swiotlb_unencrypted_base = ms_hyperv.shared_gpa_boundary;
 #endif
 		}
+
+		pr_info("Hyper-V: Isolation Config: Group A 0x%x, Group B 0x%x\n",
+			ms_hyperv.isolation_config_a, ms_hyperv.isolation_config_b);
+
 		/* Isolation VMs are unenlightened SEV-based VMs, thus this check: */
 		if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT)) {
-			if (hv_get_isolation_type() != HV_ISOLATION_TYPE_NONE)
+			if (hv_get_isolation_type() != HV_ISOLATION_TYPE_NONE
+			    && !cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
 				cc_set_vendor(CC_VENDOR_HYPERV);
 		}
 	}
drivers/hv/hv_common.c +7 -0
index ae68298c0dca..2c6602571c47 100644
@@ -268,6 +268,13 @@ bool __weak hv_isolation_type_snp(void)
 }
 EXPORT_SYMBOL_GPL(hv_isolation_type_snp);
 
+bool __weak hv_isolation_type_en_snp(void)
+{
+	return false;
+}
+EXPORT_SYMBOL_GPL(hv_isolation_type_en_snp);
+
+
 void __weak hv_setup_vmbus_handler(void (*handler)(void))
 {
 }
Signature · 1 line
2.25.1
author
#11 lore raw

RFC PATCH v2 3/18 x86/hyperv: apic change for sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

Hyperv sev-snp enlightened guest doesn't support x2apic and apic page read/write operation. Bypass these requests. ipi request maybe returned with timeout error code and add retry mechanism.

---
 arch/x86/hyperv/hv_apic.c         | 79 ++++++++++++++++++++++++-------
 include/asm-generic/hyperv-tlfs.h |  1 +
 2 files changed, 63 insertions(+), 17 deletions(-)
arch/x86/hyperv/hv_apic.c +62 -17
index fb8b2c088681..214354d20833 100644
@@ -66,9 +66,15 @@ static u32 hv_apic_read(u32 reg)
 		rdmsr(HV_X64_MSR_TPR, reg_val, hi);
 		(void)hi;
 		return reg_val;
-
+	case APIC_ID:
+		if (hv_isolation_type_en_snp())
+			return smp_processor_id();
+		fallthrough;
 	default:
-		return native_apic_mem_read(reg);
+		if (!hv_isolation_type_en_snp())
+			return native_apic_mem_read(reg);
+		else
+			return 0;
 	}
 }
 
@@ -82,7 +88,8 @@ static void hv_apic_write(u32 reg, u32 val)
 		wrmsr(HV_X64_MSR_TPR, val, 0);
 		break;
 	default:
-		native_apic_mem_write(reg, val);
+		if (!hv_isolation_type_en_snp())
+			native_apic_mem_write(reg, val);
 	}
 }
 
@@ -106,6 +113,7 @@ static bool __send_ipi_mask_ex(const struct cpumask *mask, int vector,
 	struct hv_send_ipi_ex *ipi_arg;
 	unsigned long flags;
 	int nr_bank = 0;
+	int retry = 5;
 	u64 status = HV_STATUS_INVALID_PARAMETER;
 
 	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
@@ -144,8 +152,10 @@ static bool __send_ipi_mask_ex(const struct cpumask *mask, int vector,
 		ipi_arg->vp_set.format = HV_GENERIC_SET_ALL;
 	}
 
-	status = hv_do_rep_hypercall(HVCALL_SEND_IPI_EX, 0, nr_bank,
+	do {
+		status = hv_do_rep_hypercall(HVCALL_SEND_IPI_EX, 0, nr_bank,
 			      ipi_arg, NULL);
+	} while (status == HV_STATUS_TIME_OUT && retry--);
 
 ipi_mask_ex_done:
 	local_irq_restore(flags);
@@ -159,6 +169,7 @@ static bool __send_ipi_mask(const struct cpumask *mask, int vector,
 	struct hv_send_ipi ipi_arg;
 	u64 status;
 	unsigned int weight;
+	int retry = 5;
 
 	trace_hyperv_send_ipi_mask(mask, vector);
 
@@ -212,8 +223,11 @@ static bool __send_ipi_mask(const struct cpumask *mask, int vector,
 		__set_bit(vcpu, (unsigned long *)&ipi_arg.cpu_mask);
 	}
 
-	status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, ipi_arg.vector,
-				     ipi_arg.cpu_mask);
+	do {
+		status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, ipi_arg.vector,
+						ipi_arg.cpu_mask);
+	} while (status == HV_STATUS_TIME_OUT && retry--);
+
 	return hv_result_success(status);
 
 do_ex_hypercall:
@@ -224,6 +238,7 @@ static bool __send_ipi_one(int cpu, int vector)
 {
 	int vp = hv_cpu_number_to_vp_number(cpu);
 	u64 status;
+	int retry = 5;
 
 	trace_hyperv_send_ipi_one(cpu, vector);
 
@@ -236,26 +251,48 @@ static bool __send_ipi_one(int cpu, int vector)
 	if (vp >= 64)
 		return __send_ipi_mask_ex(cpumask_of(cpu), vector, false);
 
-	status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, vector, BIT_ULL(vp));
+	do {
+		status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, vector, BIT_ULL(vp));
+	} while (status == HV_STATUS_TIME_OUT || retry--);
+
 	return hv_result_success(status);
 }
 
 static void hv_send_ipi(int cpu, int vector)
 {
-	if (!__send_ipi_one(cpu, vector))
-		orig_apic.send_IPI(cpu, vector);
+	if (!__send_ipi_one(cpu, vector)) {
+		if (!hv_isolation_type_en_snp())
+			orig_apic.send_IPI(cpu, vector);
+		else
+			WARN_ON_ONCE(1);
+	}
 }
 
 static void hv_send_ipi_mask(const struct cpumask *mask, int vector)
 {
-	if (!__send_ipi_mask(mask, vector, false))
-		orig_apic.send_IPI_mask(mask, vector);
+	if (!__send_ipi_mask(mask, vector, false)) {
+		if (!hv_isolation_type_en_snp())
+			orig_apic.send_IPI_mask(mask, vector);
+		else
+			WARN_ON_ONCE(1);
+	}
 }
 
 static void hv_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
 {
-	if (!__send_ipi_mask(mask, vector, true))
-		orig_apic.send_IPI_mask_allbutself(mask, vector);
+	unsigned int this_cpu = smp_processor_id();
+	struct cpumask new_mask;
+	const struct cpumask *local_mask;
+
+	cpumask_copy(&new_mask, mask);
+	cpumask_clear_cpu(this_cpu, &new_mask);
+	local_mask = &new_mask;
+	if (!__send_ipi_mask(local_mask, vector, true)) {
+		if (!hv_isolation_type_en_snp())
+			orig_apic.send_IPI_mask_allbutself(mask, vector);
+		else
+			WARN_ON_ONCE(1);
+	}
 }
 
 static void hv_send_ipi_allbutself(int vector)
@@ -265,14 +302,22 @@ static void hv_send_ipi_allbutself(int vector)
 
 static void hv_send_ipi_all(int vector)
 {
-	if (!__send_ipi_mask(cpu_online_mask, vector, false))
-		orig_apic.send_IPI_all(vector);
+	if (!__send_ipi_mask(cpu_online_mask, vector, false)) {
+		if (!hv_isolation_type_en_snp())
+			orig_apic.send_IPI_all(vector);
+		else
+			WARN_ON_ONCE(1);
+	}
 }
 
 static void hv_send_ipi_self(int vector)
 {
-	if (!__send_ipi_one(smp_processor_id(), vector))
-		orig_apic.send_IPI_self(vector);
+	if (!__send_ipi_one(smp_processor_id(), vector)) {
+		if (!hv_isolation_type_en_snp())
+			orig_apic.send_IPI_self(vector);
+		else
+			WARN_ON_ONCE(1);
+	}
 }
 
 void __init hv_apic_init(void)
include/asm-generic/hyperv-tlfs.h +1 -0
index fdce7a4cfc6f..6e2a090e2649 100644
@@ -208,6 +208,7 @@ enum HV_GENERIC_SET_FORMAT {
 #define HV_STATUS_INVALID_PORT_ID		17
 #define HV_STATUS_INVALID_CONNECTION_ID		18
 #define HV_STATUS_INSUFFICIENT_BUFFERS		19
+#define HV_STATUS_TIME_OUT			0x78
 
 /*
  * The Hyper-V TimeRefCount register and the TSC
Signature · 1 line
2.25.1
author
#13 lore raw

RFC PATCH v2 4/18 x86/hyperv: Decrypt hv vp assist page in sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

hv vp assist page is shared between sev snp guest and hyperv. Decrypt the page when use it.

---
 arch/x86/hyperv/hv_init.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
arch/x86/hyperv/hv_init.c +12 -0
index 29774126e931..4600c5941957 100644
@@ -30,6 +30,7 @@
 #include <clocksource/hyperv_timer.h>
 #include <linux/highmem.h>
 #include <linux/swiotlb.h>
+#include <linux/set_memory.h>
 
 int hyperv_init_cpuhp;
 u64 hv_current_partition_id = ~0ull;
@@ -112,6 +113,11 @@ static int hv_cpu_init(unsigned int cpu)
 		}
 		WARN_ON(!(*hvp));
 		if (*hvp) {
+			if (hv_isolation_type_en_snp()) {
+				WARN_ON_ONCE(set_memory_decrypted((unsigned long)(*hvp), 1) != 0);
+				memset(*hvp, 0, PAGE_SIZE);
+			}
+
 			msr.enable = 1;
 			wrmsrl(HV_X64_MSR_VP_ASSIST_PAGE, msr.as_uint64);
 		}
@@ -228,6 +234,12 @@ static int hv_cpu_die(unsigned int cpu)
 
 	if (hv_vp_assist_page && hv_vp_assist_page[cpu]) {
 		union hv_vp_assist_msr_contents msr = { 0 };
+
+		if (hv_isolation_type_en_snp())
+			WARN_ON_ONCE(set_memory_encrypted(
+				    (unsigned long)hv_vp_assist_page[cpu],
+				    1) != 0);
+
 		if (hv_root_partition) {
 			/*
 			 * For root partition the VP assist page is mapped to
Signature · 1 line
2.25.1
author
#18 lore raw

RFC PATCH v2 6/18 x86/hyperv: Use vmmcall to implement hvcall in sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

In sev-snp enlightened guest, hvcall needs to use vmmcall to trigger vmexit and notify hypervisor to handle hypercall request.

---
 arch/x86/include/asm/mshyperv.h | 66 ++++++++++++++++++++++-----------
 1 file changed, 45 insertions(+), 21 deletions(-)
arch/x86/include/asm/mshyperv.h +45 -21
index 9b8c3f638845..28d5429e33c9 100644
@@ -45,16 +45,25 @@ static inline u64 hv_do_hypercall(u64 control, void *input, void *output)
 	u64 hv_status;
 
 #ifdef CONFIG_X86_64
-	if (!hv_hypercall_pg)
-		return U64_MAX;
+	if (hv_isolation_type_en_snp()) {
+		__asm__ __volatile__("mov %4, %%r8\n"
+				"vmmcall"
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input_address)
+				:  "r" (output_address)
+				: "cc", "memory", "r8", "r9", "r10", "r11");
+	} else {
+		if (!hv_hypercall_pg)
+			return U64_MAX;
 
-	__asm__ __volatile__("mov %4, %%r8\n"
-			     CALL_NOSPEC
-			     : "=a" (hv_status), ASM_CALL_CONSTRAINT,
-			       "+c" (control), "+d" (input_address)
-			     :  "r" (output_address),
-				THUNK_TARGET(hv_hypercall_pg)
-			     : "cc", "memory", "r8", "r9", "r10", "r11");
+		__asm__ __volatile__("mov %4, %%r8\n"
+				CALL_NOSPEC
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input_address)
+				:  "r" (output_address),
+					THUNK_TARGET(hv_hypercall_pg)
+				: "cc", "memory", "r8", "r9", "r10", "r11");
+	}
 #else
 	u32 input_address_hi = upper_32_bits(input_address);
 	u32 input_address_lo = lower_32_bits(input_address);
@@ -82,12 +91,18 @@ static inline u64 hv_do_fast_hypercall8(u16 code, u64 input1)
 	u64 hv_status, control = (u64)code | HV_HYPERCALL_FAST_BIT;
 
 #ifdef CONFIG_X86_64
-	{
+	if (hv_isolation_type_en_snp()) {
+		__asm__ __volatile__(
+				"vmmcall"
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)
+				:: "cc", "r8", "r9", "r10", "r11");
+	} else {
 		__asm__ __volatile__(CALL_NOSPEC
-				     : "=a" (hv_status), ASM_CALL_CONSTRAINT,
-				       "+c" (control), "+d" (input1)
-				     : THUNK_TARGET(hv_hypercall_pg)
-				     : "cc", "r8", "r9", "r10", "r11");
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)
+				: THUNK_TARGET(hv_hypercall_pg)
+				: "cc", "r8", "r9", "r10", "r11");
 	}
 #else
 	{
@@ -113,14 +128,21 @@ static inline u64 hv_do_fast_hypercall16(u16 code, u64 input1, u64 input2)
 	u64 hv_status, control = (u64)code | HV_HYPERCALL_FAST_BIT;
 
 #ifdef CONFIG_X86_64
-	{
+	if (hv_isolation_type_en_snp()) {
 		__asm__ __volatile__("mov %4, %%r8\n"
-				     CALL_NOSPEC
-				     : "=a" (hv_status), ASM_CALL_CONSTRAINT,
-				       "+c" (control), "+d" (input1)
-				     : "r" (input2),
-				       THUNK_TARGET(hv_hypercall_pg)
-				     : "cc", "r8", "r9", "r10", "r11");
+				"vmmcall"
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)
+				: "r" (input2)
+				: "cc", "r8", "r9", "r10", "r11");
+	} else {
+		__asm__ __volatile__("mov %4, %%r8\n"
+				CALL_NOSPEC
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)
+				: "r" (input2),
+				THUNK_TARGET(hv_hypercall_pg)
+				: "cc", "r8", "r9", "r10", "r11");
 	}
 #else
 	{
@@ -177,6 +199,7 @@ int hv_map_ioapic_interrupt(int ioapic_id, bool level, int vcpu, int vector,
 		struct hv_interrupt_entry *entry);
 int hv_unmap_ioapic_interrupt(int ioapic_id, struct hv_interrupt_entry *entry);
 int hv_set_mem_host_visibility(unsigned long addr, int numpages, bool visible);
+int hv_snp_boot_ap(int cpu, unsigned long start_ip);
 
 #ifdef CONFIG_AMD_MEM_ENCRYPT
 void hv_ghcb_msr_write(u64 msr, u64 value);
@@ -191,6 +214,7 @@ static inline void hv_ghcb_terminate(unsigned int set, unsigned int reason) {}
 #endif
 
 extern bool hv_isolation_type_snp(void);
+extern bool hv_isolation_type_en_snp(void);
 
 static inline bool hv_is_synic_reg(unsigned int reg)
 {
Signature · 1 line
2.25.1
author
#22 lore raw

RFC PATCH v2 7/18 clocksource: hyper-v: decrypt hyperv tsc page in sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

Hyper-V tsc page is shared with hypervisor and it should be decrypted in sev-snp enlightened guest when it's used.

---
 drivers/clocksource/hyperv_timer.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/clocksource/hyperv_timer.c +1 -1
index bb47610bbd1c..aa68eebed5ee 100644
@@ -364,7 +364,7 @@ EXPORT_SYMBOL_GPL(hv_stimer_global_cleanup);
 static union {
 	struct ms_hyperv_tsc_page page;
 	u8 reserved[PAGE_SIZE];
-} tsc_pg __aligned(PAGE_SIZE);
+} tsc_pg __bss_decrypted __aligned(PAGE_SIZE);
 
 struct ms_hyperv_tsc_page *hv_get_tsc_page(void)
 {
Signature · 1 line
2.25.1
author
#28 lore raw

RFC PATCH v2 9/18 x86/hyperv: set target vtl in the vmbus init message

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

Set target vtl (Virtual Trust Level) in the vmbus init message.

---
 drivers/hv/connection.c | 1 +
 include/linux/hyperv.h  | 4 ++--
 2 files changed, 3 insertions(+), 2 deletions(-)
drivers/hv/connection.c +1 -0
index 43141225ea15..09a1253b539a 100644
@@ -98,6 +98,7 @@ int vmbus_negotiate_version(struct vmbus_channel_msginfo *msginfo, u32 version)
 	 */
 	if (version >= VERSION_WIN10_V5) {
 		msg->msg_sint = VMBUS_MESSAGE_SINT;
+		msg->msg_vtl = ms_hyperv.vtl;
 		vmbus_connection.msg_conn_id = VMBUS_MESSAGE_CONNECTION_ID_4;
 	} else {
 		msg->interrupt_page = virt_to_phys(vmbus_connection.int_page);
include/linux/hyperv.h +2 -2
index 3b42264333ef..2be0b5efd1ea 100644
@@ -665,8 +665,8 @@ struct vmbus_channel_initiate_contact {
 		u64 interrupt_page;
 		struct {
 			u8	msg_sint;
-			u8	padding1[3];
-			u32	padding2;
+			u8	msg_vtl;
+			u8	reserved[6];
 		};
 	};
 	u64 monitor_page1;
Signature · 1 line
2.25.1
author
#25 lore raw

RFC PATCH v2 8/18 x86/hyperv: decrypt vmbus pages for sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

Vmbus int, synic and post message pages are shared with hypervisor and so decrypt these pages in the sev-snp guest.

---
 drivers/hv/connection.c | 13 +++++++++++++
 drivers/hv/hv.c         | 32 +++++++++++++++++++++++++++++++-
 2 files changed, 44 insertions(+), 1 deletion(-)
drivers/hv/connection.c +13 -0
index 9dc27e5d367a..43141225ea15 100644
@@ -215,6 +215,15 @@ int vmbus_connect(void)
 		(void *)((unsigned long)vmbus_connection.int_page +
 			(HV_HYP_PAGE_SIZE >> 1));
 
+	if (hv_isolation_type_snp() || hv_isolation_type_en_snp()) {
+		ret = set_memory_decrypted((unsigned long)
+				vmbus_connection.int_page, 1);
+		if (ret)
+			goto cleanup;
+
+		memset(vmbus_connection.int_page, 0, PAGE_SIZE);
+	}
+
 	/*
 	 * Setup the monitor notification facility. The 1st page for
 	 * parent->child and the 2nd page for child->parent
@@ -372,6 +381,10 @@ void vmbus_disconnect(void)
 		destroy_workqueue(vmbus_connection.work_queue);
 
 	if (vmbus_connection.int_page) {
+		if (hv_isolation_type_en_snp())
+			set_memory_encrypted((unsigned long)
+				vmbus_connection.int_page, 1);
+
 		hv_free_hyperv_page((unsigned long)vmbus_connection.int_page);
 		vmbus_connection.int_page = NULL;
 	}
drivers/hv/hv.c +31 -1
index 4d6480d57546..f9111eb32739 100644
@@ -20,6 +20,7 @@
 #include <linux/interrupt.h>
 #include <clocksource/hyperv_timer.h>
 #include <asm/mshyperv.h>
+#include <linux/set_memory.h>
 #include "hyperv_vmbus.h"
 
 /* The one and only */
@@ -117,7 +118,7 @@ int hv_post_message(union hv_connection_id connection_id,
 
 int hv_synic_alloc(void)
 {
-	int cpu;
+	int cpu, ret;
 	struct hv_per_cpu_context *hv_cpu;
 
 	/*
@@ -168,6 +169,29 @@ int hv_synic_alloc(void)
 			pr_err("Unable to allocate post msg page\n");
 			goto err;
 		}
+
+		if (hv_isolation_type_en_snp()) {
+			ret = set_memory_decrypted((unsigned long)
+				hv_cpu->synic_message_page, 1);
+			ret |= set_memory_decrypted((unsigned long)
+				hv_cpu->synic_event_page, 1);
+			ret |= set_memory_decrypted((unsigned long)
+				hv_cpu->post_msg_page, 1);
+
+			if (ret) {
+				set_memory_encrypted((unsigned long)
+					hv_cpu->synic_message_page, 1);
+				set_memory_encrypted((unsigned long)
+					hv_cpu->synic_event_page, 1);
+				set_memory_encrypted((unsigned long)
+					hv_cpu->post_msg_page, 1);
+				goto err;
+			}
+
+			memset(hv_cpu->synic_message_page, 0, PAGE_SIZE);
+			memset(hv_cpu->synic_event_page, 0, PAGE_SIZE);
+			memset(hv_cpu->post_msg_page, 0, PAGE_SIZE);
+		}
 	}
 
 	return 0;
@@ -188,6 +212,12 @@ void hv_synic_free(void)
 		struct hv_per_cpu_context *hv_cpu
 			= per_cpu_ptr(hv_context.cpu_context, cpu);
 
+		if (hv_isolation_type_en_snp()) {
+			set_memory_encrypted((unsigned long)hv_cpu->synic_message_page, 1);
+			set_memory_encrypted((unsigned long)hv_cpu->synic_event_page, 1);
+			set_memory_encrypted((unsigned long)hv_cpu->post_msg_page, 1);
+		}
+
 		free_page((unsigned long)hv_cpu->synic_event_page);
 		free_page((unsigned long)hv_cpu->synic_message_page);
 		free_page((unsigned long)hv_cpu->post_msg_page);
Signature · 1 line
2.25.1
author
#16 lore raw

RFC PATCH v2 5/18 x86/hyperv: Get Virtual Trust Level via hvcall

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

sev-snp guest provides vtl(Virtual Trust Level) and get it from hyperv hvcall via HVCALL_GET_VP_REGISTERS.

---
 arch/x86/hyperv/hv_init.c      | 35 ++++++++++++++++++++++++++++++++++
 include/asm-generic/mshyperv.h |  2 ++
 2 files changed, 37 insertions(+)
arch/x86/hyperv/hv_init.c +35 -0
index 4600c5941957..5b919d4d24c0 100644
@@ -390,6 +390,39 @@ static void __init hv_get_partition_id(void)
 	local_irq_restore(flags);
 }
 
+static u8 __init get_current_vtl(void)
+{
+	u64 control = ((u64)1 << HV_HYPERCALL_REP_COMP_OFFSET) | HVCALL_GET_VP_REGISTERS;
+	struct hv_get_vp_registers_input *input = NULL;
+	struct hv_get_vp_registers_output *output = NULL;
+	u8 vtl = 0;
+	int ret;
+	unsigned long flags;
+
+	local_irq_save(flags);
+	input = *(struct hv_get_vp_registers_input **)this_cpu_ptr(hyperv_pcpu_input_arg);
+	output = (struct hv_get_vp_registers_output *)input;
+	if (!input || !output) {
+		pr_err("Hyper-V: cannot allocate a shared page!");
+		goto done;
+	}
+
+	memset(input, 0, sizeof(*input) + sizeof(input->element[0]));
+	input->header.partitionid = HV_PARTITION_ID_SELF;
+	input->header.inputvtl = 0;
+	input->element[0].name0 = 0x000D0003;
+
+	ret = hv_do_hypercall(control, input, output);
+	if (ret == 0)
+		vtl = output->as64.low & 0xf;
+	else
+		pr_err("Hyper-V: failed to get the current VTL!");
+	local_irq_restore(flags);
+
+done:
+	return vtl;
+}
+
 /*
  * This function is to be invoked early in the boot sequence after the
  * hypervisor has been detected.
@@ -527,6 +560,8 @@ void __init hyperv_init(void)
 	if (hv_is_isolation_supported())
 		swiotlb_update_mem_attributes();
 #endif
+	/* Find the current VTL */
+	ms_hyperv.vtl = get_current_vtl();
 
 	return;
 
include/asm-generic/mshyperv.h +2 -0
index bfb9eb9d7215..68133de044ec 100644
@@ -46,6 +46,7 @@ struct ms_hyperv_info {
 		};
 	};
 	u64 shared_gpa_boundary;
+	u8 vtl;
 };
 extern struct ms_hyperv_info ms_hyperv;
 
@@ -55,6 +56,7 @@ extern void * __percpu *hyperv_pcpu_output_arg;
 extern u64 hv_do_hypercall(u64 control, void *inputaddr, void *outputaddr);
 extern u64 hv_do_fast_hypercall8(u16 control, u64 input8);
 extern bool hv_isolation_type_snp(void);
+extern bool hv_isolation_type_en_snp(void);
 
 /* Helper functions that provide a consistent pattern for checking Hyper-V hypercall status. */
 static inline int hv_result(u64 status)
Signature · 1 line
2.25.1
author
#30 lore raw

RFC PATCH v2 10/18 drivers: hv: Decrypt percpu hvcall input arg page in sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

Hypervisor needs to access iput arg page and guest should decrypt the page.

---
 drivers/hv/hv_common.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)
drivers/hv/hv_common.c +15 -0
index 2c6602571c47..c16961e686a0 100644
@@ -21,6 +21,7 @@
 #include <linux/ptrace.h>
 #include <linux/slab.h>
 #include <linux/dma-map-ops.h>
+#include <linux/set_memory.h>
 #include <asm/hyperv-tlfs.h>
 #include <asm/mshyperv.h>
 
@@ -125,6 +126,7 @@ int hv_common_cpu_init(unsigned int cpu)
 	u64 msr_vp_index;
 	gfp_t flags;
 	int pgcount = hv_root_partition ? 2 : 1;
+	int ret;
 
 	/* hv_cpu_init() can be called with IRQs disabled from hv_resume() */
 	flags = irqs_disabled() ? GFP_ATOMIC : GFP_KERNEL;
@@ -134,6 +136,16 @@ int hv_common_cpu_init(unsigned int cpu)
 	if (!(*inputarg))
 		return -ENOMEM;
 
+	if (hv_isolation_type_en_snp()) {
+		ret = set_memory_decrypted((unsigned long)*inputarg, 1);
+		if (ret) {
+			kfree(*inputarg);
+			return ret;
+		}
+
+		memset(*inputarg, 0x00, PAGE_SIZE);
+	}
+
 	if (hv_root_partition) {
 		outputarg = (void **)this_cpu_ptr(hyperv_pcpu_output_arg);
 		*outputarg = (char *)(*inputarg) + HV_HYP_PAGE_SIZE;
@@ -168,6 +180,9 @@ int hv_common_cpu_die(unsigned int cpu)
 
 	local_irq_restore(flags);
 
+	if (hv_isolation_type_en_snp())
+		set_memory_encrypted((unsigned long)mem, 1);
+
 	kfree(mem);
 
 	return 0;
Signature · 1 line
2.25.1
author
#35 lore raw

RFC PATCH v2 11/18 Drivers: hv: vmbus: Decrypt vmbus ring buffer

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

The ring buffer is remapped in the hv_ringbuffer_init() and it should be with decrypt flag in order to share it with hypervisor in sev-snp enlightened guest.

---
 drivers/hv/ring_buffer.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/hv/ring_buffer.c +6 -1
index 59a4aa86d1f3..391995c76be7 100644
@@ -18,6 +18,7 @@
 #include <linux/slab.h>
 #include <linux/prefetch.h>
 #include <linux/io.h>
+#include <linux/set_memory.h>
 #include <asm/mshyperv.h>
 
 #include "hyperv_vmbus.h"
@@ -233,14 +234,18 @@ int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,
 
 		ring_info->ring_buffer = (struct hv_ring_buffer *)
 			vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP,
+				hv_isolation_type_en_snp() ?
+				pgprot_decrypted(PAGE_KERNEL_NOENC) :
 				PAGE_KERNEL);
 
+		if (hv_isolation_type_en_snp())
+			memset(ring_info->ring_buffer, 0x00, page_cnt * PAGE_SIZE);
+
 		kfree(pages_wraparound);
 		if (!ring_info->ring_buffer)
 			return -ENOMEM;
 	}
 
-
 	ring_info->ring_buffer->read_index =
 		ring_info->ring_buffer->write_index = 0;
 
Signature · 1 line
2.25.1
author
#38 lore raw

RFC PATCH v2 12/18 x86/hyperv: Initialize cpu and memory for sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

Read processor amd memory info from specific address which are populated by Hyper-V. Initialize smp cpu related ops, pvalidate system memory and add it into e820 table.

---
 arch/x86/kernel/cpu/mshyperv.c | 75 ++++++++++++++++++++++++++++++++++
 1 file changed, 75 insertions(+)
arch/x86/kernel/cpu/mshyperv.c +75 -0
index 2ea4f21c6172..f0c97210c64a 100644
@@ -34,6 +34,12 @@
 #include <clocksource/hyperv_timer.h>
 #include <asm/numa.h>
 #include <asm/coco.h>
+#include <asm/io_apic.h>
+#include <asm/svm.h>
+#include <asm/sev.h>
+#include <asm/sev-snp.h>
+#include <asm/realmode.h>
+#include <asm/e820/api.h>
 
 /* Is Linux running as the root partition? */
 bool hv_root_partition;
@@ -253,6 +259,33 @@ static void __init hv_smp_prepare_cpus(unsigned int max_cpus)
 }
 #endif
 
+static __init int hv_snp_set_rtc_noop(const struct timespec64 *now) { return -EINVAL; }
+static __init void hv_snp_get_rtc_noop(struct timespec64 *now) { }
+
+static u32 processor_count;
+
+static __init void hv_snp_get_smp_config(unsigned int early)
+{
+	if (!early) {
+		while (num_processors < processor_count) {
+			early_per_cpu(x86_cpu_to_apicid, num_processors) = num_processors;
+			early_per_cpu(x86_bios_cpu_apicid, num_processors) = num_processors;
+			physid_set(num_processors, phys_cpu_present_map);
+			set_cpu_possible(num_processors, true);
+			set_cpu_present(num_processors, true);
+			num_processors++;
+		}
+	}
+}
+
+struct memory_map_entry {
+	u64 starting_gpn;
+	u64 numpages;
+	u16 type;
+	u16 flags;
+	u32 reserved;
+};
+
 static void __init ms_hyperv_init_platform(void)
 {
 	int hv_max_functions_eax;
@@ -260,6 +293,11 @@ static void __init ms_hyperv_init_platform(void)
 	int hv_host_info_ebx;
 	int hv_host_info_ecx;
 	int hv_host_info_edx;
+	struct memory_map_entry *entry;
+	struct e820_entry *e820_entry;
+	u64 e820_end;
+	u64 ram_end;
+	u64 page;
 
 #ifdef CONFIG_PARAVIRT
 	pv_info.name = "Hyper-V";
@@ -477,6 +515,43 @@ static void __init ms_hyperv_init_platform(void)
 	if (!(ms_hyperv.features & HV_ACCESS_TSC_INVARIANT))
 		mark_tsc_unstable("running on Hyper-V");
 
+	if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+		x86_platform.legacy.reserve_bios_regions = 0;
+		x86_platform.set_wallclock = hv_snp_set_rtc_noop;
+		x86_platform.get_wallclock = hv_snp_get_rtc_noop;
+		x86_init.resources.probe_roms = x86_init_noop;
+		x86_init.resources.reserve_resources = x86_init_noop;
+		x86_init.mpparse.find_smp_config = x86_init_noop;
+		x86_init.mpparse.get_smp_config = hv_snp_get_smp_config;
+
+		/*
+		 * Hyper-V SEV-SNP enlightened guest doesn't support ioapic
+		 * and legacy APIC page read/write. Switch to hv apic here.
+		 */
+		disable_ioapic_support();
+		hv_apic_init();
+
+		processor_count = *(u32 *)__va(EN_SEV_SNP_PROCESSOR_INFO_ADDR);
+
+		entry = (struct memory_map_entry *)(__va(EN_SEV_SNP_PROCESSOR_INFO_ADDR)
+				+ sizeof(struct memory_map_entry));
+
+		for (; entry->numpages != 0; entry++) {
+			e820_entry = &e820_table->entries[e820_table->nr_entries - 1];
+			e820_end = e820_entry->addr + e820_entry->size;
+			ram_end = (entry->starting_gpn + entry->numpages) * PAGE_SIZE;
+
+			if (e820_end < entry->starting_gpn * PAGE_SIZE)
+				e820_end = entry->starting_gpn * PAGE_SIZE;
+			if (e820_end < ram_end) {
+				pr_info("Hyper-V: add [mem %#018Lx-%#018Lx]\n", e820_end, ram_end - 1);
+				e820__range_add(e820_end, ram_end - e820_end, E820_TYPE_RAM);
+				for (page = e820_end; page < ram_end; page += PAGE_SIZE)
+					pvalidate((unsigned long)__va(page), RMP_PG_SIZE_4K, true);
+			}
+		}
+	}
+
 	hardlockup_detector_disable();
 }
 
Signature · 1 line
2.25.1
author
#40 lore raw

RFC PATCH v2 13/18 x86/hyperv: Add smp support for sev-snp guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

The wakeup_secondary_cpu callback was populated with wakeup_ cpu_via_vmgexit() which doesn't work for Hyper-V. Override it with Hyper-V specific hook which uses HVCALL_START_VIRTUAL_ PROCESSOR hvcall to start AP with vmsa data structure.

---
 arch/x86/include/asm/sev.h        |  13 +++
 arch/x86/include/asm/svm.h        |  55 ++++++++++-
 arch/x86/kernel/cpu/mshyperv.c    | 147 +++++++++++++++++++++++++++++-
 include/asm-generic/hyperv-tlfs.h |  18 ++++
 4 files changed, 230 insertions(+), 3 deletions(-)
arch/x86/include/asm/sev.h +13 -0
index ebc271bb6d8e..e34aaf730220 100644
@@ -86,6 +86,19 @@ extern bool handle_vc_boot_ghcb(struct pt_regs *regs);
 
 #define RMPADJUST_VMSA_PAGE_BIT		BIT(16)
 
+union sev_rmp_adjust {
+	u64 as_uint64;
+	struct {
+		unsigned long target_vmpl : 8;
+		unsigned long enable_read : 1;
+		unsigned long enable_write : 1;
+		unsigned long enable_user_execute : 1;
+		unsigned long enable_kernel_execute : 1;
+		unsigned long reserved1 : 4;
+		unsigned long vmsa : 1;
+	};
+};
+
 /* SNP Guest message request */
 struct snp_req_data {
 	unsigned long req_gpa;
arch/x86/include/asm/svm.h +54 -1

62 lines not on this page · raw message

arch/x86/kernel/cpu/mshyperv.c +145 -2

171 lines not on this page · raw message

include/asm-generic/hyperv-tlfs.h +18 -0
index 6e2a090e2649..7072adbf5540 100644
@@ -139,6 +139,7 @@ struct ms_hyperv_tsc_page {
 #define HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST	0x0003
 #define HVCALL_NOTIFY_LONG_SPIN_WAIT		0x0008
 #define HVCALL_SEND_IPI				0x000b
+#define HVCALL_ENABLE_VP_VTL			0x000f
 #define HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX	0x0013
 #define HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX	0x0014
 #define HVCALL_SEND_IPI_EX			0x0015
@@ -156,6 +157,7 @@ struct ms_hyperv_tsc_page {
 #define HVCALL_MAP_DEVICE_INTERRUPT		0x007c
 #define HVCALL_UNMAP_DEVICE_INTERRUPT		0x007d
 #define HVCALL_RETARGET_INTERRUPT		0x007e
+#define HVCALL_START_VIRTUAL_PROCESSOR		0x0099
 #define HVCALL_FLUSH_GUEST_PHYSICAL_ADDRESS_SPACE 0x00af
 #define HVCALL_FLUSH_GUEST_PHYSICAL_ADDRESS_LIST 0x00b0
 #define HVCALL_MODIFY_SPARSE_GPA_PAGE_HOST_VISIBILITY 0x00db
@@ -763,6 +765,22 @@ struct hv_input_unmap_device_interrupt {
 	struct hv_interrupt_entry interrupt_entry;
 } __packed;
 
+struct hv_enable_vp_vtl_input {
+	u64 partitionid;
+	u32 vpindex;
+	u8 targetvtl;
+	u8 padding[3];
+	u8 context[0xe0];
+} __packed;
+
+struct hv_start_virtual_processor_input {
+	u64 partitionid;
+	u32 vpindex;
+	u8 targetvtl;
+	u8 padding[3];
+	u8 context[0xe0];
+} __packed;
+
 #define HV_SOURCE_SHADOW_NONE               0x0
 #define HV_SOURCE_SHADOW_BRIDGE_BUS_RANGE   0x1
 
Signature · 1 line
2.25.1
author
#42 lore raw

RFC PATCH v2 14/18 x86/hyperv: Add hyperv-specific hadling for VMMCALL under SEV-ES

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

Add Hyperv-specific handling for faults caused by VMMCALL instructions.

---
 arch/x86/kernel/cpu/mshyperv.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
arch/x86/kernel/cpu/mshyperv.c +16 -0
index b266f648e5cd..a4e526378603 100644
@@ -725,6 +725,20 @@ static bool __init ms_hyperv_msi_ext_dest_id(void)
 	return eax & HYPERV_VS_PROPERTIES_EAX_EXTENDED_IOAPIC_RTE;
 }
 
+static void hv_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
+{
+	/* RAX and CPL are already in the GHCB */
+	ghcb_set_rcx(ghcb, regs->cx);
+	ghcb_set_rdx(ghcb, regs->dx);
+	ghcb_set_r8(ghcb, regs->r8);
+}
+
+static bool hv_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
+{
+	/* No checking of the return state needed */
+	return true;
+}
+
 const __initconst struct hypervisor_x86 x86_hyper_ms_hyperv = {
 	.name			= "Microsoft Hyper-V",
 	.detect			= ms_hyperv_platform,
@@ -732,4 +746,6 @@ const __initconst struct hypervisor_x86 x86_hyper_ms_hyperv = {
 	.init.x2apic_available	= ms_hyperv_x2apic_available,
 	.init.msi_ext_dest_id	= ms_hyperv_msi_ext_dest_id,
 	.init.init_platform	= ms_hyperv_init_platform,
+	.runtime.sev_es_hcall_prepare = hv_sev_es_hcall_prepare,
+	.runtime.sev_es_hcall_finish = hv_sev_es_hcall_finish,
 };
Signature · 1 line
2.25.1
author
#47 lore raw

RFC PATCH v2 16/18 x86/sev: Initialize #HV doorbell and handle interrupt requests

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

Enable #HV exception to handle interrupt requests from hypervisor.

---
 arch/x86/entry/entry_64.S          |  18 ++
 arch/x86/include/asm/irqflags.h    |  19 ++
 arch/x86/include/asm/mem_encrypt.h |   2 +
 arch/x86/include/asm/msr-index.h   |   6 +
 arch/x86/include/uapi/asm/svm.h    |   4 +
 arch/x86/kernel/sev.c              | 354 ++++++++++++++++++++++++-----
 arch/x86/kernel/traps.c            |  50 ++++
 7 files changed, 400 insertions(+), 53 deletions(-)
arch/x86/entry/entry_64.S +18 -0
index b2059df43c57..fe460cf44ab5 100644
@@ -1058,6 +1058,15 @@ SYM_CODE_END(paranoid_entry)
  * R15 - old SPEC_CTRL
  */
 SYM_CODE_START_LOCAL(paranoid_exit)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
 	UNWIND_HINT_REGS
 
 	/*
@@ -1183,6 +1192,15 @@ SYM_CODE_START_LOCAL(error_entry)
 SYM_CODE_END(error_entry)
 
 SYM_CODE_START_LOCAL(error_return)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
 	UNWIND_HINT_REGS
 	DEBUG_ENTRY_ASSERT_IRQS_OFF
 	testb	$3, CS(%rsp)
arch/x86/include/asm/irqflags.h +19 -0
index 7793e52d6237..e0730d8bc0ac 100644
@@ -14,6 +14,9 @@
 /*
  * Interrupt control:
  */
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+void check_hv_pending(struct pt_regs *regs);
+#endif
 
 /* Declaration required for gcc < 4.9 to prevent -Werror=missing-prototypes */
 extern inline unsigned long native_save_fl(void);
@@ -35,6 +38,19 @@ extern __always_inline unsigned long native_save_fl(void)
 	return flags;
 }
 
+extern inline void native_restore_fl(unsigned long flags)
+{
+	asm volatile("push %0 ; popf"
+		     : /* no output */
+		     : "g" (flags)
+		     : "memory", "cc");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	if ((flags & X86_EFLAGS_IF)) {
+		check_hv_pending(NULL);
+	}
+#endif
+}
+
 static __always_inline void native_irq_disable(void)
 {
 	asm volatile("cli": : :"memory");
@@ -43,6 +59,9 @@ static __always_inline void native_irq_disable(void)
 static __always_inline void native_irq_enable(void)
 {
 	asm volatile("sti": : :"memory");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	check_hv_pending(NULL);
+#endif
 }
 
 static inline __cpuidle void native_safe_halt(void)
arch/x86/include/asm/mem_encrypt.h +2 -0
index 72ca90552b6a..7264ca5f5b2d 100644
@@ -50,6 +50,7 @@ void __init early_set_mem_enc_dec_hypercall(unsigned long vaddr, int npages,
 void __init mem_encrypt_free_decrypted_mem(void);
 
 void __init sev_es_init_vc_handling(void);
+void __init sev_snp_init_hv_handling(void);
 
 #define __bss_decrypted __section(".bss..decrypted")
 
@@ -72,6 +73,7 @@ static inline void __init sme_encrypt_kernel(struct boot_params *bp) { }
 static inline void __init sme_enable(struct boot_params *bp) { }
 
 static inline void sev_es_init_vc_handling(void) { }
+static inline void sev_snp_init_hv_handling(void) { }
 
 static inline int __init
 early_set_memory_decrypted(unsigned long vaddr, unsigned long size) { return 0; }
arch/x86/include/asm/msr-index.h +6 -0
index 10ac52705892..6fe25a6e325f 100644
@@ -562,10 +562,16 @@
 #define MSR_AMD64_SEV_ENABLED_BIT	0
 #define MSR_AMD64_SEV_ES_ENABLED_BIT	1
 #define MSR_AMD64_SEV_SNP_ENABLED_BIT	2
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT		4
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT	5
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT	6
 #define MSR_AMD64_SEV_ENABLED		BIT_ULL(MSR_AMD64_SEV_ENABLED_BIT)
 #define MSR_AMD64_SEV_ES_ENABLED	BIT_ULL(MSR_AMD64_SEV_ES_ENABLED_BIT)
 #define MSR_AMD64_SEV_SNP_ENABLED	BIT_ULL(MSR_AMD64_SEV_SNP_ENABLED_BIT)
 
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED			BIT_ULL(MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT)
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT)
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT)
 #define MSR_AMD64_VIRT_SPEC_CTRL	0xc001011f
 
 /* AMD Collaborative Processor Performance Control MSRs */
arch/x86/include/uapi/asm/svm.h +4 -0
index f69c168391aa..85d6882262e7 100644
@@ -115,6 +115,10 @@
 #define SVM_VMGEXIT_AP_CREATE_ON_INIT		0
 #define SVM_VMGEXIT_AP_CREATE			1
 #define SVM_VMGEXIT_AP_DESTROY			2
+#define SVM_VMGEXIT_HV_DOORBELL_PAGE		0x80000014
+#define SVM_VMGEXIT_GET_PREFERRED_HV_DOORBELL_PAGE	0
+#define SVM_VMGEXIT_SET_HV_DOORBELL_PAGE		1
+#define SVM_VMGEXIT_QUERY_HV_DOORBELL_PAGE		2
 #define SVM_VMGEXIT_HV_FEATURES			0x8000fffd
 #define SVM_VMGEXIT_UNSUPPORTED_EVENT		0x8000ffff
 
arch/x86/kernel/sev.c +301 -53

414 lines not on this page · raw message

arch/x86/kernel/traps.c +50 -0

60 lines not on this page · raw message

Signature · 1 line
2.25.1
author
#43 lore raw

RFC PATCH v2 15/18 x86/sev: Add a #HV exception handler

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]>

Add a #HV exception handler that uses IST stack.

---
 arch/x86/entry/entry_64.S             | 58 +++++++++++++++++++++++++++
 arch/x86/include/asm/cpu_entry_area.h |  6 +++
 arch/x86/include/asm/idtentry.h       | 39 +++++++++++++++++-
 arch/x86/include/asm/page_64_types.h  |  1 +
 arch/x86/include/asm/trapnr.h         |  1 +
 arch/x86/include/asm/traps.h          |  1 +
 arch/x86/kernel/cpu/common.c          |  1 +
 arch/x86/kernel/dumpstack_64.c        |  9 ++++-
 arch/x86/kernel/idt.c                 |  1 +
 arch/x86/kernel/sev.c                 | 53 ++++++++++++++++++++++++
 arch/x86/mm/cpu_entry_area.c          |  2 +
 11 files changed, 169 insertions(+), 3 deletions(-)
arch/x86/entry/entry_64.S +58 -0
index 9953d966d124..b2059df43c57 100644
@@ -560,6 +560,64 @@ SYM_CODE_START(\asmsym)
 .Lfrom_usermode_switch_stack_\@:
 	idtentry_body user_\cfunc, has_error_code=1
 
+_ASM_NOKPROBE(\asmsym)
+SYM_CODE_END(\asmsym)
+.endm
+/*
+ * idtentry_hv - Macro to generate entry stub for #HV
+ * @vector:		Vector number
+ * @asmsym:		ASM symbol for the entry point
+ * @cfunc:		C function to be called
+ *
+ * The macro emits code to set up the kernel context for #HV. The #HV handler
+ * runs on an IST stack and needs to be able to support nested #HV exceptions.
+ *
+ * To make this work the #HV entry code tries its best to pretend it doesn't use
+ * an IST stack by switching to the task stack if coming from user-space (which
+ * includes early SYSCALL entry path) or back to the stack in the IRET frame if
+ * entered from kernel-mode.
+ *
+ * If entered from kernel-mode the return stack is validated first, and if it is
+ * not safe to use (e.g. because it points to the entry stack) the #HV handler
+ * will switch to a fall-back stack (HV2) and call a special handler function.
+ *
+ * The macro is only used for one vector, but it is planned to be extended in
+ * the future for the #HV exception.
+ */
+.macro idtentry_hv vector asmsym cfunc
+SYM_CODE_START(\asmsym)
+	UNWIND_HINT_IRET_REGS
+	ASM_CLAC
+	pushq	$-1			/* ORIG_RAX: no syscall to restart */
+
+	testb	$3, CS-ORIG_RAX(%rsp)
+	jnz	.Lfrom_usermode_switch_stack_\@
+
+	call	paranoid_entry
+
+	UNWIND_HINT_REGS
+
+	/*
+	 * Switch off the IST stack to make it free for nested exceptions.
+	 */
+	movq	%rsp, %rdi		/* pt_regs pointer */
+	call	hv_switch_off_ist
+	movq	%rax, %rsp		/* Switch to new stack */
+
+	UNWIND_HINT_REGS
+
+	/* Update pt_regs */
+	movq	ORIG_RAX(%rsp), %rsi	/* get error code into 2nd argument*/
+	movq	$-1, ORIG_RAX(%rsp)	/* no syscall to restart */
+
+	movq	%rsp, %rdi		/* pt_regs pointer */
+	call	kernel_\cfunc
+
+	jmp	paranoid_exit
+
+.Lfrom_usermode_switch_stack_\@:
+	idtentry_body user_\cfunc, has_error_code=1
+
 _ASM_NOKPROBE(\asmsym)
 SYM_CODE_END(\asmsym)
 .endm
arch/x86/include/asm/cpu_entry_area.h +6 -0
index 75efc4c6f076..f173a16cfc59 100644
@@ -30,6 +30,10 @@
 	char	VC_stack[optional_stack_size];			\
 	char	VC2_stack_guard[guardsize];			\
 	char	VC2_stack[optional_stack_size];			\
+	char	HV_stack_guard[guardsize];			\
+	char	HV_stack[optional_stack_size];			\
+	char	HV2_stack_guard[guardsize];			\
+	char	HV2_stack[optional_stack_size];			\
 	char	IST_top_guard[guardsize];			\
 
 /* The exception stacks' physical storage. No guard pages required */
@@ -52,6 +56,8 @@ enum exception_stack_ordering {
 	ESTACK_MCE,
 	ESTACK_VC,
 	ESTACK_VC2,
+	ESTACK_HV,
+	ESTACK_HV2,
 	N_EXCEPTION_STACKS
 };
 
arch/x86/include/asm/idtentry.h +38 -1
index 72184b0b2219..652fea10d377 100644
@@ -317,6 +317,19 @@ static __always_inline void __##func(struct pt_regs *regs)
 	__visible noinstr void kernel_##func(struct pt_regs *regs, unsigned long error_code);	\
 	__visible noinstr void   user_##func(struct pt_regs *regs, unsigned long error_code)
 
+
+/**
+ * DECLARE_IDTENTRY_HV - Declare functions for the HV entry point
+ * @vector:	Vector number (ignored for C)
+ * @func:	Function name of the entry point
+ *
+ * Maps to DECLARE_IDTENTRY_RAW, but declares also the user C handler.
+ */
+#define DECLARE_IDTENTRY_HV(vector, func)				\
+	DECLARE_IDTENTRY_RAW_ERRORCODE(vector, func);			\
+	__visible noinstr void kernel_##func(struct pt_regs *regs);	\
+	__visible noinstr void   user_##func(struct pt_regs *regs)
+
 /**
  * DEFINE_IDTENTRY_IST - Emit code for IST entry points
  * @func:	Function name of the entry point
@@ -376,6 +389,26 @@ static __always_inline void __##func(struct pt_regs *regs)
 #define DEFINE_IDTENTRY_VC_USER(func)				\
 	DEFINE_IDTENTRY_RAW_ERRORCODE(user_##func)
 
+/**
+ * DEFINE_IDTENTRY_HV_KERNEL - Emit code for HV injection handler
+ *			       when raised from kernel mode
+ * @func:	Function name of the entry point
+ *
+ * Maps to DEFINE_IDTENTRY_RAW
+ */
+#define DEFINE_IDTENTRY_HV_KERNEL(func)					\
+	DEFINE_IDTENTRY_RAW(kernel_##func)
+
+/**
+ * DEFINE_IDTENTRY_HV_USER - Emit code for HV injection handler
+ *			     when raised from user mode
+ * @func:	Function name of the entry point
+ *
+ * Maps to DEFINE_IDTENTRY_RAW
+ */
+#define DEFINE_IDTENTRY_HV_USER(func)					\
+	DEFINE_IDTENTRY_RAW(user_##func)
+
 #else	/* CONFIG_X86_64 */
 
 /**
@@ -465,6 +498,9 @@ __visible noinstr void func(struct pt_regs *regs,			\
 # define DECLARE_IDTENTRY_VC(vector, func)				\
 	idtentry_vc vector asm_##func func
 
+# define DECLARE_IDTENTRY_HV(vector, func)				\
+	idtentry_hv vector asm_##func func
+
 #else
 # define DECLARE_IDTENTRY_MCE(vector, func)				\
 	DECLARE_IDTENTRY(vector, func)
@@ -622,9 +658,10 @@ DECLARE_IDTENTRY_RAW_ERRORCODE(X86_TRAP_DF,	xenpv_exc_double_fault);
 DECLARE_IDTENTRY_ERRORCODE(X86_TRAP_CP,	exc_control_protection);
 #endif
 
-/* #VC */
+/* #VC & #HV */
 #ifdef CONFIG_AMD_MEM_ENCRYPT
 DECLARE_IDTENTRY_VC(X86_TRAP_VC,	exc_vmm_communication);
+DECLARE_IDTENTRY_HV(X86_TRAP_HV,	exc_hv_injection);
 #endif
 
 #ifdef CONFIG_XEN_PV
arch/x86/include/asm/page_64_types.h +1 -0
index e9e2c3ba5923..0bd7dab676c5 100644
@@ -29,6 +29,7 @@
 #define	IST_INDEX_DB		2
 #define	IST_INDEX_MCE		3
 #define	IST_INDEX_VC		4
+#define	IST_INDEX_HV		5
 
 /*
  * Set __PAGE_OFFSET to the most negative possible address +
arch/x86/include/asm/trapnr.h +1 -0
index f5d2325aa0b7..c6583631cecb 100644
@@ -26,6 +26,7 @@
 #define X86_TRAP_XF		19	/* SIMD Floating-Point Exception */
 #define X86_TRAP_VE		20	/* Virtualization Exception */
 #define X86_TRAP_CP		21	/* Control Protection Exception */
+#define X86_TRAP_HV		28	/* HV injected exception in SNP restricted mode */
 #define X86_TRAP_VC		29	/* VMM Communication Exception */
 #define X86_TRAP_IRET		32	/* IRET Exception */
 
arch/x86/include/asm/traps.h +1 -0
index 47ecfff2c83d..6795d3e517d6 100644
@@ -16,6 +16,7 @@ asmlinkage __visible notrace
 struct pt_regs *fixup_bad_iret(struct pt_regs *bad_regs);
 void __init trap_init(void);
 asmlinkage __visible noinstr struct pt_regs *vc_switch_off_ist(struct pt_regs *eregs);
+asmlinkage __visible noinstr struct pt_regs *hv_switch_off_ist(struct pt_regs *eregs);
 #endif
 
 extern bool ibt_selftest(void);
arch/x86/kernel/cpu/common.c +1 -0
index 3e508f239098..87afa3a4c8b1 100644
@@ -2165,6 +2165,7 @@ static inline void tss_setup_ist(struct tss_struct *tss)
 	tss->x86_tss.ist[IST_INDEX_MCE] = __this_cpu_ist_top_va(MCE);
 	/* Only mapped when SEV-ES is active */
 	tss->x86_tss.ist[IST_INDEX_VC] = __this_cpu_ist_top_va(VC);
+	tss->x86_tss.ist[IST_INDEX_HV] = __this_cpu_ist_top_va(HV);
 }
 
 #else /* CONFIG_X86_64 */
arch/x86/kernel/dumpstack_64.c +7 -2
index 6c5defd6569a..23aa5912c87a 100644
@@ -26,11 +26,14 @@ static const char * const exception_stack_names[] = {
 		[ ESTACK_MCE	]	= "#MC",
 		[ ESTACK_VC	]	= "#VC",
 		[ ESTACK_VC2	]	= "#VC2",
+		[ ESTACK_HV	]	= "#HV",
+		[ ESTACK_HV2	]	= "#HV2",
+		
 };
 
 const char *stack_type_name(enum stack_type type)
 {
-	BUILD_BUG_ON(N_EXCEPTION_STACKS != 6);
+	BUILD_BUG_ON(N_EXCEPTION_STACKS != 8);
 
 	if (type == STACK_TYPE_TASK)
 		return "TASK";
@@ -89,6 +92,8 @@ struct estack_pages estack_pages[CEA_ESTACK_PAGES] ____cacheline_aligned = {
 	EPAGERANGE(MCE),
 	EPAGERANGE(VC),
 	EPAGERANGE(VC2),
+	EPAGERANGE(HV),
+	EPAGERANGE(HV2),
 };
 
 static __always_inline bool in_exception_stack(unsigned long *stack, struct stack_info *info)
@@ -98,7 +103,7 @@ static __always_inline bool in_exception_stack(unsigned long *stack, struct stac
 	struct pt_regs *regs;
 	unsigned int k;
 
-	BUILD_BUG_ON(N_EXCEPTION_STACKS != 6);
+	BUILD_BUG_ON(N_EXCEPTION_STACKS != 8);
 
 	begin = (unsigned long)__this_cpu_read(cea_exception_stacks);
 	/*
arch/x86/kernel/idt.c +1 -0
index a58c6bc1cd68..48c0a7e1dbcb 100644
@@ -113,6 +113,7 @@ static const __initconst struct idt_data def_idts[] = {
 
 #ifdef CONFIG_AMD_MEM_ENCRYPT
 	ISTG(X86_TRAP_VC,		asm_exc_vmm_communication, IST_INDEX_VC),
+	ISTG(X86_TRAP_HV,		asm_exc_hv_injection, IST_INDEX_HV),
 #endif
 
 	SYSG(X86_TRAP_OF,		asm_exc_overflow),
arch/x86/kernel/sev.c +53 -0

59 lines not on this page · raw message

arch/x86/mm/cpu_entry_area.c +2 -0
index 6c2f1b76a0b6..8b1670fa2b81 100644
@@ -115,6 +115,8 @@ static void __init percpu_setup_exception_stacks(unsigned int cpu)
 		if (cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT)) {
 			cea_map_stack(VC);
 			cea_map_stack(VC2);
+			cea_map_stack(HV);
+			cea_map_stack(HV2);
 		}
 	}
 }
Signature · 1 line
2.25.1
author
#60 lore raw

RFC PATCH v2 17/18 x86/sev: optimize system vector processing invoked from #HV exception

In reply to Tianyu Lan

From: Ashish Kalra <[email protected]>

Construct system vector table and dispatch system vector exceptions through sysvec_table from #HV exception handler instead of explicitly calling each system vector. The system vector table is created dynamically and is placed in a new named ELF section.

---
 arch/x86/entry/entry_64.S     |  6 +++
 arch/x86/kernel/sev.c         | 70 +++++++++++++----------------------
 arch/x86/kernel/vmlinux.lds.S |  7 ++++
 3 files changed, 38 insertions(+), 45 deletions(-)
arch/x86/entry/entry_64.S +6 -0
index fe460cf44ab5..05f4fcc60652 100644
@@ -409,6 +409,12 @@ SYM_CODE_START(\asmsym)
 
 _ASM_NOKPROBE(\asmsym)
 SYM_CODE_END(\asmsym)
+	.if \vector >= FIRST_SYSTEM_VECTOR && \vector < NR_VECTORS
+		.section .system_vectors, "aw"
+		.byte \vector
+		.quad \cfunc
+		.previous
+	.endif
 .endm
 
 /*
arch/x86/kernel/sev.c +25 -45

94 lines not on this page · raw message

arch/x86/kernel/vmlinux.lds.S +7 -0
index 15f29053cec4..160c7cdaa3de 100644
@@ -322,6 +322,13 @@ SECTIONS
 		*(.altinstr_replacement)
 	}
 
+	. = ALIGN(8);
+	.system_vectors : AT(ADDR(.system_vectors) - LOAD_OFFSET) {
+		__system_vectors = .;
+		*(.system_vectors)
+		__system_vectors_end = .;
+	}
+
 	. = ALIGN(8);
 	.apicdrivers : AT(ADDR(.apicdrivers) - LOAD_OFFSET) {
 		__apicdrivers = .;
Signature · 1 line
2.25.1
author
#61 lore raw

RFC PATCH v2 18/18 x86/sev: Fix interrupt exit code paths from #HV exception

In reply to Tianyu Lan

From: Ashish Kalra <[email protected]>

Add checks in interrupt exit code paths in case of returns to user mode to check if currently executing the #HV handler then don't follow the irqentry_exit_to_user_mode path as that can potentially cause the #HV handler to be preempted and rescheduled on another CPU. Rescheduled #HV handler on another cpu will cause interrupts to be handled on a different cpu than the injected one, causing invalid EOIs and missed/lost guest interrupts and corresponding hangs and/or per-cpu IRQs handled on non-intended cpu.

---
 arch/x86/include/asm/idtentry.h | 66 +++++++++++++++++++++++++++++++++
 arch/x86/kernel/sev.c           | 30 +++++++++++++++
 2 files changed, 96 insertions(+)
arch/x86/include/asm/idtentry.h +66 -0
index 652fea10d377..45b47132be7c 100644
@@ -13,6 +13,10 @@
 
 #include <asm/irq_stack.h>
 
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+noinstr void irqentry_exit_hv_cond(struct pt_regs *regs, irqentry_state_t state);
+#endif
+
 /**
  * DECLARE_IDTENTRY - Declare functions for simple IDT entry points
  *		      No error code pushed by hardware
@@ -176,6 +180,7 @@ __visible noinstr void func(struct pt_regs *regs, unsigned long error_code)
 #define DECLARE_IDTENTRY_IRQ(vector, func)				\
 	DECLARE_IDTENTRY_ERRORCODE(vector, func)
 
+#ifndef CONFIG_AMD_MEM_ENCRYPT
 /**
  * DEFINE_IDTENTRY_IRQ - Emit code for device interrupt IDT entry points
  * @func:	Function name of the entry point
@@ -205,6 +210,26 @@ __visible noinstr void func(struct pt_regs *regs,			\
 }									\
 									\
 static noinline void __##func(struct pt_regs *regs, u32 vector)
+#else
+
+#define DEFINE_IDTENTRY_IRQ(func)					\
+static void __##func(struct pt_regs *regs, u32 vector);		\
+									\
+__visible noinstr void func(struct pt_regs *regs,			\
+			    unsigned long error_code)			\
+{									\
+	irqentry_state_t state = irqentry_enter(regs);			\
+	u32 vector = (u32)(u8)error_code;				\
+									\
+	instrumentation_begin();					\
+	kvm_set_cpu_l1tf_flush_l1d();					\
+	run_irq_on_irqstack_cond(__##func, regs, vector);		\
+	instrumentation_end();						\
+	irqentry_exit_hv_cond(regs, state);				\
+}									\
+									\
+static noinline void __##func(struct pt_regs *regs, u32 vector)
+#endif
 
 /**
  * DECLARE_IDTENTRY_SYSVEC - Declare functions for system vector entry points
@@ -221,6 +246,7 @@ static noinline void __##func(struct pt_regs *regs, u32 vector)
 #define DECLARE_IDTENTRY_SYSVEC(vector, func)				\
 	DECLARE_IDTENTRY(vector, func)
 
+#ifndef CONFIG_AMD_MEM_ENCRYPT
 /**
  * DEFINE_IDTENTRY_SYSVEC - Emit code for system vector IDT entry points
  * @func:	Function name of the entry point
@@ -245,6 +271,26 @@ __visible noinstr void func(struct pt_regs *regs)			\
 }									\
 									\
 static noinline void __##func(struct pt_regs *regs)
+#else
+
+#define DEFINE_IDTENTRY_SYSVEC(func)					\
+static void __##func(struct pt_regs *regs);				\
+									\
+__visible noinstr void func(struct pt_regs *regs)			\
+{									\
+	irqentry_state_t state = irqentry_enter(regs);			\
+									\
+	instrumentation_begin();					\
+	kvm_set_cpu_l1tf_flush_l1d();					\
+	run_sysvec_on_irqstack_cond(__##func, regs);			\
+	instrumentation_end();						\
+	irqentry_exit_hv_cond(regs, state);				\
+}									\
+									\
+static noinline void __##func(struct pt_regs *regs)
+#endif
+
+#ifndef CONFIG_AMD_MEM_ENCRYPT
 
 /**
  * DEFINE_IDTENTRY_SYSVEC_SIMPLE - Emit code for simple system vector IDT
@@ -274,6 +320,26 @@ __visible noinstr void func(struct pt_regs *regs)			\
 }									\
 									\
 static __always_inline void __##func(struct pt_regs *regs)
+#else
+
+#define DEFINE_IDTENTRY_SYSVEC_SIMPLE(func)				\
+static __always_inline void __##func(struct pt_regs *regs);		\
+									\
+__visible noinstr void func(struct pt_regs *regs)			\
+{									\
+	irqentry_state_t state = irqentry_enter(regs);			\
+									\
+	instrumentation_begin();					\
+	__irq_enter_raw();						\
+	kvm_set_cpu_l1tf_flush_l1d();					\
+	__##func(regs);						\
+	__irq_exit_raw();						\
+	instrumentation_end();						\
+	irqentry_exit_hv_cond(regs, state);				\
+}									\
+									\
+static __always_inline void __##func(struct pt_regs *regs)
+#endif
 
 /**
  * DECLARE_IDTENTRY_XENCB - Declare functions for XEN HV callback entry point
arch/x86/kernel/sev.c +30 -0
index 5a2f59022c98..ef6a123c50fe 100644
@@ -153,6 +153,10 @@ struct sev_hv_doorbell_page {
 
 struct sev_snp_runtime_data {
 	struct sev_hv_doorbell_page hv_doorbell_page;
+	/*
+	 * Indication that we are currently handling #HV events.
+	 */
+	bool hv_handling_events;
 };
 
 static DEFINE_PER_CPU(struct sev_snp_runtime_data*, snp_runtime_data);
@@ -206,6 +210,8 @@ static void do_exc_hv(struct pt_regs *regs)
 	union hv_pending_events pending_events;
 	u8 vector;
 
+	this_cpu_read(snp_runtime_data)->hv_handling_events = true;
+
 	while (sev_hv_pending()) {
 		asm volatile("cli" : : : "memory");
 
@@ -244,6 +250,8 @@ static void do_exc_hv(struct pt_regs *regs)
 
 		asm volatile("sti" : : : "memory");
 	}
+
+	this_cpu_read(snp_runtime_data)->hv_handling_events = false;
 }
 
 void check_hv_pending(struct pt_regs *regs)
@@ -2541,3 +2549,25 @@ static int __init snp_init_platform_device(void)
 	return 0;
 }
 device_initcall(snp_init_platform_device);
+
+noinstr void irqentry_exit_hv_cond(struct pt_regs *regs, irqentry_state_t state)
+{
+	/*
+	 * Check whether this returns to user mode, if so and if
+	 * we are currently executing the #HV handler then we don't
+	 * want to follow the irqentry_exit_to_user_mode path as
+	 * that can potentially cause the #HV handler to be
+	 * preempted and rescheduled on another CPU. Rescheduled #HV
+	 * handler on another cpu will cause interrupts to be handled
+	 * on a different cpu than the injected one, causing
+	 * invalid EOIs and missed/lost guest interrupts and
+	 * corresponding hangs and/or per-cpu IRQs handled on
+	 * non-intended cpu.
+	 */
+	if (user_mode(regs) &&
+	    this_cpu_read(snp_runtime_data)->hv_handling_events)
+		return;
+
+	/* follow normal interrupt return/exit path */
+	irqentry_exit(regs, state);
+}
Signature · 1 line
2.25.1
#48 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

In reply to Tianyu Lan

Hello Tianyu,

Quoted 235 lines · Tianyu Lan

From: Tianyu Lan <[email protected]>

Enable #HV exception to handle interrupt requests from hypervisor.

---
  arch/x86/entry/entry_64.S          |  18 ++
  arch/x86/include/asm/irqflags.h    |  19 ++
  arch/x86/include/asm/mem_encrypt.h |   2 +
  arch/x86/include/asm/msr-index.h   |   6 +
  arch/x86/include/uapi/asm/svm.h    |   4 +
  arch/x86/kernel/sev.c              | 354 ++++++++++++++++++++++++-----
  arch/x86/kernel/traps.c            |  50 ++++
  7 files changed, 400 insertions(+), 53 deletions(-)
arch/x86/entry/entry_64.S +18 -0
index b2059df43c57..fe460cf44ab5 100644
@@ -1058,6 +1058,15 @@ SYM_CODE_END(paranoid_entry)
   * R15 - old SPEC_CTRL
   */
  SYM_CODE_START_LOCAL(paranoid_exit)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
  	UNWIND_HINT_REGS
 
  	/*
@@ -1183,6 +1192,15 @@ SYM_CODE_START_LOCAL(error_entry)
  SYM_CODE_END(error_entry)
 
  SYM_CODE_START_LOCAL(error_return)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
  	UNWIND_HINT_REGS
  	DEBUG_ENTRY_ASSERT_IRQS_OFF
  	testb	$3, CS(%rsp)
arch/x86/include/asm/irqflags.h +19 -0
index 7793e52d6237..e0730d8bc0ac 100644
@@ -14,6 +14,9 @@
  /*
   * Interrupt control:
   */
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+void check_hv_pending(struct pt_regs *regs);
+#endif
 
  /* Declaration required for gcc < 4.9 to prevent -Werror=missing-prototypes */
  extern inline unsigned long native_save_fl(void);
@@ -35,6 +38,19 @@ extern __always_inline unsigned long native_save_fl(void)
  	return flags;
  }
 
+extern inline void native_restore_fl(unsigned long flags)
+{
+	asm volatile("push %0 ; popf"
+		     : /* no output */
+		     : "g" (flags)
+		     : "memory", "cc");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	if ((flags & X86_EFLAGS_IF)) {
+		check_hv_pending(NULL);
+	}
+#endif
+}
+
  static __always_inline void native_irq_disable(void)
  {
  	asm volatile("cli": : :"memory");
@@ -43,6 +59,9 @@ static __always_inline void native_irq_disable(void)
  static __always_inline void native_irq_enable(void)
  {
  	asm volatile("sti": : :"memory");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	check_hv_pending(NULL);
+#endif
  }
 
  static inline __cpuidle void native_safe_halt(void)
arch/x86/include/asm/mem_encrypt.h +2 -0
index 72ca90552b6a..7264ca5f5b2d 100644
@@ -50,6 +50,7 @@ void __init early_set_mem_enc_dec_hypercall(unsigned long vaddr, int npages,
  void __init mem_encrypt_free_decrypted_mem(void);
 
  void __init sev_es_init_vc_handling(void);
+void __init sev_snp_init_hv_handling(void);
 
  #define __bss_decrypted __section(".bss..decrypted")
 
@@ -72,6 +73,7 @@ static inline void __init sme_encrypt_kernel(struct boot_params *bp) { }
  static inline void __init sme_enable(struct boot_params *bp) { }
 
  static inline void sev_es_init_vc_handling(void) { }
+static inline void sev_snp_init_hv_handling(void) { }
 
  static inline int __init
  early_set_memory_decrypted(unsigned long vaddr, unsigned long size) { return 0; }
arch/x86/include/asm/msr-index.h +6 -0
index 10ac52705892..6fe25a6e325f 100644
@@ -562,10 +562,16 @@
  #define MSR_AMD64_SEV_ENABLED_BIT	0
  #define MSR_AMD64_SEV_ES_ENABLED_BIT	1
  #define MSR_AMD64_SEV_SNP_ENABLED_BIT	2
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT		4
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT	5
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT	6
  #define MSR_AMD64_SEV_ENABLED		BIT_ULL(MSR_AMD64_SEV_ENABLED_BIT)
  #define MSR_AMD64_SEV_ES_ENABLED	BIT_ULL(MSR_AMD64_SEV_ES_ENABLED_BIT)
  #define MSR_AMD64_SEV_SNP_ENABLED	BIT_ULL(MSR_AMD64_SEV_SNP_ENABLED_BIT)
 
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED			BIT_ULL(MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT)
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT)
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT)
  #define MSR_AMD64_VIRT_SPEC_CTRL	0xc001011f
 
  /* AMD Collaborative Processor Performance Control MSRs */
arch/x86/include/uapi/asm/svm.h +4 -0
index f69c168391aa..85d6882262e7 100644
@@ -115,6 +115,10 @@
  #define SVM_VMGEXIT_AP_CREATE_ON_INIT		0
  #define SVM_VMGEXIT_AP_CREATE			1
  #define SVM_VMGEXIT_AP_DESTROY			2
+#define SVM_VMGEXIT_HV_DOORBELL_PAGE		0x80000014
+#define SVM_VMGEXIT_GET_PREFERRED_HV_DOORBELL_PAGE	0
+#define SVM_VMGEXIT_SET_HV_DOORBELL_PAGE		1
+#define SVM_VMGEXIT_QUERY_HV_DOORBELL_PAGE		2
  #define SVM_VMGEXIT_HV_FEATURES			0x8000fffd
  #define SVM_VMGEXIT_UNSUPPORTED_EVENT		0x8000ffff
 
arch/x86/kernel/sev.c +67 -0
index b54ee3ba37b0..23cd025f97dc 100644
@@ -104,6 +104,12 @@ struct sev_es_runtime_data {
  	 * is currently unsupported in SEV-ES guests.
  	 */
  	unsigned long dr7;
+	/*
+	 * SEV-SNP requires that the GHCB must be registered before using it.
+	 * The flag below will indicate whether the GHCB is registered, if its
+	 * not registered then sev_es_get_ghcb() will perform the registration.
+	 */
+	bool ghcb_registered;
  };
 
  struct ghcb_state {
@@ -122,6 +128,183 @@ struct sev_config {
 
  static struct sev_config sev_cfg __read_mostly;
 
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state);
+static noinstr void __sev_put_ghcb(struct ghcb_state *state);
+static int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa);
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb);
+
+union hv_pending_events {
+	u16 events;
+	struct {
+		u8 vector;
+		u8 nmi : 1;
+		u8 mc : 1;
+		u8 reserved1 : 5;
+		u8 no_further_signal : 1;
+	};
+};
+
+struct sev_hv_doorbell_page {
+	union hv_pending_events pending_events;
+	u8 no_eoi_required;
+	u8 reserved2[61];
+	u8 padding[4032];
+};
+
+struct sev_snp_runtime_data {
+	struct sev_hv_doorbell_page hv_doorbell_page;
+};
+
+static DEFINE_PER_CPU(struct sev_snp_runtime_data*, snp_runtime_data);
+
+static inline u64 sev_es_rd_ghcb_msr(void)
+{
+	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
+}
+
+static __always_inline void sev_es_wr_ghcb_msr(u64 val)
+{
+	u32 low, high;
+
+	low  = (u32)(val);
+	high = (u32)(val >> 32);
+
+	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+}
+
+struct sev_hv_doorbell_page *sev_snp_current_doorbell_page(void)
+{
+	return &this_cpu_read(snp_runtime_data)->hv_doorbell_page;
+}
+
+static u8 sev_hv_pending(void)
+{
+	return sev_snp_current_doorbell_page()->pending_events.events;
+}
+
+static void hv_doorbell_apic_eoi_write(u32 reg, u32 val)
+{
+	if (xchg(&sev_snp_current_doorbell_page()->no_eoi_required, 0) & 0x1)
+		return;
+
+	BUG_ON(reg != APIC_EOI);
+	apic->write(reg, val);

arch/x86/kernel/sev.c struct sev_config { · line 201

Quoted 10 lines
arch/x86/kernel/sev.c +10 -0
+}
+
+static void do_exc_hv(struct pt_regs *regs)
+{
+	union hv_pending_events pending_events;
+	u8 vector;
+
+	while (sev_hv_pending()) {
+		asm volatile("cli" : : : "memory");
+

Do we really need to disable interrupts here, #HV exception will be dispatched via an interrupt gate in the IDT, so interrupts should be implicitly disabled, right ?

Quoted 64 lines
arch/x86/kernel/sev.c +64 -0
+		pending_events.events = xchg(
+			&sev_snp_current_doorbell_page()->pending_events.events,
+			0);
+
+		if (pending_events.nmi)
+			exc_nmi(regs);
+
+#ifdef CONFIG_X86_MCE
+		if (pending_events.mc)
+			exc_machine_check(regs);
+#endif
+
+		if (!pending_events.vector)
+			return;
+
+		if (pending_events.vector < FIRST_EXTERNAL_VECTOR) {
+			/* Exception vectors */
+			WARN(1, "exception shouldn't happen\n");
+		} else if (pending_events.vector == FIRST_EXTERNAL_VECTOR) {
+			sysvec_irq_move_cleanup(regs);
+		} else if (pending_events.vector == IA32_SYSCALL_VECTOR) {
+			WARN(1, "syscall shouldn't happen\n");
+		} else if (pending_events.vector >= FIRST_SYSTEM_VECTOR) {
+			switch (pending_events.vector) {
+#if IS_ENABLED(CONFIG_HYPERV)
+			case HYPERV_STIMER0_VECTOR:
+				sysvec_hyperv_stimer0(regs);
+				break;
+			case HYPERVISOR_CALLBACK_VECTOR:
+				sysvec_hyperv_callback(regs);
+				break;
+#endif
+#ifdef CONFIG_SMP
+			case RESCHEDULE_VECTOR:
+				sysvec_reschedule_ipi(regs);
+				break;
+			case IRQ_MOVE_CLEANUP_VECTOR:
+				sysvec_irq_move_cleanup(regs);
+				break;
+			case REBOOT_VECTOR:
+				sysvec_reboot(regs);
+				break;
+			case CALL_FUNCTION_SINGLE_VECTOR:
+				sysvec_call_function_single(regs);
+				break;
+			case CALL_FUNCTION_VECTOR:
+				sysvec_call_function(regs);
+				break;
+#endif
+#ifdef CONFIG_X86_LOCAL_APIC
+			case ERROR_APIC_VECTOR:
+				sysvec_error_interrupt(regs);
+				break;
+			case SPURIOUS_APIC_VECTOR:
+				sysvec_spurious_apic_interrupt(regs);
+				break;
+			case LOCAL_TIMER_VECTOR:
+				sysvec_apic_timer_interrupt(regs);
+				break;
+			case X86_PLATFORM_IPI_VECTOR:
+				sysvec_x86_platform_ipi(regs);
+				break;
+#endif
+			case 0x0:

arch/x86/kernel/sev.c struct sev_config { · line 275

Quoted 10 lines
arch/x86/kernel/sev.c +10 -0
+				break;
+			default:
+				panic("Unexpected vector %d\n", vector);
+				unreachable();
+			}
+		} else {
+			common_interrupt(regs, pending_events.vector);
+		}
+
+		asm volatile("sti" : : : "memory");

Again, why do we need to re-enable interrupts here (in this loop), interrupts will get re-enabled in the irqentry_exit() code path ?

Thanks,
Ashish

Quoted 327 lines
arch/x86/kernel/sev.c +150 -53
+	}
+}
+
+void check_hv_pending(struct pt_regs *regs)
+{
+	struct pt_regs local_regs;
+
+	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
+		return;
+
+	if (regs) {
+		if ((regs->flags & X86_EFLAGS_IF) == 0)
+			return;
+
+		if (!sev_hv_pending())
+			return;
+
+		do_exc_hv(regs);
+	} else {
+		if (sev_hv_pending()) {
+			memset(&local_regs, 0, sizeof(struct pt_regs));
+			regs = &local_regs;
+			asm volatile("movl %%cs, %%eax;" : "=a" (regs->cs));
+			asm volatile("movl %%ss, %%eax;" : "=a" (regs->ss));
+			regs->orig_ax = 0xffffffff;
+			regs->flags = native_save_fl();
+			do_exc_hv(regs);
+		}
+	}
+}
+EXPORT_SYMBOL_GPL(check_hv_pending);
+
  static __always_inline bool on_vc_stack(struct pt_regs *regs)
  {
  	unsigned long sp = regs->sp;
@@ -193,68 +376,35 @@ void noinstr __sev_es_ist_exit(void)
  	this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], *(unsigned long *)ist);
  }
 
-/*
- * Nothing shall interrupt this code path while holding the per-CPU
- * GHCB. The backup GHCB is only for NMIs interrupting this path.
- *
- * Callers must disable local interrupts around it.
- */
-static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+static bool sev_restricted_injection_enabled(void)
  {
-	struct sev_es_runtime_data *data;
+	return sev_status & MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED;
+}
+
+void __init sev_snp_init_hv_handling(void)
+{
+	struct sev_snp_runtime_data *snp_data;
+	struct ghcb_state state;
  	struct ghcb *ghcb;
+	unsigned long flags;
+	int cpu;
+	int err;
 
  	WARN_ON(!irqs_disabled());
+	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP) || !sev_restricted_injection_enabled())
+		return;
 
-	data = this_cpu_read(runtime_data);
-	ghcb = &data->ghcb_page;
-
-	if (unlikely(data->ghcb_active)) {
-		/* GHCB is already in use - save its contents */
-
-		if (unlikely(data->backup_ghcb_active)) {
-			/*
-			 * Backup-GHCB is also already in use. There is no way
-			 * to continue here so just kill the machine. To make
-			 * panic() work, mark GHCBs inactive so that messages
-			 * can be printed out.
-			 */
-			data->ghcb_active        = false;
-			data->backup_ghcb_active = false;
-
-			instrumentation_begin();
-			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
-			instrumentation_end();
-		}
-
-		/* Mark backup_ghcb active before writing to it */
-		data->backup_ghcb_active = true;
-
-		state->ghcb = &data->backup_ghcb;
-
-		/* Backup GHCB content */
-		*state->ghcb = *ghcb;
-	} else {
-		state->ghcb = NULL;
-		data->ghcb_active = true;
-	}
+	local_irq_save(flags);
 
-	return ghcb;
-}
+	ghcb = __sev_get_ghcb(&state);
 
-static inline u64 sev_es_rd_ghcb_msr(void)
-{
-	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
-}
+	sev_snp_setup_hv_doorbell_page(ghcb);
 
-static __always_inline void sev_es_wr_ghcb_msr(u64 val)
-{
-	u32 low, high;
+	__sev_put_ghcb(&state);
 
-	low  = (u32)(val);
-	high = (u32)(val >> 32);
+	apic_set_eoi_write(hv_doorbell_apic_eoi_write);
 
-	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+	local_irq_restore(flags);
  }
 
  static int vc_fetch_insn_kernel(struct es_em_ctxt *ctxt,
@@ -515,6 +665,79 @@ static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt
  /* Include code shared with pre-decompression boot stage */
  #include "sev-shared.c"
 
+/*
+ * Nothing shall interrupt this code path while holding the per-CPU
+ * GHCB. The backup GHCB is only for NMIs interrupting this path.
+ *
+ * Callers must disable local interrupts around it.
+ */
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+{
+	struct sev_es_runtime_data *data;
+	struct ghcb *ghcb;
+
+	WARN_ON(!irqs_disabled());
+
+	data = this_cpu_read(runtime_data);
+	ghcb = &data->ghcb_page;
+
+	if (unlikely(data->ghcb_active)) {
+		/* GHCB is already in use - save its contents */
+
+		if (unlikely(data->backup_ghcb_active)) {
+			/*
+			 * Backup-GHCB is also already in use. There is no way
+			 * to continue here so just kill the machine. To make
+			 * panic() work, mark GHCBs inactive so that messages
+			 * can be printed out.
+			 */
+			data->ghcb_active        = false;
+			data->backup_ghcb_active = false;
+
+			instrumentation_begin();
+			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
+			instrumentation_end();
+		}
+
+		/* Mark backup_ghcb active before writing to it */
+		data->backup_ghcb_active = true;
+
+		state->ghcb = &data->backup_ghcb;
+
+		/* Backup GHCB content */
+		*state->ghcb = *ghcb;
+	} else {
+		state->ghcb = NULL;
+		data->ghcb_active = true;
+	}
+
+	/* SEV-SNP guest requires that GHCB must be registered before using it. */
+	if (!data->ghcb_registered) {
+		if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+			snp_register_ghcb_early(__pa(ghcb));
+			sev_snp_setup_hv_doorbell_page(ghcb);
+		} else {
+			sev_es_wr_ghcb_msr(__pa(ghcb));
+		}
+		data->ghcb_registered = true;
+	}
+
+	return ghcb;
+}
+
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb)
+{
+	u64 pa;
+	enum es_result ret;
+
+	pa = __pa(sev_snp_current_doorbell_page());
+	vc_ghcb_invalidate(ghcb);
+	ret = vmgexit_hv_doorbell_page(ghcb,
+			SVM_VMGEXIT_SET_HV_DOORBELL_PAGE, pa);
+	if (ret != ES_OK)
+		panic("SEV-SNP: failed to set up #HV doorbell page");
+}
+
  static noinstr void __sev_put_ghcb(struct ghcb_state *state)
  {
  	struct sev_es_runtime_data *data;
@@ -1282,6 +1505,11 @@ void setup_ghcb(void)
  		snp_register_ghcb_early(__pa(&boot_ghcb_page));
  }
 
+int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa)
+{
+	return sev_es_ghcb_hv_call(ghcb, NULL, SVM_VMGEXIT_HV_DOORBELL_PAGE, op, pa);
+}
+
  #ifdef CONFIG_HOTPLUG_CPU
  static void sev_es_ap_hlt_loop(void)
  {
@@ -1355,6 +1583,7 @@ static void __init alloc_runtime_data(int cpu)
  static void __init init_ghcb(int cpu)
  {
  	struct sev_es_runtime_data *data;
+	struct sev_snp_runtime_data *snp_data;
  	int err;
 
  	data = per_cpu(runtime_data, cpu);
@@ -1366,8 +1595,22 @@ static void __init init_ghcb(int cpu)
 
  	memset(&data->ghcb_page, 0, sizeof(data->ghcb_page));
 
+	snp_data = memblock_alloc(sizeof(*snp_data), PAGE_SIZE);
+	if (!snp_data)
+		panic("Can't allocate SEV-SNP runtime data");
+
+	err = early_set_memory_decrypted((unsigned long)&snp_data->hv_doorbell_page,
+					 sizeof(snp_data->hv_doorbell_page));
+	if (err)
+		panic("Can't map #HV doorbell pages unencrypted");
+
+	memset(&snp_data->hv_doorbell_page, 0, sizeof(snp_data->hv_doorbell_page));
+
+	per_cpu(snp_runtime_data, cpu) = snp_data;
+
  	data->ghcb_active = false;
  	data->backup_ghcb_active = false;
+	data->ghcb_registered = false;
  }
 
  void __init sev_es_init_vc_handling(void)
@@ -2006,7 +2249,12 @@ DEFINE_IDTENTRY_VC_USER(exc_vmm_communication)
 
  static bool hv_raw_handle_exception(struct pt_regs *regs)
  {
-	return false;
+	/* Clear the no_further_signal bit */
+	sev_snp_current_doorbell_page()->pending_events.events &= 0x7fff;
+
+	check_hv_pending(regs);
+
+	return true;
  }
 
  static __always_inline bool on_hv_fallback_stack(struct pt_regs *regs)
arch/x86/kernel/traps.c +50 -0
index 178015a820f0..af97e6610fbb 100644
@@ -898,6 +898,53 @@ asmlinkage __visible noinstr struct pt_regs *vc_switch_off_ist(struct pt_regs *r
 
  	return regs_ret;
  }
+
+asmlinkage __visible noinstr struct pt_regs *hv_switch_off_ist(struct pt_regs *regs)
+{
+	unsigned long sp, *stack;
+	struct stack_info info;
+	struct pt_regs *regs_ret;
+
+	/*
+	 * A malicious hypervisor can inject 2 HVs in a row, which will corrupt
+	 * the trap frame on our IST stack.  We add a defensive check here to
+	 * catch such behavior.
+	 */
+	BUG_ON(regs->sp >= __this_cpu_ist_bottom_va(HV) && regs->sp < __this_cpu_ist_top_va(HV));
+
+	/*
+	 * In the SYSCALL entry path the RSP value comes from user-space - don't
+	 * trust it and switch to the current kernel stack
+	 */
+	if (ip_within_syscall_gap(regs)) {
+		sp = this_cpu_read(cpu_current_top_of_stack);
+		goto sync;
+	}
+
+	/*
+	 * From here on the RSP value is trusted. Now check whether entry
+	 * happened from a safe stack. Not safe are the entry or unknown stacks,
+	 * use the fall-back stack instead in this case.
+	 */
+	sp    = regs->sp;
+	stack = (unsigned long *)sp;
+
+	if (!get_stack_info_noinstr(stack, current, &info) || info.type == STACK_TYPE_ENTRY ||
+	    info.type > STACK_TYPE_EXCEPTION_LAST)
+		sp = __this_cpu_ist_top_va(HV2);
+sync:
+	/*
+	 * Found a safe stack - switch to it as if the entry didn't happen via
+	 * IST stack. The code below only copies pt_regs, the real switch happens
+	 * in assembly code.
+	 */
+	sp = ALIGN_DOWN(sp, 8) - sizeof(*regs_ret);
+
+	regs_ret = (struct pt_regs *)sp;
+	*regs_ret = *regs;
+
+	return regs_ret;
+}
  #endif
 
  asmlinkage __visible noinstr struct pt_regs *fixup_bad_iret(struct pt_regs *bad_regs)
@@ -1457,4 +1504,7 @@ void __init trap_init(void)
  	/* Setup traps as cpu_init() might #GP */
  	idt_setup_traps();
  	cpu_init();
+
+	/* Init #HV doorbell pages when running as an SEV-SNP guest */
+	sev_snp_init_hv_handling();
  }
author
#49 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

Quoted 19 lines · Ashish

arch/x86/kernel/sev.c struct sev_config { · line 201

Quoted 8 lines
arch/x86/kernel/sev.c +8 -0
+static void do_exc_hv(struct pt_regs *regs)
+{
+    union hv_pending_events pending_events;
+    u8 vector;
+
+    while (sev_hv_pending()) {
+        asm volatile("cli" : : : "memory");
+

Do we really need to disable interrupts here, #HV exception will be dispatched via an interrupt gate in the IDT, so interrupts should be implicitly disabled, right ?

Quoted 8 lines

   panic("Unexpected vector %d\n", vector);
+                unreachable();
+            }
+        } else {
+            common_interrupt(regs, pending_events.vector); +        }
+
+        asm volatile("sti" : : : "memory");

Quoted 1 line

Again, why do we need to re-enable interrupts here (in this loop), interrupts will get re-enabled in the irqentry_exit() code path ?

Hi Ashish:
Thanks for your review. check_hv_pending() is also called in the
native_irq_enable() to handle some pending interrupt requests after re -enabling interrupt. For such case, disables irq when handle exception or interrupt event.

#50 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

Quoted 20 lines · Tianyu Lan
Quoted 19 lines · Ashish

arch/x86/kernel/sev.c struct sev_config { · line 201

Quoted 8 lines
arch/x86/kernel/sev.c +8 -0
+static void do_exc_hv(struct pt_regs *regs)
+{
+    union hv_pending_events pending_events;
+    u8 vector;
+
+    while (sev_hv_pending()) {
+        asm volatile("cli" : : : "memory");
+

Do we really need to disable interrupts here, #HV exception will be dispatched via an interrupt gate in the IDT, so interrupts should be implicitly disabled, right ?

Quoted 8 lines

   panic("Unexpected vector %d\n", vector);
+                unreachable();
+            }
+        } else {
+            common_interrupt(regs, pending_events.vector); +        }
+
+        asm volatile("sti" : : : "memory");

Quoted 8 lines
Quoted 1 line

Again, why do we need to re-enable interrupts here (in this loop), interrupts will get re-enabled in the irqentry_exit() code path ?

Hi Ashish:
    Thanks for your review.    check_hv_pending() is also called in the
native_irq_enable() to handle some pending interrupt requests after re -enabling interrupt. For such case, disables irq when handle exception or interrupt event.

Then probably add the interrupt disable/enable in the caller
function like native_irq_enable() which seems logical as check_hv_pending() is always called with interrupts disabled (either via HW or SW) and also interrupt disable/enable seems redundant in check_hv_pending().

Thanks,
Ashish

#51 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

In reply to Tianyu Lan
Quoted 7 lines · Tianyu Lan

From: Tianyu Lan <[email protected]>

Enable #HV exception to handle interrupt requests from hypervisor.

Quoted 9 lines
---
  arch/x86/entry/entry_64.S          |  18 ++
  arch/x86/include/asm/irqflags.h    |  19 ++
  arch/x86/include/asm/mem_encrypt.h |   2 +
  arch/x86/include/asm/msr-index.h   |   6 +
  arch/x86/include/uapi/asm/svm.h    |   4 +
  arch/x86/kernel/sev.c              | 354 ++++++++++++++++++++++++-----
  arch/x86/kernel/traps.c            |  50 ++++
  7 files changed, 400 insertions(+), 53 deletions(-)

This diff is confusing because I can't tell if functions were moved or it's just the way the diff was processed. I wonder if this couldn't be broken out into more than one patch to improve the diff and make it easier to review.

Otherwise, comments below.

Quoted 142 lines
arch/x86/entry/entry_64.S +18 -0
index b2059df43c57..fe460cf44ab5 100644
@@ -1058,6 +1058,15 @@ SYM_CODE_END(paranoid_entry)
   * R15 - old SPEC_CTRL
   */
  SYM_CODE_START_LOCAL(paranoid_exit)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
  	UNWIND_HINT_REGS
 
  	/*
@@ -1183,6 +1192,15 @@ SYM_CODE_START_LOCAL(error_entry)
  SYM_CODE_END(error_entry)
 
  SYM_CODE_START_LOCAL(error_return)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
  	UNWIND_HINT_REGS
  	DEBUG_ENTRY_ASSERT_IRQS_OFF
  	testb	$3, CS(%rsp)
arch/x86/include/asm/irqflags.h +19 -0
index 7793e52d6237..e0730d8bc0ac 100644
@@ -14,6 +14,9 @@
  /*
   * Interrupt control:
   */
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+void check_hv_pending(struct pt_regs *regs);
+#endif
 
  /* Declaration required for gcc < 4.9 to prevent -Werror=missing-prototypes */
  extern inline unsigned long native_save_fl(void);
@@ -35,6 +38,19 @@ extern __always_inline unsigned long native_save_fl(void)
  	return flags;
  }
 
+extern inline void native_restore_fl(unsigned long flags)
+{
+	asm volatile("push %0 ; popf"
+		     : /* no output */
+		     : "g" (flags)
+		     : "memory", "cc");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	if ((flags & X86_EFLAGS_IF)) {
+		check_hv_pending(NULL);
+	}
+#endif
+}
+
  static __always_inline void native_irq_disable(void)
  {
  	asm volatile("cli": : :"memory");
@@ -43,6 +59,9 @@ static __always_inline void native_irq_disable(void)
  static __always_inline void native_irq_enable(void)
  {
  	asm volatile("sti": : :"memory");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	check_hv_pending(NULL);
+#endif
  }
 
  static inline __cpuidle void native_safe_halt(void)
arch/x86/include/asm/mem_encrypt.h +2 -0
index 72ca90552b6a..7264ca5f5b2d 100644
@@ -50,6 +50,7 @@ void __init early_set_mem_enc_dec_hypercall(unsigned long vaddr, int npages,
  void __init mem_encrypt_free_decrypted_mem(void);
 
  void __init sev_es_init_vc_handling(void);
+void __init sev_snp_init_hv_handling(void);
 
  #define __bss_decrypted __section(".bss..decrypted")
 
@@ -72,6 +73,7 @@ static inline void __init sme_encrypt_kernel(struct boot_params *bp) { }
  static inline void __init sme_enable(struct boot_params *bp) { }
 
  static inline void sev_es_init_vc_handling(void) { }
+static inline void sev_snp_init_hv_handling(void) { }
 
  static inline int __init
  early_set_memory_decrypted(unsigned long vaddr, unsigned long size) { return 0; }
arch/x86/include/asm/msr-index.h +6 -0
index 10ac52705892..6fe25a6e325f 100644
@@ -562,10 +562,16 @@
  #define MSR_AMD64_SEV_ENABLED_BIT	0
  #define MSR_AMD64_SEV_ES_ENABLED_BIT	1
  #define MSR_AMD64_SEV_SNP_ENABLED_BIT	2
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT		4
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT	5
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT	6
  #define MSR_AMD64_SEV_ENABLED		BIT_ULL(MSR_AMD64_SEV_ENABLED_BIT)
  #define MSR_AMD64_SEV_ES_ENABLED	BIT_ULL(MSR_AMD64_SEV_ES_ENABLED_BIT)
  #define MSR_AMD64_SEV_SNP_ENABLED	BIT_ULL(MSR_AMD64_SEV_SNP_ENABLED_BIT)
 
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED			BIT_ULL(MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT)
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT)
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT)
  #define MSR_AMD64_VIRT_SPEC_CTRL	0xc001011f
 
  /* AMD Collaborative Processor Performance Control MSRs */
arch/x86/include/uapi/asm/svm.h +4 -0
index f69c168391aa..85d6882262e7 100644
@@ -115,6 +115,10 @@
  #define SVM_VMGEXIT_AP_CREATE_ON_INIT		0
  #define SVM_VMGEXIT_AP_CREATE			1
  #define SVM_VMGEXIT_AP_DESTROY			2
+#define SVM_VMGEXIT_HV_DOORBELL_PAGE		0x80000014
+#define SVM_VMGEXIT_GET_PREFERRED_HV_DOORBELL_PAGE	0
+#define SVM_VMGEXIT_SET_HV_DOORBELL_PAGE		1
+#define SVM_VMGEXIT_QUERY_HV_DOORBELL_PAGE		2
  #define SVM_VMGEXIT_HV_FEATURES			0x8000fffd
  #define SVM_VMGEXIT_UNSUPPORTED_EVENT		0x8000ffff
 
arch/x86/kernel/sev.c +0 -0
index b54ee3ba37b0..23cd025f97dc 100644

arch/x86/kernel/sev.c struct sev_es_runtime_data { · line 112

Quoted 9 lines
arch/x86/kernel/sev.c +6 -0
@@ -104,6 +104,12 @@ struct sev_es_runtime_data {
  	 * is currently unsupported in SEV-ES guests.
  	 */
  	unsigned long dr7;
+	/*
+	 * SEV-SNP requires that the GHCB must be registered before using it.
+	 * The flag below will indicate whether the GHCB is registered, if its
+	 * not registered then sev_es_get_ghcb() will perform the registration.
+	 */
+	bool ghcb_registered;

So this wasn't need for the base SNP support, what path is being taken that this is now required?

Quoted 57 lines
arch/x86/kernel/sev.c +50 -0
  };
 
  struct ghcb_state {
@@ -122,6 +128,183 @@ struct sev_config {
 
  static struct sev_config sev_cfg __read_mostly;
 
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state);
+static noinstr void __sev_put_ghcb(struct ghcb_state *state);
+static int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa);
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb);
+
+union hv_pending_events {
+	u16 events;
+	struct {
+		u8 vector;
+		u8 nmi : 1;
+		u8 mc : 1;
+		u8 reserved1 : 5;
+		u8 no_further_signal : 1;
+	};
+};
+
+struct sev_hv_doorbell_page {
+	union hv_pending_events pending_events;
+	u8 no_eoi_required;
+	u8 reserved2[61];
+	u8 padding[4032];
+};
+
+struct sev_snp_runtime_data {
+	struct sev_hv_doorbell_page hv_doorbell_page;
+};
+
+static DEFINE_PER_CPU(struct sev_snp_runtime_data*, snp_runtime_data);
+
+static inline u64 sev_es_rd_ghcb_msr(void)
+{
+	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
+}
+
+static __always_inline void sev_es_wr_ghcb_msr(u64 val)
+{
+	u32 low, high;
+
+	low  = (u32)(val);
+	high = (u32)(val >> 32);
+
+	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+}
+
+struct sev_hv_doorbell_page *sev_snp_current_doorbell_page(void)
+{
+	return &this_cpu_read(snp_runtime_data)->hv_doorbell_page;
+}
+
+static u8 sev_hv_pending(void)

arch/x86/kernel/sev.c struct sev_config { · line 190

Quoted 10 lines
arch/x86/kernel/sev.c +10 -0
+{
+	return sev_snp_current_doorbell_page()->pending_events.events;
+}
+
+static void hv_doorbell_apic_eoi_write(u32 reg, u32 val)
+{
+	if (xchg(&sev_snp_current_doorbell_page()->no_eoi_required, 0) & 0x1)
+		return;
+
+	BUG_ON(reg != APIC_EOI);

Should this be a WARNing instead and just exit without doing anything if the reg != APIC_EOI?

Quoted 258 lines
arch/x86/kernel/sev.c +182 -52
+	apic->write(reg, val);
+}
+
+static void do_exc_hv(struct pt_regs *regs)
+{
+	union hv_pending_events pending_events;
+	u8 vector;
+
+	while (sev_hv_pending()) {
+		asm volatile("cli" : : : "memory");
+
+		pending_events.events = xchg(
+			&sev_snp_current_doorbell_page()->pending_events.events,
+			0);
+
+		if (pending_events.nmi)
+			exc_nmi(regs);
+
+#ifdef CONFIG_X86_MCE
+		if (pending_events.mc)
+			exc_machine_check(regs);
+#endif
+
+		if (!pending_events.vector)
+			return;
+
+		if (pending_events.vector < FIRST_EXTERNAL_VECTOR) {
+			/* Exception vectors */
+			WARN(1, "exception shouldn't happen\n");
+		} else if (pending_events.vector == FIRST_EXTERNAL_VECTOR) {
+			sysvec_irq_move_cleanup(regs);
+		} else if (pending_events.vector == IA32_SYSCALL_VECTOR) {
+			WARN(1, "syscall shouldn't happen\n");
+		} else if (pending_events.vector >= FIRST_SYSTEM_VECTOR) {
+			switch (pending_events.vector) {
+#if IS_ENABLED(CONFIG_HYPERV)
+			case HYPERV_STIMER0_VECTOR:
+				sysvec_hyperv_stimer0(regs);
+				break;
+			case HYPERVISOR_CALLBACK_VECTOR:
+				sysvec_hyperv_callback(regs);
+				break;
+#endif
+#ifdef CONFIG_SMP
+			case RESCHEDULE_VECTOR:
+				sysvec_reschedule_ipi(regs);
+				break;
+			case IRQ_MOVE_CLEANUP_VECTOR:
+				sysvec_irq_move_cleanup(regs);
+				break;
+			case REBOOT_VECTOR:
+				sysvec_reboot(regs);
+				break;
+			case CALL_FUNCTION_SINGLE_VECTOR:
+				sysvec_call_function_single(regs);
+				break;
+			case CALL_FUNCTION_VECTOR:
+				sysvec_call_function(regs);
+				break;
+#endif
+#ifdef CONFIG_X86_LOCAL_APIC
+			case ERROR_APIC_VECTOR:
+				sysvec_error_interrupt(regs);
+				break;
+			case SPURIOUS_APIC_VECTOR:
+				sysvec_spurious_apic_interrupt(regs);
+				break;
+			case LOCAL_TIMER_VECTOR:
+				sysvec_apic_timer_interrupt(regs);
+				break;
+			case X86_PLATFORM_IPI_VECTOR:
+				sysvec_x86_platform_ipi(regs);
+				break;
+#endif
+			case 0x0:
+				break;
+			default:
+				panic("Unexpected vector %d\n", vector);
+				unreachable();
+			}
+		} else {
+			common_interrupt(regs, pending_events.vector);
+		}
+
+		asm volatile("sti" : : : "memory");
+	}
+}
+
+void check_hv_pending(struct pt_regs *regs)
+{
+	struct pt_regs local_regs;
+
+	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
+		return;
+
+	if (regs) {
+		if ((regs->flags & X86_EFLAGS_IF) == 0)
+			return;
+
+		if (!sev_hv_pending())
+			return;
+
+		do_exc_hv(regs);
+	} else {
+		if (sev_hv_pending()) {
+			memset(&local_regs, 0, sizeof(struct pt_regs));
+			regs = &local_regs;
+			asm volatile("movl %%cs, %%eax;" : "=a" (regs->cs));
+			asm volatile("movl %%ss, %%eax;" : "=a" (regs->ss));
+			regs->orig_ax = 0xffffffff;
+			regs->flags = native_save_fl();
+			do_exc_hv(regs);
+		}
+	}
+}
+EXPORT_SYMBOL_GPL(check_hv_pending);
+
  static __always_inline bool on_vc_stack(struct pt_regs *regs)
  {
  	unsigned long sp = regs->sp;
@@ -193,68 +376,35 @@ void noinstr __sev_es_ist_exit(void)
  	this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], *(unsigned long *)ist);
  }
 
-/*
- * Nothing shall interrupt this code path while holding the per-CPU
- * GHCB. The backup GHCB is only for NMIs interrupting this path.
- *
- * Callers must disable local interrupts around it.
- */
-static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+static bool sev_restricted_injection_enabled(void)
  {
-	struct sev_es_runtime_data *data;
+	return sev_status & MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED;
+}
+
+void __init sev_snp_init_hv_handling(void)
+{
+	struct sev_snp_runtime_data *snp_data;
+	struct ghcb_state state;
  	struct ghcb *ghcb;
+	unsigned long flags;
+	int cpu;
+	int err;
 
  	WARN_ON(!irqs_disabled());
+	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP) || !sev_restricted_injection_enabled())
+		return;
 
-	data = this_cpu_read(runtime_data);
-	ghcb = &data->ghcb_page;
-
-	if (unlikely(data->ghcb_active)) {
-		/* GHCB is already in use - save its contents */
-
-		if (unlikely(data->backup_ghcb_active)) {
-			/*
-			 * Backup-GHCB is also already in use. There is no way
-			 * to continue here so just kill the machine. To make
-			 * panic() work, mark GHCBs inactive so that messages
-			 * can be printed out.
-			 */
-			data->ghcb_active        = false;
-			data->backup_ghcb_active = false;
-
-			instrumentation_begin();
-			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
-			instrumentation_end();
-		}
-
-		/* Mark backup_ghcb active before writing to it */
-		data->backup_ghcb_active = true;
-
-		state->ghcb = &data->backup_ghcb;
-
-		/* Backup GHCB content */
-		*state->ghcb = *ghcb;
-	} else {
-		state->ghcb = NULL;
-		data->ghcb_active = true;
-	}
+	local_irq_save(flags);
 
-	return ghcb;
-}
+	ghcb = __sev_get_ghcb(&state);
 
-static inline u64 sev_es_rd_ghcb_msr(void)
-{
-	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
-}
+	sev_snp_setup_hv_doorbell_page(ghcb);
 
-static __always_inline void sev_es_wr_ghcb_msr(u64 val)
-{
-	u32 low, high;
+	__sev_put_ghcb(&state);
 
-	low  = (u32)(val);
-	high = (u32)(val >> 32);
+	apic_set_eoi_write(hv_doorbell_apic_eoi_write);
 
-	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+	local_irq_restore(flags);
  }
 
  static int vc_fetch_insn_kernel(struct es_em_ctxt *ctxt,
@@ -515,6 +665,79 @@ static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt
  /* Include code shared with pre-decompression boot stage */
  #include "sev-shared.c"
 
+/*
+ * Nothing shall interrupt this code path while holding the per-CPU
+ * GHCB. The backup GHCB is only for NMIs interrupting this path.
+ *
+ * Callers must disable local interrupts around it.
+ */
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+{
+	struct sev_es_runtime_data *data;
+	struct ghcb *ghcb;
+
+	WARN_ON(!irqs_disabled());
+
+	data = this_cpu_read(runtime_data);
+	ghcb = &data->ghcb_page;
+
+	if (unlikely(data->ghcb_active)) {
+		/* GHCB is already in use - save its contents */
+
+		if (unlikely(data->backup_ghcb_active)) {
+			/*
+			 * Backup-GHCB is also already in use. There is no way
+			 * to continue here so just kill the machine. To make
+			 * panic() work, mark GHCBs inactive so that messages
+			 * can be printed out.
+			 */
+			data->ghcb_active        = false;
+			data->backup_ghcb_active = false;
+
+			instrumentation_begin();
+			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
+			instrumentation_end();
+		}
+
+		/* Mark backup_ghcb active before writing to it */
+		data->backup_ghcb_active = true;
+
+		state->ghcb = &data->backup_ghcb;
+
+		/* Backup GHCB content */
+		*state->ghcb = *ghcb;
+	} else {
+		state->ghcb = NULL;
+		data->ghcb_active = true;
+	}
+

arch/x86/kernel/sev.c static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt · line 723

Quoted 10 lines
arch/x86/kernel/sev.c +10 -0
+	/* SEV-SNP guest requires that GHCB must be registered before using it. */
+	if (!data->ghcb_registered) {
+		if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+			snp_register_ghcb_early(__pa(ghcb));
+			sev_snp_setup_hv_doorbell_page(ghcb);
+		} else {
+			sev_es_wr_ghcb_msr(__pa(ghcb));
+		}
+		data->ghcb_registered = true;
+	}

This kind of goes with the previous comment about why this check for a registered GHCB is now needed. The path needs to be reworked so this code doesn't have to be added.

arch/x86/kernel/sev.c static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt · line 736

Quoted 13 lines
arch/x86/kernel/sev.c +13 -0
+
+	return ghcb;
+}
+
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb)
+{
+	u64 pa;
+	enum es_result ret;
+
+	pa = __pa(sev_snp_current_doorbell_page());
+	vc_ghcb_invalidate(ghcb);
+	ret = vmgexit_hv_doorbell_page(ghcb,
+			SVM_VMGEXIT_SET_HV_DOORBELL_PAGE, pa);

This can go on one line or else you'll need to properly line everything up.

Quoted 123 lines
arch/x86/kernel/sev.c +30 -1
+	if (ret != ES_OK)
+		panic("SEV-SNP: failed to set up #HV doorbell page");
+}
+
  static noinstr void __sev_put_ghcb(struct ghcb_state *state)
  {
  	struct sev_es_runtime_data *data;
@@ -1282,6 +1505,11 @@ void setup_ghcb(void)
  		snp_register_ghcb_early(__pa(&boot_ghcb_page));
  }
 
+int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa)
+{
+	return sev_es_ghcb_hv_call(ghcb, NULL, SVM_VMGEXIT_HV_DOORBELL_PAGE, op, pa);
+}
+
  #ifdef CONFIG_HOTPLUG_CPU
  static void sev_es_ap_hlt_loop(void)
  {
@@ -1355,6 +1583,7 @@ static void __init alloc_runtime_data(int cpu)
  static void __init init_ghcb(int cpu)
  {
  	struct sev_es_runtime_data *data;
+	struct sev_snp_runtime_data *snp_data;
  	int err;
 
  	data = per_cpu(runtime_data, cpu);
@@ -1366,8 +1595,22 @@ static void __init init_ghcb(int cpu)
 
  	memset(&data->ghcb_page, 0, sizeof(data->ghcb_page));
 
+	snp_data = memblock_alloc(sizeof(*snp_data), PAGE_SIZE);
+	if (!snp_data)
+		panic("Can't allocate SEV-SNP runtime data");
+
+	err = early_set_memory_decrypted((unsigned long)&snp_data->hv_doorbell_page,
+					 sizeof(snp_data->hv_doorbell_page));
+	if (err)
+		panic("Can't map #HV doorbell pages unencrypted");
+
+	memset(&snp_data->hv_doorbell_page, 0, sizeof(snp_data->hv_doorbell_page));
+
+	per_cpu(snp_runtime_data, cpu) = snp_data;
+
  	data->ghcb_active = false;
  	data->backup_ghcb_active = false;
+	data->ghcb_registered = false;
  }
 
  void __init sev_es_init_vc_handling(void)
@@ -2006,7 +2249,12 @@ DEFINE_IDTENTRY_VC_USER(exc_vmm_communication)
 
  static bool hv_raw_handle_exception(struct pt_regs *regs)
  {
-	return false;
+	/* Clear the no_further_signal bit */
+	sev_snp_current_doorbell_page()->pending_events.events &= 0x7fff;
+
+	check_hv_pending(regs);
+
+	return true;
  }
 
  static __always_inline bool on_hv_fallback_stack(struct pt_regs *regs)
arch/x86/kernel/traps.c +47 -0
index 178015a820f0..af97e6610fbb 100644
@@ -898,6 +898,53 @@ asmlinkage __visible noinstr struct pt_regs *vc_switch_off_ist(struct pt_regs *r
 
  	return regs_ret;
  }
+
+asmlinkage __visible noinstr struct pt_regs *hv_switch_off_ist(struct pt_regs *regs)
+{
+	unsigned long sp, *stack;
+	struct stack_info info;
+	struct pt_regs *regs_ret;
+
+	/*
+	 * A malicious hypervisor can inject 2 HVs in a row, which will corrupt
+	 * the trap frame on our IST stack.  We add a defensive check here to
+	 * catch such behavior.
+	 */
+	BUG_ON(regs->sp >= __this_cpu_ist_bottom_va(HV) && regs->sp < __this_cpu_ist_top_va(HV));
+
+	/*
+	 * In the SYSCALL entry path the RSP value comes from user-space - don't
+	 * trust it and switch to the current kernel stack
+	 */
+	if (ip_within_syscall_gap(regs)) {
+		sp = this_cpu_read(cpu_current_top_of_stack);
+		goto sync;
+	}
+
+	/*
+	 * From here on the RSP value is trusted. Now check whether entry
+	 * happened from a safe stack. Not safe are the entry or unknown stacks,
+	 * use the fall-back stack instead in this case.
+	 */
+	sp    = regs->sp;
+	stack = (unsigned long *)sp;
+
+	if (!get_stack_info_noinstr(stack, current, &info) || info.type == STACK_TYPE_ENTRY ||
+	    info.type > STACK_TYPE_EXCEPTION_LAST)
+		sp = __this_cpu_ist_top_va(HV2);
+sync:
+	/*
+	 * Found a safe stack - switch to it as if the entry didn't happen via
+	 * IST stack. The code below only copies pt_regs, the real switch happens
+	 * in assembly code.
+	 */
+	sp = ALIGN_DOWN(sp, 8) - sizeof(*regs_ret);
+
+	regs_ret = (struct pt_regs *)sp;
+	*regs_ret = *regs;
+
+	return regs_ret;
+}
  #endif
 
  asmlinkage __visible noinstr struct pt_regs *fixup_bad_iret(struct pt_regs *bad_regs)

arch/x86/kernel/traps.c void __init trap_init(void) · line 1509

Quoted 6 lines
arch/x86/kernel/traps.c +3 -0
@@ -1457,4 +1504,7 @@ void __init trap_init(void)
  	/* Setup traps as cpu_init() might #GP */
  	idt_setup_traps();
  	cpu_init();
+
+	/* Init #HV doorbell pages when running as an SEV-SNP guest */
+	sev_snp_init_hv_handling();

Can this be moved to right after the sev_es_init_vc_handling() call to keep things grouped together better?

Thanks,
Tom

Quoted 1 line
arch/x86/kernel/traps.c +0 -0
  }
author
#52 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

Quoted 18 lines · Tom Lendacky
Quoted 15 lines · Tianyu Lan

From: Tianyu Lan <[email protected]>

Enable #HV exception to handle interrupt requests from hypervisor.

---
  arch/x86/entry/entry_64.S          |  18 ++ arch/x86/include/asm/irqflags.h    |  19 ++ arch/x86/include/asm/mem_encrypt.h |   2 + arch/x86/include/asm/msr-index.h   |   6 + arch/x86/include/uapi/asm/svm.h    |   4 + arch/x86/kernel/sev.c              | 354 ++++++++++++++++++++++++----- arch/x86/kernel/traps.c            |  50 ++++

Quoted 1 line
  7 files changed, 400 insertions(+), 53 deletions(-)
Quoted 3 lines

This diff is confusing because I can't tell if functions were moved or it's just the way the diff was processed. I wonder if this couldn't be broken out into more than one patch to improve the diff and make it easier to review.

Otherwise, comments below.

Hi Thomas:
Thanks for your review. Sure. Will split this into more patches.

Quoted 477 lines
Quoted 142 lines
arch/x86/entry/entry_64.S +9 -0
index b2059df43c57..fe460cf44ab5 100644
@@ -1058,6 +1058,15 @@ SYM_CODE_END(paranoid_entry)
    * R15 - old SPEC_CTRL
    */
   SYM_CODE_START_LOCAL(paranoid_exit)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+    /*
+     * If a #HV was delivered during execution and interrupts were
+     * disabled, then check if it can be handled before the iret
+     * (which may re-enable interrupts).
+     */
+    mov     %rsp, %rdi
+    call    check_hv_pending
+#endif

      UNWIND_HINT_REGS
      /*

arch/x86/entry/entry_64.S +9 -0
@@ -1183,6 +1192,15 @@ SYM_CODE_START_LOCAL(error_entry)
   SYM_CODE_END(error_entry)
   SYM_CODE_START_LOCAL(error_return)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+    /*
+     * If a #HV was delivered during execution and interrupts were
+     * disabled, then check if it can be handled before the iret
+     * (which may re-enable interrupts).
+     */
+    mov     %rsp, %rdi
+    call    check_hv_pending
+#endif

      UNWIND_HINT_REGS
      DEBUG_ENTRY_ASSERT_IRQS_OFF
      testb    $3, CS(%rsp)

arch/x86/include/asm/irqflags.h +19 -1
index 7793e52d6237..e0730d8bc0ac 100644
@@ -14,6 +14,9 @@
   /*
    * Interrupt control:
    */
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+void check_hv_pending(struct pt_regs *regs);
+#endif
   /* Declaration required for gcc < 4.9 to prevent 
-Werror=missing-prototypes */
   extern inline unsigned long native_save_fl(void);
@@ -35,6 +38,19 @@ extern __always_inline unsigned long 
 native_save_fl(void)
       return flags;
   }
+extern inline void native_restore_fl(unsigned long flags)
+{
+    asm volatile("push %0 ; popf"
+             : /* no output */
+             : "g" (flags)
+             : "memory", "cc");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+    if ((flags & X86_EFLAGS_IF)) {
+        check_hv_pending(NULL);
+    }
+#endif
+}
+
   static __always_inline void native_irq_disable(void)
   {
       asm volatile("cli": : :"memory");
@@ -43,6 +59,9 @@ static __always_inline void native_irq_disable(void)
   static __always_inline void native_irq_enable(void)
   {
       asm volatile("sti": : :"memory");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+    check_hv_pending(NULL);
+#endif
   }
   static inline __cpuidle void native_safe_halt(void)
arch/x86/include/asm/mem_encrypt.h +2 -0
index 72ca90552b6a..7264ca5f5b2d 100644
@@ -50,6 +50,7 @@ void __init early_set_mem_enc_dec_hypercall(unsigned 
 long vaddr, int npages,
   void __init mem_encrypt_free_decrypted_mem(void);
   void __init sev_es_init_vc_handling(void);
+void __init sev_snp_init_hv_handling(void);
   #define __bss_decrypted __section(".bss..decrypted")
@@ -72,6 +73,7 @@ static inline void __init sme_encrypt_kernel(struct 
 boot_params *bp) { }
   static inline void __init sme_enable(struct boot_params *bp) { }
   static inline void sev_es_init_vc_handling(void) { }
+static inline void sev_snp_init_hv_handling(void) { }

  static inline int __init
  early_set_memory_decrypted(unsigned long vaddr, unsigned long size) { return 0; }

arch/x86/include/asm/msr-index.h +6 -0
index 10ac52705892..6fe25a6e325f 100644
@@ -562,10 +562,16 @@
   #define MSR_AMD64_SEV_ENABLED_BIT    0
   #define MSR_AMD64_SEV_ES_ENABLED_BIT    1
   #define MSR_AMD64_SEV_SNP_ENABLED_BIT    2
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT        4
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT    5
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT    6
   #define MSR_AMD64_SEV_ENABLED        BIT_ULL(MSR_AMD64_SEV_ENABLED_BIT)
   #define MSR_AMD64_SEV_ES_ENABLED    BIT_ULL(MSR_AMD64_SEV_ES_ENABLED_BIT)
   #define MSR_AMD64_SEV_SNP_ENABLED    BIT_ULL(MSR_AMD64_SEV_SNP_ENABLED_BIT)
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED            
 BIT_ULL(MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT)
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED    
 BIT_ULL(MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT)
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED    
 BIT_ULL(MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT)
   #define MSR_AMD64_VIRT_SPEC_CTRL    0xc001011f

  /* AMD Collaborative Processor Performance Control MSRs */

arch/x86/include/uapi/asm/svm.h +4 -0
index f69c168391aa..85d6882262e7 100644
@@ -115,6 +115,10 @@
   #define SVM_VMGEXIT_AP_CREATE_ON_INIT        0
   #define SVM_VMGEXIT_AP_CREATE            1
   #define SVM_VMGEXIT_AP_DESTROY            2
+#define SVM_VMGEXIT_HV_DOORBELL_PAGE        0x80000014
+#define SVM_VMGEXIT_GET_PREFERRED_HV_DOORBELL_PAGE    0
+#define SVM_VMGEXIT_SET_HV_DOORBELL_PAGE        1
+#define SVM_VMGEXIT_QUERY_HV_DOORBELL_PAGE        2
   #define SVM_VMGEXIT_HV_FEATURES            0x8000fffd
   #define SVM_VMGEXIT_UNSUPPORTED_EVENT        0x8000ffff
arch/x86/kernel/sev.c +2 -0
index b54ee3ba37b0..23cd025f97dc 100644
@@ -104,6 +104,12 @@ struct sev_es_runtime_data {
        * is currently unsupported in SEV-ES guests.
        */
       unsigned long dr7;
+    /*
+     * SEV-SNP requires that the GHCB must be registered before using 
Quoted 7 lines

it.
+     * The flag below will indicate whether the GHCB is registered, if its
+     * not registered then sev_es_get_ghcb() will perform the registration.
+     */
+    bool ghcb_registered;

So this wasn't need for the base SNP support, what path is being taken that this is now required?

Quoted 55 lines

  };
  struct ghcb_state {

arch/x86/kernel/sev.c +51 -0
@@ -122,6 +128,183 @@ struct sev_config {
   static struct sev_config sev_cfg __read_mostly;
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state);
+static noinstr void __sev_put_ghcb(struct ghcb_state *state);
+static int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa);
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb);
+
+union hv_pending_events {
+    u16 events;
+    struct {
+        u8 vector;
+        u8 nmi : 1;
+        u8 mc : 1;
+        u8 reserved1 : 5;
+        u8 no_further_signal : 1;
+    };
+};
+
+struct sev_hv_doorbell_page {
+    union hv_pending_events pending_events;
+    u8 no_eoi_required;
+    u8 reserved2[61];
+    u8 padding[4032];
+};
+
+struct sev_snp_runtime_data {
+    struct sev_hv_doorbell_page hv_doorbell_page;
+};
+
+static DEFINE_PER_CPU(struct sev_snp_runtime_data*, snp_runtime_data);
+
+static inline u64 sev_es_rd_ghcb_msr(void)
+{
+    return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
+}
+
+static __always_inline void sev_es_wr_ghcb_msr(u64 val)
+{
+    u32 low, high;
+
+    low  = (u32)(val);
+    high = (u32)(val >> 32);
+
+    native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+}
+
+struct sev_hv_doorbell_page *sev_snp_current_doorbell_page(void)
+{
+    return &this_cpu_read(snp_runtime_data)->hv_doorbell_page;
+}
+
+static u8 sev_hv_pending(void)
+{

arch/x86/kernel/sev.c struct sev_config { · line 189

Quoted 10 lines
arch/x86/kernel/sev.c +9 -0
+    return sev_snp_current_doorbell_page()->pending_events.events;
+}
+
+static void hv_doorbell_apic_eoi_write(u32 reg, u32 val)
+{
+    if (xchg(&sev_snp_current_doorbell_page()->no_eoi_required, 0) & 
 0x1)
+        return;
+
+    BUG_ON(reg != APIC_EOI);

Should this be a WARNing instead and just exit without doing anything if the reg != APIC_EOI?

Quoted 257 lines
arch/x86/kernel/sev.c +136 -52
+    apic->write(reg, val);
+}
+
+static void do_exc_hv(struct pt_regs *regs)
+{
+    union hv_pending_events pending_events;
+    u8 vector;
+
+    while (sev_hv_pending()) {
+        asm volatile("cli" : : : "memory");
+
+        pending_events.events = xchg(
+            &sev_snp_current_doorbell_page()->pending_events.events,
+            0);
+
+        if (pending_events.nmi)
+            exc_nmi(regs);
+
+#ifdef CONFIG_X86_MCE
+        if (pending_events.mc)
+            exc_machine_check(regs);
+#endif
+
+        if (!pending_events.vector)
+            return;
+
+        if (pending_events.vector < FIRST_EXTERNAL_VECTOR) {
+            /* Exception vectors */
+            WARN(1, "exception shouldn't happen\n");
+        } else if (pending_events.vector == FIRST_EXTERNAL_VECTOR) {
+            sysvec_irq_move_cleanup(regs);
+        } else if (pending_events.vector == IA32_SYSCALL_VECTOR) {
+            WARN(1, "syscall shouldn't happen\n");
+        } else if (pending_events.vector >= FIRST_SYSTEM_VECTOR) {
+            switch (pending_events.vector) {
+#if IS_ENABLED(CONFIG_HYPERV)
+            case HYPERV_STIMER0_VECTOR:
+                sysvec_hyperv_stimer0(regs);
+                break;
+            case HYPERVISOR_CALLBACK_VECTOR:
+                sysvec_hyperv_callback(regs);
+                break;
+#endif
+#ifdef CONFIG_SMP
+            case RESCHEDULE_VECTOR:
+                sysvec_reschedule_ipi(regs);
+                break;
+            case IRQ_MOVE_CLEANUP_VECTOR:
+                sysvec_irq_move_cleanup(regs);
+                break;
+            case REBOOT_VECTOR:
+                sysvec_reboot(regs);
+                break;
+            case CALL_FUNCTION_SINGLE_VECTOR:
+                sysvec_call_function_single(regs);
+                break;
+            case CALL_FUNCTION_VECTOR:
+                sysvec_call_function(regs);
+                break;
+#endif
+#ifdef CONFIG_X86_LOCAL_APIC
+            case ERROR_APIC_VECTOR:
+                sysvec_error_interrupt(regs);
+                break;
+            case SPURIOUS_APIC_VECTOR:
+                sysvec_spurious_apic_interrupt(regs);
+                break;
+            case LOCAL_TIMER_VECTOR:
+                sysvec_apic_timer_interrupt(regs);
+                break;
+            case X86_PLATFORM_IPI_VECTOR:
+                sysvec_x86_platform_ipi(regs);
+                break;
+#endif
+            case 0x0:
+                break;
+            default:
+                panic("Unexpected vector %d\n", vector);
+                unreachable();
+            }
+        } else {
+            common_interrupt(regs, pending_events.vector);
+        }
+
+        asm volatile("sti" : : : "memory");
+    }
+}
+
+void check_hv_pending(struct pt_regs *regs)
+{
+    struct pt_regs local_regs;
+
+    if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
+        return;
+
+    if (regs) {
+        if ((regs->flags & X86_EFLAGS_IF) == 0)
+            return;
+
+        if (!sev_hv_pending())
+            return;
+
+        do_exc_hv(regs);
+    } else {
+        if (sev_hv_pending()) {
+            memset(&local_regs, 0, sizeof(struct pt_regs));
+            regs = &local_regs;
+            asm volatile("movl %%cs, %%eax;" : "=a" (regs->cs));
+            asm volatile("movl %%ss, %%eax;" : "=a" (regs->ss));
+            regs->orig_ax = 0xffffffff;
+            regs->flags = native_save_fl();
+            do_exc_hv(regs);
+        }
+    }
+}
+EXPORT_SYMBOL_GPL(check_hv_pending);
+
   static __always_inline bool on_vc_stack(struct pt_regs *regs)
   {
       unsigned long sp = regs->sp;
@@ -193,68 +376,35 @@ void noinstr __sev_es_ist_exit(void)
       this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], *(unsigned long *)ist);
   }
-/*
- * Nothing shall interrupt this code path while holding the per-CPU
- * GHCB. The backup GHCB is only for NMIs interrupting this path.
- *
- * Callers must disable local interrupts around it.
- */
-static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+static bool sev_restricted_injection_enabled(void)
   {
-    struct sev_es_runtime_data *data;
+    return sev_status & MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED;
+}
+
+void __init sev_snp_init_hv_handling(void)
+{
+    struct sev_snp_runtime_data *snp_data;
+    struct ghcb_state state;
       struct ghcb *ghcb;
+    unsigned long flags;
+    int cpu;
+    int err;
       WARN_ON(!irqs_disabled());
+    if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP) || 
 !sev_restricted_injection_enabled())
+        return;
-    data = this_cpu_read(runtime_data);
-    ghcb = &data->ghcb_page;
-
-    if (unlikely(data->ghcb_active)) {
-        /* GHCB is already in use - save its contents */
-
-        if (unlikely(data->backup_ghcb_active)) {
-            /*
-             * Backup-GHCB is also already in use. There is no way
-             * to continue here so just kill the machine. To make
-             * panic() work, mark GHCBs inactive so that messages
-             * can be printed out.
-             */
-            data->ghcb_active        = false;
-            data->backup_ghcb_active = false;
-
-            instrumentation_begin();
-            panic("Unable to handle #VC exception! GHCB and Backup 
 GHCB are already in use");
-            instrumentation_end();
-        }
-
-        /* Mark backup_ghcb active before writing to it */
-        data->backup_ghcb_active = true;
-
-        state->ghcb = &data->backup_ghcb;
-
-        /* Backup GHCB content */
-        *state->ghcb = *ghcb;
-    } else {
-        state->ghcb = NULL;
-        data->ghcb_active = true;
-    }
+    local_irq_save(flags);
-    return ghcb;
-}
+    ghcb = __sev_get_ghcb(&state);
-static inline u64 sev_es_rd_ghcb_msr(void)
-{
-    return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
-}
+    sev_snp_setup_hv_doorbell_page(ghcb);
-static __always_inline void sev_es_wr_ghcb_msr(u64 val)
-{
-    u32 low, high;
+    __sev_put_ghcb(&state);
-    low  = (u32)(val);
-    high = (u32)(val >> 32);
+    apic_set_eoi_write(hv_doorbell_apic_eoi_write);
-    native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+    local_irq_restore(flags);
   }
   static int vc_fetch_insn_kernel(struct es_em_ctxt *ctxt,
@@ -515,6 +665,79 @@ static enum es_result vc_slow_virt_to_phys(struct 

ghcb *ghcb, struct es_em_ctxt
  /* Include code shared with pre-decompression boot stage */ #include "sev-shared.c"

arch/x86/kernel/sev.c +31 -0
+/*
+ * Nothing shall interrupt this code path while holding the per-CPU
+ * GHCB. The backup GHCB is only for NMIs interrupting this path.
+ *
+ * Callers must disable local interrupts around it.
+ */
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+{
+    struct sev_es_runtime_data *data;
+    struct ghcb *ghcb;
+
+    WARN_ON(!irqs_disabled());
+
+    data = this_cpu_read(runtime_data);
+    ghcb = &data->ghcb_page;
+
+    if (unlikely(data->ghcb_active)) {
+        /* GHCB is already in use - save its contents */
+
+        if (unlikely(data->backup_ghcb_active)) {
+            /*
+             * Backup-GHCB is also already in use. There is no way
+             * to continue here so just kill the machine. To make
+             * panic() work, mark GHCBs inactive so that messages
+             * can be printed out.
+             */
+            data->ghcb_active        = false;
+            data->backup_ghcb_active = false;
+
+            instrumentation_begin();
+            panic("Unable to handle #VC exception! GHCB and Backup 

GHCB are already in use");
+            instrumentation_end();
+        }

arch/x86/kernel/sev.c +14 -0
+
+        /* Mark backup_ghcb active before writing to it */
+        data->backup_ghcb_active = true;
+
+        state->ghcb = &data->backup_ghcb;
+
+        /* Backup GHCB content */
+        *state->ghcb = *ghcb;
+    } else {
+        state->ghcb = NULL;
+        data->ghcb_active = true;
+    }
+
+    /* SEV-SNP guest requires that GHCB must be registered before 

using it. */
+    if (!data->ghcb_registered) {
+        if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) { +            snp_register_ghcb_early(__pa(ghcb));

Quoted 8 lines
Quoted 6 lines

+            sev_snp_setup_hv_doorbell_page(ghcb); +        } else {
+            sev_es_wr_ghcb_msr(__pa(ghcb)); +        }
+        data->ghcb_registered = true;
+    }

This kind of goes with the previous comment about why this check for a registered GHCB is now needed. The path needs to be reworked so this code doesn't have to be added.

Sure. I will have a try to remove ghcb_registered.

Quoted 16 lines
Quoted 4 lines
arch/x86/kernel/sev.c +3 -0
+
+    return ghcb;
+}

arch/x86/kernel/sev.c static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt · line 736

Quoted 10 lines
arch/x86/kernel/sev.c +10 -0
+
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb)
+{
+    u64 pa;
+    enum es_result ret;
+
+    pa = __pa(sev_snp_current_doorbell_page());
+    vc_ghcb_invalidate(ghcb);
+    ret = vmgexit_hv_doorbell_page(ghcb,
+            SVM_VMGEXIT_SET_HV_DOORBELL_PAGE, pa);

This can go on one line or else you'll need to properly line everything up.

OK. Will update to line everythin up otherwise checkpatch.pl will warn "more than 80 chars in one line".

Quoted 128 lines
Quoted 128 lines

+    if (ret != ES_OK)
+        panic("SEV-SNP: failed to set up #HV doorbell page");

arch/x86/kernel/sev.c +2 -0
+}
+

  static noinstr void __sev_put_ghcb(struct ghcb_state *state) {
      struct sev_es_runtime_data *data;

arch/x86/kernel/sev.c +26 -1
@@ -1282,6 +1505,11 @@ void setup_ghcb(void)
           snp_register_ghcb_early(__pa(&boot_ghcb_page));
   }
+int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa)
+{
+    return sev_es_ghcb_hv_call(ghcb, NULL, 
 SVM_VMGEXIT_HV_DOORBELL_PAGE, op, pa);
+}
+
   #ifdef CONFIG_HOTPLUG_CPU
   static void sev_es_ap_hlt_loop(void)
   {
@@ -1355,6 +1583,7 @@ static void __init alloc_runtime_data(int cpu)
   static void __init init_ghcb(int cpu)
   {
       struct sev_es_runtime_data *data;
+    struct sev_snp_runtime_data *snp_data;
       int err;
       data = per_cpu(runtime_data, cpu);
@@ -1366,8 +1595,22 @@ static void __init init_ghcb(int cpu)
       memset(&data->ghcb_page, 0, sizeof(data->ghcb_page));
+    snp_data = memblock_alloc(sizeof(*snp_data), PAGE_SIZE);
+    if (!snp_data)
+        panic("Can't allocate SEV-SNP runtime data");
+
+    err = early_set_memory_decrypted((unsigned 
 long)&snp_data->hv_doorbell_page,
+                     sizeof(snp_data->hv_doorbell_page));
+    if (err)
+        panic("Can't map #HV doorbell pages unencrypted");
+
+    memset(&snp_data->hv_doorbell_page, 0, 
 sizeof(snp_data->hv_doorbell_page));
+
+    per_cpu(snp_runtime_data, cpu) = snp_data;
+
       data->ghcb_active = false;
       data->backup_ghcb_active = false;
+    data->ghcb_registered = false;
   }
   void __init sev_es_init_vc_handling(void)
@@ -2006,7 +2249,12 @@ DEFINE_IDTENTRY_VC_USER(exc_vmm_communication)
   static bool hv_raw_handle_exception(struct pt_regs *regs)
   {
-    return false;
+    /* Clear the no_further_signal bit */
+    sev_snp_current_doorbell_page()->pending_events.events &= 0x7fff;
+
+    check_hv_pending(regs);
+
+    return true;
   }
   static __always_inline bool on_hv_fallback_stack(struct pt_regs *regs)
arch/x86/kernel/traps.c +9 -0
index 178015a820f0..af97e6610fbb 100644
@@ -898,6 +898,53 @@ asmlinkage __visible noinstr struct pt_regs 
 *vc_switch_off_ist(struct pt_regs *r
       return regs_ret;
   }
+
+asmlinkage __visible noinstr struct pt_regs *hv_switch_off_ist(struct 
 pt_regs *regs)
+{
+    unsigned long sp, *stack;
+    struct stack_info info;
+    struct pt_regs *regs_ret;
+
+    /*
+     * A malicious hypervisor can inject 2 HVs in a row, which will 

corrupt
+     * the trap frame on our IST stack.  We add a defensive check here to
+     * catch such behavior.
+     */
+    BUG_ON(regs->sp >= __this_cpu_ist_bottom_va(HV) && regs->sp < __this_cpu_ist_top_va(HV));

arch/x86/kernel/traps.c +13 -1
+
+    /*
+     * In the SYSCALL entry path the RSP value comes from user-space 
- don't
+     * trust it and switch to the current kernel stack
+     */
+    if (ip_within_syscall_gap(regs)) {
+        sp = this_cpu_read(cpu_current_top_of_stack);
+        goto sync;
+    }
+
+    /*
+     * From here on the RSP value is trusted. Now check whether entry
+     * happened from a safe stack. Not safe are the entry or unknown 

stacks,
+     * use the fall-back stack instead in this case.
+     */
+    sp    = regs->sp;
+    stack = (unsigned long *)sp;
+
+    if (!get_stack_info_noinstr(stack, current, &info) || info.type == STACK_TYPE_ENTRY ||
+        info.type > STACK_TYPE_EXCEPTION_LAST)
+        sp = __this_cpu_ist_top_va(HV2);
+sync:
+    /*
+     * Found a safe stack - switch to it as if the entry didn't happen via
+     * IST stack. The code below only copies pt_regs, the real switch happens
+     * in assembly code.
+     */
+    sp = ALIGN_DOWN(sp, 8) - sizeof(*regs_ret);

arch/x86/kernel/traps.c +6 -0
+
+    regs_ret = (struct pt_regs *)sp;
+    *regs_ret = *regs;
+
+    return regs_ret;
+}

  #endif
  asmlinkage __visible noinstr struct pt_regs *fixup_bad_iret(struct pt_regs *bad_regs)

arch/x86/kernel/traps.c +0 -0
@@ -1457,4 +1504,7 @@ void __init trap_init(void)
Quoted 8 lines

arch/x86/kernel/traps.c void __init trap_init(void) · line 1509

Quoted 6 lines
arch/x86/kernel/traps.c +3 -0
       /* Setup traps as cpu_init() might #GP */
       idt_setup_traps();
       cpu_init();
+
+    /* Init #HV doorbell pages when running as an SEV-SNP guest */
+    sev_snp_init_hv_handling();

Can this be moved to right after the sev_es_init_vc_handling() call to keep things grouped together better?

Sure. I will update in the next version.

Quoted 2 lines
Quoted 2 lines
arch/x86/kernel/traps.c +0 -0
   }
#53 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

In reply to Tianyu Lan

On 2022-11-18 at 22:46 -05, Tianyu Lan <[email protected]> wrote...

Quoted 313 lines

From: Tianyu Lan <[email protected]>

Enable #HV exception to handle interrupt requests from hypervisor.

---
 arch/x86/entry/entry_64.S          |  18 ++
 arch/x86/include/asm/irqflags.h    |  19 ++
 arch/x86/include/asm/mem_encrypt.h |   2 +
 arch/x86/include/asm/msr-index.h   |   6 +
 arch/x86/include/uapi/asm/svm.h    |   4 +
 arch/x86/kernel/sev.c              | 354 ++++++++++++++++++++++++-----
 arch/x86/kernel/traps.c            |  50 ++++
 7 files changed, 400 insertions(+), 53 deletions(-)
arch/x86/entry/entry_64.S +18 -0
index b2059df43c57..fe460cf44ab5 100644
@@ -1058,6 +1058,15 @@ SYM_CODE_END(paranoid_entry)
  * R15 - old SPEC_CTRL
  */
 SYM_CODE_START_LOCAL(paranoid_exit)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
 	UNWIND_HINT_REGS
 
 	/*
@@ -1183,6 +1192,15 @@ SYM_CODE_START_LOCAL(error_entry)
 SYM_CODE_END(error_entry)
 
 SYM_CODE_START_LOCAL(error_return)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
 	UNWIND_HINT_REGS
 	DEBUG_ENTRY_ASSERT_IRQS_OFF
 	testb	$3, CS(%rsp)
arch/x86/include/asm/irqflags.h +19 -0
index 7793e52d6237..e0730d8bc0ac 100644
@@ -14,6 +14,9 @@
 /*
  * Interrupt control:
  */
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+void check_hv_pending(struct pt_regs *regs);
+#endif
 
 /* Declaration required for gcc < 4.9 to prevent -Werror=missing-prototypes */
 extern inline unsigned long native_save_fl(void);
@@ -35,6 +38,19 @@ extern __always_inline unsigned long native_save_fl(void)
 	return flags;
 }
 
+extern inline void native_restore_fl(unsigned long flags)
+{
+	asm volatile("push %0 ; popf"
+		     : /* no output */
+		     : "g" (flags)
+		     : "memory", "cc");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	if ((flags & X86_EFLAGS_IF)) {
+		check_hv_pending(NULL);
+	}
+#endif
+}
+
 static __always_inline void native_irq_disable(void)
 {
 	asm volatile("cli": : :"memory");
@@ -43,6 +59,9 @@ static __always_inline void native_irq_disable(void)
 static __always_inline void native_irq_enable(void)
 {
 	asm volatile("sti": : :"memory");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	check_hv_pending(NULL);
+#endif
 }
 
 static inline __cpuidle void native_safe_halt(void)
arch/x86/include/asm/mem_encrypt.h +2 -0
index 72ca90552b6a..7264ca5f5b2d 100644
@@ -50,6 +50,7 @@ void __init early_set_mem_enc_dec_hypercall(unsigned long vaddr, int npages,
 void __init mem_encrypt_free_decrypted_mem(void);
 
 void __init sev_es_init_vc_handling(void);
+void __init sev_snp_init_hv_handling(void);
 
 #define __bss_decrypted __section(".bss..decrypted")
 
@@ -72,6 +73,7 @@ static inline void __init sme_encrypt_kernel(struct boot_params *bp) { }
 static inline void __init sme_enable(struct boot_params *bp) { }
 
 static inline void sev_es_init_vc_handling(void) { }
+static inline void sev_snp_init_hv_handling(void) { }
 
 static inline int __init
 early_set_memory_decrypted(unsigned long vaddr, unsigned long size) { return 0; }
arch/x86/include/asm/msr-index.h +6 -0
index 10ac52705892..6fe25a6e325f 100644
@@ -562,10 +562,16 @@
 #define MSR_AMD64_SEV_ENABLED_BIT	0
 #define MSR_AMD64_SEV_ES_ENABLED_BIT	1
 #define MSR_AMD64_SEV_SNP_ENABLED_BIT	2
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT		4
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT	5
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT	6
 #define MSR_AMD64_SEV_ENABLED		BIT_ULL(MSR_AMD64_SEV_ENABLED_BIT)
 #define MSR_AMD64_SEV_ES_ENABLED	BIT_ULL(MSR_AMD64_SEV_ES_ENABLED_BIT)
 #define MSR_AMD64_SEV_SNP_ENABLED	BIT_ULL(MSR_AMD64_SEV_SNP_ENABLED_BIT)
 
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED			BIT_ULL(MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT)
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT)
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT)
 #define MSR_AMD64_VIRT_SPEC_CTRL	0xc001011f
 
 /* AMD Collaborative Processor Performance Control MSRs */
arch/x86/include/uapi/asm/svm.h +4 -0
index f69c168391aa..85d6882262e7 100644
@@ -115,6 +115,10 @@
 #define SVM_VMGEXIT_AP_CREATE_ON_INIT		0
 #define SVM_VMGEXIT_AP_CREATE			1
 #define SVM_VMGEXIT_AP_DESTROY			2
+#define SVM_VMGEXIT_HV_DOORBELL_PAGE		0x80000014
+#define SVM_VMGEXIT_GET_PREFERRED_HV_DOORBELL_PAGE	0
+#define SVM_VMGEXIT_SET_HV_DOORBELL_PAGE		1
+#define SVM_VMGEXIT_QUERY_HV_DOORBELL_PAGE		2
 #define SVM_VMGEXIT_HV_FEATURES			0x8000fffd
 #define SVM_VMGEXIT_UNSUPPORTED_EVENT		0x8000ffff
 
arch/x86/kernel/sev.c +145 -0
index b54ee3ba37b0..23cd025f97dc 100644
@@ -104,6 +104,12 @@ struct sev_es_runtime_data {
 	 * is currently unsupported in SEV-ES guests.
 	 */
 	unsigned long dr7;
+	/*
+	 * SEV-SNP requires that the GHCB must be registered before using it.
+	 * The flag below will indicate whether the GHCB is registered, if its
+	 * not registered then sev_es_get_ghcb() will perform the registration.
+	 */
+	bool ghcb_registered;
 };
 
 struct ghcb_state {
@@ -122,6 +128,183 @@ struct sev_config {
 
 static struct sev_config sev_cfg __read_mostly;
 
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state);
+static noinstr void __sev_put_ghcb(struct ghcb_state *state);
+static int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa);
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb);
+
+union hv_pending_events {
+	u16 events;
+	struct {
+		u8 vector;
+		u8 nmi : 1;
+		u8 mc : 1;
+		u8 reserved1 : 5;
+		u8 no_further_signal : 1;
+	};
+};
+
+struct sev_hv_doorbell_page {
+	union hv_pending_events pending_events;
+	u8 no_eoi_required;
+	u8 reserved2[61];
+	u8 padding[4032];
+};
+
+struct sev_snp_runtime_data {
+	struct sev_hv_doorbell_page hv_doorbell_page;
+};
+
+static DEFINE_PER_CPU(struct sev_snp_runtime_data*, snp_runtime_data);
+
+static inline u64 sev_es_rd_ghcb_msr(void)
+{
+	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
+}
+
+static __always_inline void sev_es_wr_ghcb_msr(u64 val)
+{
+	u32 low, high;
+
+	low  = (u32)(val);
+	high = (u32)(val >> 32);
+
+	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+}
+
+struct sev_hv_doorbell_page *sev_snp_current_doorbell_page(void)
+{
+	return &this_cpu_read(snp_runtime_data)->hv_doorbell_page;
+}
+
+static u8 sev_hv_pending(void)
+{
+	return sev_snp_current_doorbell_page()->pending_events.events;
+}
+
+static void hv_doorbell_apic_eoi_write(u32 reg, u32 val)
+{
+	if (xchg(&sev_snp_current_doorbell_page()->no_eoi_required, 0) & 0x1)
+		return;
+
+	BUG_ON(reg != APIC_EOI);
+	apic->write(reg, val);
+}
+
+static void do_exc_hv(struct pt_regs *regs)
+{
+	union hv_pending_events pending_events;
+	u8 vector;
+
+	while (sev_hv_pending()) {
+		asm volatile("cli" : : : "memory");
+
+		pending_events.events = xchg(
+			&sev_snp_current_doorbell_page()->pending_events.events,
+			0);
+
+		if (pending_events.nmi)
+			exc_nmi(regs);
+
+#ifdef CONFIG_X86_MCE
+		if (pending_events.mc)
+			exc_machine_check(regs);
+#endif
+
+		if (!pending_events.vector)
+			return;
+
+		if (pending_events.vector < FIRST_EXTERNAL_VECTOR) {
+			/* Exception vectors */
+			WARN(1, "exception shouldn't happen\n");
+		} else if (pending_events.vector == FIRST_EXTERNAL_VECTOR) {
+			sysvec_irq_move_cleanup(regs);
+		} else if (pending_events.vector == IA32_SYSCALL_VECTOR) {
+			WARN(1, "syscall shouldn't happen\n");
+		} else if (pending_events.vector >= FIRST_SYSTEM_VECTOR) {
+			switch (pending_events.vector) {
+#if IS_ENABLED(CONFIG_HYPERV)
+			case HYPERV_STIMER0_VECTOR:
+				sysvec_hyperv_stimer0(regs);
+				break;
+			case HYPERVISOR_CALLBACK_VECTOR:
+				sysvec_hyperv_callback(regs);
+				break;
+#endif
+#ifdef CONFIG_SMP
+			case RESCHEDULE_VECTOR:
+				sysvec_reschedule_ipi(regs);
+				break;
+			case IRQ_MOVE_CLEANUP_VECTOR:
+				sysvec_irq_move_cleanup(regs);
+				break;
+			case REBOOT_VECTOR:
+				sysvec_reboot(regs);
+				break;
+			case CALL_FUNCTION_SINGLE_VECTOR:
+				sysvec_call_function_single(regs);
+				break;
+			case CALL_FUNCTION_VECTOR:
+				sysvec_call_function(regs);
+				break;
+#endif
+#ifdef CONFIG_X86_LOCAL_APIC
+			case ERROR_APIC_VECTOR:
+				sysvec_error_interrupt(regs);
+				break;
+			case SPURIOUS_APIC_VECTOR:
+				sysvec_spurious_apic_interrupt(regs);
+				break;
+			case LOCAL_TIMER_VECTOR:
+				sysvec_apic_timer_interrupt(regs);
+				break;
+			case X86_PLATFORM_IPI_VECTOR:
+				sysvec_x86_platform_ipi(regs);
+				break;
+#endif
+			case 0x0:
+				break;
+			default:
+				panic("Unexpected vector %d\n", vector);
+				unreachable();

arch/x86/kernel/sev.c struct sev_config { · line 279

Quoted 10 lines
arch/x86/kernel/sev.c +10 -0
+			}
+		} else {
+			common_interrupt(regs, pending_events.vector);
+		}
+
+		asm volatile("sti" : : : "memory");
+	}
+}
+
+void check_hv_pending(struct pt_regs *regs)

This looks like two functions, one with regs == NULL and one with regs, different internal logic, different call sites. Would you consider splitting into two?

Quoted 317 lines
arch/x86/kernel/sev.c +28 -0
+{
+	struct pt_regs local_regs;
+
+	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
+		return;
 
 
+
+	if (regs) {
+		if ((regs->flags & X86_EFLAGS_IF) == 0)
+			return;
+
+		if (!sev_hv_pending())
+			return;
+
+		do_exc_hv(regs);
+	} else {
+		if (sev_hv_pending()) {
+			memset(&local_regs, 0, sizeof(struct pt_regs));
+			regs = &local_regs;
+			asm volatile("movl %%cs, %%eax;" : "=a" (regs->cs));
+			asm volatile("movl %%ss, %%eax;" : "=a" (regs->ss));
+			regs->orig_ax = 0xffffffff;
+			regs->flags = native_save_fl();
+			do_exc_hv(regs);
+		}
+	}
+}
+EXPORT_SYMBOL_GPL(check_hv_pending);
+
 static __always_inline bool on_vc_stack(struct pt_regs *regs)

{

arch/x86/kernel/sev.c +118 -53
 	unsigned long sp = regs->sp;
@@ -193,68 +376,35 @@ void noinstr __sev_es_ist_exit(void)
 	this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], *(unsigned long *)ist);
 }
 
-/*
- * Nothing shall interrupt this code path while holding the per-CPU
- * GHCB. The backup GHCB is only for NMIs interrupting this path.
- *
- * Callers must disable local interrupts around it.
- */
-static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+static bool sev_restricted_injection_enabled(void)
 {
-	struct sev_es_runtime_data *data;
+	return sev_status & MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED;
+}
+
+void __init sev_snp_init_hv_handling(void)
+{
+	struct sev_snp_runtime_data *snp_data;
+	struct ghcb_state state;
 	struct ghcb *ghcb;
+	unsigned long flags;
+	int cpu;
+	int err;
 
 	WARN_ON(!irqs_disabled());
+	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP) || !sev_restricted_injection_enabled())
+		return;
 
-	data = this_cpu_read(runtime_data);
-	ghcb = &data->ghcb_page;
-
-	if (unlikely(data->ghcb_active)) {
-		/* GHCB is already in use - save its contents */
-
-		if (unlikely(data->backup_ghcb_active)) {
-			/*
-			 * Backup-GHCB is also already in use. There is no way
-			 * to continue here so just kill the machine. To make
-			 * panic() work, mark GHCBs inactive so that messages
-			 * can be printed out.
-			 */
-			data->ghcb_active        = false;
-			data->backup_ghcb_active = false;
-
-			instrumentation_begin();
-			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
-			instrumentation_end();
-		}
-
-		/* Mark backup_ghcb active before writing to it */
-		data->backup_ghcb_active = true;
-
-		state->ghcb = &data->backup_ghcb;
-
-		/* Backup GHCB content */
-		*state->ghcb = *ghcb;
-	} else {
-		state->ghcb = NULL;
-		data->ghcb_active = true;
-	}
+	local_irq_save(flags);
 
-	return ghcb;
-}
+	ghcb = __sev_get_ghcb(&state);
 
-static inline u64 sev_es_rd_ghcb_msr(void)
-{
-	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
-}
+	sev_snp_setup_hv_doorbell_page(ghcb);
 
-static __always_inline void sev_es_wr_ghcb_msr(u64 val)
-{
-	u32 low, high;
+	__sev_put_ghcb(&state);
 
-	low  = (u32)(val);
-	high = (u32)(val >> 32);
+	apic_set_eoi_write(hv_doorbell_apic_eoi_write);
 
-	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+	local_irq_restore(flags);
 }
 
 static int vc_fetch_insn_kernel(struct es_em_ctxt *ctxt,
@@ -515,6 +665,79 @@ static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt
 /* Include code shared with pre-decompression boot stage */
 #include "sev-shared.c"
 
+/*
+ * Nothing shall interrupt this code path while holding the per-CPU
+ * GHCB. The backup GHCB is only for NMIs interrupting this path.
+ *
+ * Callers must disable local interrupts around it.
+ */
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+{
+	struct sev_es_runtime_data *data;
+	struct ghcb *ghcb;
+
+	WARN_ON(!irqs_disabled());
+
+	data = this_cpu_read(runtime_data);
+	ghcb = &data->ghcb_page;
+
+	if (unlikely(data->ghcb_active)) {
+		/* GHCB is already in use - save its contents */
+
+		if (unlikely(data->backup_ghcb_active)) {
+			/*
+			 * Backup-GHCB is also already in use. There is no way
+			 * to continue here so just kill the machine. To make
+			 * panic() work, mark GHCBs inactive so that messages
+			 * can be printed out.
+			 */
+			data->ghcb_active        = false;
+			data->backup_ghcb_active = false;
+
+			instrumentation_begin();
+			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
+			instrumentation_end();
+		}
+
+		/* Mark backup_ghcb active before writing to it */
+		data->backup_ghcb_active = true;
+
+		state->ghcb = &data->backup_ghcb;
+
+		/* Backup GHCB content */
+		*state->ghcb = *ghcb;
+	} else {
+		state->ghcb = NULL;
+		data->ghcb_active = true;
+	}
+
+	/* SEV-SNP guest requires that GHCB must be registered before using it. */
+	if (!data->ghcb_registered) {
+		if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+			snp_register_ghcb_early(__pa(ghcb));
+			sev_snp_setup_hv_doorbell_page(ghcb);
+		} else {
+			sev_es_wr_ghcb_msr(__pa(ghcb));
+		}
+		data->ghcb_registered = true;
+	}
+
+	return ghcb;
+}
+
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb)
+{
+	u64 pa;
+	enum es_result ret;
+
+	pa = __pa(sev_snp_current_doorbell_page());
+	vc_ghcb_invalidate(ghcb);
+	ret = vmgexit_hv_doorbell_page(ghcb,
+			SVM_VMGEXIT_SET_HV_DOORBELL_PAGE, pa);
+	if (ret != ES_OK)
+		panic("SEV-SNP: failed to set up #HV doorbell page");
+}
+
 static noinstr void __sev_put_ghcb(struct ghcb_state *state)
 {
 	struct sev_es_runtime_data *data;
@@ -1282,6 +1505,11 @@ void setup_ghcb(void)
 		snp_register_ghcb_early(__pa(&boot_ghcb_page));
 }
 
+int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa)
+{
+	return sev_es_ghcb_hv_call(ghcb, NULL, SVM_VMGEXIT_HV_DOORBELL_PAGE, op, pa);
+}
+
 #ifdef CONFIG_HOTPLUG_CPU
 static void sev_es_ap_hlt_loop(void)
 {
@@ -1355,6 +1583,7 @@ static void __init alloc_runtime_data(int cpu)
 static void __init init_ghcb(int cpu)
 {
 	struct sev_es_runtime_data *data;
+	struct sev_snp_runtime_data *snp_data;
 	int err;
 
 	data = per_cpu(runtime_data, cpu);
@@ -1366,8 +1595,22 @@ static void __init init_ghcb(int cpu)
 
 	memset(&data->ghcb_page, 0, sizeof(data->ghcb_page));
 
+	snp_data = memblock_alloc(sizeof(*snp_data), PAGE_SIZE);
+	if (!snp_data)
+		panic("Can't allocate SEV-SNP runtime data");
+
+	err = early_set_memory_decrypted((unsigned long)&snp_data->hv_doorbell_page,
+					 sizeof(snp_data->hv_doorbell_page));
+	if (err)
+		panic("Can't map #HV doorbell pages unencrypted");
+
+	memset(&snp_data->hv_doorbell_page, 0, sizeof(snp_data->hv_doorbell_page));
+
+	per_cpu(snp_runtime_data, cpu) = snp_data;
+
 	data->ghcb_active = false;
 	data->backup_ghcb_active = false;
+	data->ghcb_registered = false;
 }
 
 void __init sev_es_init_vc_handling(void)
@@ -2006,7 +2249,12 @@ DEFINE_IDTENTRY_VC_USER(exc_vmm_communication)
 
 static bool hv_raw_handle_exception(struct pt_regs *regs)
 {
-	return false;
+	/* Clear the no_further_signal bit */
+	sev_snp_current_doorbell_page()->pending_events.events &= 0x7fff;
+
+	check_hv_pending(regs);
+
+	return true;
 }
 
 static __always_inline bool on_hv_fallback_stack(struct pt_regs *regs)
arch/x86/kernel/traps.c +47 -0
index 178015a820f0..af97e6610fbb 100644
@@ -898,6 +898,53 @@ asmlinkage __visible noinstr struct pt_regs *vc_switch_off_ist(struct pt_regs *r
 
 	return regs_ret;
 }
+
+asmlinkage __visible noinstr struct pt_regs *hv_switch_off_ist(struct pt_regs *regs)
+{
+	unsigned long sp, *stack;
+	struct stack_info info;
+	struct pt_regs *regs_ret;
+
+	/*
+	 * A malicious hypervisor can inject 2 HVs in a row, which will corrupt
+	 * the trap frame on our IST stack.  We add a defensive check here to
+	 * catch such behavior.
+	 */
+	BUG_ON(regs->sp >= __this_cpu_ist_bottom_va(HV) && regs->sp < __this_cpu_ist_top_va(HV));
+
+	/*
+	 * In the SYSCALL entry path the RSP value comes from user-space - don't
+	 * trust it and switch to the current kernel stack
+	 */
+	if (ip_within_syscall_gap(regs)) {
+		sp = this_cpu_read(cpu_current_top_of_stack);
+		goto sync;
+	}
+
+	/*
+	 * From here on the RSP value is trusted. Now check whether entry
+	 * happened from a safe stack. Not safe are the entry or unknown stacks,
+	 * use the fall-back stack instead in this case.
+	 */
+	sp    = regs->sp;
+	stack = (unsigned long *)sp;
+
+	if (!get_stack_info_noinstr(stack, current, &info) || info.type == STACK_TYPE_ENTRY ||
+	    info.type > STACK_TYPE_EXCEPTION_LAST)
+		sp = __this_cpu_ist_top_va(HV2);
+sync:
+	/*
+	 * Found a safe stack - switch to it as if the entry didn't happen via
+	 * IST stack. The code below only copies pt_regs, the real switch happens
+	 * in assembly code.
+	 */
+	sp = ALIGN_DOWN(sp, 8) - sizeof(*regs_ret);
+
+	regs_ret = (struct pt_regs *)sp;
+	*regs_ret = *regs;
+
+	return regs_ret;
+}
 #endif
 
 asmlinkage __visible noinstr struct pt_regs *fixup_bad_iret(struct pt_regs *bad_regs)

arch/x86/kernel/traps.c void __init trap_init(void) · line 1510

Quoted 7 lines
arch/x86/kernel/traps.c +3 -0
@@ -1457,4 +1504,7 @@ void __init trap_init(void)
 	/* Setup traps as cpu_init() might #GP */
 	idt_setup_traps();
 	cpu_init();
+
+	/* Init #HV doorbell pages when running as an SEV-SNP guest */
+	sev_snp_init_hv_handling();
 }

--
Cheers,
Christophe de Dinechin (https://c3d.github.io)
Theory of Incomplete Measurements (https://c3d.github.io/TIM)

author
#54 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

Quoted 5 lines · Christophe de Dinechin
Quoted 1 line

+void check_hv_pending(struct pt_regs *regs)

This looks like two functions, one with regs == NULL and one with regs, different internal logic, different call sites. Would you consider splitting into two?

Good suggestion. Will update in the next version.

maintainer
#2 lore raw

Re: [RFC PATCH V2 01/18] x86/sev: Pvalidate memory gab for decompressing kernel

In reply to Tianyu Lan
Quoted 1 line · Tianyu Lan

Subject: Re: [RFC PATCH V2 01/18] x86/sev: Pvalidate memory gab for decompressing kernel

"gab"?

As in gabber? :-)

Quoted 3 lines

From: Tianyu Lan <[email protected]>

Pvalidate needed pages for decompressing kernel. The E820_TYPE_RAM

"Validate" - let's not start inventing new words. We're barely handling the existing ones. :)

Quoted 2 lines

entry includes only validated memory. The kernel expects that the RAM entry's addr is fixed while the entry size is to be extended

"addr"?

Commit message needs to be english - not a code/english hybrid. Pls be more diligent here. Commit messages are not write-only.

Quoted 1 line

to cover addresses to the start of next entry. This patch increases

Avoid having "This patch" or "This commit" in the commit message. It is tautologically useless.

Also, do

$ git grep 'This patch' Documentation/process

for more details.

Pls check your whole set.

Also, to the tone, from Documentation/process/submitting-patches.rst:

"Describe your changes in imperative mood, e.g. "make xyzzy do frotz"
instead of "[This patch] makes xyzzy do frotz" or "[I] changed xyzzy to do frotz", as if you are giving orders to the codebase to change its behaviour."

Quoted 1 line

the RAM entry size to cover all possilble memory addresses until

Unknown word [possilble] in commit message.
Suggestions: ['possible', 'possibly', 'passable', 'plausible', 'assailable', 'pliable', 'passably']

Please introduce a spellchecker into your patch creation workflow.

Quoted 1 line

init_size.

This whole commit message doesn't tell me a whole lot. Please try structuring it this way:

Problem is A.

It happens because of B.

Fix it by doing C.

(Potentially do D).

For more detailed info, see Documentation/process/submitting-patches.rst, Section "2) Describe your changes".

Quoted 11 lines
---
 arch/x86/boot/compressed/head_64.S |  8 +++
 arch/x86/boot/compressed/sev.c     | 84 ++++++++++++++++++++++++++++++
 2 files changed, 92 insertions(+)
arch/x86/boot/compressed/head_64.S +0 -0
index d33f060900d2..818edaf5d0cf 100644

arch/x86/boot/compressed/head_64.S SYM_CODE_START(startup_64) · line 352

Quoted 5 lines
arch/x86/boot/compressed/head_64.S +2 -0
@@ -348,6 +348,14 @@ SYM_CODE_START(startup_64)
 	cld
 	cli
 
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/* pvalidate memory on demand if SNP is enabled. */

So this is going to be executed unconditionally on *every* SNP guest - not only Hyper-V ones.

Why is that ok?

Quoted 33 lines
arch/x86/boot/compressed/head_64.S +6 -0
+	pushq	%rsi
+	movq    %rsi, %rdi
+	call 	pvalidate_for_startup_64
+	popq	%rsi
+#endif
+
 	/* Setup data segments. */
 	xorl	%eax, %eax
 	movl	%eax, %ds
arch/x86/boot/compressed/sev.c +3 -0
index 960968f8bf75..3a5a1ab16095 100644
@@ -12,8 +12,10 @@
  */
 #include "misc.h"
 
+#include <asm/msr-index.h>
 #include <asm/pgtable_types.h>
 #include <asm/sev.h>
+#include <asm/svm.h>
 #include <asm/trapnr.h>
 #include <asm/trap_pf.h>
 #include <asm/msr-index.h>
@@ -21,6 +23,7 @@
 #include <asm/ptrace.h>
 #include <asm/svm.h>
 #include <asm/cpuid.h>
+#include <asm/e820/types.h>
 
 #include "error.h"
 #include "../msr.h"

arch/x86/boot/compressed/sev.c static enum es_result vc_read_mem(struct es_em_ctxt *ctxt, · line 124

Quoted 5 lines
arch/x86/boot/compressed/sev.c +2 -0
@@ -117,6 +120,22 @@ static enum es_result vc_read_mem(struct es_em_ctxt *ctxt,
 /* Include code for early handlers */
 #include "../../kernel/sev-shared.c"
 
+/* Check SEV-SNP via MSR */
+static bool sev_snp_runtime_check(void)

Functions need to have a verb in the name.

arch/x86/boot/compressed/sev.c static enum es_result vc_read_mem(struct es_em_ctxt *ctxt, · line 134

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+{
+	unsigned long low, high;
+	u64 val;
+
+	asm volatile("rdmsr\n" : "=a" (low), "=d" (high) :
+			"c" (MSR_AMD64_SEV));
+
+	val = (high << 32) | low;
+	if (val & MSR_AMD64_SEV_SNP_ENABLED)
+		return true;

There already is a sev_snp_enabled() in that very same file. Did you not see it?

Why are you even adding such a function?!

Quoted 31 lines
arch/x86/boot/compressed/sev.c +24 -0
+	return false;
+}
+
 static inline bool sev_snp_enabled(void)
 {
 	return sev_status & MSR_AMD64_SEV_SNP_ENABLED;
@@ -456,3 +475,68 @@ void sev_prep_identity_maps(unsigned long top_level_pgt)
 
 	sev_verify_cbit(top_level_pgt);
 }
+
+static void extend_e820_on_demand(struct boot_e820_entry *e820_entry,
+				  u64 needed_ram_end)
+{
+	u64 end, paddr;
+	unsigned long eflags;
+	int rc;
+
+	if (!e820_entry)
+		return;
+
+	/* Validated memory must be aligned by PAGE_SIZE. */
+	end = ALIGN(e820_entry->addr + e820_entry->size, PAGE_SIZE);
+	if (needed_ram_end > end && e820_entry->type == E820_TYPE_RAM) {
+		for (paddr = end; paddr < needed_ram_end; paddr += PAGE_SIZE) {
+			rc = pvalidate(paddr, RMP_PG_SIZE_4K, true);
+			if (rc) {
+				error("Failed to validate address.n");
+				return;
+			}
+		}

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 508

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+		e820_entry->size = needed_ram_end - e820_entry->addr;
+	}
+}
+
+/*
+ * Explicitly pvalidate needed pages for decompressing the kernel.
+ * The E820_TYPE_RAM entry includes only validated memory. The kernel
+ * expects that the RAM entry's addr is fixed while the entry size is to be
+ * extended to cover addresses to the start of next entry.
+ * The function increases the RAM entry size to cover all possible memory

Similar issue as above: you don't need to say "this function" above this function. IOW, it should say:

"Increase the RAM entry size..."

I.e., imperative mood above.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 516

Quoted 8 lines
arch/x86/boot/compressed/sev.c +8 -0
+ * addresses until init_size.
+ * For example,  init_end = 0x4000000,
+ * [RAM: 0x0 - 0x0],                       M[RAM: 0x0 - 0xa0000]
+ * [RSVD: 0xa0000 - 0x10000]                [RSVD: 0xa0000 - 0x10000]
+ * [ACPI: 0x10000 - 0x20000]      ==>       [ACPI: 0x10000 - 0x20000]
+ * [RSVD: 0x800000 - 0x900000]              [RSVD: 0x800000 - 0x900000]
+ * [RAM: 0x1000000 - 0x2000000]            M[RAM: 0x1000000 - 0x2001000]
+ * [RAM: 0x2001000 - 0x2007000]            M[RAM: 0x2001000 - 0x4000000]

What is this trying to tell me?

That the end range 0x2007000 gets raised to 0x4000000?

Why?

This all sounds like there is some requirement somewhere but nothing says what that requirement is and why.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 519

Quoted 3 lines
arch/x86/boot/compressed/sev.c +3 -0
+ * Other RAM memory after init_end is pvalidated by ms_hyperv_init_platform
+ */
+__visible void pvalidate_for_startup_64(struct boot_params *boot_params)

This doesn't do any validation. And yet it has "pvalidate" in the name.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 523

Quoted 4 lines
arch/x86/boot/compressed/sev.c +4 -0
+{
+	struct boot_e820_entry *e820_entry;
+	u64 init_end =
+		boot_params->hdr.pref_address + boot_params->hdr.init_size;

Nope, we never break lines like that.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 525

Quoted 2 lines
arch/x86/boot/compressed/sev.c +2 -0
+	u8 i, nr_entries = boot_params->e820_entries;
+	u64 needed_end;

The tip-tree preferred ordering of variable declarations at the beginning of a function is reverse fir tree order::

	struct long_struct_name *descriptive_name;
	unsigned long foo, bar;
	unsigned int tmp;
	int ret;

The above is faster to parse than the reverse ordering::

	int ret;
	unsigned int tmp;
	unsigned long foo, bar;
	struct long_struct_name *descriptive_name;

And even more so than random ordering::

	unsigned long foo, bar;
	int ret;
	struct long_struct_name *descriptive_name;
	unsigned int tmp;
Quoted 4 lines
arch/x86/boot/compressed/sev.c +4 -0
+	if (!sev_snp_runtime_check())
+		return;
+
+	for (i = 0; i < nr_entries; ++i) {

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 540

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+		/* Pvalidate memory holes in e820 RAM entries. */
+		e820_entry = &boot_params->e820_table[i];
+		if (i < nr_entries - 1) {
+			needed_end = boot_params->e820_table[i + 1].addr;
+			if (needed_end < e820_entry->addr)
+				error("e820 table is not sorted.\n");
+		} else {
+			needed_end = init_end;
+		}
+		extend_e820_on_demand(e820_entry, needed_end);

Now *this* function does call pvalidate() and yet it doesn't have "pvalidate" in the name. This all looks real confused.

So first of all, you need to explain *why* you're doing this.

It looks like it is because the guest needs to do the memory validation by itself because nobody else does that.

If so, this needs to be explained in detail in the commit message.

Also, why is that ok for SNP guests on other hypervisors which get the memory validated by the boot loader or firmware?

And so on and so on.

Thx.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
author
#3 lore raw

Re: [RFC PATCH V2 01/18] x86/sev: Pvalidate memory gab for decompressing kernel

Quoted 17 lines · Borislav Petkov

arch/x86/boot/compressed/sev.c static enum es_result vc_read_mem(struct es_em_ctxt *ctxt, · line 124

Quoted 2 lines
arch/x86/boot/compressed/sev.c +2 -0
+/* Check SEV-SNP via MSR */
+static bool sev_snp_runtime_check(void)

Functions need to have a verb in the name.

arch/x86/boot/compressed/sev.c static enum es_result vc_read_mem(struct es_em_ctxt *ctxt, · line 134

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+{
+	unsigned long low, high;
+	u64 val;
+
+	asm volatile("rdmsr\n" : "=a" (low), "=d" (high) :
+			"c" (MSR_AMD64_SEV));
+
+	val = (high << 32) | low;
+	if (val & MSR_AMD64_SEV_SNP_ENABLED)
+		return true;

There already is a sev_snp_enabled() in that very same file. Did you not see it?

Quoted 1 line

Why are you even adding such a function?!

Hi Boris:
Thanks for your review. sev_snp_enabled() is used after
sev_status was initialized in sev_enable() while pvalidate_for_startup_
64() is called before sev_enable(). So add sev_snp_runtime_check() to check sev snp capability before calling sev_enable().

Quoted 118 lines
Quoted 32 lines
arch/x86/boot/compressed/sev.c +24 -0
+	return false;
+}
+
  static inline bool sev_snp_enabled(void)
  {
  	return sev_status & MSR_AMD64_SEV_SNP_ENABLED;
@@ -456,3 +475,68 @@ void sev_prep_identity_maps(unsigned long top_level_pgt)
 
  	sev_verify_cbit(top_level_pgt);
  }
+
+static void extend_e820_on_demand(struct boot_e820_entry *e820_entry,
+				  u64 needed_ram_end)
+{
+	u64 end, paddr;
+	unsigned long eflags;
+	int rc;
+
+	if (!e820_entry)
+		return;
+
+	/* Validated memory must be aligned by PAGE_SIZE. */
+	end = ALIGN(e820_entry->addr + e820_entry->size, PAGE_SIZE);
+	if (needed_ram_end > end && e820_entry->type == E820_TYPE_RAM) {
+		for (paddr = end; paddr < needed_ram_end; paddr += PAGE_SIZE) {
+			rc = pvalidate(paddr, RMP_PG_SIZE_4K, true);
+			if (rc) {
+				error("Failed to validate address.n");
+				return;
+			}
+		}

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 508

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+		e820_entry->size = needed_ram_end - e820_entry->addr;
+	}
+}
+
+/*
+ * Explicitly pvalidate needed pages for decompressing the kernel.
+ * The E820_TYPE_RAM entry includes only validated memory. The kernel
+ * expects that the RAM entry's addr is fixed while the entry size is to be
+ * extended to cover addresses to the start of next entry.
+ * The function increases the RAM entry size to cover all possible memory

Similar issue as above: you don't need to say "this function" above this function. IOW, it should say:

"Increase the RAM entry size..."

I.e., imperative mood above.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 516

Quoted 8 lines
arch/x86/boot/compressed/sev.c +8 -0
+ * addresses until init_size.
+ * For example,  init_end = 0x4000000,
+ * [RAM: 0x0 - 0x0],                       M[RAM: 0x0 - 0xa0000]
+ * [RSVD: 0xa0000 - 0x10000]                [RSVD: 0xa0000 - 0x10000]
+ * [ACPI: 0x10000 - 0x20000]      ==>       [ACPI: 0x10000 - 0x20000]
+ * [RSVD: 0x800000 - 0x900000]              [RSVD: 0x800000 - 0x900000]
+ * [RAM: 0x1000000 - 0x2000000]            M[RAM: 0x1000000 - 0x2001000]
+ * [RAM: 0x2001000 - 0x2007000]            M[RAM: 0x2001000 - 0x4000000]

What is this trying to tell me?

That the end range 0x2007000 gets raised to 0x4000000?

Why?

This all sounds like there is some requirement somewhere but nothing says what that requirement is and why.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 519

Quoted 3 lines
arch/x86/boot/compressed/sev.c +3 -0
+ * Other RAM memory after init_end is pvalidated by ms_hyperv_init_platform
+ */
+__visible void pvalidate_for_startup_64(struct boot_params *boot_params)

This doesn't do any validation. And yet it has "pvalidate" in the name.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 523

Quoted 4 lines
arch/x86/boot/compressed/sev.c +4 -0
+{
+	struct boot_e820_entry *e820_entry;
+	u64 init_end =
+		boot_params->hdr.pref_address + boot_params->hdr.init_size;

Nope, we never break lines like that.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 525

Quoted 2 lines
arch/x86/boot/compressed/sev.c +2 -0
+	u8 i, nr_entries = boot_params->e820_entries;
+	u64 needed_end;

The tip-tree preferred ordering of variable declarations at the beginning of a function is reverse fir tree order::

	struct long_struct_name *descriptive_name;
	unsigned long foo, bar;
	unsigned int tmp;
	int ret;

The above is faster to parse than the reverse ordering::

	int ret;
	unsigned int tmp;
	unsigned long foo, bar;
	struct long_struct_name *descriptive_name;

And even more so than random ordering::

	unsigned long foo, bar;
	int ret;
	struct long_struct_name *descriptive_name;
	unsigned int tmp;
Quoted 4 lines
arch/x86/boot/compressed/sev.c +4 -0
+	if (!sev_snp_runtime_check())
+		return;
+
+	for (i = 0; i < nr_entries; ++i) {

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 540

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+		/* Pvalidate memory holes in e820 RAM entries. */
+		e820_entry = &boot_params->e820_table[i];
+		if (i < nr_entries - 1) {
+			needed_end = boot_params->e820_table[i + 1].addr;
+			if (needed_end < e820_entry->addr)
+				error("e820 table is not sorted.\n");
+		} else {
+			needed_end = init_end;
+		}
+		extend_e820_on_demand(e820_entry, needed_end);

Now*this* function does call pvalidate() and yet it doesn't have "pvalidate" in the name. This all looks real confused.

Quoted 6 lines

So first of all, you need to explain*why* you're doing this.

It looks like it is because the guest needs to do the memory validation by itself because nobody else does that.

If so, this needs to be explained in detail in the commit message.

Yes, I will update in the next version. Thanks for suggestion.

Quoted 3 lines

Also, why is that ok for SNP guests on other hypervisors which get the memory validated by the boot loader or firmware?

This is for Linux direct boot mode and so it needs to do such check here. Will fix this in the next version.

maintainer
#4 lore raw

Re: [RFC PATCH V2 01/18] x86/sev: Pvalidate memory gab for decompressing kernel

Quoted 3 lines · Tianyu Lan

Thanks for your review. sev_snp_enabled() is used after sev_status was initialized in sev_enable() while pvalidate_for_startup_ 64() is called before sev_enable().

Then you're going to have to change the code so that sev_status is initialized before you need it. And not break others in the process.

And lemme save you some time - I won't accept sloppy code. You need to integrate the functionality you need in the code paths properly - not bolt it on in complete disregard of the flow just because it is easier.

Quoted 2 lines

This is for Linux direct boot mode and so it needs to do such check here.

I don't know what "Linux direct boot mode" is so until you define it properly and explain everything in detail, this is not going anywhere.

Thx.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
#6 lore raw

Re: [RFC PATCH V2 01/18] x86/sev: Pvalidate memory gab for decompressing kernel

In reply to Tianyu Lan

Hi Tianyu,

Some minor nits below.

Quoted 5 lines

Pvalidate needed pages for decompressing kernel. The E820_TYPE_RAM entry includes only validated memory. The kernel expects that the RAM entry's addr is fixed while the entry size is to be extended to cover addresses to the start of next entry. This patch increases the RAM entry size to cover all possilble memory addresses until

s/possilble/possible

Quoted 81 lines

init_size.

---
  arch/x86/boot/compressed/head_64.S |  8 +++
  arch/x86/boot/compressed/sev.c     | 84 ++++++++++++++++++++++++++++++
  2 files changed, 92 insertions(+)
arch/x86/boot/compressed/head_64.S +8 -0
index d33f060900d2..818edaf5d0cf 100644
@@ -348,6 +348,14 @@ SYM_CODE_START(startup_64)
  	cld
  	cli
 
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/* pvalidate memory on demand if SNP is enabled. */
+	pushq	%rsi
+	movq    %rsi, %rdi
+	call 	pvalidate_for_startup_64
+	popq	%rsi
+#endif
+
  	/* Setup data segments. */
  	xorl	%eax, %eax
  	movl	%eax, %ds
arch/x86/boot/compressed/sev.c +23 -0
index 960968f8bf75..3a5a1ab16095 100644
@@ -12,8 +12,10 @@
   */
  #include "misc.h"
 
+#include <asm/msr-index.h>
  #include <asm/pgtable_types.h>
  #include <asm/sev.h>
+#include <asm/svm.h>
  #include <asm/trapnr.h>
  #include <asm/trap_pf.h>
  #include <asm/msr-index.h>
@@ -21,6 +23,7 @@
  #include <asm/ptrace.h>
  #include <asm/svm.h>
  #include <asm/cpuid.h>
+#include <asm/e820/types.h>
 
  #include "error.h"
  #include "../msr.h"
@@ -117,6 +120,22 @@ static enum es_result vc_read_mem(struct es_em_ctxt *ctxt,
  /* Include code for early handlers */
  #include "../../kernel/sev-shared.c"
 
+/* Check SEV-SNP via MSR */
+static bool sev_snp_runtime_check(void)
+{
+	unsigned long low, high;
+	u64 val;
+
+	asm volatile("rdmsr\n" : "=a" (low), "=d" (high) :
+			"c" (MSR_AMD64_SEV));
+
+	val = (high << 32) | low;
+	if (val & MSR_AMD64_SEV_SNP_ENABLED)
+		return true;
+
+	return false;
+}
+
  static inline bool sev_snp_enabled(void)
  {
  	return sev_status & MSR_AMD64_SEV_SNP_ENABLED;
@@ -456,3 +475,68 @@ void sev_prep_identity_maps(unsigned long top_level_pgt)
 
  	sev_verify_cbit(top_level_pgt);
  }
+
+static void extend_e820_on_demand(struct boot_e820_entry *e820_entry,
+				  u64 needed_ram_end)
+{

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 491

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+	u64 end, paddr;
+	unsigned long eflags;
+	int rc;
+
+	if (!e820_entry)
+		return;
+
+	/* Validated memory must be aligned by PAGE_SIZE. */
+	end = ALIGN(e820_entry->addr + e820_entry->size, PAGE_SIZE);
+	if (needed_ram_end > end && e820_entry->type == E820_TYPE_RAM) {

maybe "if check" can be used to return from here, would read code better?

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 492

Quoted 1 line
arch/x86/boot/compressed/sev.c +1 -0
+		for (paddr = end; paddr < needed_ram_end; paddr += PAGE_SIZE) 

Maybe we could reuse 'pvalidate_pages' here?

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 495

Quoted 3 lines
arch/x86/boot/compressed/sev.c +3 -0
+			rc = pvalidate(paddr, RMP_PG_SIZE_4K, true);
+			if (rc) {
+				error("Failed to validate address.n");

Would it make sense to print the cause of failure here?

Quoted 20 lines
arch/x86/boot/compressed/sev.c +20 -0
+				return;
+			}
+		}
+		e820_entry->size = needed_ram_end - e820_entry->addr;
+	}
+}
+
+/*
+ * Explicitly pvalidate needed pages for decompressing the kernel.
+ * The E820_TYPE_RAM entry includes only validated memory. The kernel
+ * expects that the RAM entry's addr is fixed while the entry size is to be
+ * extended to cover addresses to the start of next entry.
+ * The function increases the RAM entry size to cover all possible memory
+ * addresses until init_size.
+ * For example,  init_end = 0x4000000,
+ * [RAM: 0x0 - 0x0],                       M[RAM: 0x0 - 0xa0000]
+ * [RSVD: 0xa0000 - 0x10000]                [RSVD: 0xa0000 - 0x10000]
+ * [ACPI: 0x10000 - 0x20000]      ==>       [ACPI: 0x10000 - 0x20000]
+ * [RSVD: 0x800000 - 0x900000]              [RSVD: 0x800000 - 0x900000]
+ * [RAM: 0x1000000 - 0x2000000]            M[RAM: 0x1000000 - 0x2001000]

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 525

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+ * [RAM: 0x2001000 - 0x2007000]            M[RAM: 0x2001000 - 0x4000000]
+ * Other RAM memory after init_end is pvalidated by ms_hyperv_init_platform
+ */
+__visible void pvalidate_for_startup_64(struct boot_params *boot_params)
+{
+	struct boot_e820_entry *e820_entry;
+	u64 init_end =
+		boot_params->hdr.pref_address + boot_params->hdr.init_size;
+	u8 i, nr_entries = boot_params->e820_entries;
+	u64 needed_end;

Could not very well interpret the name 'needed_end'.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 531

Quoted 6 lines
arch/x86/boot/compressed/sev.c +6 -0
+
+	if (!sev_snp_runtime_check())
+		return;
+
+	for (i = 0; i < nr_entries; ++i) {
+		/* Pvalidate memory holes in e820 RAM entries. */

Pvalidate and pvalidate names are used interchangeably in comments.

Quoted 11 lines
arch/x86/boot/compressed/sev.c +11 -0
+		e820_entry = &boot_params->e820_table[i];
+		if (i < nr_entries - 1) {
+			needed_end = boot_params->e820_table[i + 1].addr;
+			if (needed_end < e820_entry->addr)
+				error("e820 table is not sorted.\n");
+		} else {
+			needed_end = init_end;
+		}
+		extend_e820_on_demand(e820_entry, needed_end);
+	}
+}
#55 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

In reply to Tianyu Lan
Quoted 70 lines

From: Tianyu Lan <[email protected]>

Enable #HV exception to handle interrupt requests from hypervisor.

---
  arch/x86/entry/entry_64.S          |  18 ++
  arch/x86/include/asm/irqflags.h    |  19 ++
  arch/x86/include/asm/mem_encrypt.h |   2 +
  arch/x86/include/asm/msr-index.h   |   6 +
  arch/x86/include/uapi/asm/svm.h    |   4 +
  arch/x86/kernel/sev.c              | 354 ++++++++++++++++++++++++-----
  arch/x86/kernel/traps.c            |  50 ++++
  7 files changed, 400 insertions(+), 53 deletions(-)
arch/x86/entry/entry_64.S +18 -0
index b2059df43c57..fe460cf44ab5 100644
@@ -1058,6 +1058,15 @@ SYM_CODE_END(paranoid_entry)
   * R15 - old SPEC_CTRL
   */
  SYM_CODE_START_LOCAL(paranoid_exit)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
  	UNWIND_HINT_REGS
 
  	/*
@@ -1183,6 +1192,15 @@ SYM_CODE_START_LOCAL(error_entry)
  SYM_CODE_END(error_entry)
 
  SYM_CODE_START_LOCAL(error_return)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
  	UNWIND_HINT_REGS
  	DEBUG_ENTRY_ASSERT_IRQS_OFF
  	testb	$3, CS(%rsp)
arch/x86/include/asm/irqflags.h +3 -0
index 7793e52d6237..e0730d8bc0ac 100644
@@ -14,6 +14,9 @@
  /*
   * Interrupt control:
   */
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+void check_hv_pending(struct pt_regs *regs);
+#endif
 
  /* Declaration required for gcc < 4.9 to prevent -Werror=missing-prototypes */
  extern inline unsigned long native_save_fl(void);
@@ -35,6 +38,19 @@ extern __always_inline unsigned long native_save_fl(void)
  	return flags;

arch/x86/include/asm/irqflags.h extern __always_inline unsigned long native_save_fl(void) · line 48

Quoted 10 lines
arch/x86/include/asm/irqflags.h +8 -0
  }
 
+extern inline void native_restore_fl(unsigned long flags)
+{
+	asm volatile("push %0 ; popf"
+		     : /* no output */
+		     : "g" (flags)
+		     : "memory", "cc");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	if ((flags & X86_EFLAGS_IF)) {

Single '()' pair is enough.

Quoted 9 lines
arch/x86/include/asm/irqflags.h +5 -0
+		check_hv_pending(NULL);
+	}
+#endif
+}
+
  static __always_inline void native_irq_disable(void)
  {
  	asm volatile("cli": : :"memory");

arch/x86/include/asm/irqflags.h static __always_inline void native_irq_disable(void) · line 63

Quoted 5 lines
arch/x86/include/asm/irqflags.h +2 -0
@@ -43,6 +59,9 @@ static __always_inline void native_irq_disable(void)
  static __always_inline void native_irq_enable(void)
  {
  	asm volatile("sti": : :"memory");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	check_hv_pending(NULL);

Just trying to understand when regs will be NULL?

Quoted 228 lines
arch/x86/include/asm/irqflags.h +1 -0
+#endif
  }
 
  static inline __cpuidle void native_safe_halt(void)
arch/x86/include/asm/mem_encrypt.h +2 -0
index 72ca90552b6a..7264ca5f5b2d 100644
@@ -50,6 +50,7 @@ void __init early_set_mem_enc_dec_hypercall(unsigned long vaddr, int npages,
  void __init mem_encrypt_free_decrypted_mem(void);
 
  void __init sev_es_init_vc_handling(void);
+void __init sev_snp_init_hv_handling(void);
 
  #define __bss_decrypted __section(".bss..decrypted")
 
@@ -72,6 +73,7 @@ static inline void __init sme_encrypt_kernel(struct boot_params *bp) { }
  static inline void __init sme_enable(struct boot_params *bp) { }
 
  static inline void sev_es_init_vc_handling(void) { }
+static inline void sev_snp_init_hv_handling(void) { }
 
  static inline int __init
  early_set_memory_decrypted(unsigned long vaddr, unsigned long size) { return 0; }
arch/x86/include/asm/msr-index.h +6 -0
index 10ac52705892..6fe25a6e325f 100644
@@ -562,10 +562,16 @@
  #define MSR_AMD64_SEV_ENABLED_BIT	0
  #define MSR_AMD64_SEV_ES_ENABLED_BIT	1
  #define MSR_AMD64_SEV_SNP_ENABLED_BIT	2
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT		4
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT	5
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT	6
  #define MSR_AMD64_SEV_ENABLED		BIT_ULL(MSR_AMD64_SEV_ENABLED_BIT)
  #define MSR_AMD64_SEV_ES_ENABLED	BIT_ULL(MSR_AMD64_SEV_ES_ENABLED_BIT)
  #define MSR_AMD64_SEV_SNP_ENABLED	BIT_ULL(MSR_AMD64_SEV_SNP_ENABLED_BIT)
 
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED			BIT_ULL(MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT)
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT)
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT)
  #define MSR_AMD64_VIRT_SPEC_CTRL	0xc001011f
 
  /* AMD Collaborative Processor Performance Control MSRs */
arch/x86/include/uapi/asm/svm.h +4 -0
index f69c168391aa..85d6882262e7 100644
@@ -115,6 +115,10 @@
  #define SVM_VMGEXIT_AP_CREATE_ON_INIT		0
  #define SVM_VMGEXIT_AP_CREATE			1
  #define SVM_VMGEXIT_AP_DESTROY			2
+#define SVM_VMGEXIT_HV_DOORBELL_PAGE		0x80000014
+#define SVM_VMGEXIT_GET_PREFERRED_HV_DOORBELL_PAGE	0
+#define SVM_VMGEXIT_SET_HV_DOORBELL_PAGE		1
+#define SVM_VMGEXIT_QUERY_HV_DOORBELL_PAGE		2
  #define SVM_VMGEXIT_HV_FEATURES			0x8000fffd
  #define SVM_VMGEXIT_UNSUPPORTED_EVENT		0x8000ffff
 
arch/x86/kernel/sev.c +153 -0
index b54ee3ba37b0..23cd025f97dc 100644
@@ -104,6 +104,12 @@ struct sev_es_runtime_data {
  	 * is currently unsupported in SEV-ES guests.
  	 */
  	unsigned long dr7;
+	/*
+	 * SEV-SNP requires that the GHCB must be registered before using it.
+	 * The flag below will indicate whether the GHCB is registered, if its
+	 * not registered then sev_es_get_ghcb() will perform the registration.
+	 */
+	bool ghcb_registered;
  };
 
  struct ghcb_state {
@@ -122,6 +128,183 @@ struct sev_config {
 
  static struct sev_config sev_cfg __read_mostly;
 
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state);
+static noinstr void __sev_put_ghcb(struct ghcb_state *state);
+static int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa);
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb);
+
+union hv_pending_events {
+	u16 events;
+	struct {
+		u8 vector;
+		u8 nmi : 1;
+		u8 mc : 1;
+		u8 reserved1 : 5;
+		u8 no_further_signal : 1;
+	};
+};
+
+struct sev_hv_doorbell_page {
+	union hv_pending_events pending_events;
+	u8 no_eoi_required;
+	u8 reserved2[61];
+	u8 padding[4032];
+};
+
+struct sev_snp_runtime_data {
+	struct sev_hv_doorbell_page hv_doorbell_page;
+};
+
+static DEFINE_PER_CPU(struct sev_snp_runtime_data*, snp_runtime_data);
+
+static inline u64 sev_es_rd_ghcb_msr(void)
+{
+	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
+}
+
+static __always_inline void sev_es_wr_ghcb_msr(u64 val)
+{
+	u32 low, high;
+
+	low  = (u32)(val);
+	high = (u32)(val >> 32);
+
+	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+}
+
+struct sev_hv_doorbell_page *sev_snp_current_doorbell_page(void)
+{
+	return &this_cpu_read(snp_runtime_data)->hv_doorbell_page;
+}
+
+static u8 sev_hv_pending(void)
+{
+	return sev_snp_current_doorbell_page()->pending_events.events;
+}
+
+static void hv_doorbell_apic_eoi_write(u32 reg, u32 val)
+{
+	if (xchg(&sev_snp_current_doorbell_page()->no_eoi_required, 0) & 0x1)
+		return;
+
+	BUG_ON(reg != APIC_EOI);
+	apic->write(reg, val);
+}
+
+static void do_exc_hv(struct pt_regs *regs)
+{
+	union hv_pending_events pending_events;
+	u8 vector;
+
+	while (sev_hv_pending()) {
+		asm volatile("cli" : : : "memory");
+
+		pending_events.events = xchg(
+			&sev_snp_current_doorbell_page()->pending_events.events,
+			0);
+
+		if (pending_events.nmi)
+			exc_nmi(regs);
+
+#ifdef CONFIG_X86_MCE
+		if (pending_events.mc)
+			exc_machine_check(regs);
+#endif
+
+		if (!pending_events.vector)
+			return;
+
+		if (pending_events.vector < FIRST_EXTERNAL_VECTOR) {
+			/* Exception vectors */
+			WARN(1, "exception shouldn't happen\n");
+		} else if (pending_events.vector == FIRST_EXTERNAL_VECTOR) {
+			sysvec_irq_move_cleanup(regs);
+		} else if (pending_events.vector == IA32_SYSCALL_VECTOR) {
+			WARN(1, "syscall shouldn't happen\n");
+		} else if (pending_events.vector >= FIRST_SYSTEM_VECTOR) {
+			switch (pending_events.vector) {
+#if IS_ENABLED(CONFIG_HYPERV)
+			case HYPERV_STIMER0_VECTOR:
+				sysvec_hyperv_stimer0(regs);
+				break;
+			case HYPERVISOR_CALLBACK_VECTOR:
+				sysvec_hyperv_callback(regs);
+				break;
+#endif
+#ifdef CONFIG_SMP
+			case RESCHEDULE_VECTOR:
+				sysvec_reschedule_ipi(regs);
+				break;
+			case IRQ_MOVE_CLEANUP_VECTOR:
+				sysvec_irq_move_cleanup(regs);
+				break;
+			case REBOOT_VECTOR:
+				sysvec_reboot(regs);
+				break;
+			case CALL_FUNCTION_SINGLE_VECTOR:
+				sysvec_call_function_single(regs);
+				break;
+			case CALL_FUNCTION_VECTOR:
+				sysvec_call_function(regs);
+				break;
+#endif
+#ifdef CONFIG_X86_LOCAL_APIC
+			case ERROR_APIC_VECTOR:
+				sysvec_error_interrupt(regs);
+				break;
+			case SPURIOUS_APIC_VECTOR:
+				sysvec_spurious_apic_interrupt(regs);
+				break;
+			case LOCAL_TIMER_VECTOR:
+				sysvec_apic_timer_interrupt(regs);
+				break;
+			case X86_PLATFORM_IPI_VECTOR:
+				sysvec_x86_platform_ipi(regs);
+				break;
+#endif
+			case 0x0:
+				break;
+			default:
+				panic("Unexpected vector %d\n", vector);
+				unreachable();
+			}
+		} else {
+			common_interrupt(regs, pending_events.vector);
+		}
+
+		asm volatile("sti" : : : "memory");
+	}
+}

arch/x86/kernel/sev.c struct sev_config { · line 287

Quoted 10 lines
arch/x86/kernel/sev.c +10 -0
+
+void check_hv_pending(struct pt_regs *regs)
+{
+	struct pt_regs local_regs;
+
+	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
+		return;
+
+	if (regs) {
+		if ((regs->flags & X86_EFLAGS_IF) == 0)

Mostly saw !(regs->flags & X86_EFLAGS_IF)

arch/x86/kernel/sev.c struct sev_config { · line 290

Quoted 3 lines
arch/x86/kernel/sev.c +3 -0
+			return;
+
+		if (!sev_hv_pending())

Maybe we can have a common sev_hv_pending() check for if & else.

arch/x86/kernel/sev.c struct sev_config { · line 302

Quoted 12 lines
arch/x86/kernel/sev.c +12 -0
+			return;
+
+		do_exc_hv(regs);
+	} else {
+		if (sev_hv_pending()) {
+			memset(&local_regs, 0, sizeof(struct pt_regs));
+			regs = &local_regs;
+			asm volatile("movl %%cs, %%eax;" : "=a" (regs->cs));
+			asm volatile("movl %%ss, %%eax;" : "=a" (regs->ss));
+			regs->orig_ax = 0xffffffff;
+			regs->flags = native_save_fl();
+			do_exc_hv(regs);

This also looks redundant for both if-else.

Quoted 299 lines
arch/x86/kernel/sev.c +123 -53
+		}
+	}
+}
+EXPORT_SYMBOL_GPL(check_hv_pending);
+
  static __always_inline bool on_vc_stack(struct pt_regs *regs)
  {
  	unsigned long sp = regs->sp;
@@ -193,68 +376,35 @@ void noinstr __sev_es_ist_exit(void)
  	this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], *(unsigned long *)ist);
  }
 
-/*
- * Nothing shall interrupt this code path while holding the per-CPU
- * GHCB. The backup GHCB is only for NMIs interrupting this path.
- *
- * Callers must disable local interrupts around it.
- */
-static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+static bool sev_restricted_injection_enabled(void)
  {
-	struct sev_es_runtime_data *data;
+	return sev_status & MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED;
+}
+
+void __init sev_snp_init_hv_handling(void)
+{
+	struct sev_snp_runtime_data *snp_data;
+	struct ghcb_state state;
  	struct ghcb *ghcb;
+	unsigned long flags;
+	int cpu;
+	int err;
 
  	WARN_ON(!irqs_disabled());
+	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP) || !sev_restricted_injection_enabled())
+		return;
 
-	data = this_cpu_read(runtime_data);
-	ghcb = &data->ghcb_page;
-
-	if (unlikely(data->ghcb_active)) {
-		/* GHCB is already in use - save its contents */
-
-		if (unlikely(data->backup_ghcb_active)) {
-			/*
-			 * Backup-GHCB is also already in use. There is no way
-			 * to continue here so just kill the machine. To make
-			 * panic() work, mark GHCBs inactive so that messages
-			 * can be printed out.
-			 */
-			data->ghcb_active        = false;
-			data->backup_ghcb_active = false;
-
-			instrumentation_begin();
-			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
-			instrumentation_end();
-		}
-
-		/* Mark backup_ghcb active before writing to it */
-		data->backup_ghcb_active = true;
-
-		state->ghcb = &data->backup_ghcb;
-
-		/* Backup GHCB content */
-		*state->ghcb = *ghcb;
-	} else {
-		state->ghcb = NULL;
-		data->ghcb_active = true;
-	}
+	local_irq_save(flags);
 
-	return ghcb;
-}
+	ghcb = __sev_get_ghcb(&state);
 
-static inline u64 sev_es_rd_ghcb_msr(void)
-{
-	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
-}
+	sev_snp_setup_hv_doorbell_page(ghcb);
 
-static __always_inline void sev_es_wr_ghcb_msr(u64 val)
-{
-	u32 low, high;
+	__sev_put_ghcb(&state);
 
-	low  = (u32)(val);
-	high = (u32)(val >> 32);
+	apic_set_eoi_write(hv_doorbell_apic_eoi_write);
 
-	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+	local_irq_restore(flags);
  }
 
  static int vc_fetch_insn_kernel(struct es_em_ctxt *ctxt,
@@ -515,6 +665,79 @@ static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt
  /* Include code shared with pre-decompression boot stage */
  #include "sev-shared.c"
 
+/*
+ * Nothing shall interrupt this code path while holding the per-CPU
+ * GHCB. The backup GHCB is only for NMIs interrupting this path.
+ *
+ * Callers must disable local interrupts around it.
+ */
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+{
+	struct sev_es_runtime_data *data;
+	struct ghcb *ghcb;
+
+	WARN_ON(!irqs_disabled());
+
+	data = this_cpu_read(runtime_data);
+	ghcb = &data->ghcb_page;
+
+	if (unlikely(data->ghcb_active)) {
+		/* GHCB is already in use - save its contents */
+
+		if (unlikely(data->backup_ghcb_active)) {
+			/*
+			 * Backup-GHCB is also already in use. There is no way
+			 * to continue here so just kill the machine. To make
+			 * panic() work, mark GHCBs inactive so that messages
+			 * can be printed out.
+			 */
+			data->ghcb_active        = false;
+			data->backup_ghcb_active = false;
+
+			instrumentation_begin();
+			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
+			instrumentation_end();
+		}
+
+		/* Mark backup_ghcb active before writing to it */
+		data->backup_ghcb_active = true;
+
+		state->ghcb = &data->backup_ghcb;
+
+		/* Backup GHCB content */
+		*state->ghcb = *ghcb;
+	} else {
+		state->ghcb = NULL;
+		data->ghcb_active = true;
+	}
+
+	/* SEV-SNP guest requires that GHCB must be registered before using it. */
+	if (!data->ghcb_registered) {
+		if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+			snp_register_ghcb_early(__pa(ghcb));
+			sev_snp_setup_hv_doorbell_page(ghcb);
+		} else {
+			sev_es_wr_ghcb_msr(__pa(ghcb));
+		}
+		data->ghcb_registered = true;
+	}
+
+	return ghcb;
+}
+
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb)
+{
+	u64 pa;
+	enum es_result ret;
+
+	pa = __pa(sev_snp_current_doorbell_page());
+	vc_ghcb_invalidate(ghcb);
+	ret = vmgexit_hv_doorbell_page(ghcb,
+			SVM_VMGEXIT_SET_HV_DOORBELL_PAGE, pa);
+	if (ret != ES_OK)
+		panic("SEV-SNP: failed to set up #HV doorbell page");
+}
+
  static noinstr void __sev_put_ghcb(struct ghcb_state *state)
  {
  	struct sev_es_runtime_data *data;
@@ -1282,6 +1505,11 @@ void setup_ghcb(void)
  		snp_register_ghcb_early(__pa(&boot_ghcb_page));
  }
 
+int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa)
+{
+	return sev_es_ghcb_hv_call(ghcb, NULL, SVM_VMGEXIT_HV_DOORBELL_PAGE, op, pa);
+}
+
  #ifdef CONFIG_HOTPLUG_CPU
  static void sev_es_ap_hlt_loop(void)
  {
@@ -1355,6 +1583,7 @@ static void __init alloc_runtime_data(int cpu)
  static void __init init_ghcb(int cpu)
  {
  	struct sev_es_runtime_data *data;
+	struct sev_snp_runtime_data *snp_data;
  	int err;
 
  	data = per_cpu(runtime_data, cpu);
@@ -1366,8 +1595,22 @@ static void __init init_ghcb(int cpu)
 
  	memset(&data->ghcb_page, 0, sizeof(data->ghcb_page));
 
+	snp_data = memblock_alloc(sizeof(*snp_data), PAGE_SIZE);
+	if (!snp_data)
+		panic("Can't allocate SEV-SNP runtime data");
+
+	err = early_set_memory_decrypted((unsigned long)&snp_data->hv_doorbell_page,
+					 sizeof(snp_data->hv_doorbell_page));
+	if (err)
+		panic("Can't map #HV doorbell pages unencrypted");
+
+	memset(&snp_data->hv_doorbell_page, 0, sizeof(snp_data->hv_doorbell_page));
+
+	per_cpu(snp_runtime_data, cpu) = snp_data;
+
  	data->ghcb_active = false;
  	data->backup_ghcb_active = false;
+	data->ghcb_registered = false;
  }
 
  void __init sev_es_init_vc_handling(void)
@@ -2006,7 +2249,12 @@ DEFINE_IDTENTRY_VC_USER(exc_vmm_communication)
 
  static bool hv_raw_handle_exception(struct pt_regs *regs)
  {
-	return false;
+	/* Clear the no_further_signal bit */
+	sev_snp_current_doorbell_page()->pending_events.events &= 0x7fff;
+
+	check_hv_pending(regs);
+
+	return true;
  }
 
  static __always_inline bool on_hv_fallback_stack(struct pt_regs *regs)
arch/x86/kernel/traps.c +50 -0
index 178015a820f0..af97e6610fbb 100644
@@ -898,6 +898,53 @@ asmlinkage __visible noinstr struct pt_regs *vc_switch_off_ist(struct pt_regs *r
 
  	return regs_ret;
  }
+
+asmlinkage __visible noinstr struct pt_regs *hv_switch_off_ist(struct pt_regs *regs)
+{
+	unsigned long sp, *stack;
+	struct stack_info info;
+	struct pt_regs *regs_ret;
+
+	/*
+	 * A malicious hypervisor can inject 2 HVs in a row, which will corrupt
+	 * the trap frame on our IST stack.  We add a defensive check here to
+	 * catch such behavior.
+	 */
+	BUG_ON(regs->sp >= __this_cpu_ist_bottom_va(HV) && regs->sp < __this_cpu_ist_top_va(HV));
+
+	/*
+	 * In the SYSCALL entry path the RSP value comes from user-space - don't
+	 * trust it and switch to the current kernel stack
+	 */
+	if (ip_within_syscall_gap(regs)) {
+		sp = this_cpu_read(cpu_current_top_of_stack);
+		goto sync;
+	}
+
+	/*
+	 * From here on the RSP value is trusted. Now check whether entry
+	 * happened from a safe stack. Not safe are the entry or unknown stacks,
+	 * use the fall-back stack instead in this case.
+	 */
+	sp    = regs->sp;
+	stack = (unsigned long *)sp;
+
+	if (!get_stack_info_noinstr(stack, current, &info) || info.type == STACK_TYPE_ENTRY ||
+	    info.type > STACK_TYPE_EXCEPTION_LAST)
+		sp = __this_cpu_ist_top_va(HV2);
+sync:
+	/*
+	 * Found a safe stack - switch to it as if the entry didn't happen via
+	 * IST stack. The code below only copies pt_regs, the real switch happens
+	 * in assembly code.
+	 */
+	sp = ALIGN_DOWN(sp, 8) - sizeof(*regs_ret);
+
+	regs_ret = (struct pt_regs *)sp;
+	*regs_ret = *regs;
+
+	return regs_ret;
+}
  #endif
 
  asmlinkage __visible noinstr struct pt_regs *fixup_bad_iret(struct pt_regs *bad_regs)
@@ -1457,4 +1504,7 @@ void __init trap_init(void)
  	/* Setup traps as cpu_init() might #GP */
  	idt_setup_traps();
  	cpu_init();
+
+	/* Init #HV doorbell pages when running as an SEV-SNP guest */
+	sev_snp_init_hv_handling();
  }
#58 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

In reply to Tianyu Lan
Quoted 68 lines · Tianyu Lan

From: Tianyu Lan <[email protected]>

Enable #HV exception to handle interrupt requests from hypervisor.

---
  arch/x86/entry/entry_64.S          |  18 ++
  arch/x86/include/asm/irqflags.h    |  19 ++
  arch/x86/include/asm/mem_encrypt.h |   2 +
  arch/x86/include/asm/msr-index.h   |   6 +
  arch/x86/include/uapi/asm/svm.h    |   4 +
  arch/x86/kernel/sev.c              | 354 ++++++++++++++++++++++++-----
  arch/x86/kernel/traps.c            |  50 ++++
  7 files changed, 400 insertions(+), 53 deletions(-)
arch/x86/entry/entry_64.S +18 -0
index b2059df43c57..fe460cf44ab5 100644
@@ -1058,6 +1058,15 @@ SYM_CODE_END(paranoid_entry)
   * R15 - old SPEC_CTRL
   */
  SYM_CODE_START_LOCAL(paranoid_exit)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
  	UNWIND_HINT_REGS
 
  	/*
@@ -1183,6 +1192,15 @@ SYM_CODE_START_LOCAL(error_entry)
  SYM_CODE_END(error_entry)
 
  SYM_CODE_START_LOCAL(error_return)
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	/*
+	 * If a #HV was delivered during execution and interrupts were
+	 * disabled, then check if it can be handled before the iret
+	 * (which may re-enable interrupts).
+	 */
+	mov     %rsp, %rdi
+	call    check_hv_pending
+#endif
  	UNWIND_HINT_REGS
  	DEBUG_ENTRY_ASSERT_IRQS_OFF
  	testb	$3, CS(%rsp)
arch/x86/include/asm/irqflags.h +3 -0
index 7793e52d6237..e0730d8bc0ac 100644
@@ -14,6 +14,9 @@
  /*
   * Interrupt control:
   */
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+void check_hv_pending(struct pt_regs *regs);
+#endif
 
  /* Declaration required for gcc < 4.9 to prevent -Werror=missing-prototypes */
  extern inline unsigned long native_save_fl(void);

arch/x86/include/asm/irqflags.h extern __always_inline unsigned long native_save_fl(void) · line 41

Quoted 4 lines
arch/x86/include/asm/irqflags.h +1 -0
@@ -35,6 +38,19 @@ extern __always_inline unsigned long native_save_fl(void)
  	return flags;
  }
 
+extern inline void native_restore_fl(unsigned long flags)

Don't know if want to re-introduce this again? as it was removed with ab234a260b1f ("x86/pv: Rework arch_local_irq_restore() to not use popf")

Quoted 114 lines
arch/x86/include/asm/irqflags.h +15 -0
+{
+	asm volatile("push %0 ; popf"
+		     : /* no output */
+		     : "g" (flags)
+		     : "memory", "cc");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	if ((flags & X86_EFLAGS_IF)) {
+		check_hv_pending(NULL);
+	}
+#endif
+}
+
  static __always_inline void native_irq_disable(void)
  {
  	asm volatile("cli": : :"memory");
@@ -43,6 +59,9 @@ static __always_inline void native_irq_disable(void)
  static __always_inline void native_irq_enable(void)
  {
  	asm volatile("sti": : :"memory");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+	check_hv_pending(NULL);
+#endif
  }
 
  static inline __cpuidle void native_safe_halt(void)
arch/x86/include/asm/mem_encrypt.h +2 -0
index 72ca90552b6a..7264ca5f5b2d 100644
@@ -50,6 +50,7 @@ void __init early_set_mem_enc_dec_hypercall(unsigned long vaddr, int npages,
  void __init mem_encrypt_free_decrypted_mem(void);
 
  void __init sev_es_init_vc_handling(void);
+void __init sev_snp_init_hv_handling(void);
 
  #define __bss_decrypted __section(".bss..decrypted")
 
@@ -72,6 +73,7 @@ static inline void __init sme_encrypt_kernel(struct boot_params *bp) { }
  static inline void __init sme_enable(struct boot_params *bp) { }
 
  static inline void sev_es_init_vc_handling(void) { }
+static inline void sev_snp_init_hv_handling(void) { }
 
  static inline int __init
  early_set_memory_decrypted(unsigned long vaddr, unsigned long size) { return 0; }
arch/x86/include/asm/msr-index.h +6 -0
index 10ac52705892..6fe25a6e325f 100644
@@ -562,10 +562,16 @@
  #define MSR_AMD64_SEV_ENABLED_BIT	0
  #define MSR_AMD64_SEV_ES_ENABLED_BIT	1
  #define MSR_AMD64_SEV_SNP_ENABLED_BIT	2
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT		4
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT	5
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT	6
  #define MSR_AMD64_SEV_ENABLED		BIT_ULL(MSR_AMD64_SEV_ENABLED_BIT)
  #define MSR_AMD64_SEV_ES_ENABLED	BIT_ULL(MSR_AMD64_SEV_ES_ENABLED_BIT)
  #define MSR_AMD64_SEV_SNP_ENABLED	BIT_ULL(MSR_AMD64_SEV_SNP_ENABLED_BIT)
 
+#define MSR_AMD64_SEV_REFLECTVC_ENABLED			BIT_ULL(MSR_AMD64_SEV_REFLECTVC_ENABLED_BIT)
+#define MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED_BIT)
+#define MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED	BIT_ULL(MSR_AMD64_SEV_ALTERNATE_INJECTION_ENABLED_BIT)
  #define MSR_AMD64_VIRT_SPEC_CTRL	0xc001011f
 
  /* AMD Collaborative Processor Performance Control MSRs */
arch/x86/include/uapi/asm/svm.h +4 -0
index f69c168391aa..85d6882262e7 100644
@@ -115,6 +115,10 @@
  #define SVM_VMGEXIT_AP_CREATE_ON_INIT		0
  #define SVM_VMGEXIT_AP_CREATE			1
  #define SVM_VMGEXIT_AP_DESTROY			2
+#define SVM_VMGEXIT_HV_DOORBELL_PAGE		0x80000014
+#define SVM_VMGEXIT_GET_PREFERRED_HV_DOORBELL_PAGE	0
+#define SVM_VMGEXIT_SET_HV_DOORBELL_PAGE		1
+#define SVM_VMGEXIT_QUERY_HV_DOORBELL_PAGE		2
  #define SVM_VMGEXIT_HV_FEATURES			0x8000fffd
  #define SVM_VMGEXIT_UNSUPPORTED_EVENT		0x8000ffff
 
arch/x86/kernel/sev.c +18 -0
index b54ee3ba37b0..23cd025f97dc 100644
@@ -104,6 +104,12 @@ struct sev_es_runtime_data {
  	 * is currently unsupported in SEV-ES guests.
  	 */
  	unsigned long dr7;
+	/*
+	 * SEV-SNP requires that the GHCB must be registered before using it.
+	 * The flag below will indicate whether the GHCB is registered, if its
+	 * not registered then sev_es_get_ghcb() will perform the registration.
+	 */
+	bool ghcb_registered;
  };
 
  struct ghcb_state {
@@ -122,6 +128,183 @@ struct sev_config {
 
  static struct sev_config sev_cfg __read_mostly;
 
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state);
+static noinstr void __sev_put_ghcb(struct ghcb_state *state);
+static int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa);
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb);
+
+union hv_pending_events {
+	u16 events;
+	struct {
+		u8 vector;
+		u8 nmi : 1;
+		u8 mc : 1;
+		u8 reserved1 : 5;

arch/x86/kernel/sev.c struct sev_config { · line 152

Quoted 10 lines
arch/x86/kernel/sev.c +10 -0
+		u8 no_further_signal : 1;
+	};
+};
+
+struct sev_hv_doorbell_page {
+	union hv_pending_events pending_events;
+	u8 no_eoi_required;
+	u8 reserved2[61];
+	u8 padding[4032];
+};

Does the introduce/initialize doorbell part better to be in a separate preparatory patch?

Quoted 449 lines
arch/x86/kernel/sev.c +273 -53
+
+struct sev_snp_runtime_data {
+	struct sev_hv_doorbell_page hv_doorbell_page;
+};
+
+static DEFINE_PER_CPU(struct sev_snp_runtime_data*, snp_runtime_data);
+
+static inline u64 sev_es_rd_ghcb_msr(void)
+{
+	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
+}
+
+static __always_inline void sev_es_wr_ghcb_msr(u64 val)
+{
+	u32 low, high;
+
+	low  = (u32)(val);
+	high = (u32)(val >> 32);
+
+	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+}
+
+struct sev_hv_doorbell_page *sev_snp_current_doorbell_page(void)
+{
+	return &this_cpu_read(snp_runtime_data)->hv_doorbell_page;
+}
+
+static u8 sev_hv_pending(void)
+{
+	return sev_snp_current_doorbell_page()->pending_events.events;
+}
+
+static void hv_doorbell_apic_eoi_write(u32 reg, u32 val)
+{
+	if (xchg(&sev_snp_current_doorbell_page()->no_eoi_required, 0) & 0x1)
+		return;
+
+	BUG_ON(reg != APIC_EOI);
+	apic->write(reg, val);
+}
+
+static void do_exc_hv(struct pt_regs *regs)
+{
+	union hv_pending_events pending_events;
+	u8 vector;
+
+	while (sev_hv_pending()) {
+		asm volatile("cli" : : : "memory");
+
+		pending_events.events = xchg(
+			&sev_snp_current_doorbell_page()->pending_events.events,
+			0);
+
+		if (pending_events.nmi)
+			exc_nmi(regs);
+
+#ifdef CONFIG_X86_MCE
+		if (pending_events.mc)
+			exc_machine_check(regs);
+#endif
+
+		if (!pending_events.vector)
+			return;
+
+		if (pending_events.vector < FIRST_EXTERNAL_VECTOR) {
+			/* Exception vectors */
+			WARN(1, "exception shouldn't happen\n");
+		} else if (pending_events.vector == FIRST_EXTERNAL_VECTOR) {
+			sysvec_irq_move_cleanup(regs);
+		} else if (pending_events.vector == IA32_SYSCALL_VECTOR) {
+			WARN(1, "syscall shouldn't happen\n");
+		} else if (pending_events.vector >= FIRST_SYSTEM_VECTOR) {
+			switch (pending_events.vector) {
+#if IS_ENABLED(CONFIG_HYPERV)
+			case HYPERV_STIMER0_VECTOR:
+				sysvec_hyperv_stimer0(regs);
+				break;
+			case HYPERVISOR_CALLBACK_VECTOR:
+				sysvec_hyperv_callback(regs);
+				break;
+#endif
+#ifdef CONFIG_SMP
+			case RESCHEDULE_VECTOR:
+				sysvec_reschedule_ipi(regs);
+				break;
+			case IRQ_MOVE_CLEANUP_VECTOR:
+				sysvec_irq_move_cleanup(regs);
+				break;
+			case REBOOT_VECTOR:
+				sysvec_reboot(regs);
+				break;
+			case CALL_FUNCTION_SINGLE_VECTOR:
+				sysvec_call_function_single(regs);
+				break;
+			case CALL_FUNCTION_VECTOR:
+				sysvec_call_function(regs);
+				break;
+#endif
+#ifdef CONFIG_X86_LOCAL_APIC
+			case ERROR_APIC_VECTOR:
+				sysvec_error_interrupt(regs);
+				break;
+			case SPURIOUS_APIC_VECTOR:
+				sysvec_spurious_apic_interrupt(regs);
+				break;
+			case LOCAL_TIMER_VECTOR:
+				sysvec_apic_timer_interrupt(regs);
+				break;
+			case X86_PLATFORM_IPI_VECTOR:
+				sysvec_x86_platform_ipi(regs);
+				break;
+#endif
+			case 0x0:
+				break;
+			default:
+				panic("Unexpected vector %d\n", vector);
+				unreachable();
+			}
+		} else {
+			common_interrupt(regs, pending_events.vector);
+		}
+
+		asm volatile("sti" : : : "memory");
+	}
+}
+
+void check_hv_pending(struct pt_regs *regs)
+{
+	struct pt_regs local_regs;
+
+	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
+		return;
+
+	if (regs) {
+		if ((regs->flags & X86_EFLAGS_IF) == 0)
+			return;
+
+		if (!sev_hv_pending())
+			return;
+
+		do_exc_hv(regs);
+	} else {
+		if (sev_hv_pending()) {
+			memset(&local_regs, 0, sizeof(struct pt_regs));
+			regs = &local_regs;
+			asm volatile("movl %%cs, %%eax;" : "=a" (regs->cs));
+			asm volatile("movl %%ss, %%eax;" : "=a" (regs->ss));
+			regs->orig_ax = 0xffffffff;
+			regs->flags = native_save_fl();
+			do_exc_hv(regs);
+		}
+	}
+}
+EXPORT_SYMBOL_GPL(check_hv_pending);
+
  static __always_inline bool on_vc_stack(struct pt_regs *regs)
  {
  	unsigned long sp = regs->sp;
@@ -193,68 +376,35 @@ void noinstr __sev_es_ist_exit(void)
  	this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], *(unsigned long *)ist);
  }
 
-/*
- * Nothing shall interrupt this code path while holding the per-CPU
- * GHCB. The backup GHCB is only for NMIs interrupting this path.
- *
- * Callers must disable local interrupts around it.
- */
-static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+static bool sev_restricted_injection_enabled(void)
  {
-	struct sev_es_runtime_data *data;
+	return sev_status & MSR_AMD64_SEV_RESTRICTED_INJECTION_ENABLED;
+}
+
+void __init sev_snp_init_hv_handling(void)
+{
+	struct sev_snp_runtime_data *snp_data;
+	struct ghcb_state state;
  	struct ghcb *ghcb;
+	unsigned long flags;
+	int cpu;
+	int err;
 
  	WARN_ON(!irqs_disabled());
+	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP) || !sev_restricted_injection_enabled())
+		return;
 
-	data = this_cpu_read(runtime_data);
-	ghcb = &data->ghcb_page;
-
-	if (unlikely(data->ghcb_active)) {
-		/* GHCB is already in use - save its contents */
-
-		if (unlikely(data->backup_ghcb_active)) {
-			/*
-			 * Backup-GHCB is also already in use. There is no way
-			 * to continue here so just kill the machine. To make
-			 * panic() work, mark GHCBs inactive so that messages
-			 * can be printed out.
-			 */
-			data->ghcb_active        = false;
-			data->backup_ghcb_active = false;
-
-			instrumentation_begin();
-			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
-			instrumentation_end();
-		}
-
-		/* Mark backup_ghcb active before writing to it */
-		data->backup_ghcb_active = true;
-
-		state->ghcb = &data->backup_ghcb;
-
-		/* Backup GHCB content */
-		*state->ghcb = *ghcb;
-	} else {
-		state->ghcb = NULL;
-		data->ghcb_active = true;
-	}
+	local_irq_save(flags);
 
-	return ghcb;
-}
+	ghcb = __sev_get_ghcb(&state);
 
-static inline u64 sev_es_rd_ghcb_msr(void)
-{
-	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
-}
+	sev_snp_setup_hv_doorbell_page(ghcb);
 
-static __always_inline void sev_es_wr_ghcb_msr(u64 val)
-{
-	u32 low, high;
+	__sev_put_ghcb(&state);
 
-	low  = (u32)(val);
-	high = (u32)(val >> 32);
+	apic_set_eoi_write(hv_doorbell_apic_eoi_write);
 
-	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
+	local_irq_restore(flags);
  }
 
  static int vc_fetch_insn_kernel(struct es_em_ctxt *ctxt,
@@ -515,6 +665,79 @@ static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt
  /* Include code shared with pre-decompression boot stage */
  #include "sev-shared.c"
 
+/*
+ * Nothing shall interrupt this code path while holding the per-CPU
+ * GHCB. The backup GHCB is only for NMIs interrupting this path.
+ *
+ * Callers must disable local interrupts around it.
+ */
+static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
+{
+	struct sev_es_runtime_data *data;
+	struct ghcb *ghcb;
+
+	WARN_ON(!irqs_disabled());
+
+	data = this_cpu_read(runtime_data);
+	ghcb = &data->ghcb_page;
+
+	if (unlikely(data->ghcb_active)) {
+		/* GHCB is already in use - save its contents */
+
+		if (unlikely(data->backup_ghcb_active)) {
+			/*
+			 * Backup-GHCB is also already in use. There is no way
+			 * to continue here so just kill the machine. To make
+			 * panic() work, mark GHCBs inactive so that messages
+			 * can be printed out.
+			 */
+			data->ghcb_active        = false;
+			data->backup_ghcb_active = false;
+
+			instrumentation_begin();
+			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
+			instrumentation_end();
+		}
+
+		/* Mark backup_ghcb active before writing to it */
+		data->backup_ghcb_active = true;
+
+		state->ghcb = &data->backup_ghcb;
+
+		/* Backup GHCB content */
+		*state->ghcb = *ghcb;
+	} else {
+		state->ghcb = NULL;
+		data->ghcb_active = true;
+	}
+
+	/* SEV-SNP guest requires that GHCB must be registered before using it. */
+	if (!data->ghcb_registered) {
+		if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+			snp_register_ghcb_early(__pa(ghcb));
+			sev_snp_setup_hv_doorbell_page(ghcb);
+		} else {
+			sev_es_wr_ghcb_msr(__pa(ghcb));
+		}
+		data->ghcb_registered = true;
+	}
+
+	return ghcb;
+}
+
+static void sev_snp_setup_hv_doorbell_page(struct ghcb *ghcb)
+{
+	u64 pa;
+	enum es_result ret;
+
+	pa = __pa(sev_snp_current_doorbell_page());
+	vc_ghcb_invalidate(ghcb);
+	ret = vmgexit_hv_doorbell_page(ghcb,
+			SVM_VMGEXIT_SET_HV_DOORBELL_PAGE, pa);
+	if (ret != ES_OK)
+		panic("SEV-SNP: failed to set up #HV doorbell page");
+}
+
  static noinstr void __sev_put_ghcb(struct ghcb_state *state)
  {
  	struct sev_es_runtime_data *data;
@@ -1282,6 +1505,11 @@ void setup_ghcb(void)
  		snp_register_ghcb_early(__pa(&boot_ghcb_page));
  }
 
+int vmgexit_hv_doorbell_page(struct ghcb *ghcb, u64 op, u64 pa)
+{
+	return sev_es_ghcb_hv_call(ghcb, NULL, SVM_VMGEXIT_HV_DOORBELL_PAGE, op, pa);
+}
+
  #ifdef CONFIG_HOTPLUG_CPU
  static void sev_es_ap_hlt_loop(void)
  {
@@ -1355,6 +1583,7 @@ static void __init alloc_runtime_data(int cpu)
  static void __init init_ghcb(int cpu)
  {
  	struct sev_es_runtime_data *data;
+	struct sev_snp_runtime_data *snp_data;
  	int err;
 
  	data = per_cpu(runtime_data, cpu);
@@ -1366,8 +1595,22 @@ static void __init init_ghcb(int cpu)
 
  	memset(&data->ghcb_page, 0, sizeof(data->ghcb_page));
 
+	snp_data = memblock_alloc(sizeof(*snp_data), PAGE_SIZE);
+	if (!snp_data)
+		panic("Can't allocate SEV-SNP runtime data");
+
+	err = early_set_memory_decrypted((unsigned long)&snp_data->hv_doorbell_page,
+					 sizeof(snp_data->hv_doorbell_page));
+	if (err)
+		panic("Can't map #HV doorbell pages unencrypted");
+
+	memset(&snp_data->hv_doorbell_page, 0, sizeof(snp_data->hv_doorbell_page));
+
+	per_cpu(snp_runtime_data, cpu) = snp_data;
+
  	data->ghcb_active = false;
  	data->backup_ghcb_active = false;
+	data->ghcb_registered = false;
  }
 
  void __init sev_es_init_vc_handling(void)
@@ -2006,7 +2249,12 @@ DEFINE_IDTENTRY_VC_USER(exc_vmm_communication)
 
  static bool hv_raw_handle_exception(struct pt_regs *regs)
  {
-	return false;
+	/* Clear the no_further_signal bit */
+	sev_snp_current_doorbell_page()->pending_events.events &= 0x7fff;
+
+	check_hv_pending(regs);
+
+	return true;
  }
 
  static __always_inline bool on_hv_fallback_stack(struct pt_regs *regs)
arch/x86/kernel/traps.c +50 -0
index 178015a820f0..af97e6610fbb 100644
@@ -898,6 +898,53 @@ asmlinkage __visible noinstr struct pt_regs *vc_switch_off_ist(struct pt_regs *r
 
  	return regs_ret;
  }
+
+asmlinkage __visible noinstr struct pt_regs *hv_switch_off_ist(struct pt_regs *regs)
+{
+	unsigned long sp, *stack;
+	struct stack_info info;
+	struct pt_regs *regs_ret;
+
+	/*
+	 * A malicious hypervisor can inject 2 HVs in a row, which will corrupt
+	 * the trap frame on our IST stack.  We add a defensive check here to
+	 * catch such behavior.
+	 */
+	BUG_ON(regs->sp >= __this_cpu_ist_bottom_va(HV) && regs->sp < __this_cpu_ist_top_va(HV));
+
+	/*
+	 * In the SYSCALL entry path the RSP value comes from user-space - don't
+	 * trust it and switch to the current kernel stack
+	 */
+	if (ip_within_syscall_gap(regs)) {
+		sp = this_cpu_read(cpu_current_top_of_stack);
+		goto sync;
+	}
+
+	/*
+	 * From here on the RSP value is trusted. Now check whether entry
+	 * happened from a safe stack. Not safe are the entry or unknown stacks,
+	 * use the fall-back stack instead in this case.
+	 */
+	sp    = regs->sp;
+	stack = (unsigned long *)sp;
+
+	if (!get_stack_info_noinstr(stack, current, &info) || info.type == STACK_TYPE_ENTRY ||
+	    info.type > STACK_TYPE_EXCEPTION_LAST)
+		sp = __this_cpu_ist_top_va(HV2);
+sync:
+	/*
+	 * Found a safe stack - switch to it as if the entry didn't happen via
+	 * IST stack. The code below only copies pt_regs, the real switch happens
+	 * in assembly code.
+	 */
+	sp = ALIGN_DOWN(sp, 8) - sizeof(*regs_ret);
+
+	regs_ret = (struct pt_regs *)sp;
+	*regs_ret = *regs;
+
+	return regs_ret;
+}
  #endif
 
  asmlinkage __visible noinstr struct pt_regs *fixup_bad_iret(struct pt_regs *bad_regs)
@@ -1457,4 +1504,7 @@ void __init trap_init(void)
  	/* Setup traps as cpu_init() might #GP */
  	idt_setup_traps();
  	cpu_init();
+
+	/* Init #HV doorbell pages when running as an SEV-SNP guest */
+	sev_snp_init_hv_handling();
  }
author
#7 lore raw

Re: [RFC PATCH V2 01/18] x86/sev: Pvalidate memory gab for decompressing kernel

In reply to Gupta, Pankaj
Quoted 4 lines · Pankaj

Hi Tianyu,

Some minor nits below.

Thanks a lot to review. I will change these in the next version.

Quoted 165 lines
Quoted 5 lines

Pvalidate needed pages for decompressing kernel. The E820_TYPE_RAM entry includes only validated memory. The kernel expects that the RAM entry's addr is fixed while the entry size is to be extended to cover addresses to the start of next entry. This patch increases the RAM entry size to cover all possilble memory addresses until

s/possilble/possible

Quoted 78 lines

init_size.

---
  arch/x86/boot/compressed/head_64.S |  8 +++ arch/x86/boot/compressed/sev.c     | 84 ++++++++++++++++++++++++++++++

  2 files changed, 92 insertions(+)
arch/x86/boot/compressed/head_64.S +0 -0
index d33f060900d2..818edaf5d0cf 100644
@@ -348,6 +348,14 @@ SYM_CODE_START(startup_64)

      cld
      cli

arch/x86/boot/compressed/head_64.S +8 -0
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+    /* pvalidate memory on demand if SNP is enabled. */
+    pushq    %rsi
+    movq    %rsi, %rdi
+    call     pvalidate_for_startup_64
+    popq    %rsi
+#endif
+

      /* Setup data segments. */
      xorl    %eax, %eax
      movl    %eax, %ds

arch/x86/boot/compressed/sev.c +23 -0
index 960968f8bf75..3a5a1ab16095 100644
@@ -12,8 +12,10 @@
    */
   #include "misc.h"
+#include <asm/msr-index.h>
   #include <asm/pgtable_types.h>
   #include <asm/sev.h>
+#include <asm/svm.h>
   #include <asm/trapnr.h>
   #include <asm/trap_pf.h>
   #include <asm/msr-index.h>
@@ -21,6 +23,7 @@
   #include <asm/ptrace.h>
   #include <asm/svm.h>
   #include <asm/cpuid.h>
+#include <asm/e820/types.h>
   #include "error.h"
   #include "../msr.h"
@@ -117,6 +120,22 @@ static enum es_result vc_read_mem(struct 
 es_em_ctxt *ctxt,
   /* Include code for early handlers */
   #include "../../kernel/sev-shared.c"
+/* Check SEV-SNP via MSR */
+static bool sev_snp_runtime_check(void)
+{
+    unsigned long low, high;
+    u64 val;
+
+    asm volatile("rdmsr\n" : "=a" (low), "=d" (high) :
+            "c" (MSR_AMD64_SEV));
+
+    val = (high << 32) | low;
+    if (val & MSR_AMD64_SEV_SNP_ENABLED)
+        return true;
+
+    return false;
+}
+
   static inline bool sev_snp_enabled(void)
   {
       return sev_status & MSR_AMD64_SEV_SNP_ENABLED;
@@ -456,3 +475,68 @@ void sev_prep_identity_maps(unsigned long 
 top_level_pgt)
       sev_verify_cbit(top_level_pgt);
   }
+
+static void extend_e820_on_demand(struct boot_e820_entry *e820_entry,
+                  u64 needed_ram_end)
+{

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long · line 491

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+    u64 end, paddr;
+    unsigned long eflags;
+    int rc;
+
+    if (!e820_entry)
+        return;
+
+    /* Validated memory must be aligned by PAGE_SIZE. */
+    end = ALIGN(e820_entry->addr + e820_entry->size, PAGE_SIZE);
+    if (needed_ram_end > end && e820_entry->type == E820_TYPE_RAM) {

maybe "if check" can be used to return from here, would read code better?

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long · line 492

Quoted 1 line
arch/x86/boot/compressed/sev.c +1 -0
+        for (paddr = end; paddr < needed_ram_end; paddr += PAGE_SIZE) 

Maybe we could reuse 'pvalidate_pages' here?

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 495

Quoted 3 lines
arch/x86/boot/compressed/sev.c +3 -0
+            rc = pvalidate(paddr, RMP_PG_SIZE_4K, true);
+            if (rc) {
+                error("Failed to validate address.n");

Would it make sense to print the cause of failure here?

Quoted 26 lines
arch/x86/boot/compressed/sev.c +11 -0
+                return;
+            }
+        }
+        e820_entry->size = needed_ram_end - e820_entry->addr;
+    }
+}
+
+/*
+ * Explicitly pvalidate needed pages for decompressing the kernel.
+ * The E820_TYPE_RAM entry includes only validated memory. The kernel
+ * expects that the RAM entry's addr is fixed while the entry size is 

to be
+ * extended to cover addresses to the start of next entry.
+ * The function increases the RAM entry size to cover all possible memory
+ * addresses until init_size.
+ * For example,  init_end = 0x4000000,
+ * [RAM: 0x0 - 0x0],                       M[RAM: 0x0 - 0xa0000]
+ * [RSVD: 0xa0000 - 0x10000]                [RSVD: 0xa0000 - 0x10000]
+ * [ACPI: 0x10000 - 0x20000]      ==>       [ACPI: 0x10000 - 0x20000]
+ * [RSVD: 0x800000 - 0x900000]              [RSVD: 0x800000 - 0x900000]
+ * [RAM: 0x1000000 - 0x2000000]            M[RAM: 0x1000000 - 0x2001000]
+ * [RAM: 0x2001000 - 0x2007000]            M[RAM: 0x2001000 - 0x4000000]
+ * Other RAM memory after init_end is pvalidated by ms_hyperv_init_platform
+ */

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 525

Quoted 7 lines
arch/x86/boot/compressed/sev.c +7 -0
+__visible void pvalidate_for_startup_64(struct boot_params *boot_params)
+{
+    struct boot_e820_entry *e820_entry;
+    u64 init_end =
+        boot_params->hdr.pref_address + boot_params->hdr.init_size;
+    u8 i, nr_entries = boot_params->e820_entries;
+    u64 needed_end;

Could not very well interpret the name 'needed_end'.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 531

Quoted 6 lines
arch/x86/boot/compressed/sev.c +6 -0
+
+    if (!sev_snp_runtime_check())
+        return;
+
+    for (i = 0; i < nr_entries; ++i) {
+        /* Pvalidate memory holes in e820 RAM entries. */

Pvalidate and pvalidate names are used interchangeably in comments.

Quoted 11 lines

+        e820_entry = &boot_params->e820_table[i];
+        if (i < nr_entries - 1) {
+            needed_end = boot_params->e820_table[i + 1].addr; +            if (needed_end < e820_entry->addr)
+                error("e820 table is not sorted.\n"); +        } else {
+            needed_end = init_end;
+        }
+        extend_e820_on_demand(e820_entry, needed_end); +    }
+}

author
#56 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

In reply to Gupta, Pankaj
Quoted 6 lines · Pankaj
Quoted 6 lines
arch/x86/include/asm/irqflags.h +3 -0
+#endif
+}
+

  static __always_inline void native_irq_disable(void) {
      asm volatile("cli": : :"memory");

Quoted 8 lines

arch/x86/include/asm/irqflags.h static __always_inline void native_irq_disable(void) · line 63

Quoted 6 lines
arch/x86/include/asm/irqflags.h +2 -0
@@ -43,6 +59,9 @@ static __always_inline void native_irq_disable(void)
   static __always_inline void native_irq_enable(void)
   {
       asm volatile("sti": : :"memory");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+    check_hv_pending(NULL);

Just trying to understand when regs will be NULL?

check_hv_pending() will be divided into two functions.

The one handles #hv event in the #HV exception code path. The other one handles pending irq event in the irq re-enable code path。 In this version, the "regs = NULL" for check_hv_pending() is used in the irq re-enable code path.

author
#59 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

In reply to Gupta, Pankaj
Quoted 13 lines · Pankaj
Quoted 13 lines
arch/x86/include/asm/irqflags.h +3 -1
index 7793e52d6237..e0730d8bc0ac 100644
@@ -14,6 +14,9 @@
   /*
    * Interrupt control:
    */
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+void check_hv_pending(struct pt_regs *regs);
+#endif
   /* Declaration required for gcc < 4.9 to prevent 
-Werror=missing-prototypes */
Quoted 8 lines

arch/x86/include/asm/irqflags.h extern __always_inline unsigned long · line 41

Quoted 6 lines
arch/x86/include/asm/irqflags.h +1 -0
   extern inline unsigned long native_save_fl(void);
@@ -35,6 +38,19 @@ extern __always_inline unsigned long 
 native_save_fl(void)
       return flags;
   }
+extern inline void native_restore_fl(unsigned long flags)

Don't know if want to re-introduce this again? as it was removed with ab234a260b1f ("x86/pv: Rework arch_local_irq_restore() to not use popf")

Nice catch! This should be here again. Will remove it in the next version.

Thanks.

#57 lore raw

Re: [RFC PATCH V2 16/18] x86/sev: Initialize #HV doorbell and handle interrupt requests

In reply to Tianyu Lan
Quoted 16 lines · Tianyu Lan
Quoted 6 lines · Pankaj
Quoted 6 lines
arch/x86/include/asm/irqflags.h +3 -0
+#endif
+}
+

  static __always_inline void native_irq_disable(void) {
      asm volatile("cli": : :"memory");

Quoted 8 lines

arch/x86/include/asm/irqflags.h static __always_inline void native_irq_disable(void) · line 63

Quoted 6 lines
arch/x86/include/asm/irqflags.h +2 -0
@@ -43,6 +59,9 @@ static __always_inline void native_irq_disable(void)
   static __always_inline void native_irq_enable(void)
   {
       asm volatile("sti": : :"memory");
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+    check_hv_pending(NULL);

Just trying to understand when regs will be NULL?

Quoted 6 lines

check_hv_pending() will be divided into two functions.

The one handles #hv event in the #HV exception code path. The other one handles pending irq event in the irq re-enable code path。 In this version, the "regs = NULL" for check_hv_pending() is used in the irq re-enable code path.

Got it.

Thanks,
Pankaj

maintainer
#31 lore raw

RE: [RFC PATCH V2 10/18] drivers: hv: Decrypt percpu hvcall input arg page in sev-snp enlightened guest

In reply to Tianyu Lan
Quoted 13 lines

From: Tianyu Lan <[email protected]>
Sent: Friday, November 18, 2022 7:46 PM
[...]

drivers/hv/hv_common.c +1 -0
@@ -125,6 +126,7 @@ int hv_common_cpu_init(unsigned int cpu)
 	u64 msr_vp_index;
 	gfp_t flags;
 	int pgcount = hv_root_partition ? 2 : 1;
+	int ret;
 
 	/* hv_cpu_init() can be called with IRQs disabled from hv_resume() */
 	flags = irqs_disabled() ? GFP_ATOMIC : GFP_KERNEL;

drivers/hv/hv_common.c int hv_common_cpu_init(unsigned int cpu) · line 140

Quoted 5 lines
drivers/hv/hv_common.c +2 -0
@@ -134,6 +136,16 @@ int hv_common_cpu_init(unsigned int cpu)
 	if (!(*inputarg))
 		return -ENOMEM;
 
+	if (hv_isolation_type_en_snp()) {
+		ret = set_memory_decrypted((unsigned long)*inputarg, 1);

Is it possible hv_root_partition==1 here? If yes, the pgcount is 2.

Quoted 12 lines
drivers/hv/hv_common.c +8 -0
+		if (ret) {
+			kfree(*inputarg);
+			return ret;
+		}
+
+		memset(*inputarg, 0x00, PAGE_SIZE);
+	}
+
 	if (hv_root_partition) {
 		outputarg = (void **)this_cpu_ptr(hyperv_pcpu_output_arg);
 		*outputarg = (char *)(*inputarg) + HV_HYP_PAGE_SIZE;

drivers/hv/hv_common.c int hv_common_cpu_die(unsigned int cpu) · line 184

Quoted 5 lines
drivers/hv/hv_common.c +2 -0
@@ -168,6 +180,9 @@ int hv_common_cpu_die(unsigned int cpu)
 
 	local_irq_restore(flags);
 
+	if (hv_isolation_type_en_snp())
+		set_memory_encrypted((unsigned long)mem, 1);

If set_memory_encrypted() fails, we should not free the 'mem'.

Quoted 4 lines
drivers/hv/hv_common.c +1 -0
+
 	kfree(mem);
 
 	return 0;
author
#32 lore raw

Re: [RFC PATCH V2 10/18] drivers: hv: Decrypt percpu hvcall input arg page in sev-snp enlightened guest

Quoted 8 lines · Dexuan Cui

drivers/hv/hv_common.c int hv_common_cpu_init(unsigned int cpu) · line 140

Quoted 6 lines
drivers/hv/hv_common.c +2 -0
@@ -134,6 +136,16 @@ int hv_common_cpu_init(unsigned int cpu)
  	if (!(*inputarg))
  		return -ENOMEM;
 
+	if (hv_isolation_type_en_snp()) {
+		ret = set_memory_decrypted((unsigned long)*inputarg, 1);

Is it possible hv_root_partition==1 here? If yes, the pgcount is 2.

Hi Dexuan:
Thanks for review. So far, root partition doesn't support sev
enlightened guest and so here assume pgcount is always 1. We may use pgcount variable here instead of the number.

Quoted 13 lines
Quoted 13 lines
drivers/hv/hv_common.c +8 -0
+		if (ret) {
+			kfree(*inputarg);
+			return ret;
+		}
+
+		memset(*inputarg, 0x00, PAGE_SIZE);
+	}
+
  	if (hv_root_partition) {
  		outputarg = (void **)this_cpu_ptr(hyperv_pcpu_output_arg);
  		*outputarg = (char *)(*inputarg) + HV_HYP_PAGE_SIZE;
@@ -168,6 +180,9 @@ int hv_common_cpu_die(unsigned int cpu)
 
Quoted 5 lines

drivers/hv/hv_common.c int hv_common_cpu_die(unsigned int cpu) · line 184

Quoted 4 lines
drivers/hv/hv_common.c +2 -0
  	local_irq_restore(flags);
 
+	if (hv_isolation_type_en_snp())
+		set_memory_encrypted((unsigned long)mem, 1);

If set_memory_encrypted() fails, we should not free the 'mem'.

Good point. Will update in the next version.

#9 lore raw

RE: [RFC PATCH V2 02/18] x86/hyperv: Add sev-snp enlightened guest specific config

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 52 lines

Introduce static key isolation_type_en_snp for enlightened guest check and add some specific options in ms_hyperv_init_ platform().

---
 arch/x86/hyperv/ivm.c           | 12 +++++++++++-
 arch/x86/include/asm/mshyperv.h |  2 ++
 arch/x86/kernel/cpu/mshyperv.c  | 29 ++++++++++++++++++++++++-----
 drivers/hv/hv_common.c          |  7 +++++++
 4 files changed, 44 insertions(+), 6 deletions(-)
arch/x86/hyperv/ivm.c +11 -1
index 1dbcbd9da74d..e9c30dad3419 100644
@@ -259,10 +259,20 @@ bool hv_is_isolation_supported(void)
 }
 
 DEFINE_STATIC_KEY_FALSE(isolation_type_snp);
+DEFINE_STATIC_KEY_FALSE(isolation_type_en_snp);
+
+/*
+ * hv_isolation_type_en_snp - Check system runs in the AMD SEV-SNP based
+ * isolation enlightened VM.
+ */
+bool hv_isolation_type_en_snp(void)
+{
+	return static_branch_unlikely(&isolation_type_en_snp);
+}
 
 /*
  * hv_isolation_type_snp - Check system runs in the AMD SEV-SNP based
- * isolation VM.
+ * isolation VM with vTOM support.
  */
 bool hv_isolation_type_snp(void)
 {
arch/x86/include/asm/mshyperv.h +1 -0
index 61f0c206bff0..9b8c3f638845 100644
@@ -14,6 +14,7 @@
 union hv_ghcb;
 
 DECLARE_STATIC_KEY_FALSE(isolation_type_snp);
+DECLARE_STATIC_KEY_FALSE(isolation_type_en_snp);
 
 typedef int (*hyperv_fill_flush_list_func)(
 		struct hv_guest_mapping_flush_list *flush,

arch/x86/include/asm/mshyperv.h extern u64 hv_current_partition_id; · line 36

Quoted 4 lines
arch/x86/include/asm/mshyperv.h +1 -0
@@ -32,6 +33,7 @@ extern u64 hv_current_partition_id;
 
 extern union hv_ghcb * __percpu *hv_ghcb_pg;
 
+extern bool hv_isolation_type_en_snp(void);

This file also has a declaration for hv_isolation_type_snp(). I think this new declaration is near the beginning of this file so that it can be referenced by hv_do_hypercall() and related functions in Patch 6 of this series. So maybe move the declaration of hv_isolation_type_snp() up so it is adjacent to this one? It would make sense for the two to be together.

Quoted 16 lines
arch/x86/include/asm/mshyperv.h +0 -0
 int hv_call_deposit_pages(int node, u64 partition_id, u32 num_pages);
 int hv_call_add_logical_proc(int node, u32 lp_index, u32 acpi_id);
 int hv_call_create_vp(int node, u64 partition_id, u32 vp_index, u32 flags);
arch/x86/kernel/cpu/mshyperv.c +5 -0
index 831613959a92..2ea4f21c6172 100644
@@ -273,6 +273,21 @@ static void __init ms_hyperv_init_platform(void)
 	ms_hyperv.misc_features = cpuid_edx(HYPERV_CPUID_FEATURES);
 	ms_hyperv.hints    = cpuid_eax(HYPERV_CPUID_ENLIGHTMENT_INFO);
 
+	/*
+	 * Add custom configuration for SEV-SNP Enlightened guest
+	 */
+	if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+		ms_hyperv.features |= HV_ACCESS_FREQUENCY_MSRS;

arch/x86/kernel/cpu/mshyperv.c static void __init ms_hyperv_init_platform(void) · line 290

Quoted 10 lines
arch/x86/kernel/cpu/mshyperv.c +10 -0
+		ms_hyperv.misc_features |= HV_FEATURE_FREQUENCY_MSRS_AVAILABLE;
+		ms_hyperv.misc_features &= ~HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE;
+		ms_hyperv.hints |= HV_DEPRECATING_AEOI_RECOMMENDED;
+		ms_hyperv.hints |= HV_X64_APIC_ACCESS_RECOMMENDED;
+		ms_hyperv.hints |= HV_X64_CLUSTER_IPI_RECOMMENDED;
+	}
+
+	pr_info("Hyper-V: enlightment features 0x%x, hints 0x%x, misc 0x%x\n",
+		ms_hyperv.features, ms_hyperv.hints, ms_hyperv.misc_features);
+

What's the reason for this additional call to pr_info()? There's a call to pr_info() a couple of lines below that displays the same information, and a little bit more. It seems like the above call should be deleted, as I think we should try to be as consistent as possible in the output.

Quoted 13 lines
arch/x86/kernel/cpu/mshyperv.c +2 -4
 	hv_max_functions_eax = cpuid_eax(HYPERV_CPUID_VENDOR_AND_MAX_FUNCTIONS);
 
 	pr_info("Hyper-V: privilege flags low 0x%x, high 0x%x, hints 0x%x, misc 0x%x\n",
@@ -328,18 +343,22 @@ static void __init ms_hyperv_init_platform(void)
 		ms_hyperv.shared_gpa_boundary =
 			BIT_ULL(ms_hyperv.shared_gpa_boundary_bits);
 
-		pr_info("Hyper-V: Isolation Config: Group A 0x%x, Group B 0x%x\n",
-			ms_hyperv.isolation_config_a, ms_hyperv.isolation_config_b);
-
-		if (hv_get_isolation_type() == HV_ISOLATION_TYPE_SNP) {
+		if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+			static_branch_enable(&isolation_type_en_snp);

arch/x86/kernel/cpu/mshyperv.c static void __init ms_hyperv_init_platform(void) · line 357

Quoted 10 lines
arch/x86/kernel/cpu/mshyperv.c +5 -0
+		} else if (hv_get_isolation_type() == HV_ISOLATION_TYPE_SNP) {
 			static_branch_enable(&isolation_type_snp);
 #ifdef CONFIG_SWIOTLB
 			swiotlb_unencrypted_base = ms_hyperv.shared_gpa_boundary;
 #endif
 		}
+
+		pr_info("Hyper-V: Isolation Config: Group A 0x%x, Group B 0x%x\n",
+			ms_hyperv.isolation_config_a, ms_hyperv.isolation_config_b);
+

Is there a reason for moving this pr_info() down a few lines? I can't see that the intervening code changes any of the settings that are displayed, so it should be good in the original location. Just trying to minimize changes that don't add value ...

Quoted 13 lines
arch/x86/kernel/cpu/mshyperv.c +2 -1
 		/* Isolation VMs are unenlightened SEV-based VMs, thus this check: */
 		if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT)) {
-			if (hv_get_isolation_type() != HV_ISOLATION_TYPE_NONE)
+			if (hv_get_isolation_type() != HV_ISOLATION_TYPE_NONE
+			    && !cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
 				cc_set_vendor(CC_VENDOR_HYPERV);
 		}
 	}
drivers/hv/hv_common.c +0 -0
index ae68298c0dca..2c6602571c47 100644

drivers/hv/hv_common.c bool __weak hv_isolation_type_snp(void) · line 277

Quoted 10 lines
drivers/hv/hv_common.c +7 -0
@@ -268,6 +268,13 @@ bool __weak hv_isolation_type_snp(void)
 }
 EXPORT_SYMBOL_GPL(hv_isolation_type_snp);
 
+bool __weak hv_isolation_type_en_snp(void)
+{
+	return false;
+}
+EXPORT_SYMBOL_GPL(hv_isolation_type_en_snp);
+
+

Nit: One blank line is sufficient. Don't need to add two.

Quoted 5 lines
drivers/hv/hv_common.c +0 -1
 void __weak hv_setup_vmbus_handler(void (*handler)(void))
 {
 }
--

2.25.1

#12 lore raw

RE: [RFC PATCH V2 03/18] x86/hyperv: apic change for sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 2 lines

Hyperv sev-snp enlightened guest doesn't support x2apic and

How does lack of support for x2apic factor into the code below? I didn't see anything specific to the x2apic, but maybe I'm just not knowledgeable about the association.

Quoted 15 lines

apic page read/write operation. Bypass these requests. ipi request maybe returned with timeout error code and add retry mechanism.

---
 arch/x86/hyperv/hv_apic.c         | 79 ++++++++++++++++++++++++-------
 include/asm-generic/hyperv-tlfs.h |  1 +
 2 files changed, 63 insertions(+), 17 deletions(-)
arch/x86/hyperv/hv_apic.c +0 -0
index fb8b2c088681..214354d20833 100644

arch/x86/hyperv/hv_apic.c static u32 hv_apic_read(u32 reg) · line 71

Quoted 7 lines
arch/x86/hyperv/hv_apic.c +3 -1
@@ -66,9 +66,15 @@ static u32 hv_apic_read(u32 reg)
 		rdmsr(HV_X64_MSR_TPR, reg_val, hi);
 		(void)hi;
 		return reg_val;
-
+	case APIC_ID:
+		if (hv_isolation_type_en_snp())
+			return smp_processor_id();

Hmmm. The Linux processor ID is not always equal to the APIC ID. The specific case I'm aware of is a VM with multiple NUMA nodes, where the number of vCPUs in a NUMA node is not a power of 2. In that case, there's a gap in the APIC IDs, but not in the Linux processor IDs. But even outside that specific case, there's no guarantee that the Linux processor IDs match the APIC IDs.

What specific code is trying to read the APIC ID this way? That use case may influence choosing an alternate method of getting the APIC ID that will be correct in all cases.

arch/x86/hyperv/hv_apic.c static u32 hv_apic_read(u32 reg) · line 77

Quoted 7 lines
arch/x86/hyperv/hv_apic.c +5 -1
+		fallthrough;
 	default:
-		return native_apic_mem_read(reg);
+		if (!hv_isolation_type_en_snp())
+			return native_apic_mem_read(reg);
+		else
+			return 0;

At first glance, just silently returning zero seems dangerous. Is it a scenario that we never expect this to happen? Or is there a known set of scenarios where we know it is OK to return zero? If the former, what about using WARN_ON_ONCE(1) to catch any unexpected uses?

In any case, a comment with an explanation would help folks in the future who look at this code.

arch/x86/hyperv/hv_apic.c static void hv_apic_write(u32 reg, u32 val) · line 92

Quoted 10 lines
arch/x86/hyperv/hv_apic.c +2 -1
 	}
 }
 
@@ -82,7 +88,8 @@ static void hv_apic_write(u32 reg, u32 val)
 		wrmsr(HV_X64_MSR_TPR, val, 0);
 		break;
 	default:
-		native_apic_mem_write(reg, val);
+		if (!hv_isolation_type_en_snp())
+			native_apic_mem_write(reg, val);

Same here. Doing nothing seems dangerous.

Quoted 12 lines
arch/x86/hyperv/hv_apic.c +1 -0
 	}
 }
 
@@ -106,6 +113,7 @@ static bool __send_ipi_mask_ex(const struct cpumask *mask, int vector,
 	struct hv_send_ipi_ex *ipi_arg;
 	unsigned long flags;
 	int nr_bank = 0;
+	int retry = 5;
 	u64 status = HV_STATUS_INVALID_PARAMETER;
 
 	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))

arch/x86/hyperv/hv_apic.c static bool __send_ipi_mask_ex(const struct cpumask *mask, int vector, · line 158

Quoted 8 lines
arch/x86/hyperv/hv_apic.c +3 -1
@@ -144,8 +152,10 @@ static bool __send_ipi_mask_ex(const struct cpumask *mask, int vector,
 		ipi_arg->vp_set.format = HV_GENERIC_SET_ALL;
 	}
 
-	status = hv_do_rep_hypercall(HVCALL_SEND_IPI_EX, 0, nr_bank,
+	do {
+		status = hv_do_rep_hypercall(HVCALL_SEND_IPI_EX, 0, nr_bank,
 			      ipi_arg, NULL);
+	} while (status == HV_STATUS_TIME_OUT && retry--);

Since the u64 status returned by hv_do_hypercall contains other fields besides just the hypercall result, test "hv_result(status)" instead of just "status"

Quoted 12 lines
arch/x86/hyperv/hv_apic.c +1 -0
 ipi_mask_ex_done:
 	local_irq_restore(flags);
@@ -159,6 +169,7 @@ static bool __send_ipi_mask(const struct cpumask *mask, int vector,
 	struct hv_send_ipi ipi_arg;
 	u64 status;
 	unsigned int weight;
+	int retry = 5;
 
 	trace_hyperv_send_ipi_mask(mask, vector);
 

arch/x86/hyperv/hv_apic.c static bool __send_ipi_mask(const struct cpumask *mask, int vector, · line 230

Quoted 10 lines
arch/x86/hyperv/hv_apic.c +5 -2
@@ -212,8 +223,11 @@ static bool __send_ipi_mask(const struct cpumask *mask, int vector,
 		__set_bit(vcpu, (unsigned long *)&ipi_arg.cpu_mask);
 	}
 
-	status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, ipi_arg.vector,
-				     ipi_arg.cpu_mask);
+	do {
+		status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, ipi_arg.vector,
+						ipi_arg.cpu_mask);
+	} while (status == HV_STATUS_TIME_OUT && retry--);
+

Same here. Test "hv_result(status)".

Quoted 12 lines
arch/x86/hyperv/hv_apic.c +1 -0
 	return hv_result_success(status);
 
 do_ex_hypercall:
@@ -224,6 +238,7 @@ static bool __send_ipi_one(int cpu, int vector)
 {
 	int vp = hv_cpu_number_to_vp_number(cpu);
 	u64 status;
+	int retry = 5;
 
 	trace_hyperv_send_ipi_one(cpu, vector);
 

arch/x86/hyperv/hv_apic.c static bool __send_ipi_one(int cpu, int vector) · line 257

Quoted 8 lines
arch/x86/hyperv/hv_apic.c +4 -1
@@ -236,26 +251,48 @@ static bool __send_ipi_one(int cpu, int vector)
 	if (vp >= 64)
 		return __send_ipi_mask_ex(cpumask_of(cpu), vector, false);
 
-	status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, vector, BIT_ULL(vp));
+	do {
+		status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, vector, BIT_ULL(vp));
+	} while (status == HV_STATUS_TIME_OUT || retry--);
+

Same here. And I think you want "&&" like in the previous two cases instead of "||".

Quoted 88 lines
arch/x86/hyperv/hv_apic.c +37 -10
 	return hv_result_success(status);
 }
 
 static void hv_send_ipi(int cpu, int vector)
 {
-	if (!__send_ipi_one(cpu, vector))
-		orig_apic.send_IPI(cpu, vector);
+	if (!__send_ipi_one(cpu, vector)) {
+		if (!hv_isolation_type_en_snp())
+			orig_apic.send_IPI(cpu, vector);
+		else
+			WARN_ON_ONCE(1);
+	}
 }
 
 static void hv_send_ipi_mask(const struct cpumask *mask, int vector)
 {
-	if (!__send_ipi_mask(mask, vector, false))
-		orig_apic.send_IPI_mask(mask, vector);
+	if (!__send_ipi_mask(mask, vector, false)) {
+		if (!hv_isolation_type_en_snp())
+			orig_apic.send_IPI_mask(mask, vector);
+		else
+			WARN_ON_ONCE(1);
+	}
 }
 
 static void hv_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
 {
-	if (!__send_ipi_mask(mask, vector, true))
-		orig_apic.send_IPI_mask_allbutself(mask, vector);
+	unsigned int this_cpu = smp_processor_id();
+	struct cpumask new_mask;
+	const struct cpumask *local_mask;
+
+	cpumask_copy(&new_mask, mask);
+	cpumask_clear_cpu(this_cpu, &new_mask);
+	local_mask = &new_mask;
+	if (!__send_ipi_mask(local_mask, vector, true)) {
+		if (!hv_isolation_type_en_snp())
+			orig_apic.send_IPI_mask_allbutself(mask, vector);
+		else
+			WARN_ON_ONCE(1);
+	}
 }
 
 static void hv_send_ipi_allbutself(int vector)
@@ -265,14 +302,22 @@ static void hv_send_ipi_allbutself(int vector)
 
 static void hv_send_ipi_all(int vector)
 {
-	if (!__send_ipi_mask(cpu_online_mask, vector, false))
-		orig_apic.send_IPI_all(vector);
+	if (!__send_ipi_mask(cpu_online_mask, vector, false)) {
+		if (!hv_isolation_type_en_snp())
+			orig_apic.send_IPI_all(vector);
+		else
+			WARN_ON_ONCE(1);
+	}
 }
 
 static void hv_send_ipi_self(int vector)
 {
-	if (!__send_ipi_one(smp_processor_id(), vector))
-		orig_apic.send_IPI_self(vector);
+	if (!__send_ipi_one(smp_processor_id(), vector)) {
+		if (!hv_isolation_type_en_snp())
+			orig_apic.send_IPI_self(vector);
+		else
+			WARN_ON_ONCE(1);
+	}
 }
 
 void __init hv_apic_init(void)
include/asm-generic/hyperv-tlfs.h +1 -1
index fdce7a4cfc6f..6e2a090e2649 100644
@@ -208,6 +208,7 @@ enum HV_GENERIC_SET_FORMAT {
 #define HV_STATUS_INVALID_PORT_ID		17
 #define HV_STATUS_INVALID_CONNECTION_ID		18
 #define HV_STATUS_INSUFFICIENT_BUFFERS		19
+#define HV_STATUS_TIME_OUT			0x78
 
 /*
  * The Hyper-V TimeRefCount register and the TSC
--

2.25.1

#14 lore raw

RE: [RFC PATCH V2 04/18] x86/hyperv: Decrypt hv vp assist page in sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 22 lines

hv vp assist page is shared between sev snp guest and hyperv. Decrypt the page when use it.

---
 arch/x86/hyperv/hv_init.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
arch/x86/hyperv/hv_init.c +1 -0
index 29774126e931..4600c5941957 100644
@@ -30,6 +30,7 @@
 #include <clocksource/hyperv_timer.h>
 #include <linux/highmem.h>
 #include <linux/swiotlb.h>
+#include <linux/set_memory.h>
 
 int hyperv_init_cpuhp;
 u64 hv_current_partition_id = ~0ull;

arch/x86/hyperv/hv_init.c static int hv_cpu_init(unsigned int cpu) · line 117

Quoted 5 lines
arch/x86/hyperv/hv_init.c +2 -0
@@ -112,6 +113,11 @@ static int hv_cpu_init(unsigned int cpu)
 		}
 		WARN_ON(!(*hvp));
 		if (*hvp) {
+			if (hv_isolation_type_en_snp()) {
+				WARN_ON_ONCE(set_memory_decrypted((unsigned long)(*hvp), 1) != 0);

The "!= 0" isn't needed.

Quoted 7 lines
arch/x86/hyperv/hv_init.c +3 -0
+				memset(*hvp, 0, PAGE_SIZE);
+			}
+
 			msr.enable = 1;
 			wrmsrl(HV_X64_MSR_VP_ASSIST_PAGE, msr.as_uint64);
 		}

arch/x86/hyperv/hv_init.c static int hv_cpu_die(unsigned int cpu) · line 242

Quoted 9 lines
arch/x86/hyperv/hv_init.c +6 -0
@@ -228,6 +234,12 @@ static int hv_cpu_die(unsigned int cpu)
 
 	if (hv_vp_assist_page && hv_vp_assist_page[cpu]) {
 		union hv_vp_assist_msr_contents msr = { 0 };
+
+		if (hv_isolation_type_en_snp())
+			WARN_ON_ONCE(set_memory_encrypted(
+				    (unsigned long)hv_vp_assist_page[cpu],
+				    1) != 0);
+

The re-encryption should not be done here (or anywhere else, for
that matter) since the VP assist pages are never freed. The Hyper-V
synthetic MSR pointing to the page gets cleared, but the memory isn't freed. If the CPU should come back online later, the previously allocated
VP assist page is reused. The decryption in hv_cpu_init() is done only
when a new page is allocated to use as a VP assist page. So just leave the page decrypted here, and it will get reused in its decrypted state.

This handling of the VP assist page is admittedly a bit weird. But it is needed. See discussion with Vitaly Kuznetsov here:
https://lore.kernel.org/linux-hyperv/[email protected]/

Michael

Quoted 5 lines
arch/x86/hyperv/hv_init.c +0 -1
 		if (hv_root_partition) {
 			/*
 			 * For root partition the VP assist page is mapped to
--

2.25.1

#17 lore raw

RE: [RFC PATCH V2 05/18] x86/hyperv: Get Virtual Trust Level via hvcall

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 3 lines

sev-snp guest provides vtl(Virtual Trust Level) and get it from hyperv hvcall via HVCALL_GET_VP_REGISTERS.

Two general comments:

1) Could this patch be combined with Patch 9 of the series? It seems like they go together since Patch 9 is the consumer of the VTL.

2) What is the bigger picture motivation for this patch and Patch 9 being part of the patch series for support fully enlightened SEV-SNP guests? Won't the VTL always be 0 in such a guest? The code currently assumes VTL 0, so it seems like this patch doesn't change anything. Or maybe there's a scenario that I'm not aware of where the VTL is other than 0.

Quoted 12 lines
---
 arch/x86/hyperv/hv_init.c      | 35 ++++++++++++++++++++++++++++++++++
 include/asm-generic/mshyperv.h |  2 ++
 2 files changed, 37 insertions(+)
arch/x86/hyperv/hv_init.c +0 -0
index 4600c5941957..5b919d4d24c0 100644

arch/x86/hyperv/hv_init.c static void __init hv_get_partition_id(void) · line 393

Quoted 4 lines
arch/x86/hyperv/hv_init.c +1 -0
@@ -390,6 +390,39 @@ static void __init hv_get_partition_id(void)
 	local_irq_restore(flags);
 }
 
+static u8 __init get_current_vtl(void)

The name get_current_vtl() seems to imply that there might be a "previous" VTL, or that the VTL might change over time. I'm not aware that either is the case. Couldn't this just be get_vtl()?

arch/x86/hyperv/hv_init.c static void __init hv_get_partition_id(void) · line 395

Quoted 2 lines
arch/x86/hyperv/hv_init.c +2 -0
+{
+	u64 control = ((u64)1 << HV_HYPERCALL_REP_COMP_OFFSET) | HVCALL_GET_VP_REGISTERS;

Simplify by using HV_HYPERCALL_REP_COMP_1.

arch/x86/hyperv/hv_init.c static void __init hv_get_partition_id(void) · line 397

Quoted 2 lines
arch/x86/hyperv/hv_init.c +2 -0
+	struct hv_get_vp_registers_input *input = NULL;
+	struct hv_get_vp_registers_output *output = NULL;

It doesn't seem like the above two initializations to NULL are needed.

arch/x86/hyperv/hv_init.c static void __init hv_get_partition_id(void) · line 399

Quoted 2 lines
arch/x86/hyperv/hv_init.c +2 -0
+	u8 vtl = 0;
+	int ret;

The result of hv_do_hypercall() should always be a u64.

arch/x86/hyperv/hv_init.c static void __init hv_get_partition_id(void) · line 405

Quoted 6 lines
arch/x86/hyperv/hv_init.c +6 -0
+	unsigned long flags;
+
+	local_irq_save(flags);
+	input = *(struct hv_get_vp_registers_input **)this_cpu_ptr(hyperv_pcpu_input_arg);
+	output = (struct hv_get_vp_registers_output *)input;
+	if (!input || !output) {

Don't need to check both values since one is assigned from the other. :-)

arch/x86/hyperv/hv_init.c static void __init hv_get_partition_id(void) · line 406

Quoted 1 line
arch/x86/hyperv/hv_init.c +1 -0
+		pr_err("Hyper-V: cannot allocate a shared page!");

Error message text isn't correct.

arch/x86/hyperv/hv_init.c static void __init hv_get_partition_id(void) · line 407

Quoted 1 line
arch/x86/hyperv/hv_init.c +1 -0
+		goto done;

Need to do local_irq_restore() before goto done.

arch/x86/hyperv/hv_init.c static void __init hv_get_partition_id(void) · line 413

Quoted 6 lines
arch/x86/hyperv/hv_init.c +6 -0
+	}
+
+	memset(input, 0, sizeof(*input) + sizeof(input->element[0]));
+	input->header.partitionid = HV_PARTITION_ID_SELF;
+	input->header.inputvtl = 0;
+	input->element[0].name0 = 0x000D0003;

This constant should go in one of the hyperv-tlfs.h header files. If I recall correctly, we're currently treating VTLs as x86-specific, so should go in arch/x86/include/asm/hyperv-tlfs.h.

arch/x86/hyperv/hv_init.c static void __init hv_get_partition_id(void) · line 416

Quoted 3 lines
arch/x86/hyperv/hv_init.c +3 -0
+
+	ret = hv_do_hypercall(control, input, output);
+	if (ret == 0)

Use hv_result_success(ret).

arch/x86/hyperv/hv_init.c static void __init hv_get_partition_id(void) · line 417

Quoted 1 line
arch/x86/hyperv/hv_init.c +1 -0
+		vtl = output->as64.low & 0xf;

The 0xF mask should be defined in the hyperv-tlfs.h per above.

arch/x86/hyperv/hv_init.c static void __init hv_get_partition_id(void) · line 419

Quoted 2 lines
arch/x86/hyperv/hv_init.c +2 -0
+	else
+		pr_err("Hyper-V: failed to get the current VTL!");

Again, drop the word "current". And the exclamation mark isn't needed. :-)

Quoted 10 lines
arch/x86/hyperv/hv_init.c +6 -0
+	local_irq_restore(flags);
+
+done:
+	return vtl;
+}
+
 /*
  * This function is to be invoked early in the boot sequence after the
  * hypervisor has been detected.

arch/x86/hyperv/hv_init.c void __init hyperv_init(void) · line 564

Quoted 5 lines
arch/x86/hyperv/hv_init.c +2 -0
@@ -527,6 +560,8 @@ void __init hyperv_init(void)
 	if (hv_is_isolation_supported())
 		swiotlb_update_mem_attributes();
 #endif
+	/* Find the current VTL */
+	ms_hyperv.vtl = get_current_vtl();

Drop "current" in the comment and function name.

Quoted 16 lines
arch/x86/hyperv/hv_init.c +0 -0
 	return;
 
include/asm-generic/mshyperv.h +1 -0
index bfb9eb9d7215..68133de044ec 100644
@@ -46,6 +46,7 @@ struct ms_hyperv_info {
 		};
 	};
 	u64 shared_gpa_boundary;
+	u8 vtl;
 };
 extern struct ms_hyperv_info ms_hyperv;
 

include/asm-generic/mshyperv.h extern void * __percpu *hyperv_pcpu_output_arg; · line 59

Quoted 4 lines
include/asm-generic/mshyperv.h +1 -0
@@ -55,6 +56,7 @@ extern void * __percpu *hyperv_pcpu_output_arg;
 extern u64 hv_do_hypercall(u64 control, void *inputaddr, void *outputaddr);
 extern u64 hv_do_fast_hypercall8(u16 control, u64 input8);
 extern bool hv_isolation_type_snp(void);
+extern bool hv_isolation_type_en_snp(void);

This declaration of hv_isolation_type_en_snp() shouldn't be needed here as it has already been added to arch/x86/include/asm/mshyperv.h.

The declaration of hv_isolation_type_snp() occurs both places, but I think that's some sloppiness from the past that could be fixed. In fact, hv_isolation_type_snp() occurs *twice* in include/asm-generic/mshyperv.h.

Quoted 6 lines
include/asm-generic/mshyperv.h +0 -0
 /* Helper functions that provide a consistent pattern for checking Hyper-V hypercall

status. */
static inline int hv_result(u64 status)
--
2.25.1

#62 lore raw

Re: [RFC PATCH V2 18/18] x86/sev: Fix interrupt exit code paths from #HV exception

In reply to Tianyu Lan
Quoted 23 lines

Add checks in interrupt exit code paths in case of returns to user mode to check if currently executing the #HV handler then don't follow the irqentry_exit_to_user_mode path as that can potentially cause the #HV handler to be preempted and rescheduled on another CPU. Rescheduled #HV handler on another cpu will cause interrupts to be handled on a different cpu than the injected one, causing invalid EOIs and missed/lost guest interrupts and corresponding hangs and/or per-cpu IRQs handled on non-intended cpu.

---
  arch/x86/include/asm/idtentry.h | 66 +++++++++++++++++++++++++++++++++
  arch/x86/kernel/sev.c           | 30 +++++++++++++++
  2 files changed, 96 insertions(+)
arch/x86/include/asm/idtentry.h +0 -0
index 652fea10d377..45b47132be7c 100644

arch/x86/include/asm/idtentry.h line 17

Quoted 5 lines
arch/x86/include/asm/idtentry.h +2 -0
@@ -13,6 +13,10 @@
 
  #include <asm/irq_stack.h>
 
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+noinstr void irqentry_exit_hv_cond(struct pt_regs *regs, irqentry_state_t state);

For linux host, CONFIG_AMD_MEM_ENCRYPT also gets enabled at host side (for SME) and 'irqentry_exit_hv_cond' gets called. So, we need to handle the below cases even when host CONFIG_AMD_MEM_ENCRYPT is enabled?

Thanks,
Pankaj

Quoted 161 lines
arch/x86/include/asm/idtentry.h +64 -0
+#endif
+
  /**
   * DECLARE_IDTENTRY - Declare functions for simple IDT entry points
   *		      No error code pushed by hardware
@@ -176,6 +180,7 @@ __visible noinstr void func(struct pt_regs *regs, unsigned long error_code)
  #define DECLARE_IDTENTRY_IRQ(vector, func)				\
  	DECLARE_IDTENTRY_ERRORCODE(vector, func)
 
+#ifndef CONFIG_AMD_MEM_ENCRYPT
  /**
   * DEFINE_IDTENTRY_IRQ - Emit code for device interrupt IDT entry points
   * @func:	Function name of the entry point
@@ -205,6 +210,26 @@ __visible noinstr void func(struct pt_regs *regs,			\
  }									\
  									\
  static noinline void __##func(struct pt_regs *regs, u32 vector)
+#else
+
+#define DEFINE_IDTENTRY_IRQ(func)					\
+static void __##func(struct pt_regs *regs, u32 vector);		\
+									\
+__visible noinstr void func(struct pt_regs *regs,			\
+			    unsigned long error_code)			\
+{									\
+	irqentry_state_t state = irqentry_enter(regs);			\
+	u32 vector = (u32)(u8)error_code;				\
+									\
+	instrumentation_begin();					\
+	kvm_set_cpu_l1tf_flush_l1d();					\
+	run_irq_on_irqstack_cond(__##func, regs, vector);		\
+	instrumentation_end();						\
+	irqentry_exit_hv_cond(regs, state);				\
+}									\
+									\
+static noinline void __##func(struct pt_regs *regs, u32 vector)
+#endif
 
  /**
   * DECLARE_IDTENTRY_SYSVEC - Declare functions for system vector entry points
@@ -221,6 +246,7 @@ static noinline void __##func(struct pt_regs *regs, u32 vector)
  #define DECLARE_IDTENTRY_SYSVEC(vector, func)				\
  	DECLARE_IDTENTRY(vector, func)
 
+#ifndef CONFIG_AMD_MEM_ENCRYPT
  /**
   * DEFINE_IDTENTRY_SYSVEC - Emit code for system vector IDT entry points
   * @func:	Function name of the entry point
@@ -245,6 +271,26 @@ __visible noinstr void func(struct pt_regs *regs)			\
  }									\
  									\
  static noinline void __##func(struct pt_regs *regs)
+#else
+
+#define DEFINE_IDTENTRY_SYSVEC(func)					\
+static void __##func(struct pt_regs *regs);				\
+									\
+__visible noinstr void func(struct pt_regs *regs)			\
+{									\
+	irqentry_state_t state = irqentry_enter(regs);			\
+									\
+	instrumentation_begin();					\
+	kvm_set_cpu_l1tf_flush_l1d();					\
+	run_sysvec_on_irqstack_cond(__##func, regs);			\
+	instrumentation_end();						\
+	irqentry_exit_hv_cond(regs, state);				\
+}									\
+									\
+static noinline void __##func(struct pt_regs *regs)
+#endif
+
+#ifndef CONFIG_AMD_MEM_ENCRYPT
 
  /**
   * DEFINE_IDTENTRY_SYSVEC_SIMPLE - Emit code for simple system vector IDT
@@ -274,6 +320,26 @@ __visible noinstr void func(struct pt_regs *regs)			\
  }									\
  									\
  static __always_inline void __##func(struct pt_regs *regs)
+#else
+
+#define DEFINE_IDTENTRY_SYSVEC_SIMPLE(func)				\
+static __always_inline void __##func(struct pt_regs *regs);		\
+									\
+__visible noinstr void func(struct pt_regs *regs)			\
+{									\
+	irqentry_state_t state = irqentry_enter(regs);			\
+									\
+	instrumentation_begin();					\
+	__irq_enter_raw();						\
+	kvm_set_cpu_l1tf_flush_l1d();					\
+	__##func(regs);						\
+	__irq_exit_raw();						\
+	instrumentation_end();						\
+	irqentry_exit_hv_cond(regs, state);				\
+}									\
+									\
+static __always_inline void __##func(struct pt_regs *regs)
+#endif
 
  /**
   * DECLARE_IDTENTRY_XENCB - Declare functions for XEN HV callback entry point
arch/x86/kernel/sev.c +30 -0
index 5a2f59022c98..ef6a123c50fe 100644
@@ -153,6 +153,10 @@ struct sev_hv_doorbell_page {
 
  struct sev_snp_runtime_data {
  	struct sev_hv_doorbell_page hv_doorbell_page;
+	/*
+	 * Indication that we are currently handling #HV events.
+	 */
+	bool hv_handling_events;
  };
 
  static DEFINE_PER_CPU(struct sev_snp_runtime_data*, snp_runtime_data);
@@ -206,6 +210,8 @@ static void do_exc_hv(struct pt_regs *regs)
  	union hv_pending_events pending_events;
  	u8 vector;
 
+	this_cpu_read(snp_runtime_data)->hv_handling_events = true;
+
  	while (sev_hv_pending()) {
  		asm volatile("cli" : : : "memory");
 
@@ -244,6 +250,8 @@ static void do_exc_hv(struct pt_regs *regs)
 
  		asm volatile("sti" : : : "memory");
  	}
+
+	this_cpu_read(snp_runtime_data)->hv_handling_events = false;
  }
 
  void check_hv_pending(struct pt_regs *regs)
@@ -2541,3 +2549,25 @@ static int __init snp_init_platform_device(void)
  	return 0;
  }
  device_initcall(snp_init_platform_device);
+
+noinstr void irqentry_exit_hv_cond(struct pt_regs *regs, irqentry_state_t state)
+{
+	/*
+	 * Check whether this returns to user mode, if so and if
+	 * we are currently executing the #HV handler then we don't
+	 * want to follow the irqentry_exit_to_user_mode path as
+	 * that can potentially cause the #HV handler to be
+	 * preempted and rescheduled on another CPU. Rescheduled #HV
+	 * handler on another cpu will cause interrupts to be handled
+	 * on a different cpu than the injected one, causing
+	 * invalid EOIs and missed/lost guest interrupts and
+	 * corresponding hangs and/or per-cpu IRQs handled on
+	 * non-intended cpu.
+	 */
+	if (user_mode(regs) &&
+	    this_cpu_read(snp_runtime_data)->hv_handling_events)
+		return;
+
+	/* follow normal interrupt return/exit path */
+	irqentry_exit(regs, state);
+}
author
#10 lore raw

Re: [RFC PATCH V2 02/18] x86/hyperv: Add sev-snp enlightened guest specific config

In reply to Michael Kelley (LINUX)
Quoted 11 lines · Michael Kelley (LINUX)

arch/x86/include/asm/mshyperv.h extern u64 hv_current_partition_id; · line 36

Quoted 5 lines
arch/x86/include/asm/mshyperv.h +1 -0
@@ -32,6 +33,7 @@ extern u64 hv_current_partition_id;
 
  extern union hv_ghcb * __percpu *hv_ghcb_pg;
 
+extern bool hv_isolation_type_en_snp(void);

This file also has a declaration for hv_isolation_type_snp(). I think this new declaration is near the beginning of this file so that it can be referenced by hv_do_hypercall() and related functions in Patch 6 of this series. So maybe move the declaration of hv_isolation_type_snp() up so it is adjacent to this one? It would make sense for the two to be together.

Agree. Will update in the next version.

Quoted 20 lines
Quoted 20 lines
arch/x86/kernel/cpu/mshyperv.c +11 -0
index 831613959a92..2ea4f21c6172 100644
@@ -273,6 +273,21 @@ static void __init ms_hyperv_init_platform(void)
  	ms_hyperv.misc_features = cpuid_edx(HYPERV_CPUID_FEATURES);
  	ms_hyperv.hints    = cpuid_eax(HYPERV_CPUID_ENLIGHTMENT_INFO);
 
+	/*
+	 * Add custom configuration for SEV-SNP Enlightened guest
+	 */
+	if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+		ms_hyperv.features |= HV_ACCESS_FREQUENCY_MSRS;
+		ms_hyperv.misc_features |= HV_FEATURE_FREQUENCY_MSRS_AVAILABLE;
+		ms_hyperv.misc_features &= ~HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE;
+		ms_hyperv.hints |= HV_DEPRECATING_AEOI_RECOMMENDED;
+		ms_hyperv.hints |= HV_X64_APIC_ACCESS_RECOMMENDED;
+		ms_hyperv.hints |= HV_X64_CLUSTER_IPI_RECOMMENDED;
+	}
Quoted 8 lines

arch/x86/kernel/cpu/mshyperv.c static void __init ms_hyperv_init_platform(void) · line 290

Quoted 4 lines
arch/x86/kernel/cpu/mshyperv.c +4 -0
+
+	pr_info("Hyper-V: enlightment features 0x%x, hints 0x%x, misc 0x%x\n",
+		ms_hyperv.features, ms_hyperv.hints, ms_hyperv.misc_features);
+

What's the reason for this additional call to pr_info()? There's a call to pr_info() a couple of lines below that displays the same information, and a little bit more. It seems like the above call should be deleted, as I think we should try to be as consistent as possible in the output.

Sorry for noise. This one should be redundant. Will remove in the next version.

Quoted 16 lines
Quoted 16 lines
arch/x86/kernel/cpu/mshyperv.c +3 -4
@@ -328,18 +343,22 @@ static void __init ms_hyperv_init_platform(void)
  		ms_hyperv.shared_gpa_boundary =
  			BIT_ULL(ms_hyperv.shared_gpa_boundary_bits);
 
-		pr_info("Hyper-V: Isolation Config: Group A 0x%x, Group B 0x%x\n",
-			ms_hyperv.isolation_config_a, ms_hyperv.isolation_config_b);
-
-		if (hv_get_isolation_type() == HV_ISOLATION_TYPE_SNP) {
+		if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+			static_branch_enable(&isolation_type_en_snp);
+		} else if (hv_get_isolation_type() == HV_ISOLATION_TYPE_SNP) {
  			static_branch_enable(&isolation_type_snp);
  #ifdef CONFIG_SWIOTLB
  			swiotlb_unencrypted_base = ms_hyperv.shared_gpa_boundary;
  #endif
Quoted 9 lines

arch/x86/kernel/cpu/mshyperv.c static void __init ms_hyperv_init_platform(void) · line 357

Quoted 5 lines
arch/x86/kernel/cpu/mshyperv.c +4 -0
  		}
+
+		pr_info("Hyper-V: Isolation Config: Group A 0x%x, Group B 0x%x\n",
+			ms_hyperv.isolation_config_a, ms_hyperv.isolation_config_b);
+

Is there a reason for moving this pr_info() down a few lines? I can't see that the intervening code changes any of the settings that are displayed, so it should be good in the original location. Just trying to minimize changes that don't add value ...

Agree. Will keep previous order. Thanks.

author
#15 lore raw

Re: [RFC PATCH V2 04/18] x86/hyperv: Decrypt hv vp assist page in sev-snp enlightened guest

In reply to Michael Kelley (LINUX)
Quoted 18 lines · Michael Kelley (LINUX)

arch/x86/hyperv/hv_init.c static int hv_cpu_die(unsigned int cpu) · line 242

Quoted 10 lines
arch/x86/hyperv/hv_init.c +6 -0
@@ -228,6 +234,12 @@ static int hv_cpu_die(unsigned int cpu)
 
  	if (hv_vp_assist_page && hv_vp_assist_page[cpu]) {
  		union hv_vp_assist_msr_contents msr = { 0 };
+
+		if (hv_isolation_type_en_snp())
+			WARN_ON_ONCE(set_memory_encrypted(
+				    (unsigned long)hv_vp_assist_page[cpu],
+				    1) != 0);
+

The re-encryption should not be done here (or anywhere else, for
that matter) since the VP assist pages are never freed. The Hyper-V
synthetic MSR pointing to the page gets cleared, but the memory isn't freed. If the CPU should come back online later, the previously allocated
VP assist page is reused. The decryption in hv_cpu_init() is done only
when a new page is allocated to use as a VP assist page. So just leave the page decrypted here, and it will get reused in its decrypted state.

Quoted 4 lines

This handling of the VP assist page is admittedly a bit weird. But it is needed. See discussion with Vitaly Kuznetsov here:
https://lore.kernel.org/linux-hyperv/[email protected]/

Agree. Good point. Will update in the next version.

Thanks.

#19 lore raw

RE: [RFC PATCH V2 06/18] x86/hyperv: Use vmmcall to implement hvcall in sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 2 lines

In sev-snp enlightened guest, hvcall needs to use vmmcall to trigger

What does "hvcall" refer to here? Is this a Hyper-V hypercall, or just a generic hypervisor call?

Quoted 57 lines

vmexit and notify hypervisor to handle hypercall request.

---
 arch/x86/include/asm/mshyperv.h | 66 ++++++++++++++++++++++-----------
 1 file changed, 45 insertions(+), 21 deletions(-)
arch/x86/include/asm/mshyperv.h +24 -10
index 9b8c3f638845..28d5429e33c9 100644
@@ -45,16 +45,25 @@ static inline u64 hv_do_hypercall(u64 control, void *input, void *output)
 	u64 hv_status;
 
 #ifdef CONFIG_X86_64
-	if (!hv_hypercall_pg)
-		return U64_MAX;
+	if (hv_isolation_type_en_snp()) {
+		__asm__ __volatile__("mov %4, %%r8\n"
+				"vmmcall"
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input_address)
+				:  "r" (output_address)
+				: "cc", "memory", "r8", "r9", "r10", "r11");
+	} else {
+		if (!hv_hypercall_pg)
+			return U64_MAX;
 
-	__asm__ __volatile__("mov %4, %%r8\n"
-			     CALL_NOSPEC
-			     : "=a" (hv_status), ASM_CALL_CONSTRAINT,
-			       "+c" (control), "+d" (input_address)
-			     :  "r" (output_address),
-				THUNK_TARGET(hv_hypercall_pg)
-			     : "cc", "memory", "r8", "r9", "r10", "r11");
+		__asm__ __volatile__("mov %4, %%r8\n"
+				CALL_NOSPEC
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input_address)
+				:  "r" (output_address),
+					THUNK_TARGET(hv_hypercall_pg)
+				: "cc", "memory", "r8", "r9", "r10", "r11");
+	}
 #else
 	u32 input_address_hi = upper_32_bits(input_address);
 	u32 input_address_lo = lower_32_bits(input_address);
@@ -82,12 +91,18 @@ static inline u64 hv_do_fast_hypercall8(u16 code, u64 input1)
 	u64 hv_status, control = (u64)code | HV_HYPERCALL_FAST_BIT;
 
 #ifdef CONFIG_X86_64
-	{
+	if (hv_isolation_type_en_snp()) {
+		__asm__ __volatile__(
+				"vmmcall"
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)
+				:: "cc", "r8", "r9", "r10", "r11");

arch/x86/include/asm/mshyperv.h static inline u64 hv_do_fast_hypercall8(u16 code, u64 input1) · line 105

Quoted 10 lines
arch/x86/include/asm/mshyperv.h +5 -4
+	} else {
 		__asm__ __volatile__(CALL_NOSPEC
-				     : "=a" (hv_status), ASM_CALL_CONSTRAINT,
-				       "+c" (control), "+d" (input1)
-				     : THUNK_TARGET(hv_hypercall_pg)
-				     : "cc", "r8", "r9", "r10", "r11");
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)
+				: THUNK_TARGET(hv_hypercall_pg)
+				: "cc", "r8", "r9", "r10", "r11");

The above 4 lines appear to just have changes in indentation. Maybe there's value in having the same indentation as the new code you've added, so I won't object if you want to keep the changes.

Quoted 19 lines
arch/x86/include/asm/mshyperv.h +4 -7
 	}
 #else
 	{
@@ -113,14 +128,21 @@ static inline u64 hv_do_fast_hypercall16(u16 code, u64 input1, u64 input2)
 	u64 hv_status, control = (u64)code | HV_HYPERCALL_FAST_BIT;
 
 #ifdef CONFIG_X86_64
-	{
+	if (hv_isolation_type_en_snp()) {
 		__asm__ __volatile__("mov %4, %%r8\n"
-				     CALL_NOSPEC
-				     : "=a" (hv_status), ASM_CALL_CONSTRAINT,
-				       "+c" (control), "+d" (input1)
-				     : "r" (input2),
-				       THUNK_TARGET(hv_hypercall_pg)
-				     : "cc", "r8", "r9", "r10", "r11");
+				"vmmcall"
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)

arch/x86/include/asm/mshyperv.h static inline u64 hv_do_fast_hypercall16(u16 code, u64 input1, u64 input2) · line 145

Quoted 10 lines
arch/x86/include/asm/mshyperv.h +10 -0
+				: "r" (input2)
+				: "cc", "r8", "r9", "r10", "r11");
+	} else {
+		__asm__ __volatile__("mov %4, %%r8\n"
+				CALL_NOSPEC
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)
+				: "r" (input2),
+				THUNK_TARGET(hv_hypercall_pg)
+				: "cc", "r8", "r9", "r10", "r11");

Same here. Above 5 lines appear to be changes only in indentation.

arch/x86/include/asm/mshyperv.h int hv_map_ioapic_interrupt(int ioapic_id, bool level, int vcpu, · line 203

Quoted 9 lines
arch/x86/include/asm/mshyperv.h +1 -0
 	}
 #else
 	{
@@ -177,6 +199,7 @@ int hv_map_ioapic_interrupt(int ioapic_id, bool level, int vcpu,
 int vector,
 		struct hv_interrupt_entry *entry);
 int hv_unmap_ioapic_interrupt(int ioapic_id, struct hv_interrupt_entry *entry);
 int hv_set_mem_host_visibility(unsigned long addr, int numpages, bool visible);
+int hv_snp_boot_ap(int cpu, unsigned long start_ip);

This declaration doesn't seem to belong in this patch. It should be in Patch 13 of the series.

arch/x86/include/asm/mshyperv.h static inline void hv_ghcb_terminate(unsigned int set, unsigned · line 218

Quoted 9 lines
arch/x86/include/asm/mshyperv.h +1 -0
 #ifdef CONFIG_AMD_MEM_ENCRYPT
 void hv_ghcb_msr_write(u64 msr, u64 value);
@@ -191,6 +214,7 @@ static inline void hv_ghcb_terminate(unsigned int set, unsigned
 int reason) {}
 #endif
 
 extern bool hv_isolation_type_snp(void);
+extern bool hv_isolation_type_en_snp(void);

This declaration seems to be a duplicate that doesn't belong in this patch.

Quoted 5 lines
arch/x86/include/asm/mshyperv.h +0 -1
 static inline bool hv_is_synic_reg(unsigned int reg)
 {
--

2.25.1

#23 lore raw

RE: [RFC PATCH V2 07/18] clocksource: hyper-v: decrypt hyperv tsc page in sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 1 line

Previous patches to the Hyper-V clocksource driver have not been very consistent in the Subject line prefix, but let's use "clocksource/drivers/hyper-v:" since it has been used the most.

Quoted 23 lines

Hyper-V tsc page is shared with hypervisor and it should be decrypted in sev-snp enlightened guest when it's used.

---
 drivers/clocksource/hyperv_timer.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/clocksource/hyperv_timer.c +1 -2
index bb47610bbd1c..aa68eebed5ee 100644
@@ -364,7 +364,7 @@ EXPORT_SYMBOL_GPL(hv_stimer_global_cleanup);
 static union {
 	struct ms_hyperv_tsc_page page;
 	u8 reserved[PAGE_SIZE];
-} tsc_pg __aligned(PAGE_SIZE);
+} tsc_pg __bss_decrypted __aligned(PAGE_SIZE);
 
 struct ms_hyperv_tsc_page *hv_get_tsc_page(void)
 {
--

2.25.1

#26 lore raw

RE: [RFC PATCH V2 08/18] x86/hyperv: decrypt vmbus pages for sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 1 line

The Subject prefix for this patch should be "Drivers: hv: vmbus:"

Quoted 14 lines

Vmbus int, synic and post message pages are shared with hypervisor and so decrypt these pages in the sev-snp guest.

---
 drivers/hv/connection.c | 13 +++++++++++++
 drivers/hv/hv.c         | 32 +++++++++++++++++++++++++++++++-
 2 files changed, 44 insertions(+), 1 deletion(-)
drivers/hv/connection.c +0 -0
index 9dc27e5d367a..43141225ea15 100644

drivers/hv/connection.c int vmbus_connect(void) · line 218

Quoted 4 lines
drivers/hv/connection.c +1 -0
@@ -215,6 +215,15 @@ int vmbus_connect(void)
 		(void *)((unsigned long)vmbus_connection.int_page +
 			(HV_HYP_PAGE_SIZE >> 1));
 
+	if (hv_isolation_type_snp() || hv_isolation_type_en_snp()) {

This decryption should be done only for a fully enlightened SEV-SNP guest, not for a vTOM guest.

drivers/hv/connection.c int vmbus_connect(void) · line 222

Quoted 4 lines
drivers/hv/connection.c +4 -0
+		ret = set_memory_decrypted((unsigned long)
+				vmbus_connection.int_page, 1);
+		if (ret)
+			goto cleanup;

This cleanup path doesn't work correctly. It calls vmbus_disconnect(), which will try to re-encrypt the memory. But if the original decryption failed, re-encrypting is the wrong thing to do.

It looks like this same bug exists in current code if the decryption of the monitor pages fails or if just one of the original memory allocations fails. vmbus_disconnect() doesn't know whether it should re-encrypt the pages.

Quoted 50 lines
drivers/hv/connection.c +8 -0
+
+		memset(vmbus_connection.int_page, 0, PAGE_SIZE);
+	}
+
 	/*
 	 * Setup the monitor notification facility. The 1st page for
 	 * parent->child and the 2nd page for child->parent
@@ -372,6 +381,10 @@ void vmbus_disconnect(void)
 		destroy_workqueue(vmbus_connection.work_queue);
 
 	if (vmbus_connection.int_page) {
+		if (hv_isolation_type_en_snp())
+			set_memory_encrypted((unsigned long)
+				vmbus_connection.int_page, 1);
+
 		hv_free_hyperv_page((unsigned long)vmbus_connection.int_page);
 		vmbus_connection.int_page = NULL;
 	}
drivers/hv/hv.c +9 -1
index 4d6480d57546..f9111eb32739 100644
@@ -20,6 +20,7 @@
 #include <linux/interrupt.h>
 #include <clocksource/hyperv_timer.h>
 #include <asm/mshyperv.h>
+#include <linux/set_memory.h>
 #include "hyperv_vmbus.h"
 
 /* The one and only */
@@ -117,7 +118,7 @@ int hv_post_message(union hv_connection_id connection_id,
 
 int hv_synic_alloc(void)
 {
-	int cpu;
+	int cpu, ret;
 	struct hv_per_cpu_context *hv_cpu;
 
 	/*
@@ -168,6 +169,29 @@ int hv_synic_alloc(void)
 			pr_err("Unable to allocate post msg page\n");
 			goto err;
 		}
+
+		if (hv_isolation_type_en_snp()) {
+			ret = set_memory_decrypted((unsigned long)
+				hv_cpu->synic_message_page, 1);
+			ret |= set_memory_decrypted((unsigned long)
+				hv_cpu->synic_event_page, 1);
+			ret |= set_memory_decrypted((unsigned long)

drivers/hv/hv.c int hv_synic_alloc(void) · line 188

Quoted 10 lines
drivers/hv/hv.c +10 -0
+				hv_cpu->post_msg_page, 1);
+
+			if (ret) {
+				set_memory_encrypted((unsigned long)
+					hv_cpu->synic_message_page, 1);
+				set_memory_encrypted((unsigned long)
+					hv_cpu->synic_event_page, 1);
+				set_memory_encrypted((unsigned long)
+					hv_cpu->post_msg_page, 1);
+				goto err;

Same kind of cleanup problem here. Some of the memory may have been decrypted, but some may not have. Re-encrypting all three pages risks re-encrypting a page that failed to be decrypted, and that might cause problems.

Quoted 10 lines
drivers/hv/hv.c +6 -0
+			}
+
+			memset(hv_cpu->synic_message_page, 0, PAGE_SIZE);
+			memset(hv_cpu->synic_event_page, 0, PAGE_SIZE);
+			memset(hv_cpu->post_msg_page, 0, PAGE_SIZE);
+		}
 	}
 
 	return 0;

drivers/hv/hv.c void hv_synic_free(void) · line 218

Quoted 7 lines
drivers/hv/hv.c +4 -0
@@ -188,6 +212,12 @@ void hv_synic_free(void)
 		struct hv_per_cpu_context *hv_cpu
 			= per_cpu_ptr(hv_context.cpu_context, cpu);
 
+		if (hv_isolation_type_en_snp()) {
+			set_memory_encrypted((unsigned long)hv_cpu->synic_message_page, 1);
+			set_memory_encrypted((unsigned long)hv_cpu->synic_event_page, 1);
+			set_memory_encrypted((unsigned long)hv_cpu->post_msg_page, 1);

This cleanup doesn't always work correctly. There are multiple memory allocations in hv_synic_alloc(). If some succeeded, but some failed, then might get here with some memory that was allocated but not decrypted. Trying to re-encrypt that memory before freeing it could cause problems.

Quoted 7 lines
drivers/hv/hv.c +2 -1
+		}
+
 		free_page((unsigned long)hv_cpu->synic_event_page);
 		free_page((unsigned long)hv_cpu->synic_message_page);
 		free_page((unsigned long)hv_cpu->post_msg_page);
--

2.25.1

author
#20 lore raw

Re: [RFC PATCH V2 06/18] x86/hyperv: Use vmmcall to implement hvcall in sev-snp enlightened guest

In reply to Michael Kelley (LINUX)
Quoted 6 lines · Michael Kelley (LINUX)

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 2 lines

In sev-snp enlightened guest, hvcall needs to use vmmcall to trigger

What does "hvcall" refer to here? Is this a Hyper-V hypercall, or just a generic hypervisor call?

It's should be Hyper-V hypercall. Will make it accurate in the next version. Thanks for reminder.

Quoted 64 lines
Quoted 64 lines

vmexit and notify hypervisor to handle hypercall request.

---
  arch/x86/include/asm/mshyperv.h | 66 ++++++++++++++++++++++-----------
  1 file changed, 45 insertions(+), 21 deletions(-)
arch/x86/include/asm/mshyperv.h +25 -14
index 9b8c3f638845..28d5429e33c9 100644
@@ -45,16 +45,25 @@ static inline u64 hv_do_hypercall(u64 control, void *input, void *output)
  	u64 hv_status;
 
  #ifdef CONFIG_X86_64
-	if (!hv_hypercall_pg)
-		return U64_MAX;
+	if (hv_isolation_type_en_snp()) {
+		__asm__ __volatile__("mov %4, %%r8\n"
+				"vmmcall"
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input_address)
+				:  "r" (output_address)
+				: "cc", "memory", "r8", "r9", "r10", "r11");
+	} else {
+		if (!hv_hypercall_pg)
+			return U64_MAX;
 
-	__asm__ __volatile__("mov %4, %%r8\n"
-			     CALL_NOSPEC
-			     : "=a" (hv_status), ASM_CALL_CONSTRAINT,
-			       "+c" (control), "+d" (input_address)
-			     :  "r" (output_address),
-				THUNK_TARGET(hv_hypercall_pg)
-			     : "cc", "memory", "r8", "r9", "r10", "r11");
+		__asm__ __volatile__("mov %4, %%r8\n"
+				CALL_NOSPEC
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input_address)
+				:  "r" (output_address),
+					THUNK_TARGET(hv_hypercall_pg)
+				: "cc", "memory", "r8", "r9", "r10", "r11");
+	}
  #else
  	u32 input_address_hi = upper_32_bits(input_address);
  	u32 input_address_lo = lower_32_bits(input_address);
@@ -82,12 +91,18 @@ static inline u64 hv_do_fast_hypercall8(u16 code, u64 input1)
  	u64 hv_status, control = (u64)code | HV_HYPERCALL_FAST_BIT;
 
  #ifdef CONFIG_X86_64
-	{
+	if (hv_isolation_type_en_snp()) {
+		__asm__ __volatile__(
+				"vmmcall"
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)
+				:: "cc", "r8", "r9", "r10", "r11");
+	} else {
  		__asm__ __volatile__(CALL_NOSPEC
-				     : "=a" (hv_status), ASM_CALL_CONSTRAINT,
-				       "+c" (control), "+d" (input1)
-				     : THUNK_TARGET(hv_hypercall_pg)
-				     : "cc", "r8", "r9", "r10", "r11");
Quoted 8 lines

arch/x86/include/asm/mshyperv.h static inline u64 hv_do_fast_hypercall8(u16 code, u64 input1) · line 105

Quoted 4 lines
arch/x86/include/asm/mshyperv.h +4 -0
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)
+				: THUNK_TARGET(hv_hypercall_pg)
+				: "cc", "r8", "r9", "r10", "r11");

The above 4 lines appear to just have changes in indentation. Maybe there's value in having the same indentation as the new code you've added, so I won't object if you want to keep the changes.

OK. Will update in the next version.

Quoted 31 lines
Quoted 30 lines
arch/x86/include/asm/mshyperv.h +14 -7
  	}
  #else
  	{
@@ -113,14 +128,21 @@ static inline u64 hv_do_fast_hypercall16(u16 code, u64 input1, u64 input2)
  	u64 hv_status, control = (u64)code | HV_HYPERCALL_FAST_BIT;
 
  #ifdef CONFIG_X86_64
-	{
+	if (hv_isolation_type_en_snp()) {
  		__asm__ __volatile__("mov %4, %%r8\n"
-				     CALL_NOSPEC
-				     : "=a" (hv_status), ASM_CALL_CONSTRAINT,
-				       "+c" (control), "+d" (input1)
-				     : "r" (input2),
-				       THUNK_TARGET(hv_hypercall_pg)
-				     : "cc", "r8", "r9", "r10", "r11");
+				"vmmcall"
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)
+				: "r" (input2)
+				: "cc", "r8", "r9", "r10", "r11");
+	} else {
+		__asm__ __volatile__("mov %4, %%r8\n"
+				CALL_NOSPEC
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input1)
+				: "r" (input2),
+				THUNK_TARGET(hv_hypercall_pg)
+				: "cc", "r8", "r9", "r10", "r11");
author
#24 lore raw

Re: [RFC PATCH V2 07/18] clocksource: hyper-v: decrypt hyperv tsc page in sev-snp enlightened guest

In reply to Michael Kelley (LINUX)
Quoted 5 lines · Michael Kelley (LINUX)

From: Tianyu Lan<[email protected]> Sent: Friday, November 18, 2022 7:46 PM Previous patches to the Hyper-V clocksource driver have not been very consistent in the Subject line prefix, but let's use "clocksource/drivers/hyper-v:" since it has been used the most.

OK. Will update in the next version.

Thanks.

#29 lore raw

RE: [RFC PATCH V2 09/18] x86/hyperv: set target vtl in the vmbus init message

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 1 line

Subject prefix here should be "Drivers: hv: vmbus:"

Per my comments on Patch 5, I'd suggest squashing this patch into Patch 5. But also per my comments on Patch 5, I'm not clear on why the current behavior of defaulting the VTL to zero needs to be replaced by explicitly fetching and setting the VTL.

Michael

Quoted 37 lines

Set target vtl (Virtual Trust Level) in the vmbus init message.

---
 drivers/hv/connection.c | 1 +
 include/linux/hyperv.h  | 4 ++--
 2 files changed, 3 insertions(+), 2 deletions(-)
drivers/hv/connection.c +1 -0
index 43141225ea15..09a1253b539a 100644
@@ -98,6 +98,7 @@ int vmbus_negotiate_version(struct vmbus_channel_msginfo *msginfo, u32 version)
 	 */
 	if (version >= VERSION_WIN10_V5) {
 		msg->msg_sint = VMBUS_MESSAGE_SINT;
+		msg->msg_vtl = ms_hyperv.vtl;
 		vmbus_connection.msg_conn_id = VMBUS_MESSAGE_CONNECTION_ID_4;
 	} else {
 		msg->interrupt_page = virt_to_phys(vmbus_connection.int_page);
include/linux/hyperv.h +2 -3
index 3b42264333ef..2be0b5efd1ea 100644
@@ -665,8 +665,8 @@ struct vmbus_channel_initiate_contact {
 		u64 interrupt_page;
 		struct {
 			u8	msg_sint;
-			u8	padding1[3];
-			u32	padding2;
+			u8	msg_vtl;
+			u8	reserved[6];
 		};
 	};
 	u64 monitor_page1;
--

2.25.1

#33 lore raw

RE: [RFC PATCH V2 10/18] drivers: hv: Decrypt percpu hvcall input arg page in sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 30 lines

Hypervisor needs to access iput arg page and guest should decrypt the page.

---
 drivers/hv/hv_common.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)
drivers/hv/hv_common.c +2 -0
index 2c6602571c47..c16961e686a0 100644
@@ -21,6 +21,7 @@
 #include <linux/ptrace.h>
 #include <linux/slab.h>
 #include <linux/dma-map-ops.h>
+#include <linux/set_memory.h>
 #include <asm/hyperv-tlfs.h>
 #include <asm/mshyperv.h>
 
@@ -125,6 +126,7 @@ int hv_common_cpu_init(unsigned int cpu)
 	u64 msr_vp_index;
 	gfp_t flags;
 	int pgcount = hv_root_partition ? 2 : 1;
+	int ret;
 
 	/* hv_cpu_init() can be called with IRQs disabled from hv_resume() */
 	flags = irqs_disabled() ? GFP_ATOMIC : GFP_KERNEL;

drivers/hv/hv_common.c int hv_common_cpu_init(unsigned int cpu) · line 142

Quoted 7 lines
drivers/hv/hv_common.c +4 -0
@@ -134,6 +136,16 @@ int hv_common_cpu_init(unsigned int cpu)
 	if (!(*inputarg))
 		return -ENOMEM;
 
+	if (hv_isolation_type_en_snp()) {
+		ret = set_memory_decrypted((unsigned long)*inputarg, 1);
+		if (ret) {
+			kfree(*inputarg);

After the kfree(), set *inputarg back to NULL. There's other code that tests the value of *inputarg to know if the per-CPU hypercall page has been successfully allocated.

Quoted 21 lines
drivers/hv/hv_common.c +9 -1
+			return ret;
+		}
+
+		memset(*inputarg, 0x00, PAGE_SIZE);
+	}
+
 	if (hv_root_partition) {
 		outputarg = (void **)this_cpu_ptr(hyperv_pcpu_output_arg);
 		*outputarg = (char *)(*inputarg) + HV_HYP_PAGE_SIZE;
@@ -168,6 +180,9 @@ int hv_common_cpu_die(unsigned int cpu)
 
 	local_irq_restore(flags);
 
+	if (hv_isolation_type_en_snp())
+		set_memory_encrypted((unsigned long)mem, 1);
+
 	kfree(mem);
 
 	return 0;
--

2.25.1

#36 lore raw

RE: [RFC PATCH V2 11/18] Drivers: hv: vmbus: Decrypt vmbus ring buffer

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 4 lines

The ring buffer is remapped in the hv_ringbuffer_init() and it should be with decrypt flag in order to share it with hypervisor in sev-snp enlightened guest.

FWIW, the change in this patch is included in Patch 9 in my vTOM-related patch series.

Quoted 19 lines
---
 drivers/hv/ring_buffer.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/hv/ring_buffer.c +1 -0
index 59a4aa86d1f3..391995c76be7 100644
@@ -18,6 +18,7 @@
 #include <linux/slab.h>
 #include <linux/prefetch.h>
 #include <linux/io.h>
+#include <linux/set_memory.h>
 #include <asm/mshyperv.h>
 
 #include "hyperv_vmbus.h"

drivers/hv/ring_buffer.c int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info, · line 239

Quoted 6 lines
drivers/hv/ring_buffer.c +2 -0
@@ -233,14 +234,18 @@ int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,
 
 		ring_info->ring_buffer = (struct hv_ring_buffer *)
 			vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP,
+				hv_isolation_type_en_snp() ?
+				pgprot_decrypted(PAGE_KERNEL_NOENC) :
 				PAGE_KERNEL);

Note that pgprot_decrypted(PAGE_KERNEL_NOENC) can always be used, as it does nothing in a normal VM. Its use doesn't need to be conditional on the isolation type.

drivers/hv/ring_buffer.c int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info, · line 242

Quoted 3 lines
drivers/hv/ring_buffer.c +2 -0
+		if (hv_isolation_type_en_snp())
+			memset(ring_info->ring_buffer, 0x00, page_cnt * PAGE_SIZE);

My version of the change always does the zero'ing, but only of the first page, as Hyper-V expects the ring buffer header page to be clean. The initial contents of the rest of the ring buffer doesn't really matter.

drivers/hv/ring_buffer.c int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info, · line 243

Quoted 7 lines
drivers/hv/ring_buffer.c +1 -1
+
 		kfree(pages_wraparound);
 		if (!ring_info->ring_buffer)
 			return -ENOMEM;
 	}
 
-

The above looks like a spurious whitespace change.

Quoted 5 lines
drivers/hv/ring_buffer.c +0 -1
 	ring_info->ring_buffer->read_index =
 		ring_info->ring_buffer->write_index = 0;
 
--

2.25.1

author
#27 lore raw

Re: [RFC PATCH V2 08/18] x86/hyperv: decrypt vmbus pages for sev-snp enlightened guest

In reply to Michael Kelley (LINUX)
Quoted 4 lines · Michael Kelley (LINUX)

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 1 line

The Subject prefix for this patch should be "Drivers: hv: vmbus:"

Sure. Will update in the next version.

Quoted 33 lines
Quoted 15 lines

Vmbus int, synic and post message pages are shared with hypervisor and so decrypt these pages in the sev-snp guest.

---
  drivers/hv/connection.c | 13 +++++++++++++
  drivers/hv/hv.c         | 32 +++++++++++++++++++++++++++++++-
  2 files changed, 44 insertions(+), 1 deletion(-)
drivers/hv/connection.c +0 -0
index 9dc27e5d367a..43141225ea15 100644

drivers/hv/connection.c int vmbus_connect(void) · line 218

Quoted 4 lines
drivers/hv/connection.c +1 -0
@@ -215,6 +215,15 @@ int vmbus_connect(void)
  		(void *)((unsigned long)vmbus_connection.int_page +
  			(HV_HYP_PAGE_SIZE >> 1));
 
+	if (hv_isolation_type_snp() || hv_isolation_type_en_snp()) {

This decryption should be done only for a fully enlightened SEV-SNP guest, not for a vTOM guest.

drivers/hv/connection.c int vmbus_connect(void) · line 222

Quoted 4 lines
drivers/hv/connection.c +4 -0
+		ret = set_memory_decrypted((unsigned long)
+				vmbus_connection.int_page, 1);
+		if (ret)
+			goto cleanup;

This cleanup path doesn't work correctly. It calls vmbus_disconnect(), which will try to re-encrypt the memory. But if the original decryption failed, re-encrypting is the wrong thing to do.

Quoted 4 lines

It looks like this same bug exists in current code if the decryption of the monitor pages fails or if just one of the original memory allocations fails. vmbus_disconnect() doesn't know whether it should re-encrypt the pages.

Agree. It's necessary to handle decryption failure case by case in stead of re-encryting all pages. Will fix this in the next version. Thanks to point out.

author
#34 lore raw

Re: [RFC PATCH V2 10/18] drivers: hv: Decrypt percpu hvcall input arg page in sev-snp enlightened guest

In reply to Michael Kelley (LINUX)
Quoted 4 lines · Michael Kelley (LINUX)
Quoted 4 lines
drivers/hv/hv_common.c +0 -0
@@ -134,6 +136,16 @@ int hv_common_cpu_init(unsigned int cpu)
  	if (!(*inputarg))
  		return -ENOMEM;
 
Quoted 8 lines

drivers/hv/hv_common.c int hv_common_cpu_init(unsigned int cpu) · line 142

Quoted 4 lines
drivers/hv/hv_common.c +4 -0
+	if (hv_isolation_type_en_snp()) {
+		ret = set_memory_decrypted((unsigned long)*inputarg, 1);
+		if (ret) {
+			kfree(*inputarg);

After the kfree(), set *inputarg back to NULL. There's other code that tests the value of *inputarg to know if the per-CPU hypercall page has been successfully allocated.

Good point! Will add in the next version.

author
#37 lore raw

Re: [RFC PATCH V2 11/18] Drivers: hv: vmbus: Decrypt vmbus ring buffer

In reply to Michael Kelley (LINUX)
Quoted 7 lines · Michael Kelley (LINUX)

From: Tianyu Lan<[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 3 lines

The ring buffer is remapped in the hv_ringbuffer_init() and it should be with decrypt flag in order to share it with hypervisor in sev-snp enlightened guest.

FWIW, the change in this patch is included in Patch 9 in my vTOM-related patch series.

I will rebase next version on your series. Thanks for reminder.

#39 lore raw

RE: [RFC PATCH V2 12/18] x86/hyperv: Initialize cpu and memory for sev-snp enlightened guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 28 lines

Read processor amd memory info from specific address which are populated by Hyper-V. Initialize smp cpu related ops, pvalidate system memory and add it into e820 table.

---
 arch/x86/kernel/cpu/mshyperv.c | 75 ++++++++++++++++++++++++++++++++++
 1 file changed, 75 insertions(+)
arch/x86/kernel/cpu/mshyperv.c +6 -0
index 2ea4f21c6172..f0c97210c64a 100644
@@ -34,6 +34,12 @@
 #include <clocksource/hyperv_timer.h>
 #include <asm/numa.h>
 #include <asm/coco.h>
+#include <asm/io_apic.h>
+#include <asm/svm.h>
+#include <asm/sev.h>
+#include <asm/sev-snp.h>
+#include <asm/realmode.h>
+#include <asm/e820/api.h>
 
 /* Is Linux running as the root partition? */
 bool hv_root_partition;

arch/x86/kernel/cpu/mshyperv.c static void __init hv_smp_prepare_cpus(unsigned int max_cpus) · line 263

Quoted 5 lines
arch/x86/kernel/cpu/mshyperv.c +2 -0
@@ -253,6 +259,33 @@ static void __init hv_smp_prepare_cpus(unsigned int max_cpus)
 }
 #endif
 
+static __init int hv_snp_set_rtc_noop(const struct timespec64 *now) { return -EINVAL; }
+static __init void hv_snp_get_rtc_noop(struct timespec64 *now) { }

These two functions duplicate set_rtc_noop() and get_rtc_noop() in x86_init.c. Couldn't the "static" be removed from these two, and just use them, like with x86_init_noop()?
You'd also need to add extern statements in x86_init.h. But making them like the
other *_init_noop() functions seems better than duplicating them.

Also, it looks like these two functions might be called well after Linux is booted, so having them marked as "__init" seems problematic. The same is true for set_rtc_noop() and get_rtc_noop().

I'd suggesting breaking out this RTC handling into a separate patch in the series.

Quoted 41 lines
arch/x86/kernel/cpu/mshyperv.c +30 -0
+
+static u32 processor_count;
+
+static __init void hv_snp_get_smp_config(unsigned int early)
+{
+	if (!early) {
+		while (num_processors < processor_count) {
+			early_per_cpu(x86_cpu_to_apicid, num_processors) = num_processors;
+			early_per_cpu(x86_bios_cpu_apicid, num_processors) = num_processors;
+			physid_set(num_processors, phys_cpu_present_map);
+			set_cpu_possible(num_processors, true);
+			set_cpu_present(num_processors, true);
+			num_processors++;
+		}
+	}
+}
+
+struct memory_map_entry {
+	u64 starting_gpn;
+	u64 numpages;
+	u16 type;
+	u16 flags;
+	u32 reserved;
+};
+
 static void __init ms_hyperv_init_platform(void)
 {
 	int hv_max_functions_eax;
@@ -260,6 +293,11 @@ static void __init ms_hyperv_init_platform(void)
 	int hv_host_info_ebx;
 	int hv_host_info_ecx;
 	int hv_host_info_edx;
+	struct memory_map_entry *entry;
+	struct e820_entry *e820_entry;
+	u64 e820_end;
+	u64 ram_end;
+	u64 page;
 
 #ifdef CONFIG_PARAVIRT
 	pv_info.name = "Hyper-V";

arch/x86/kernel/cpu/mshyperv.c static void __init ms_hyperv_init_platform(void) · line 521

Quoted 7 lines
arch/x86/kernel/cpu/mshyperv.c +4 -0
@@ -477,6 +515,43 @@ static void __init ms_hyperv_init_platform(void)
 	if (!(ms_hyperv.features & HV_ACCESS_TSC_INVARIANT))
 		mark_tsc_unstable("running on Hyper-V");
 
+	if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
+		x86_platform.legacy.reserve_bios_regions = 0;
+		x86_platform.set_wallclock = hv_snp_set_rtc_noop;
+		x86_platform.get_wallclock = hv_snp_get_rtc_noop;

Should x86_platform.legacy.rtc be set to 0 as well so that add_rtc_cmos() does not try to register the RTC as a platform device?

arch/x86/kernel/cpu/mshyperv.c static void __init ms_hyperv_init_platform(void) · line 523

Quoted 2 lines
arch/x86/kernel/cpu/mshyperv.c +2 -0
+		x86_init.resources.probe_roms = x86_init_noop;
+		x86_init.resources.reserve_resources = x86_init_noop;

reserve_bios_regions, probe_roms, and reserve_resources are all being set to 0 or NULL. Seems like there should be a comment or text in the commit message saying why. I can work with you offline to write a concise message.

arch/x86/kernel/cpu/mshyperv.c static void __init ms_hyperv_init_platform(void) · line 534

Quoted 11 lines
arch/x86/kernel/cpu/mshyperv.c +11 -0
+		x86_init.mpparse.find_smp_config = x86_init_noop;
+		x86_init.mpparse.get_smp_config = hv_snp_get_smp_config;
+
+		/*
+		 * Hyper-V SEV-SNP enlightened guest doesn't support ioapic
+		 * and legacy APIC page read/write. Switch to hv apic here.
+		 */
+		disable_ioapic_support();
+		hv_apic_init();
+
+		processor_count = *(u32 *)__va(EN_SEV_SNP_PROCESSOR_INFO_ADDR);

EN_SEV_SNP_PROCESSOR_INFO_ADDR is not defined until Patch 13 of this series.

arch/x86/kernel/cpu/mshyperv.c static void __init ms_hyperv_init_platform(void) · line 537

Quoted 3 lines
arch/x86/kernel/cpu/mshyperv.c +3 -0
+
+		entry = (struct memory_map_entry *)(__va(EN_SEV_SNP_PROCESSOR_INFO_ADDR)
+				+ sizeof(struct memory_map_entry));

This calculation isn't what I expected. A brief summary of the layout of the memory at this fixed address would be helpful. Evidently, the first 32 bits is the processor count, and then there are multiple memory map entries. But why skip over an entire memory map entry to account for the 32-bit processor_count?

Maybe the definition of struct memory_map_entry and EN_SEV_SNP_PROCESSOR_INFO_ADDR should go in arch/x86/include/asm/mshyperv.h, along with some explanatory comments.

arch/x86/kernel/cpu/mshyperv.c static void __init ms_hyperv_init_platform(void) · line 546

Quoted 9 lines
arch/x86/kernel/cpu/mshyperv.c +9 -0
+
+		for (; entry->numpages != 0; entry++) {
+			e820_entry = &e820_table->entries[e820_table->nr_entries - 1];
+			e820_end = e820_entry->addr + e820_entry->size;
+			ram_end = (entry->starting_gpn + entry->numpages) * PAGE_SIZE;
+
+			if (e820_end < entry->starting_gpn * PAGE_SIZE)
+				e820_end = entry->starting_gpn * PAGE_SIZE;
+			if (e820_end < ram_end) {

I'm curious about any gaps or overlaps with the existing e820 map entries. What is expected? Or is this code just trying to be defensive?

arch/x86/kernel/cpu/mshyperv.c static void __init ms_hyperv_init_platform(void) · line 547

Quoted 1 line
arch/x86/kernel/cpu/mshyperv.c +1 -0
+				pr_info("Hyper-V: add [mem %#018Lx-%#018Lx]\n", e820_end, ram_end - 1);

Seems like the above message should include "e820" to make clear this is an update to the e820 map.

arch/x86/kernel/cpu/mshyperv.c static void __init ms_hyperv_init_platform(void) · line 552

Quoted 5 lines
arch/x86/kernel/cpu/mshyperv.c +5 -0
+				e820__range_add(e820_end, ram_end - e820_end, E820_TYPE_RAM);
+				for (page = e820_end; page < ram_end; page += PAGE_SIZE)
+					pvalidate((unsigned long)__va(page), RMP_PG_SIZE_4K, true);
+			}
+		}

Does e820__update_table() need to be called once any range adds are complete? I don't know, so I'm just asking.

Quoted 7 lines
arch/x86/kernel/cpu/mshyperv.c +2 -1
+	}
+
 	hardlockup_detector_disable();
 }
 
--

2.25.1

#41 lore raw

RE: [RFC PATCH V2 13/18] x86/hyperv: Add smp support for sev-snp guest

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Friday, November 18, 2022 7:46 PM

Quoted 43 lines

The wakeup_secondary_cpu callback was populated with wakeup_ cpu_via_vmgexit() which doesn't work for Hyper-V. Override it with Hyper-V specific hook which uses HVCALL_START_VIRTUAL_ PROCESSOR hvcall to start AP with vmsa data structure.

---
 arch/x86/include/asm/sev.h        |  13 +++
 arch/x86/include/asm/svm.h        |  55 ++++++++++-
 arch/x86/kernel/cpu/mshyperv.c    | 147 +++++++++++++++++++++++++++++-
 include/asm-generic/hyperv-tlfs.h |  18 ++++
 4 files changed, 230 insertions(+), 3 deletions(-)
arch/x86/include/asm/sev.h +13 -0
index ebc271bb6d8e..e34aaf730220 100644
@@ -86,6 +86,19 @@ extern bool handle_vc_boot_ghcb(struct pt_regs *regs);
 
 #define RMPADJUST_VMSA_PAGE_BIT		BIT(16)
 
+union sev_rmp_adjust {
+	u64 as_uint64;
+	struct {
+		unsigned long target_vmpl : 8;
+		unsigned long enable_read : 1;
+		unsigned long enable_write : 1;
+		unsigned long enable_user_execute : 1;
+		unsigned long enable_kernel_execute : 1;
+		unsigned long reserved1 : 4;
+		unsigned long vmsa : 1;
+	};
+};
+
 /* SNP Guest message request */
 struct snp_req_data {
 	unsigned long req_gpa;
arch/x86/include/asm/svm.h +0 -0
index 0361626841bc..fc54d3e7f817 100644

arch/x86/include/asm/svm.h struct vmcb_save_area { · line 336

Quoted 10 lines
arch/x86/include/asm/svm.h +6 -1
@@ -328,8 +328,61 @@ struct vmcb_save_area {
 	u64 br_to;
 	u64 last_excp_from;
 	u64 last_excp_to;
-	u8 reserved_6[72];
+
+	/*
+	 * The following part of the save area is valid only for
+	 * SEV-ES guests when referenced through the GHCB or for
+	 * saving to the host save area.
+	 */

It seems unexpected to add these SEV-ES specific fields to a structure with a comment that says for legacy and SEV-MEM guests. There's already a struct sev_es_save_area with a comment that says for SEV-ES and SEV_SNP guests, and that struct seems to have most or all of what is being added here. Hopefully there's a way to use struct sev_es_save_area, perhaps with some minor tweaks if necessary.

Quoted 57 lines
arch/x86/include/asm/svm.h +48 -0
+	u8 reserved_7[72];
 	u32 spec_ctrl;		/* Guest version of SPEC_CTRL at 0x2E0 */
+	u8 reserved_7b[4];
+	u32 pkru;
+	u8 reserved_7a[20];
+	u64 reserved_8;		/* rax already available at 0x01f8 */
+	u64 rcx;
+	u64 rdx;
+	u64 rbx;
+	u64 reserved_9;		/* rsp already available at 0x01d8 */
+	u64 rbp;
+	u64 rsi;
+	u64 rdi;
+	u64 r8;
+	u64 r9;
+	u64 r10;
+	u64 r11;
+	u64 r12;
+	u64 r13;
+	u64 r14;
+	u64 r15;
+	u8 reserved_10[16];
+	u64 sw_exit_code;
+	u64 sw_exit_info_1;
+	u64 sw_exit_info_2;
+	u64 sw_scratch;
+	union {
+		u64 sev_features;
+		struct {
+			u64 sev_feature_snp			: 1;
+			u64 sev_feature_vtom			: 1;
+			u64 sev_feature_reflectvc		: 1;
+			u64 sev_feature_restrict_injection	: 1;
+			u64 sev_feature_alternate_injection	: 1;
+			u64 sev_feature_full_debug		: 1;
+			u64 sev_feature_reserved1		: 1;
+			u64 sev_feature_snpbtb_isolation	: 1;
+			u64 sev_feature_resrved2		: 56;
+		};
+	};
+	u64 vintr_ctrl;
+	u64 guest_error_code;
+	u64 virtual_tom;
+	u64 tlb_id;
+	u64 pcpu_id;
+	u64 event_inject;
+	u64 xcr0;
+	u8 valid_bitmap[16];
+	u64 x87_state_gpa;
 } __packed;
 
 /* Save area definition for SEV-ES and SEV-SNP guests */
arch/x86/kernel/cpu/mshyperv.c +0 -0
index f0c97210c64a..b266f648e5cd 100644

arch/x86/kernel/cpu/mshyperv.c line 46

Quoted 6 lines
arch/x86/kernel/cpu/mshyperv.c +3 -0
@@ -41,6 +41,10 @@
 #include <asm/realmode.h>
 #include <asm/e820/api.h>
 
+#define EN_SEV_SNP_PROCESSOR_INFO_ADDR	 0x802000
+#define HV_AP_INIT_GPAT_DEFAULT		0x0007040600070406ULL
+#define HV_AP_SEGMENT_LIMIT		0xffffffff

The above three definitions would benefit from some comments explaining what they are.

Quoted 14 lines
arch/x86/kernel/cpu/mshyperv.c +7 -0
+
 /* Is Linux running as the root partition? */
 bool hv_root_partition;
 struct ms_hyperv_info ms_hyperv;
@@ -232,6 +236,136 @@ static void __init hv_smp_prepare_boot_cpu(void)
 #endif
 }
 
+static u8 ap_start_input_arg[PAGE_SIZE] __bss_decrypted __aligned(PAGE_SIZE);
+static u8 ap_start_stack[PAGE_SIZE] __aligned(PAGE_SIZE);
+
+int hv_snp_boot_ap(int cpu, unsigned long start_ip)
+{
+	struct vmcb_save_area *vmsa = (struct vmcb_save_area *)

arch/x86/kernel/cpu/mshyperv.c static void __init hv_smp_prepare_boot_cpu(void) · line 254

Quoted 10 lines
arch/x86/kernel/cpu/mshyperv.c +10 -0
+		__get_free_page(GFP_KERNEL | __GFP_ZERO);
+	struct desc_ptr gdtr;
+	u64 ret, retry = 5;
+	struct hv_enable_vp_vtl_input *enable_vtl_input;
+	struct hv_start_virtual_processor_input *start_vp_input;
+	union sev_rmp_adjust rmp_adjust;
+	void **arg;
+	unsigned long flags;
+
+	*(void **)per_cpu_ptr(hyperv_pcpu_input_arg, cpu) = ap_start_input_arg;

I don't understand the above. It seems like the hyperv_pcpu_input_arg is being set to the same static location for all APs. The static location gets overwritten in hv_common_cpu_init(), so maybe everything works. But it seems like ap_start_input_arg can just be used directly in this function without having to update hyperv_pcpu_input_arg.

arch/x86/kernel/cpu/mshyperv.c static void __init hv_smp_prepare_boot_cpu(void) · line 256

Quoted 2 lines
arch/x86/kernel/cpu/mshyperv.c +2 -0
+
+	hv_vp_index[cpu] = cpu;

The hv_vp_index[cpu] is also updated in hv_common_cpu_init(). Is there a reason to initialize the value here? This code also assumes that Linux CPU numbers and Hyper-V VP indices are the same. I've always observed that they are indeed the same, but Hyper-V doesn't guarantee that. Hence we set the value in hv_common_cpu_init() based on reading the per-CPU synthetic register that contains the VP index.

arch/x86/kernel/cpu/mshyperv.c static void __init hv_smp_prepare_boot_cpu(void) · line 259

Quoted 3 lines
arch/x86/kernel/cpu/mshyperv.c +3 -0
+
+	/* Prevent APs from entering busy calibration loop */
+	preset_lpj = lpj_fine;

I wonder if this is really needed. In a SEV-SNP guest that isn't running on Hyper-V, how is this handled?

arch/x86/kernel/cpu/mshyperv.c static void __init hv_smp_prepare_boot_cpu(void) · line 262

Quoted 3 lines
arch/x86/kernel/cpu/mshyperv.c +3 -0
+
+	/* Replace the provided real-mode start_ip */
+	start_ip = (unsigned long)secondary_startup_64_no_verify;

Any reason to update this global value? The starting IP is passed to Hyper-V via the VMSA, so it doesn't seem like a global update is needed.

arch/x86/kernel/cpu/mshyperv.c static void __init hv_smp_prepare_boot_cpu(void) · line 275

Quoted 13 lines
arch/x86/kernel/cpu/mshyperv.c +13 -0
+
+	native_store_gdt(&gdtr);
+
+	vmsa->gdtr.base = gdtr.address;
+	vmsa->gdtr.limit = gdtr.size;
+
+	asm volatile("movl %%es, %%eax;" : "=a" (vmsa->es.selector));
+	if (vmsa->es.selector) {
+		vmsa->es.base = 0;
+		vmsa->es.limit = HV_AP_SEGMENT_LIMIT;
+		vmsa->es.attrib = *(u16 *)(vmsa->gdtr.base + vmsa->es.selector + 5);
+		vmsa->es.attrib = (vmsa->es.attrib & 0xFF) | ((vmsa->es.attrib >> 4) & 0xF00);
+	}

The above "if" statement is repeated four times with different registers. Seems like a helper function could easily encapsulate it, though not the "asm volatile" statement.

Quoted 45 lines
arch/x86/kernel/cpu/mshyperv.c +45 -0
+
+	asm volatile("movl %%cs, %%eax;" : "=a" (vmsa->cs.selector));
+	if (vmsa->cs.selector) {
+		vmsa->cs.base = 0;
+		vmsa->cs.limit = HV_AP_SEGMENT_LIMIT;
+		vmsa->cs.attrib = *(u16 *)(vmsa->gdtr.base + vmsa->cs.selector + 5);
+		vmsa->cs.attrib = (vmsa->cs.attrib & 0xFF) | ((vmsa->cs.attrib >> 4) & 0xF00);
+	}
+
+	asm volatile("movl %%ss, %%eax;" : "=a" (vmsa->ss.selector));
+	if (vmsa->ss.selector) {
+		vmsa->ss.base = 0;
+		vmsa->ss.limit = HV_AP_SEGMENT_LIMIT;
+		vmsa->ss.attrib = *(u16 *)(vmsa->gdtr.base + vmsa->ss.selector + 5);
+		vmsa->ss.attrib = (vmsa->ss.attrib & 0xFF) | ((vmsa->ss.attrib >> 4) & 0xF00);
+	}
+
+	asm volatile("movl %%ds, %%eax;" : "=a" (vmsa->ds.selector));
+	if (vmsa->ds.selector) {
+		vmsa->ds.base = 0;
+		vmsa->ds.limit = HV_AP_SEGMENT_LIMIT;
+		vmsa->ds.attrib = *(u16 *)(vmsa->gdtr.base + vmsa->ds.selector + 5);
+		vmsa->ds.attrib = (vmsa->ds.attrib & 0xFF) | ((vmsa->ds.attrib >> 4) & 0xF00);
+	}
+
+	vmsa->efer = native_read_msr(MSR_EFER);
+
+	asm volatile("movq %%cr4, %%rax;" : "=a" (vmsa->cr4));
+	asm volatile("movq %%cr3, %%rax;" : "=a" (vmsa->cr3));
+	asm volatile("movq %%cr0, %%rax;" : "=a" (vmsa->cr0));
+
+	vmsa->xcr0 = 1;
+	vmsa->g_pat = HV_AP_INIT_GPAT_DEFAULT;
+	vmsa->rip = (u64)start_ip;
+	vmsa->rsp = (u64)&ap_start_stack[PAGE_SIZE];
+
+	vmsa->sev_feature_snp = 1;
+	vmsa->sev_feature_restrict_injection = 1;
+
+	rmp_adjust.as_uint64 = 0;
+	rmp_adjust.target_vmpl = 1;
+	rmp_adjust.vmsa = 1;
+	ret = rmpadjust((unsigned long)vmsa, RMP_PG_SIZE_4K,
+			rmp_adjust.as_uint64);
+	if (ret != 0) {

arch/x86/kernel/cpu/mshyperv.c static void __init hv_smp_prepare_boot_cpu(void) · line 330

Quoted 10 lines
arch/x86/kernel/cpu/mshyperv.c +10 -0
+		pr_err("RMPADJUST(%llx) failed: %llx\n", (u64)vmsa, ret);
+		return ret;
+	}
+
+	local_irq_save(flags);
+	arg = (void **)this_cpu_ptr(hyperv_pcpu_input_arg);
+	if (unlikely(!*arg)) {
+		ret = -ENOMEM;
+		goto done;
+	}

This code seems unnecessary. Just use ap_start_input_arg directly. No need to disable interrupts.

arch/x86/kernel/cpu/mshyperv.c static void __init hv_smp_prepare_boot_cpu(void) · line 341

Quoted 11 lines
arch/x86/kernel/cpu/mshyperv.c +11 -0
+
+	if (ms_hyperv.vtl != 0) {
+		enable_vtl_input = (struct hv_enable_vp_vtl_input *)*arg;
+		memset(enable_vtl_input, 0, sizeof(*enable_vtl_input));
+		enable_vtl_input->partitionid = -1;
+		enable_vtl_input->vpindex = cpu;
+		enable_vtl_input->targetvtl = ms_hyperv.vtl;
+		*(u64 *)&enable_vtl_input->context[0] = __pa(vmsa) | 1;
+
+		ret = hv_do_hypercall(HVCALL_ENABLE_VP_VTL, enable_vtl_input, NULL);
+		if (ret != 0) {

Use hv_result_success() to test the hypercall result.

Quoted 6 lines
arch/x86/kernel/cpu/mshyperv.c +6 -0
+			pr_err("HvCallEnableVpVtl failed: %llx\n", ret);
+			goto done;
+		}
+	}
+
+	start_vp_input = (struct hv_start_virtual_processor_input *)*arg;

arch/x86/kernel/cpu/mshyperv.c static void __init hv_smp_prepare_boot_cpu(void) · line 357

Quoted 10 lines
arch/x86/kernel/cpu/mshyperv.c +10 -0
+	memset(start_vp_input, 0, sizeof(*start_vp_input));
+	start_vp_input->partitionid = -1;
+	start_vp_input->vpindex = cpu;
+	start_vp_input->targetvtl = ms_hyperv.vtl;
+	*(u64 *)&start_vp_input->context[0] = __pa(vmsa) | 1;
+
+	do {
+		ret = hv_do_hypercall(HVCALL_START_VIRTUAL_PROCESSOR,
+				      start_vp_input, NULL);
+	} while (ret == HV_STATUS_TIME_OUT && retry--);

Use hv_result() to check for HV_STATUS_TIME_OUT.

arch/x86/kernel/cpu/mshyperv.c static void __init hv_smp_prepare_boot_cpu(void) · line 359

Quoted 2 lines
arch/x86/kernel/cpu/mshyperv.c +2 -0
+
+	if (ret != 0) {

Use hv_result_success().

arch/x86/kernel/cpu/mshyperv.c static void __init hv_smp_prepare_boot_cpu(void) · line 365

Quoted 6 lines
arch/x86/kernel/cpu/mshyperv.c +6 -0
+		pr_err("HvCallStartVirtualProcessor failed: %llx\n", ret);
+		goto done;
+	}
+
+done:
+	local_irq_restore(flags);

The entry to this function allocates a page for the VMSA. Does that page ever get freed?

Quoted 54 lines
arch/x86/kernel/cpu/mshyperv.c +14 -2
+	return ret;
+}
+
 static void __init hv_smp_prepare_cpus(unsigned int max_cpus)
 {
 #ifdef CONFIG_X86_64
@@ -241,6 +375,16 @@ static void __init hv_smp_prepare_cpus(unsigned int max_cpus)
 
 	native_smp_prepare_cpus(max_cpus);
 
+	/*
+	 *  Override wakeup_secondary_cpu callback for SEV-SNP
+	 *  enlightened guest.
+	 */
+	if (hv_isolation_type_en_snp())
+		apic->wakeup_secondary_cpu = hv_snp_boot_ap;
+
+	if (!hv_root_partition)
+		return;
+
 #ifdef CONFIG_X86_64
 	for_each_present_cpu(i) {
 		if (i == 0)
@@ -489,8 +633,7 @@ static void __init ms_hyperv_init_platform(void)
 
 # ifdef CONFIG_SMP
 	smp_ops.smp_prepare_boot_cpu = hv_smp_prepare_boot_cpu;
-	if (hv_root_partition)
-		smp_ops.smp_prepare_cpus = hv_smp_prepare_cpus;
+	smp_ops.smp_prepare_cpus = hv_smp_prepare_cpus;
 # endif
 
 	/*
include/asm-generic/hyperv-tlfs.h +2 -0
index 6e2a090e2649..7072adbf5540 100644
@@ -139,6 +139,7 @@ struct ms_hyperv_tsc_page {
 #define HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST	0x0003
 #define HVCALL_NOTIFY_LONG_SPIN_WAIT		0x0008
 #define HVCALL_SEND_IPI				0x000b
+#define HVCALL_ENABLE_VP_VTL			0x000f
 #define HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX	0x0013
 #define HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX	0x0014
 #define HVCALL_SEND_IPI_EX			0x0015
@@ -156,6 +157,7 @@ struct ms_hyperv_tsc_page {
 #define HVCALL_MAP_DEVICE_INTERRUPT		0x007c
 #define HVCALL_UNMAP_DEVICE_INTERRUPT		0x007d
 #define HVCALL_RETARGET_INTERRUPT		0x007e
+#define HVCALL_START_VIRTUAL_PROCESSOR		0x0099
 #define HVCALL_FLUSH_GUEST_PHYSICAL_ADDRESS_SPACE 0x00af
 #define HVCALL_FLUSH_GUEST_PHYSICAL_ADDRESS_LIST 0x00b0
 #define HVCALL_MODIFY_SPARSE_GPA_PAGE_HOST_VISIBILITY 0x00db

include/asm-generic/hyperv-tlfs.h struct hv_input_unmap_device_interrupt { · line 773

Quoted 9 lines
include/asm-generic/hyperv-tlfs.h +6 -0
@@ -763,6 +765,22 @@ struct hv_input_unmap_device_interrupt {
 	struct hv_interrupt_entry interrupt_entry;
 } __packed;
 
+struct hv_enable_vp_vtl_input {
+	u64 partitionid;
+	u32 vpindex;
+	u8 targetvtl;
+	u8 padding[3];
+	u8 context[0xe0];

It looks like the 0xe0 comes from the Hyper-V TLFS, but your code is doing something different -- it's setting the VMSA address instead of putting the context values inline.

include/asm-generic/hyperv-tlfs.h struct hv_input_unmap_device_interrupt { · line 781

Quoted 8 lines
include/asm-generic/hyperv-tlfs.h +8 -0
+} __packed;
+
+struct hv_start_virtual_processor_input {
+	u64 partitionid;
+	u32 vpindex;
+	u8 targetvtl;
+	u8 padding[3];
+	u8 context[0xe0];

Same here.

Quoted 7 lines
include/asm-generic/hyperv-tlfs.h +2 -1
+} __packed;
+
 #define HV_SOURCE_SHADOW_NONE               0x0
 #define HV_SOURCE_SHADOW_BRIDGE_BUS_RANGE   0x1
 
--

2.25.1

#5 lore raw

RE: [RFC PATCH V2 01/18] x86/sev: Pvalidate memory gab for decompressing kernel

In reply to Tianyu Lan

From: Tianyu Lan <[email protected]> Sent: Tuesday, November 29, 2022 6:43 AM

Quoted 19 lines · Borislav Petkov
Quoted 17 lines

arch/x86/boot/compressed/sev.c static enum es_result vc_read_mem(struct es_em_ctxt *ctxt, · line 124

Quoted 2 lines
arch/x86/boot/compressed/sev.c +2 -0
+/* Check SEV-SNP via MSR */
+static bool sev_snp_runtime_check(void)

Functions need to have a verb in the name.

arch/x86/boot/compressed/sev.c static enum es_result vc_read_mem(struct es_em_ctxt *ctxt, · line 134

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+{
+	unsigned long low, high;
+	u64 val;
+
+	asm volatile("rdmsr\n" : "=a" (low), "=d" (high) :
+			"c" (MSR_AMD64_SEV));
+
+	val = (high << 32) | low;
+	if (val & MSR_AMD64_SEV_SNP_ENABLED)
+		return true;

There already is a sev_snp_enabled() in that very same file. Did you not see it?

Quoted 7 lines
Quoted 1 line

Why are you even adding such a function?!

Hi Boris:
Thanks for your review. sev_snp_enabled() is used after
sev_status was initialized in sev_enable() while pvalidate_for_startup_ 64() is called before sev_enable(). So add sev_snp_runtime_check() to check sev snp capability before calling sev_enable().

I understand needing to find out if SEV-SNP is enabled before sev_enable() has set the value of sev_status. Unfortunately, just checking the SEV_SNP_ENABLED bit in MSR_AMD64_SEV isn't sufficient. sev_enable() correctly does a more complex check. The CPUID bit is checked first, and the MSR is read only if the CPUID bit indicating SEV is set. The reason for this behavior is described in the commit message for 009767dbf42a. Furthermore, even if the MSR is safe to read, just checking the SEV_SNP_ENABLED bit isn't sufficient because that bit is "1" in a vTOM VM. The MSR check would need to be that SEV_SNP_ENABLED is set, and VTOM_ENABLED is not set.

Michael

Quoted 134 lines
Quoted 119 lines
Quoted 33 lines
arch/x86/boot/compressed/sev.c +24 -0
+	return false;
+}
+
  static inline bool sev_snp_enabled(void)
  {
  	return sev_status & MSR_AMD64_SEV_SNP_ENABLED;
@@ -456,3 +475,68 @@ void sev_prep_identity_maps(unsigned long top_level_pgt)
 
  	sev_verify_cbit(top_level_pgt);
  }
+
+static void extend_e820_on_demand(struct boot_e820_entry *e820_entry,
+				  u64 needed_ram_end)
+{
+	u64 end, paddr;
+	unsigned long eflags;
+	int rc;
+
+	if (!e820_entry)
+		return;
+
+	/* Validated memory must be aligned by PAGE_SIZE. */
+	end = ALIGN(e820_entry->addr + e820_entry->size, PAGE_SIZE);
+	if (needed_ram_end > end && e820_entry->type == E820_TYPE_RAM) {
+		for (paddr = end; paddr < needed_ram_end; paddr += PAGE_SIZE) {
+			rc = pvalidate(paddr, RMP_PG_SIZE_4K, true);
+			if (rc) {
+				error("Failed to validate address.n");
+				return;
+			}
+		}

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 508

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+		e820_entry->size = needed_ram_end - e820_entry->addr;
+	}
+}
+
+/*
+ * Explicitly pvalidate needed pages for decompressing the kernel.
+ * The E820_TYPE_RAM entry includes only validated memory. The kernel
+ * expects that the RAM entry's addr is fixed while the entry size is to be
+ * extended to cover addresses to the start of next entry.
+ * The function increases the RAM entry size to cover all possible memory

Similar issue as above: you don't need to say "this function" above this function. IOW, it should say:

"Increase the RAM entry size..."

I.e., imperative mood above.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 516

Quoted 8 lines
arch/x86/boot/compressed/sev.c +8 -0
+ * addresses until init_size.
+ * For example,  init_end = 0x4000000,
+ * [RAM: 0x0 - 0x0],                       M[RAM: 0x0 - 0xa0000]
+ * [RSVD: 0xa0000 - 0x10000]                [RSVD: 0xa0000 - 0x10000]
+ * [ACPI: 0x10000 - 0x20000]      ==>       [ACPI: 0x10000 - 0x20000]
+ * [RSVD: 0x800000 - 0x900000]              [RSVD: 0x800000 - 0x900000]
+ * [RAM: 0x1000000 - 0x2000000]            M[RAM: 0x1000000 - 0x2001000]
+ * [RAM: 0x2001000 - 0x2007000]            M[RAM: 0x2001000 - 0x4000000]

What is this trying to tell me?

That the end range 0x2007000 gets raised to 0x4000000?

Why?

This all sounds like there is some requirement somewhere but nothing says what that requirement is and why.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 519

Quoted 3 lines
arch/x86/boot/compressed/sev.c +3 -0
+ * Other RAM memory after init_end is pvalidated by ms_hyperv_init_platform
+ */
+__visible void pvalidate_for_startup_64(struct boot_params *boot_params)

This doesn't do any validation. And yet it has "pvalidate" in the name.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 523

Quoted 4 lines
arch/x86/boot/compressed/sev.c +4 -0
+{
+	struct boot_e820_entry *e820_entry;
+	u64 init_end =
+		boot_params->hdr.pref_address + boot_params->hdr.init_size;

Nope, we never break lines like that.

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 525

Quoted 2 lines
arch/x86/boot/compressed/sev.c +2 -0
+	u8 i, nr_entries = boot_params->e820_entries;
+	u64 needed_end;

The tip-tree preferred ordering of variable declarations at the beginning of a function is reverse fir tree order::

	struct long_struct_name *descriptive_name;
	unsigned long foo, bar;
	unsigned int tmp;
	int ret;

The above is faster to parse than the reverse ordering::

	int ret;
	unsigned int tmp;
	unsigned long foo, bar;
	struct long_struct_name *descriptive_name;

And even more so than random ordering::

	unsigned long foo, bar;
	int ret;
	struct long_struct_name *descriptive_name;
	unsigned int tmp;
Quoted 4 lines
arch/x86/boot/compressed/sev.c +4 -0
+	if (!sev_snp_runtime_check())
+		return;
+
+	for (i = 0; i < nr_entries; ++i) {

arch/x86/boot/compressed/sev.c void sev_prep_identity_maps(unsigned long top_level_pgt) · line 540

Quoted 10 lines
arch/x86/boot/compressed/sev.c +10 -0
+		/* Pvalidate memory holes in e820 RAM entries. */
+		e820_entry = &boot_params->e820_table[i];
+		if (i < nr_entries - 1) {
+			needed_end = boot_params->e820_table[i + 1].addr;
+			if (needed_end < e820_entry->addr)
+				error("e820 table is not sorted.\n");
+		} else {
+			needed_end = init_end;
+		}
+		extend_e820_on_demand(e820_entry, needed_end);

Now*this* function does call pvalidate() and yet it doesn't have "pvalidate" in the name. This all looks real confused.

Quoted 6 lines

So first of all, you need to explain*why* you're doing this.

It looks like it is because the guest needs to do the memory validation by itself because nobody else does that.

If so, this needs to be explained in detail in the commit message.

Yes, I will update in the next version. Thanks for suggestion.

Quoted 3 lines

Also, why is that ok for SNP guests on other hypervisors which get the memory validated by the boot loader or firmware?

This is for Linux direct boot mode and so it needs to do such check here. Will fix this in the next version.

maintainer
#21 lore raw

RE: [RFC PATCH V2 06/18] x86/hyperv: Use vmmcall to implement hvcall in sev-snp enlightened guest

In reply to Tianyu Lan
Quoted 8 lines

From: Tianyu Lan <[email protected]>
Sent: Friday, November 18, 2022 7:46 PM

arch/x86/include/asm/mshyperv.h +0 -0
@@ -45,16 +45,25 @@ static inline u64 hv_do_hypercall(u64 control, void
 *input, void *output)
 	u64 hv_status;
 

arch/x86/include/asm/mshyperv.h static inline u64 hv_do_hypercall(u64 control, void · line 55

Quoted 10 lines
arch/x86/include/asm/mshyperv.h +7 -2
 #ifdef CONFIG_X86_64
-	if (!hv_hypercall_pg)
-		return U64_MAX;
+	if (hv_isolation_type_en_snp()) {
+		__asm__ __volatile__("mov %4, %%r8\n"
+				"vmmcall"
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input_address)
+				:  "r" (output_address)
+				: "cc", "memory", "r8", "r9", "r10", "r11");

Add a "return hv_status;" here to avoid chaning the indentation below?

Quoted 20 lines
arch/x86/include/asm/mshyperv.h +11 -7
+	} else {
+		if (!hv_hypercall_pg)
+			return U64_MAX;
 
-	__asm__ __volatile__("mov %4, %%r8\n"
-			     CALL_NOSPEC
-			     : "=a" (hv_status), ASM_CALL_CONSTRAINT,
-			       "+c" (control), "+d" (input_address)
-			     :  "r" (output_address),
-				THUNK_TARGET(hv_hypercall_pg)
-			     : "cc", "memory", "r8", "r9", "r10", "r11");
+		__asm__ __volatile__("mov %4, %%r8\n"
+				CALL_NOSPEC
+				: "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				"+c" (control), "+d" (input_address)
+				:  "r" (output_address),
+					THUNK_TARGET(hv_hypercall_pg)
+				: "cc", "memory", "r8", "r9", "r10", "r11");
+	}
 #else
#44 lore raw

Re: [RFC PATCH V2 15/18] x86/sev: Add a #HV exception handler

In reply to Tianyu Lan
Quoted 39 lines

Add a #HV exception handler that uses IST stack.

---
  arch/x86/entry/entry_64.S             | 58 +++++++++++++++++++++++++++
  arch/x86/include/asm/cpu_entry_area.h |  6 +++
  arch/x86/include/asm/idtentry.h       | 39 +++++++++++++++++-
  arch/x86/include/asm/page_64_types.h  |  1 +
  arch/x86/include/asm/trapnr.h         |  1 +
  arch/x86/include/asm/traps.h          |  1 +
  arch/x86/kernel/cpu/common.c          |  1 +
  arch/x86/kernel/dumpstack_64.c        |  9 ++++-
  arch/x86/kernel/idt.c                 |  1 +
  arch/x86/kernel/sev.c                 | 53 ++++++++++++++++++++++++
  arch/x86/mm/cpu_entry_area.c          |  2 +
  11 files changed, 169 insertions(+), 3 deletions(-)
arch/x86/entry/entry_64.S +13 -0
index 9953d966d124..b2059df43c57 100644
@@ -560,6 +560,64 @@ SYM_CODE_START(\asmsym)
  .Lfrom_usermode_switch_stack_\@:
  	idtentry_body user_\cfunc, has_error_code=1
 
+_ASM_NOKPROBE(\asmsym)
+SYM_CODE_END(\asmsym)
+.endm
+/*
+ * idtentry_hv - Macro to generate entry stub for #HV
+ * @vector:		Vector number
+ * @asmsym:		ASM symbol for the entry point
+ * @cfunc:		C function to be called
+ *
+ * The macro emits code to set up the kernel context for #HV. The #HV handler
+ * runs on an IST stack and needs to be able to support nested #HV exceptions.
+ *
+ * To make this work the #HV entry code tries its best to pretend it doesn't use

arch/x86/entry/entry_64.S SYM_CODE_START(\asmsym) · line 585

Quoted 10 lines
arch/x86/entry/entry_64.S +10 -0
+ * an IST stack by switching to the task stack if coming from user-space (which
+ * includes early SYSCALL entry path) or back to the stack in the IRET frame if
+ * entered from kernel-mode.
+ *
+ * If entered from kernel-mode the return stack is validated first, and if it is
+ * not safe to use (e.g. because it points to the entry stack) the #HV handler
+ * will switch to a fall-back stack (HV2) and call a special handler function.
+ *
+ * The macro is only used for one vector, but it is planned to be extended in
+ * the future for the #HV exception.

Noticed same comment line in the #VC exception handling section (macro idtentry_vc). Just wondering we need to remove this line as the patch being proposed for the #HV exception handling? unless this is planned to be extended in some other way.

Thanks,
Pankaj

Quoted 310 lines
arch/x86/entry/entry_64.S +35 -0
+ */
+.macro idtentry_hv vector asmsym cfunc
+SYM_CODE_START(\asmsym)
+	UNWIND_HINT_IRET_REGS
+	ASM_CLAC
+	pushq	$-1			/* ORIG_RAX: no syscall to restart */
+
+	testb	$3, CS-ORIG_RAX(%rsp)
+	jnz	.Lfrom_usermode_switch_stack_\@
+
+	call	paranoid_entry
+
+	UNWIND_HINT_REGS
+
+	/*
+	 * Switch off the IST stack to make it free for nested exceptions.
+	 */
+	movq	%rsp, %rdi		/* pt_regs pointer */
+	call	hv_switch_off_ist
+	movq	%rax, %rsp		/* Switch to new stack */
+
+	UNWIND_HINT_REGS
+
+	/* Update pt_regs */
+	movq	ORIG_RAX(%rsp), %rsi	/* get error code into 2nd argument*/
+	movq	$-1, ORIG_RAX(%rsp)	/* no syscall to restart */
+
+	movq	%rsp, %rdi		/* pt_regs pointer */
+	call	kernel_\cfunc
+
+	jmp	paranoid_exit
+
+.Lfrom_usermode_switch_stack_\@:
+	idtentry_body user_\cfunc, has_error_code=1
+
  _ASM_NOKPROBE(\asmsym)
  SYM_CODE_END(\asmsym)
  .endm
arch/x86/include/asm/cpu_entry_area.h +6 -0
index 75efc4c6f076..f173a16cfc59 100644
@@ -30,6 +30,10 @@
  	char	VC_stack[optional_stack_size];			\
  	char	VC2_stack_guard[guardsize];			\
  	char	VC2_stack[optional_stack_size];			\
+	char	HV_stack_guard[guardsize];			\
+	char	HV_stack[optional_stack_size];			\
+	char	HV2_stack_guard[guardsize];			\
+	char	HV2_stack[optional_stack_size];			\
  	char	IST_top_guard[guardsize];			\
 
  /* The exception stacks' physical storage. No guard pages required */
@@ -52,6 +56,8 @@ enum exception_stack_ordering {
  	ESTACK_MCE,
  	ESTACK_VC,
  	ESTACK_VC2,
+	ESTACK_HV,
+	ESTACK_HV2,
  	N_EXCEPTION_STACKS
  };
 
arch/x86/include/asm/idtentry.h +38 -1
index 72184b0b2219..652fea10d377 100644
@@ -317,6 +317,19 @@ static __always_inline void __##func(struct pt_regs *regs)
  	__visible noinstr void kernel_##func(struct pt_regs *regs, unsigned long error_code);	\
  	__visible noinstr void   user_##func(struct pt_regs *regs, unsigned long error_code)
 
+
+/**
+ * DECLARE_IDTENTRY_HV - Declare functions for the HV entry point
+ * @vector:	Vector number (ignored for C)
+ * @func:	Function name of the entry point
+ *
+ * Maps to DECLARE_IDTENTRY_RAW, but declares also the user C handler.
+ */
+#define DECLARE_IDTENTRY_HV(vector, func)				\
+	DECLARE_IDTENTRY_RAW_ERRORCODE(vector, func);			\
+	__visible noinstr void kernel_##func(struct pt_regs *regs);	\
+	__visible noinstr void   user_##func(struct pt_regs *regs)
+
  /**
   * DEFINE_IDTENTRY_IST - Emit code for IST entry points
   * @func:	Function name of the entry point
@@ -376,6 +389,26 @@ static __always_inline void __##func(struct pt_regs *regs)
  #define DEFINE_IDTENTRY_VC_USER(func)				\
  	DEFINE_IDTENTRY_RAW_ERRORCODE(user_##func)
 
+/**
+ * DEFINE_IDTENTRY_HV_KERNEL - Emit code for HV injection handler
+ *			       when raised from kernel mode
+ * @func:	Function name of the entry point
+ *
+ * Maps to DEFINE_IDTENTRY_RAW
+ */
+#define DEFINE_IDTENTRY_HV_KERNEL(func)					\
+	DEFINE_IDTENTRY_RAW(kernel_##func)
+
+/**
+ * DEFINE_IDTENTRY_HV_USER - Emit code for HV injection handler
+ *			     when raised from user mode
+ * @func:	Function name of the entry point
+ *
+ * Maps to DEFINE_IDTENTRY_RAW
+ */
+#define DEFINE_IDTENTRY_HV_USER(func)					\
+	DEFINE_IDTENTRY_RAW(user_##func)
+
  #else	/* CONFIG_X86_64 */
 
  /**
@@ -465,6 +498,9 @@ __visible noinstr void func(struct pt_regs *regs,			\
  # define DECLARE_IDTENTRY_VC(vector, func)				\
  	idtentry_vc vector asm_##func func
 
+# define DECLARE_IDTENTRY_HV(vector, func)				\
+	idtentry_hv vector asm_##func func
+
  #else
  # define DECLARE_IDTENTRY_MCE(vector, func)				\
  	DECLARE_IDTENTRY(vector, func)
@@ -622,9 +658,10 @@ DECLARE_IDTENTRY_RAW_ERRORCODE(X86_TRAP_DF,	xenpv_exc_double_fault);
  DECLARE_IDTENTRY_ERRORCODE(X86_TRAP_CP,	exc_control_protection);
  #endif
 
-/* #VC */
+/* #VC & #HV */
  #ifdef CONFIG_AMD_MEM_ENCRYPT
  DECLARE_IDTENTRY_VC(X86_TRAP_VC,	exc_vmm_communication);
+DECLARE_IDTENTRY_HV(X86_TRAP_HV,	exc_hv_injection);
  #endif
 
  #ifdef CONFIG_XEN_PV
arch/x86/include/asm/page_64_types.h +1 -0
index e9e2c3ba5923..0bd7dab676c5 100644
@@ -29,6 +29,7 @@
  #define	IST_INDEX_DB		2
  #define	IST_INDEX_MCE		3
  #define	IST_INDEX_VC		4
+#define	IST_INDEX_HV		5
 
  /*
   * Set __PAGE_OFFSET to the most negative possible address +
arch/x86/include/asm/trapnr.h +1 -0
index f5d2325aa0b7..c6583631cecb 100644
@@ -26,6 +26,7 @@
  #define X86_TRAP_XF		19	/* SIMD Floating-Point Exception */
  #define X86_TRAP_VE		20	/* Virtualization Exception */
  #define X86_TRAP_CP		21	/* Control Protection Exception */
+#define X86_TRAP_HV		28	/* HV injected exception in SNP restricted mode */
  #define X86_TRAP_VC		29	/* VMM Communication Exception */
  #define X86_TRAP_IRET		32	/* IRET Exception */
 
arch/x86/include/asm/traps.h +1 -0
index 47ecfff2c83d..6795d3e517d6 100644
@@ -16,6 +16,7 @@ asmlinkage __visible notrace
  struct pt_regs *fixup_bad_iret(struct pt_regs *bad_regs);
  void __init trap_init(void);
  asmlinkage __visible noinstr struct pt_regs *vc_switch_off_ist(struct pt_regs *eregs);
+asmlinkage __visible noinstr struct pt_regs *hv_switch_off_ist(struct pt_regs *eregs);
  #endif
 
  extern bool ibt_selftest(void);
arch/x86/kernel/cpu/common.c +1 -0
index 3e508f239098..87afa3a4c8b1 100644
@@ -2165,6 +2165,7 @@ static inline void tss_setup_ist(struct tss_struct *tss)
  	tss->x86_tss.ist[IST_INDEX_MCE] = __this_cpu_ist_top_va(MCE);
  	/* Only mapped when SEV-ES is active */
  	tss->x86_tss.ist[IST_INDEX_VC] = __this_cpu_ist_top_va(VC);
+	tss->x86_tss.ist[IST_INDEX_HV] = __this_cpu_ist_top_va(HV);
  }
 
  #else /* CONFIG_X86_64 */
arch/x86/kernel/dumpstack_64.c +7 -2
index 6c5defd6569a..23aa5912c87a 100644
@@ -26,11 +26,14 @@ static const char * const exception_stack_names[] = {
  		[ ESTACK_MCE	]	= "#MC",
  		[ ESTACK_VC	]	= "#VC",
  		[ ESTACK_VC2	]	= "#VC2",
+		[ ESTACK_HV	]	= "#HV",
+		[ ESTACK_HV2	]	= "#HV2",
+		
  };
 
  const char *stack_type_name(enum stack_type type)
  {
-	BUILD_BUG_ON(N_EXCEPTION_STACKS != 6);
+	BUILD_BUG_ON(N_EXCEPTION_STACKS != 8);
 
  	if (type == STACK_TYPE_TASK)
  		return "TASK";
@@ -89,6 +92,8 @@ struct estack_pages estack_pages[CEA_ESTACK_PAGES] ____cacheline_aligned = {
  	EPAGERANGE(MCE),
  	EPAGERANGE(VC),
  	EPAGERANGE(VC2),
+	EPAGERANGE(HV),
+	EPAGERANGE(HV2),
  };
 
  static __always_inline bool in_exception_stack(unsigned long *stack, struct stack_info *info)
@@ -98,7 +103,7 @@ static __always_inline bool in_exception_stack(unsigned long *stack, struct stac
  	struct pt_regs *regs;
  	unsigned int k;
 
-	BUILD_BUG_ON(N_EXCEPTION_STACKS != 6);
+	BUILD_BUG_ON(N_EXCEPTION_STACKS != 8);
 
  	begin = (unsigned long)__this_cpu_read(cea_exception_stacks);
  	/*
arch/x86/kernel/idt.c +1 -0
index a58c6bc1cd68..48c0a7e1dbcb 100644
@@ -113,6 +113,7 @@ static const __initconst struct idt_data def_idts[] = {
 
  #ifdef CONFIG_AMD_MEM_ENCRYPT
  	ISTG(X86_TRAP_VC,		asm_exc_vmm_communication, IST_INDEX_VC),
+	ISTG(X86_TRAP_HV,		asm_exc_hv_injection, IST_INDEX_HV),
  #endif
 
  	SYSG(X86_TRAP_OF,		asm_exc_overflow),
arch/x86/kernel/sev.c +53 -0
index a428c62330d3..b54ee3ba37b0 100644
@@ -2004,6 +2004,59 @@ DEFINE_IDTENTRY_VC_USER(exc_vmm_communication)
  	irqentry_exit_to_user_mode(regs);
  }
 
+static bool hv_raw_handle_exception(struct pt_regs *regs)
+{
+	return false;
+}
+
+static __always_inline bool on_hv_fallback_stack(struct pt_regs *regs)
+{
+	unsigned long sp = (unsigned long)regs;
+
+	return (sp >= __this_cpu_ist_bottom_va(HV2) && sp < __this_cpu_ist_top_va(HV2));
+}
+
+DEFINE_IDTENTRY_HV_USER(exc_hv_injection)
+{
+	irqentry_enter_from_user_mode(regs);
+	instrumentation_begin();
+
+	if (!hv_raw_handle_exception(regs)) {
+		/*
+		 * Do not kill the machine if user-space triggered the
+		 * exception. Send SIGBUS instead and let user-space deal
+		 * with it.
+		 */
+		force_sig_fault(SIGBUS, BUS_OBJERR, (void __user *)0);
+	}
+
+	instrumentation_end();
+	irqentry_exit_to_user_mode(regs);
+}
+
+DEFINE_IDTENTRY_HV_KERNEL(exc_hv_injection)
+{
+	irqentry_state_t irq_state;
+
+	irq_state = irqentry_nmi_enter(regs);
+	instrumentation_begin();
+
+	if (!hv_raw_handle_exception(regs)) {
+		pr_emerg("PANIC: Unhandled #HV exception in kernel space\n");
+
+		/* Show some debug info */
+		show_regs(regs);
+
+		/* Ask hypervisor to sev_es_terminate */
+		sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SEV_ES_GEN_REQ);
+
+		panic("Returned from Terminate-Request to Hypervisor\n");
+	}
+
+	instrumentation_end();
+	irqentry_nmi_exit(regs, irq_state);
+}
+
  bool __init handle_vc_boot_ghcb(struct pt_regs *regs)
  {
  	unsigned long exit_code = regs->orig_ax;
arch/x86/mm/cpu_entry_area.c +2 -0
index 6c2f1b76a0b6..8b1670fa2b81 100644
@@ -115,6 +115,8 @@ static void __init percpu_setup_exception_stacks(unsigned int cpu)
  		if (cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT)) {
  			cea_map_stack(VC);
  			cea_map_stack(VC2);
+			cea_map_stack(HV);
+			cea_map_stack(HV2);
  		}
  	}
  }
author
#45 lore raw

Re: [RFC PATCH V2 15/18] x86/sev: Add a #HV exception handler

Quoted 20 lines · Pankaj
Quoted 11 lines

+ *
+ * To make this work the #HV entry code tries its best to pretend it doesn't use
+ * an IST stack by switching to the task stack if coming from user-space (which
+ * includes early SYSCALL entry path) or back to the stack in the IRET frame if
+ * entered from kernel-mode.
+ *
+ * If entered from kernel-mode the return stack is validated first, and if it is

Quoted 8 lines

+ * not safe to use (e.g. because it points to the entry stack) the #HV handler
+ * will switch to a fall-back stack (HV2) and call a special handler function.
+ *
+ * The macro is only used for one vector, but it is planned to be extended in
+ * the future for the #HV exception.

Quoted 1 line

Noticed same comment line in the #VC exception handling section (macro idtentry_vc). Just wondering we need to remove this line as the patch being proposed for the #HV exception handling? unless this is planned to be extended in some other way.

Nice catch! Will remove this in the next version.

Thanks.

#46 lore raw

Re: [RFC PATCH V2 15/18] x86/sev: Add a #HV exception handler

Quoted 20 lines
Quoted 20 lines
Quoted 19 lines

+ * To make this work the #HV entry code tries its best to pretend it doesn't use
+ * an IST stack by switching to the task stack if coming from user-space (which
+ * includes early SYSCALL entry path) or back to the stack in the IRET frame if
+ * entered from kernel-mode.
+ *
+ * If entered from kernel-mode the return stack is validated first, and if it is
+ * not safe to use (e.g. because it points to the entry stack) the #HV handler
+ * will switch to a fall-back stack (HV2) and call a special handler function.
+ *
+ * The macro is only used for one vector, but it is planned to be extended in
+ * the future for the #HV exception.

Quoted 3 lines
Quoted 1 line

Noticed same comment line in the #VC exception handling section (macro idtentry_vc). Just wondering we need to remove this line as the patch being proposed for the #HV exception handling? unless this is planned to be extended in some other way.

Nice catch! Will remove this in the next version.

Thanks. Just a note: the referred 'idtentry_vc' macro has some instructions added/moved (e.g CLD is moved at start of IDT entry, ENDBR added) plus some additional comments added later. Don't know what all of them are required to be replicated in 'idtentry_hv', just want to share my observation.

Thanks,
Pankaj