Message-Id: <20240226143630.33643-27-jiangshanlai@gmail.com>
Date: Mon, 26 Feb 2024 22:35:43 +0800
From: Lai Jiangshan <jiangshanlai@...il.com>
To: linux-kernel@...r.kernel.org
Cc: Lai Jiangshan <jiangshan.ljs@...group.com>,
	Hou Wenlong <houwenlong.hwl@...group.com>,
	Linus Torvalds <torvalds@...ux-foundation.org>,
	Peter Zijlstra <peterz@...radead.org>,
	Sean Christopherson <seanjc@...gle.com>,
	Thomas Gleixner <tglx@...utronix.de>,
	Borislav Petkov <bp@...en8.de>,
	Ingo Molnar <mingo@...hat.com>,
	kvm@...r.kernel.org,
	Paolo Bonzini <pbonzini@...hat.com>,
	x86@...nel.org,
	Kees Cook <keescook@...omium.org>,
	Juergen Gross <jgross@...e.com>,
	Dave Hansen <dave.hansen@...ux.intel.com>,
	"H. Peter Anvin" <hpa@...or.com>
Subject: [RFC PATCH 26/73] KVM: x86/PVM: Implement event delivery flags related callbacks

From: Lai Jiangshan <jiangshan.ljs@...group.com>

To reduce the number of VM exits for modifying the X86_EFLAGS_IF bit in
guest suprvisor mode, a shared structure is used between the guest and
hypervisor in PVM. This structure is stored in the guest memory. In this
way, the guest supervisor can change its X86_EFLAGS_IF bit without
causing a VM exit, as long as there is no IRQ window request. After a VM
exit occurs, the hypervisor updates the guest's X86_EFLAGS_IF bit from
the shared structure.

Since the SRET/URET synthetic instruction always induces a VM exit,
there is nothing to do in the enable_nmi_window() callback.
Additionally, SMM mode is not supported now.

Signed-off-by: Lai Jiangshan <jiangshan.ljs@...group.com>
Signed-off-by: Hou Wenlong <houwenlong.hwl@...group.com>
---
 arch/x86/kvm/pvm/pvm.c | 194 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 194 insertions(+)

diff --git a/arch/x86/kvm/pvm/pvm.c b/arch/x86/kvm/pvm/pvm.c
index ce047d211657..3d2a3c472664 100644
--- a/arch/x86/kvm/pvm/pvm.c
+++ b/arch/x86/kvm/pvm/pvm.c
@@ -585,6 +585,143 @@ static bool pvm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
 	return false;
 }
 
+static struct pvm_vcpu_struct *pvm_get_vcpu_struct(struct vcpu_pvm *pvm)
+{
+	struct gfn_to_pfn_cache *gpc = &pvm->pvcs_gpc;
+
+	read_lock_irq(&gpc->lock);
+	while (!kvm_gpc_check(gpc, PAGE_SIZE)) {
+		read_unlock_irq(&gpc->lock);
+
+		if (kvm_gpc_refresh(gpc, PAGE_SIZE))
+			return NULL;
+
+		read_lock_irq(&gpc->lock);
+	}
+
+	return (struct pvm_vcpu_struct *)(gpc->khva);
+}
+
+static void pvm_put_vcpu_struct(struct vcpu_pvm *pvm, bool dirty)
+{
+	struct gfn_to_pfn_cache *gpc = &pvm->pvcs_gpc;
+
+	read_unlock_irq(&gpc->lock);
+	if (dirty)
+		mark_page_dirty_in_slot(pvm->vcpu.kvm, gpc->memslot,
+					gpc->gpa >> PAGE_SHIFT);
+}
+
+static void pvm_vcpu_gpc_refresh(struct kvm_vcpu *vcpu)
+{
+	struct vcpu_pvm *pvm = to_pvm(vcpu);
+	struct gfn_to_pfn_cache *gpc = &pvm->pvcs_gpc;
+
+	if (!gpc->active)
+		return;
+
+	if (pvm_get_vcpu_struct(pvm))
+		pvm_put_vcpu_struct(pvm, false);
+	else
+		kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
+}
+
+static void pvm_event_flags_update(struct kvm_vcpu *vcpu, unsigned long set,
+				   unsigned long clear)
+{
+	struct vcpu_pvm *pvm = to_pvm(vcpu);
+	static struct pvm_vcpu_struct *pvcs;
+	unsigned long old_flags, new_flags;
+
+	if (!pvm->msr_vcpu_struct)
+		return;
+
+	pvcs = pvm_get_vcpu_struct(pvm);
+	if (!pvcs)
+		return;
+
+	old_flags = pvcs->event_flags;
+	new_flags = (old_flags | set) & ~clear;
+	if (new_flags != old_flags)
+		pvcs->event_flags = new_flags;
+
+	pvm_put_vcpu_struct(pvm, new_flags != old_flags);
+}
+
+static unsigned long pvm_get_rflags(struct kvm_vcpu *vcpu)
+{
+	return to_pvm(vcpu)->rflags;
+}
+
+static void pvm_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
+{
+	struct vcpu_pvm *pvm = to_pvm(vcpu);
+	int need_update = !!((pvm->rflags ^ rflags) & X86_EFLAGS_IF);
+
+	pvm->rflags = rflags;
+
+	/*
+	 * The IF bit of 'pvcs->event_flags' should not be changed in user
+	 * mode. It is recommended for this bit to be cleared when switching to
+	 * user mode, so that when the guest switches back to supervisor mode,
+	 * the X86_EFLAGS_IF is already cleared.
+	 */
+	if (!need_update || !is_smod(pvm))
+		return;
+
+	if (rflags & X86_EFLAGS_IF)
+		pvm_event_flags_update(vcpu, X86_EFLAGS_IF, PVM_EVENT_FLAGS_IP);
+	else
+		pvm_event_flags_update(vcpu, 0, X86_EFLAGS_IF);
+}
+
+static bool pvm_get_if_flag(struct kvm_vcpu *vcpu)
+{
+	return pvm_get_rflags(vcpu) & X86_EFLAGS_IF;
+}
+
+static u32 pvm_get_interrupt_shadow(struct kvm_vcpu *vcpu)
+{
+	return to_pvm(vcpu)->int_shadow;
+}
+
+static void pvm_set_interrupt_shadow(struct kvm_vcpu *vcpu, int mask)
+{
+	/* PVM spec: ignore interrupt shadow when in PVM mode. */
+}
+
+static void enable_irq_window(struct kvm_vcpu *vcpu)
+{
+	pvm_event_flags_update(vcpu, PVM_EVENT_FLAGS_IP, 0);
+}
+
+static int pvm_interrupt_allowed(struct kvm_vcpu *vcpu, bool for_injection)
+{
+	return (pvm_get_rflags(vcpu) & X86_EFLAGS_IF) &&
+		!to_pvm(vcpu)->int_shadow;
+}
+
+static bool pvm_get_nmi_mask(struct kvm_vcpu *vcpu)
+{
+	return to_pvm(vcpu)->nmi_mask;
+}
+
+static void pvm_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked)
+{
+	to_pvm(vcpu)->nmi_mask = masked;
+}
+
+static void enable_nmi_window(struct kvm_vcpu *vcpu)
+{
+}
+
+static int pvm_nmi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
+{
+	struct vcpu_pvm *pvm = to_pvm(vcpu);
+
+	return !pvm->nmi_mask && !pvm->int_shadow;
+}
+
 static void pvm_setup_mce(struct kvm_vcpu *vcpu)
 {
 }
@@ -826,12 +963,29 @@ static fastpath_t pvm_vcpu_run(struct kvm_vcpu *vcpu)
 	pvm_vcpu_run_noinstr(vcpu);
 
 	if (is_smod(pvm)) {
+		struct pvm_vcpu_struct *pvcs = pvm->pvcs_gpc.khva;
+
+		/*
+		 * Load the X86_EFLAGS_IF bit from PVCS. In user mode, the
+		 * Interrupt Flag is considered to be set and cannot be
+		 * changed. Since it is already set in 'pvm->rflags', so
+		 * nothing to do. In supervisor mode, the Interrupt Flag is
+		 * reflected in 'pvcs->event_flags' and can be changed
+		 * directly without triggering a VM exit.
+		 */
+		pvm->rflags &= ~X86_EFLAGS_IF;
+		if (likely(pvm->msr_vcpu_struct))
+			pvm->rflags |= X86_EFLAGS_IF & pvcs->event_flags;
+
 		if (pvm->hw_cs != __USER_CS || pvm->hw_ss != __USER_DS)
 			kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
 	}
 
 	pvm_load_host_xsave_state(vcpu);
 
+	mark_page_dirty_in_slot(vcpu->kvm, pvm->pvcs_gpc.memslot,
+				pvm->pvcs_gpc.gpa >> PAGE_SHIFT);
+
 	return EXIT_FASTPATH_NONE;
 }
 
@@ -965,6 +1119,27 @@ static int pvm_check_processor_compat(void)
 	return 0;
 }
 
+#ifdef CONFIG_KVM_SMM
+static int pvm_smi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
+{
+	return 0;
+}
+
+static int pvm_enter_smm(struct kvm_vcpu *vcpu, union kvm_smram *smram)
+{
+	return 0;
+}
+
+static int pvm_leave_smm(struct kvm_vcpu *vcpu, const union kvm_smram *smram)
+{
+	return 0;
+}
+
+static void enable_smi_window(struct kvm_vcpu *vcpu)
+{
+}
+#endif
+
 /*
  * When in PVM mode, the hardware MSR_LSTAR is set to the entry point
  * provided by the host entry code (switcher), and the
@@ -1098,10 +1273,21 @@ static struct kvm_x86_ops pvm_x86_ops __initdata = {
 	.set_msr = pvm_set_msr,
 	.get_cpl = pvm_get_cpl,
 	.load_mmu_pgd = pvm_load_mmu_pgd,
+	.get_rflags = pvm_get_rflags,
+	.set_rflags = pvm_set_rflags,
+	.get_if_flag = pvm_get_if_flag,
 
 	.vcpu_pre_run = pvm_vcpu_pre_run,
 	.vcpu_run = pvm_vcpu_run,
 	.handle_exit = pvm_handle_exit,
+	.set_interrupt_shadow = pvm_set_interrupt_shadow,
+	.get_interrupt_shadow = pvm_get_interrupt_shadow,
+	.interrupt_allowed = pvm_interrupt_allowed,
+	.nmi_allowed = pvm_nmi_allowed,
+	.get_nmi_mask = pvm_get_nmi_mask,
+	.set_nmi_mask = pvm_set_nmi_mask,
+	.enable_nmi_window = enable_nmi_window,
+	.enable_irq_window = enable_irq_window,
 	.refresh_apicv_exec_ctrl = pvm_refresh_apicv_exec_ctrl,
 	.deliver_interrupt = pvm_deliver_interrupt,
 
@@ -1117,10 +1303,18 @@ static struct kvm_x86_ops pvm_x86_ops __initdata = {
 
 	.setup_mce = pvm_setup_mce,
 
+#ifdef CONFIG_KVM_SMM
+	.smi_allowed = pvm_smi_allowed,
+	.enter_smm = pvm_enter_smm,
+	.leave_smm = pvm_leave_smm,
+	.enable_smi_window = enable_smi_window,
+#endif
+
 	.apic_init_signal_blocked = pvm_apic_init_signal_blocked,
 	.msr_filter_changed = pvm_msr_filter_changed,
 	.complete_emulated_msr = kvm_complete_insn_gp,
 	.vcpu_deliver_sipi_vector = kvm_vcpu_deliver_sipi_vector,
+	.vcpu_gpc_refresh = pvm_vcpu_gpc_refresh,
 };
 
 static struct kvm_x86_init_ops pvm_init_ops __initdata = {
-- 
2.19.1.6.gb485710b

Powered by blists - more mailing lists