$ cat /proc/cpuinfo | grep vmx $ lsmod | grep kvm kvm_intel 38370 0 kvm 176424 1 kvm_intel
$ cat /sys/module/kvm_intel/parameters/ept N
$ qemu-system-i386 -enable-kvm linux-0.2.img -vnc 0.0.0.0:1
trace-cmd。需要內核支援 Debugfs 和安裝 SWIG。$ git clone git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/trace-cmd.git trace-cmd $ cd trace-cmd $ make prefix=$INSTALL; make prefix=$INSTALL install $ export PATH=$INSTALL/bin:$PATH # 可能因為 NFS 的關係，必須在 /tmp 運行 trace-cmd。 $ trace-cmd record -b 20000 -e kvm $ qemu-system-i386 -enable-kvm linux-0.2.img -vnc 0.0.0.0:1 # trace-cmd 會在當前目錄底下建立 trace.dat。 $ trace-cmd report
CPU 1 is empty CPU 3 is empty CPU 4 is empty cpus=8 # /sys/kernel/debug/tracing/trace # TASK-PID CPU# TIMESTAMP FUNCTION # | | | | | qemu-system-i38-14653 [000] 7751.718118: kvm_fpu: load qemu-system-i38-14653 [000] 7751.718122: kvm_entry: vcpu 0 qemu-system-i38-14653 [000] 7751.718125: kvm_exit: [FAILED TO PARSE] exit_reason=0 guest_rip=0xfff0 qemu-system-i38-14653 [000] 7751.718126: kvm_page_fault: address ffff0 error_code 1d qemu-system-i38-14653 [000] 7751.718141: kvm_entry: vcpu 0 qemu-system-i38-14653 [000] 7751.718142: kvm_exit: [FAILED TO PARSE] exit_reason=0 guest_rip=0xc701 qemu-system-i38-14653 [000] 7751.718143: kvm_page_fault: address feffd066 error_code 9 qemu-system-i38-14653 [000] 7751.718149: kvm_entry: vcpu 0 qemu-system-i38-14653 [000] 7751.718150: kvm_exit: [FAILED TO PARSE] exit_reason=30 guest_rip=0xc701 qemu-system-i38-14653 [000] 7751.718152: kvm_pio: pio_write at 0x70 size 1 count 1
$ git clone git://git.kernel.org/pub/scm/virt/kvm/kvm.git
KVM 初始化概觀請見 The initialization of a kvm。
/dev/kvm 取得 KVM fd，此為 KVM 暴露給上層應用程序 (QEMU) 的介面，應用程序透過 ioctl 操控 KVM fd 對 KVM 發出命令。強烈建議閱讀 KVM 实现机制 和 kvm: the Linux Virtual Machine Monitor 一文。KVM 和 QEMU 分別運行在根模式中的內核態 (kernel) 與用戶態 (user)，客戶機則是運行在非根模式 (guest)。在 KVM 文件或代碼中提到的 userspace 就是指 QEMU。KVM 实现机制 針對內存虛擬化主要著墨在影子頁表，如 "3.2.2. VTLB 实现" 一節。
kvm_dev_ioctl (virt/kvm/kvm_main.c) 負責處理對 KVM fd 下達的命令。static long kvm_dev_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg) { long r = -EINVAL; switch (ioctl) { case KVM_GET_API_VERSION: r = -EINVAL; if (arg) goto out; r = KVM_API_VERSION; break; case KVM_CREATE_VM: r = kvm_dev_ioctl_create_vm(arg); break; ... 略 ... } }
kvm_vm_ioctl (virt/kvm/kvm_main.c) 負責處理對 VM fd 下達的命令。static long kvm_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg) { switch (ioctl) { case KVM_CREATE_VCPU: r = kvm_vm_ioctl_create_vcpu(kvm, arg); if (r < 0) goto out; break; case KVM_SET_USER_MEMORY_REGION: { struct kvm_userspace_memory_region kvm_userspace_mem; r = -EFAULT; // 注意! 客戶機物理內存是從 QEMU 虛擬內存分配。底下將 KVM_SET_USER_MEMORY_REGION // 後面接的 kvm_userspace_mem 參數從用戶態拷貝到內核態。 if (copy_from_user(&kvm_userspace_mem, argp, sizeof kvm_userspace_mem)) goto out; r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem, 1); if (r) goto out; break; } ... 略 ... } }
/* * Allocate some memory and give it an address in the guest physical address * space. * * Discontiguous memory is allowed, mostly for framebuffers. * * Must be called holding mmap_sem for write. */ int __kvm_set_memory_region(struct kvm *kvm, struct kvm_userspace_memory_region *mem, int user_alloc) { }
kvm_vcpu_ioctl (virt/kvm/kvm_main.c) 負責處理對 VCPU fd 下達的命令。static long kvm_vcpu_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg) { switch (ioctl) { case KVM_RUN: r = -EINVAL; if (arg) goto out; r = kvm_arch_vcpu_ioctl_run(vcpu, vcpu->run); trace_kvm_userspace_exit(vcpu->run->exit_reason, r); break; ... 略 ... }
arch/x86/kvm/vmx.c 定義 VMCS。struct vmcs { u32 revision_id; u32 abort; char data[0]; };
vmcs_writel 或 vmcs_readl 及其包裝函式。static void vmcs_writel(unsigned long field, unsigned long value) { u8 error; // arch/x86/include/asm/vmx.h // #define ASM_VMX_VMWRITE_RAX_RDX ".byte 0x0f, 0x79, 0xd0" // 將 RAX 內容寫入用 RDX 索引到的 VMCS 欄位，RAX 和 RDX 分別由參數 value 和 field 給值。 asm volatile (__ex(ASM_VMX_VMWRITE_RAX_RDX) "; setna %0" : "=q"(error) : "a"(value), "d"(field) : "cc"); if (unlikely(error)) vmwrite_error(field, value); } static __always_inline unsigned long vmcs_readl(unsigned long field) { unsigned long value; // #define ASM_VMX_VMREAD_RDX_RAX ".byte 0x0f, 0x78, 0xd0" asm volatile (__ex_clear(ASM_VMX_VMREAD_RDX_RAX, "%0") : "=a"(value) : "d"(field) : "cc"); return value; }
include/linux/kvm_host.h 定義 kvm，此即為 VM。struct kvm { spinlock_t mmu_lock; struct mutex slots_lock; struct mm_struct *mm; /* userspace tied to this vm */ struct kvm_memslots *memslots; struct srcu_struct srcu; #ifdef CONFIG_KVM_APIC_ARCHITECTURE u32 bsp_vcpu_id; // SMP 一般支援 APIC，此時其中一個 VCPU 扮演 BSP (bootstrap processor)。 #endif struct kvm_vcpu *vcpus[KVM_MAX_VCPUS]; // VM 內可含多個 VCPU。 struct kvm_arch arch; // 不同的平台會定義自己的 kvm_arch。 ... 略 ... };
kvm_arch (arch/x86/include/asm/kvm_host.h)。struct kvm_arch { unsigned int n_used_mmu_pages; unsigned int n_requested_mmu_pages; unsigned int n_max_mmu_pages; unsigned int indirect_shadow_pages; struct hlist_head mmu_page_hash[KVM_NUM_MMU_PAGES]; /* * Hash table of struct kvm_mmu_page. */ struct list_head active_mmu_pages; struct list_head assigned_dev_head; struct iommu_domain *iommu_domain; int iommu_flags; struct kvm_pic *vpic; struct kvm_ioapic *vioapic; // x86 上的 IOAPIC。 ... 略 ... };
include/linux/kvm_host.h 定義 kvm_vcpu，此即為 VCPU。struct kvm_vcpu { struct kvm *kvm; int cpu; int vcpu_id; int srcu_idx; int mode; // VCPU 處於何種模式，如: OUTSIDE_GUEST_MODE 或 IN_GUEST_MODE。 unsigned long requests; unsigned long guest_debug; struct mutex mutex; struct kvm_run *run; // 保存諸如 KVM VMExit 相關訊息。 int fpu_active; int guest_fpu_loaded, guest_xcr0_loaded; wait_queue_head_t wq; struct pid *pid; int sigset_active; sigset_t sigset; struct kvm_vcpu_stat stat; ... 略 ... struct kvm_vcpu_arch arch; /* 不同 ISA 有自己的 kvm_vcpu_arch，其中包含平台特定的暫存器組。 */ };
arch/x86/include/asm/kvm_host.h 定義 x86 自己的 kvm_vcpu_arch。struct kvm_vcpu_arch { unsigned long regs[NR_VCPU_REGS]; // x86 SMP 平台，CPU 內有一個 LAPIC 接收 IOAPIC 送上來的中斷。 // 在此以軟體模擬。 struct kvm_lapic *apic; /* kernel irqchip context */ ... 略 ... /* * Paging state of the vcpu * * If the vcpu runs in guest mode with two level paging this still saves * the paging mode of the l1 guest. This context is always used to * handle faults. */ struct kvm_mmu mmu; ... 略 ... };
struct kvm_ioapic (virt/kvm/ioapic.h)。struct kvm_ioapic { u64 base_address; u32 ioregsel; u32 id; u32 irr; u32 pad; // IOAPIC 會將從週邊收到的中斷轉發給 CPU 的 LAPIC。 union kvm_ioapic_redirect_entry redirtbl[IOAPIC_NUM_PINS]; unsigned long irq_states[IOAPIC_NUM_PINS]; struct kvm_io_device dev; struct kvm *kvm; void (*ack_notifier)(void *opaque, int irq); spinlock_t lock; DECLARE_BITMAP(handled_vectors, 256); };
struct kvm_lapic (arch/x86/kvm/lapic.h)。struct kvm_lapic { unsigned long base_address; struct kvm_io_device dev; struct kvm_timer lapic_timer; u32 divide_count; struct kvm_vcpu *vcpu; // 指向所屬的 VCPU。 bool irr_pending; void *regs; gpa_t vapic_addr; struct page *vapic_page; };
vmx_init → kvm_init → kvm_arch_init
kvm_init 註冊平台特定的回調函式。以 VMX 為例，vmx_init (arch/x86/kvm/vmx.c) 為其入口函式。static struct kvm_x86_ops vmx_x86_ops = { ... 略 ... .vcpu_create = vmx_create_vcpu, .vcpu_free = vmx_free_vcpu, .vcpu_reset = vmx_vcpu_reset, ... 略 ... .set_cr3 = vmx_set_cr3, // 設置客戶機 CR3。 .set_tdp_cr3 = vmx_set_cr3, // 設置 EPTP。 .handle_exit = vmx_handle_exit, // 當客戶機 VMExit 時，陷入 VMM，交由此函式處理。 }; static int __init vmx_init(void) { ... 略 ... r = kvm_init(&vmx_x86_ops, sizeof(struct vcpu_vmx), __alignof__(struct vcpu_vmx), THIS_MODULE); ... 略 ... if (enable_ept) { kvm_mmu_set_mask_ptes(0ull, 0ull, 0ull, 0ull, VMX_EPT_EXECUTABLE_MASK); ept_set_mmio_spte_mask(); kvm_enable_tdp(); } else kvm_disable_tdp(); return 0; }
kvm_init 初始化 KVM。int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align, struct module *module) { int r; int cpu; // 註冊平台特定的回調函式供後續函式使用，如: kvm_arch_hardware_setup。 r = kvm_arch_init(opaque); ... 略 ... // 設置 VMCS 並檢測各項硬體支援，如: 是否支援 EPT/NPT、VPID 等。 r = kvm_arch_hardware_setup(); }
int kvm_arch_init(void *opaque) { int r; struct kvm_x86_ops *ops = (struct kvm_x86_ops *)opaque; /* 檢查硬體是否支援 KVM。 */ if (!ops->cpu_has_kvm_support()) { printk(KERN_ERR "kvm: no hardware support\n"); r = -EOPNOTSUPP; goto out; } if (ops->disabled_by_bios()) { printk(KERN_ERR "kvm: disabled by bios\n"); r = -EOPNOTSUPP; goto out; } r = kvm_mmu_module_init(); kvm_set_mmio_spte_mask(); kvm_init_msr_list(); kvm_x86_ops = ops; // 相當重要! kvm_timer_init(); perf_register_guest_info_callbacks(&kvm_guest_cbs); if (cpu_has_xsave) host_xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK); return 0; }
kvm_arch_hardware_setup (arch/x86/kvm/x86.c) 呼叫 kvm_x86_ops 中註冊的 hardware_setup 函式，設置 VMCS 並做相關硬體檢測。static __init int hardware_setup(void) { if (setup_vmcs_config(&vmcs_config) < 0) return -EIO; if (boot_cpu_has(X86_FEATURE_NX)) kvm_enable_efer_bits(EFER_NX); if (!cpu_has_vmx_vpid()) enable_vpid = 0; if (!cpu_has_vmx_ept() || !cpu_has_vmx_ept_4levels()) { enable_ept = 0; enable_unrestricted_guest = 0; } ... 略 ... // 分配 VMCS 給所有的 VCPU。 return alloc_kvm_area(); }
setup_vmcs_config (arch/x86/kvm/vmx.c) 設定傳入的 VMCS vmcs_conf 供其它函式使用。static __init int setup_vmcs_config(struct vmcs_config *vmcs_conf) { /* 設定客戶機何時需要 VMExit */ min = CPU_BASED_HLT_EXITING | ... 略 ... CPU_BASED_MONITOR_EXITING | CPU_BASED_INVLPG_EXITING | CPU_BASED_RDPMC_EXITING; /* 設定使用何種硬體加速 */ opt = CPU_BASED_TPR_SHADOW | CPU_BASED_USE_MSR_BITMAPS | CPU_BASED_ACTIVATE_SECONDARY_CONTROLS; }
arch/x86/include/asm/vmx.h 定義的上述 VMCS 控制段。alloc_kvm_area (arch/x86/kvm/vmx.c) 分配 VMCS 給所有的 VCPU。static __init int alloc_kvm_area(void) { int cpu; for_each_possible_cpu(cpu) { struct vmcs *vmcs; vmcs = alloc_vmcs_cpu(cpu); if (!vmcs) { free_kvm_area(); return -ENOMEM; } per_cpu(vmxarea, cpu) = vmcs; } return 0; } static struct vmcs *alloc_vmcs_cpu(int cpu) { int node = cpu_to_node(cpu); struct page *pages; struct vmcs *vmcs; pages = alloc_pages_exact_node(node, GFP_KERNEL, vmcs_config.order); if (!pages) return NULL; vmcs = page_address(pages); memset(vmcs, 0, vmcs_config.size); vmcs->revision_id = vmcs_config.revision_id; /* vmcs revision id */ return vmcs; }
struct vmcs { u32 revision_id; u32 abort; char data[0]; }; /* * Track a VMCS that may be loaded on a certain CPU. If it is (cpu!=-1), also * remember whether it was VMLAUNCHed, and maintain a linked list of all VMCSs * loaded on this CPU (so we can clear them if the CPU goes down). */ struct loaded_vmcs { struct vmcs *vmcs; int cpu; int launched; struct list_head loaded_vmcss_on_cpu_link; }; struct vcpu_vmx { struct kvm_vcpu vcpu; ... 略 ... /* * loaded_vmcs points to the VMCS currently used in this vcpu. For a * non-nested (L1) guest, it always points to vmcs01. For a nested * guest (L2), it points to a different VMCS. */ struct loaded_vmcs vmcs01; struct loaded_vmcs *loaded_vmcs; bool __launched; /* temporary, used in vmx_vcpu_run */ ... 略 ... };
kvm_vm_ioctl (virt/kvm/kvm_main.c)。static long kvm_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg) { switch (ioctl) { case KVM_CREATE_VCPU: r = kvm_vm_ioctl_create_vcpu(kvm, arg); if (r < 0) goto out; break; ... 略 ... } }
kvm_vm_ioctl_create_vcpu (virt/kvm/kvm_main.c) 替每一個 VCPU 建立一個 kvm_vcpu。static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id) { int r; struct kvm_vcpu *vcpu, *v; // 替每一個 VCPU 建立一個 kvm_vcpu。 vcpu = kvm_arch_vcpu_create(kvm, id); /* Now it's all set up, let userspace reach it */ kvm_get_kvm(kvm); r = create_vcpu_fd(vcpu); // 返回 VCPU fd 給 QEMU。 ... 略 ... // 在此 VM 紀錄該 VCPU。 kvm->vcpus[atomic_read(&kvm->online_vcpus)] = vcpu; smp_wmb(); atomic_inc(&kvm->online_vcpus); mutex_unlock(&kvm->lock); return r; }
kvm_arch_vcpu_create (virt/kvm/kvm_main.c)。struct kvm_vcpu *kvm_arch_vcpu_create(struct kvm *kvm, unsigned int id) { if (check_tsc_unstable() && atomic_read(&kvm->online_vcpus) != 0) printk_once(KERN_WARNING "kvm: SMP vm created on host with unstable TSC; " "guest TSC will not be reliable\n"); return kvm_x86_ops->vcpu_create(kvm, id); }
vmx_create_vcpu (arch/x86/kvm/vmx.c)。static struct kvm_vcpu *vmx_create_vcpu(struct kvm *kvm, unsigned int id) { ... 略 ... vmx->loaded_vmcs->vmcs = alloc_vmcs(); ... 略 ... } static struct vmcs *alloc_vmcs_cpu(int cpu) { int node = cpu_to_node(cpu); struct page *pages; struct vmcs *vmcs; pages = alloc_pages_exact_node(node, GFP_KERNEL, vmcs_config.order); if (!pages) return NULL; vmcs = page_address(pages); memset(vmcs, 0, vmcs_config.size); vmcs->revision_id = vmcs_config.revision_id; /* vmcs revision id */ return vmcs; } static struct vmcs *alloc_vmcs(void) { return alloc_vmcs_cpu(raw_smp_processor_id()); }
/* * Sets up the vmcs for emulated real mode. */ static int vmx_vcpu_setup(struct vcpu_vmx *vmx) { #ifdef CONFIG_X86_64 unsigned long a; #endif int i; /* I/O */ vmcs_write64(IO_BITMAP_A, __pa(vmx_io_bitmap_a)); vmcs_write64(IO_BITMAP_B, __pa(vmx_io_bitmap_b)); ... 略 ... }
kvm_arch_vcpu_setup。int kvm_arch_vcpu_setup(struct kvm_vcpu *vcpu) { int r; vcpu->arch.mtrr_state.have_fixed = 1; vcpu_load(vcpu); r = kvm_arch_vcpu_reset(vcpu); if (r == 0) r = kvm_mmu_setup(vcpu); vcpu_put(vcpu); return r; }
void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu) { /* Address WBINVD may be executed by guest */ if (need_emulate_wbinvd(vcpu)) { if (kvm_x86_ops->has_wbinvd_exit()) cpumask_set_cpu(cpu, vcpu->arch.wbinvd_dirty_mask); else if (vcpu->cpu != -1 && vcpu->cpu != cpu) smp_call_function_single(vcpu->cpu, wbinvd_ipi, NULL, 1); } kvm_x86_ops->vcpu_load(vcpu, cpu); ... 略 ... }
QEMU (用戶態) 針對 VCPU 發起 KVM_RUN 命令，KVM (內核態) 處理該命令，並切至非根模式運行客戶機。
KVM_RUN -> kvm_vcpu_ioctl (kvm_main.c) -> kvm_arch_vcpu_ioctl_run (x86.c) -> __vcpu_run (x86.c) -> vcpu_enter_guest (x86.c) -> kvm_x86_ops->run(vcpu) (vmx_vcpu_run in vmx.c)
kvm_vcpu_ioctl (virt/kvm/kvm_main.c) 處理針對 VCPU 發起的命令。static long kvm_vcpu_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg) { switch (ioctl) { case KVM_RUN: r = -EINVAL; if (arg) goto out; r = kvm_arch_vcpu_ioctl_run(vcpu, vcpu->run); trace_kvm_userspace_exit(vcpu->run->exit_reason, r); break; ... 略 ... }
kvm_arch_vcpu_ioctl_run (arch/x86/kvm/x86.c)。int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run) { ... 略 ... r = __vcpu_run(vcpu); out: post_kvm_run_save(vcpu); if (vcpu->sigset_active) sigprocmask(SIG_SETMASK, &sigsaved, NULL); return r; }
arch/x86/kvm/x86.c)。static int __vcpu_run(struct kvm_vcpu *vcpu) { ... 略 ... while (r > 0) { if (vcpu->arch.mp_state == KVM_MP_STATE_RUNNABLE && !vcpu->arch.apf.halted) r = vcpu_enter_guest(vcpu); else { ... 略 ... } ... 略 ... } srcu_read_unlock(&kvm->srcu, vcpu->srcu_idx); vapic_exit(vcpu); return r; }
vcpu_enter_guest (arch/x86/kvm/x86.c) 切入非根模式。static int vcpu_enter_guest(struct kvm_vcpu *vcpu) { // 檢查 VCPU 是否有待處理的事件。 if (vcpu->requests) { } // 注入中斷至 VCPU。 if (kvm_check_request(KVM_REQ_EVENT, vcpu) || req_int_win) { inject_pending_event(vcpu); { // 載入客戶機頁表。 r = kvm_mmu_reload(vcpu); kvm_guest_enter(); // 進入非根模式，運行客戶機。 kvm_x86_ops->run(vcpu); ... 略 ... // 處理 VMExit。 r = kvm_x86_ops->handle_exit(vcpu); out: return r; }
kvm_guest_enter1)。static inline void kvm_guest_enter(void) { BUG_ON(preemptible()); account_system_vtime(current); curre nt->flags |= PF_VCPU; /* KVM does not hold any references to rcu protected data when it * switches CPU into a guest mode. In fact switching to a guest mode * is very similar to exiting to userspase from rcu point of view. In * addition CPU may stay in a guest mode for quite a long time (up to * one time slice). Lets treat guest mode as quiescent state, just like * we do with user-mode execution. */ rcu_virt_note_context_switch(smp_processor_id()); }
rcu_virt_note_context_switch 是 rcu_note_context_switch 的包裝。有 rcu 前綴代表它是用來存取共享資料的版本。static inline void rcu_virt_note_context_switch(int cpu) { rcu_note_context_switch(cpu); } /* * Note a context switch. This is a quiescent state for RCU-sched, * and requires special handling for preemptible RCU. * The caller must have disabled preemption. */ void rcu_note_context_switch(int cpu) { trace_rcu_utilization("Start context switch"); rcu_sched_qs(cpu); trace_rcu_utilization("End context switch"); }
vmx_vcpu_run 之前已註冊在 kvm_x86_ops 中。vmx_vcpu_run (arch/x86/kvm/vmx.c) 載入必要的客戶機狀態，並發起 VMEntry 切換至客戶機模式 2)。static void __noclone vmx_vcpu_run(struct kvm_vcpu *vcpu) { vmx->__launched = vmx->loaded_vmcs->launched; asm( ... 略 ... // 硬體並不會將所有客戶機暫存器載入至 CPU，部分交由 KVM 處理。 /* Load guest registers. Don't clobber flags. */ "mov %c[rax](%0), %%"R"ax \n\t" "mov %c[rbx](%0), %%"R"bx \n\t" "mov %c[rdx](%0), %%"R"dx \n\t" "mov %c[rsi](%0), %%"R"si \n\t" "mov %c[rdi](%0), %%"R"di \n\t" "mov %c[rbp](%0), %%"R"bp \n\t" "mov %c[rcx](%0), %%"R"cx \n\t" /* kills %0 (ecx) */ // x86/include/asm/vmx.h 以 hex 定義 ASM_VMX_VMLAUNCH 和 ASM_VMX_VMRESUME， // 這是因應舊有的組譯器認不得 VMX_VMLAUNCH 和 VMX_VMRESUME 指令。 /* Enter guest mode */ "jne .Llaunched \n\t" __ex(ASM_VMX_VMLAUNCH) "\n\t" "jmp .Lkvm_vmx_return \n\t" ".Llaunched: " __ex(ASM_VMX_VMRESUME) "\n\t" ".Lkvm_vmx_return: " // 返回根模式內核態。 /* Save guest registers, load host registers, keep flags */ ... 略 ... ); ... 略 ... }
vmx_handle_exit (arch/x86/kvm/vmx.c) 之前已註冊在 kvm_x86_ops 中。/* * The guest has exited. See if we can fix it or if we need userspace * assistance. */ static int vmx_handle_exit(struct kvm_vcpu *vcpu) { ... 略 ... if (exit_reason < kvm_vmx_max_exit_handlers && kvm_vmx_exit_handlers[exit_reason]) return kvm_vmx_exit_handlers[exit_reason](vcpu); else { vcpu->run->exit_reason = KVM_EXIT_UNKNOWN; vcpu->run->hw.hardware_exit_reason = exit_reason; } return 0; }
kvm_vmx_exit_handlers 針對各種 VMExit 的來源定義對應的處理函式。/* * The exit handlers return 1 if the exit was handled fully and guest execution * may resume. Otherwise they set the kvm_run parameter to indicate what needs * to be done to userspace and return 0. */ static int (*kvm_vmx_exit_handlers[])(struct kvm_vcpu *vcpu) = { [EXIT_REASON_EXCEPTION_NMI] = handle_exception, ... 略 ... [EXIT_REASON_CR_ACCESS] = handle_cr, [EXIT_REASON_INVLPG] = handle_invlpg, [EXIT_REASON_MONITOR_INSTRUCTION] = handle_invalid_op, };
handle_invlpg。static int handle_invlpg(struct kvm_vcpu *vcpu) { // 進一步讀取 VMExit 的原因。對 INVLPG 而言，此為欲剔除的 GVA。 unsigned long exit_qualification = vmcs_readl(EXIT_QUALIFICATION); // 交由 KVM 處理。 kvm_mmu_invlpg(vcpu, exit_qualification); // 跳過此條已被處理過的 (客戶機) INVLPG，將客戶機 eip 指向下一條指令。 skip_emulated_instruction(vcpu); return 1; }
kvm_mmu_invlpg (kvm/mmu.c)。void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva) { // 依情況不同有不同實現方式。 vcpu->arch.mmu.invlpg(vcpu, gva); // 向 VCPU 注入 KVM_REQ_TLB_FLUSH 的 request。 kvm_mmu_flush_tlb(vcpu); ++vcpu->stat.invlpg; }
處於非根模式的 CPU 在執行特定指令時，會返回 KVM。客戶機作業系統亦可主動透過 VMCALL 返回 KVM 3)。
早期 VT-x 版本無法正確處理實模式的指令，因此需跳回至 KVM 模擬。此外，MMIO 也需要交由 KVM 模擬 4)。
作業系統創建頁表，由硬體搜尋頁表作位址轉換，同時以 TLB (硬體) 作快取。作業系統和硬體一同協作以維護頁表和 TLB 內容的一致性。內存虛擬化需要經過底下兩層位址轉換:
影子頁表用來加速位址轉換 (GVA → HPA)。當客戶機作業系統修改頁表時，影子頁表也需要修改。這屬於軟體上的加速。請見 KVM MMU Virtualization 第 7 頁。硬體 CR3 指向的是影子頁表。硬體支援的內存虛擬化即是多加一個擴展頁表，並由硬體來實現另一次轉換。當客戶機作業系統查找頁表時 (GVA → GPA)，硬體同時會查找擴展頁表 (GPA → HPA)。頁表和擴展頁表分別由客戶機作業系統和 VMM 創建，皆由硬體負責查找。請見 KVM MMU Virtualization 第 10 頁。 在 VMCS 裡，會分別存放客戶機和宿主機的狀態。一般的硬體 CR3 是載入客戶機 CR3 的值，EPT CR3 載入的是與其相對應頁表的值。
KVM 实现机制 一文。基於軟體實現的內存虛擬化，一般稱為影子頁表 (shadow page table)。針對客戶機中的頁表，VMM 內部皆有一份對應的影子頁表 (GVA → HPA)，硬體 CR3 指向的亦是影子頁表。這是因為客戶機頁表內存的是 GVA → GPA 的映射，無法用 GPA 存取宿主機上的物理內存。當客戶機欲存取 GVA 時，硬體會查詢 TLB 和影子頁表，如果影子頁表沒有 GVA → HPA 的映射，就會發生頁缺失。影子頁表加上硬體 TLB 被當作是客戶機的虛擬 TLB (VTLB)。客戶機可以修改其頁表，但這會導致頁表和影子頁表不一致，VMM 需要同步這兩者的差異。第一種情況是，客戶機頁表擁有 GVA → GPA 的映射，而影子頁表尚未有 GPA → HPA 的映射。此時客戶機存取 GVA 會發生頁缺失，陷入 VMM，VMM 再根據客戶機頁表修改影子頁表。第二種情況是，客戶機將某頁表項的權限降低，此時該頁表會與影子頁表不一致。但是客戶機在修改頁表項之後，會用 INVLPG 或是重載 CR3 的方式刷新 (虛擬) TLB，這會陷入 VMM，VMM 此時可同步客戶機頁表和影子頁表。
反過來說，若 VMM 修改到影子頁表，則此項修改亦須與客戶機頁表同步。實際上，硬體查找的是影子頁表，此時有可能會修改影子頁表項中的 A (Access) 或是 D (Dirty) 位。以 A 位來說，當 TLB 存在某條映射，相關的頁表項其 A 位必定為 1，這是因為硬體必定已經查詢過相關的頁表項，才會在 TLB 有該條映射。將 A 位置 1 由硬體達成，將 A 位清 0 由軟體達成。後者須用 INVLPG 使對應的 TLB 項失效。如果不這樣做，後續存取會持續使用已存在的 TLB 項，而不會再把該頁表項的 A 位置 1，這會造成該頁自 A 位清 0 之後從未被存取的假象。D 位僅在最後一級頁表項中存在，代表所指向的頁已被寫過，且 TLB 項也有該位。當第一次寫該頁時，若 TLB 項不存在，或是該頁表項 D 位為 0，此時將該頁表項 D 位置 1，這是由硬體達成。如果軟體將頁表項 D 位清 0，必須使用 INVLPG 使對應的 TLB 項失效。如果不這樣做，下一次再寫該頁時，硬體看到的是舊有 TLB 項，且其 D 位為 1，硬體不會再去將該頁表項 D 位置 1。這會導致該頁自 D 位清 0 之後未被寫過的假象。前述 A 或 D 位的修改不能自動的從影子頁表反映到客戶機頁表，VMM 必須在初次修改影子頁表 A 或 D 位時加以捕獲，進而同步客戶機頁表和影子頁表。
針對 A 位。當客戶機第一次存取 GVA 時，會因為影子頁表沒有 GVA → HPA 的映射而發生頁缺失，陷入 VMM。VMM 在建立 GVA → HPA 映射的同時，便可將客戶機頁表項的 A 位置 1。至於 D 位，VMM 在建立 GVA → HPA 映射時，可將 HPA 所指頁面設為只讀。之後客戶機寫 GVA 所指頁面時，就會發生頁缺失，陷入 VMM，VMM 再同步客戶機頁表項和影子頁表項。The Shadowy Depths of the KVM MMU 第 9 頁即是此意思，所謂 shadowed page table 是指已有相對應影子頁表的客戶機頁表。
硬體支援的內存虛擬化，EPT，是在原有 CR3 的基礎上再多加一個 EPTP。CR3 指向客戶機頁表 (GVA → GPA)，EPTP 指向擴展頁表 (GPA → HPA)。VMM 會負責填充擴展頁表的內容。注意! CR3 內存的是 GPA，指向客戶機頁表。所以從 CR3 取得的 GPA 還需要再經過 EPT 得到 HPA，方可使用該 HPA 存取客戶機頁表。再由 GVA 中的 offset 取得客戶機頁表項，得到下一級客戶機頁表的 GPA，該 GPA 仍需透過 EPT 取得 HPA，再使用該 HPA 存取下一級客戶機頁表。由前述可知，EPT 基本上每一次查詢頁表會需要查詢 M (客戶機 M 級頁表) * N (宿主機 N 級頁表) 層頁表。另外，EPT 只有在非根模式 (guest) 底下才參與地址轉換。
We can't rely on invlpg and mov cr3 to tell us when we need to invalidate shadow page table entries So, we track guest page table modifications ourselves:
every shadowed guest page is write protected against guest modifications
if the guest tries to modify, we trap and emulate the modifying instruction
because we know the address, we can clear the associated shadow page table entry(ies)
支援 EPT/NPT 的硬體上仍然只有一份 TLB，其中存放 GVA → HPA 的映射 6)。Virtualization and x86 TLB
請見 Documentation/virtual/kvm/mmu.txt。常見術語:
pfn host page frame number
hpa host physical address
hva host virtual address
gfn guest frame number
gpa guest physical address
gva guest virtual address
ngpa nested guest physical address
ngva nested guest virtual address
pte page table entry (used also to refer generically to paging structure
entries)
gpte guest pte (referring to gfns)
spte shadow pte (referring to pfns)
tdp two dimensional paging (vendor neutral term for NPT and EPT)
Guest memory (gpa) is part of the user address space of the process that is using kvm. Userspace defines the translation between guest addresses and user addresses (gpa→hva); note that two gpas may alias to the same hva, but not vice versa.
One of the initialization steps that KVM does when a virtual machine (VM) is started, is setting up the vCPU's memory management unit (MMU) to translate virtual (lineal) addresses into physical ones within the guest's domain.
init_kvm_mmu，這裡我們只關注開啟 EPT 的情況，也就是 tdp_enabled 為真。tdp 是 two dimensional paging 的縮寫。/* * When setting this variable to true it enables Two-Dimensional-Paging * where the hardware walks 2 page tables: * 1. the guest-virtual to guest-physical * 2. while doing 1. it walks guest-physical to host-physical * If the hardware supports that we don't need to do shadow paging. */ bool tdp_enabled = false; static int init_kvm_mmu(struct kvm_vcpu *vcpu) { if (mmu_is_nested(vcpu)) return init_kvm_nested_mmu(vcpu); else if (tdp_enabled) return init_kvm_tdp_mmu(vcpu); else return init_kvm_softmmu(vcpu); }
include/linux/kvm_host.h 定義 struct kvm_vcpu，這是最重要的資料結構。struct kvm_vcpu { struct kvm *kvm; #ifdef CONFIG_PREEMPT_NOTIFIERS struct preempt_notifier preempt_notifier; #endif int cpu; int vcpu_id; int srcu_idx; int mode; unsigned long requests; unsigned long guest_debug; struct mutex mutex; struct kvm_run *run; ... 略 ... struct kvm_vcpu_arch arch; /* 不同 ISA 有自己的 kvm_vcpu_arch */ };
arch/x86/include/asm/kvm_host.h 定義 x86 自己的 kvm_vcpu_arch。struct kvm_vcpu_arch { ... 略 ... /* * Paging state of the vcpu * * If the vcpu runs in guest mode with two level paging this still saves * the paging mode of the l1 guest. This context is always used to * handle faults. */ struct kvm_mmu mmu; /* * Paging state of an L2 guest (used for nested npt) * * This context will save all necessary information to walk page tables * of the an L2 guest. This context is only initialized for page table * walking and not for faulting since we never handle l2 page faults on * the host. */ struct kvm_mmu nested_mmu; /* * Pointer to the mmu context currently used for * gva_to_gpa translations. */ struct kvm_mmu *walk_mmu; ... 略 ... }
walk_mmu 這項資料結構，可以設定如何查詢客戶機頁表。arch/x86/include/asm/kvm_host.h 定義 x86 上內存虛擬化相關資料結構。struct kvm_mmu { void (*new_cr3)(struct kvm_vcpu *vcpu); void (*set_cr3)(struct kvm_vcpu *vcpu, unsigned long root); unsigned long (*get_cr3)(struct kvm_vcpu *vcpu); u64 (*get_pdptr)(struct kvm_vcpu *vcpu, int index); int (*page_fault)(struct kvm_vcpu *vcpu, gva_t gva, u32 err, bool prefault); void (*inject_page_fault)(struct kvm_vcpu *vcpu, struct x86_exception *fault); void (*free)(struct kvm_vcpu *vcpu); gpa_t (*gva_to_gpa)(struct kvm_vcpu *vcpu, gva_t gva, u32 access, struct x86_exception *exception); gpa_t (*translate_gpa)(struct kvm_vcpu *vcpu, gpa_t gpa, u32 access); int (*sync_page)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp); void (*invlpg)(struct kvm_vcpu *vcpu, gva_t gva); void (*update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp, u64 *spte, const void *pte); hpa_t root_hpa; int root_level; int shadow_root_level; union kvm_mmu_page_role base_role; bool direct_map; u64 *pae_root; u64 *lm_root; u64 rsvd_bits_mask[2][4]; bool nx; u64 pdptrs[4]; /* pae */ };
init_kvm_tdp_mmu 負責填充 walk_mmu。static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu) { struct kvm_mmu *context = vcpu->arch.walk_mmu; ... 略 ... if (!is_paging(vcpu)) { context->nx = false; context->gva_to_gpa = nonpaging_gva_to_gpa; context->root_level = 0; } else if (is_long_mode(vcpu)) { context->nx = is_nx(vcpu); context->root_level = PT64_ROOT_LEVEL; reset_rsvds_bits_mask(vcpu, context); context->gva_to_gpa = paging64_gva_to_gpa; } else if (is_pae(vcpu)) { context->nx = is_nx(vcpu); context->root_level = PT32E_ROOT_LEVEL; reset_rsvds_bits_mask(vcpu, context); context->gva_to_gpa = paging64_gva_to_gpa; } else { context->nx = false; context->root_level = PT32_ROOT_LEVEL; reset_rsvds_bits_mask(vcpu, context); context->gva_to_gpa = paging32_gva_to_gpa; } return 0; }
gva_to_gpa 被賦與什麼值，此函式負責 GVA → GPA 的轉換。FNAME 是在 arch/x86/kvm/paging_tmpl.h 裡定義的宏，會將函數名稱擴展。kvm: hardware assisted paging。static gpa_t FNAME(gva_to_gpa)(struct kvm_vcpu *vcpu, gva_t vaddr, u32 access, struct x86_exception *exception) { struct guest_walker walker; gpa_t gpa = UNMAPPED_GVA; int r; r = FNAME(walk_addr)(&walker, vcpu, vaddr, access); if (r) { gpa = gfn_to_gpa(walker.gfn); gpa |= vaddr & ~PAGE_MASK; } else if (exception) *exception = walker.fault; return gpa; }
walk_addr 會轉呼叫 walk_addr_generic，walk_addr_generic 會取得 GVA 對應的客戶頁表項。static int FNAME(walk_addr)(struct guest_walker *walker, struct kvm_vcpu *vcpu, gva_t addr, u32 access) { return FNAME(walk_addr_generic)(walker, vcpu, &vcpu->arch.mmu, addr, access); } /* * Fetch a guest pte for a guest virtual address */ static int FNAME(walk_addr_generic)(struct guest_walker *walker, struct kvm_vcpu *vcpu, struct kvm_mmu *mmu, gva_t addr, u32 access) { if (last_gpte) { int lvl = walker->level; gpa_t real_gpa; gfn_t gfn; u32 ac; gfn = gpte_to_gfn_lvl(pte, lvl); gfn += (addr & PT_LVL_OFFSET_MASK(lvl)) >> PAGE_SHIFT; if (PTTYPE == 32 && walker->level == PT_DIRECTORY_LEVEL && is_cpuid_PSE36()) gfn += pse36_gfn_delta(pte); ac = write_fault | fetch_fault | user_fault; real_gpa = mmu->translate_gpa(vcpu, gfn_to_gpa(gfn), ac); if (real_gpa == UNMAPPED_GVA) return 0; walker->gfn = real_gpa >> PAGE_SHIFT; break; } ... 略 ... error: errcode |= write_fault | user_fault; if (fetch_fault && (mmu->nx || kvm_read_cr4_bits(vcpu, X86_CR4_SMEP))) errcode |= PFERR_FETCH_MASK; /* 填充 walker->fault */ trace_kvm_mmu_walker_error(walker->fault.error_code); return 0; }
kvm_read_guest_page_mmu (arch/x86/kvm/x86.c) 負責讀取 GVA 相對應的 HPA。/* * This function will be used to read from the physical memory of the currently * running guest. The difference to kvm_read_guest_page is that this function * can read from guest physical or from the guest's guest physical memory. */ int kvm_read_guest_page_mmu(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu, gfn_t ngfn, void *data, int offset, int len, u32 access) { gfn_t real_gfn; gpa_t ngpa; ngpa = gfn_to_gpa(ngfn); real_gfn = mmu->translate_gpa(vcpu, ngpa, access); if (real_gfn == UNMAPPED_GVA) return -EFAULT; real_gfn = gpa_to_gfn(real_gfn); return kvm_read_guest_page(vcpu->kvm, real_gfn, data, offset, len); }
KVM 是將 QEMU 的虛擬內存分配給客戶機作為物理內存。KVM 將客戶機物理內存分為數個 slot7)。KVM 利用硬體內存虛擬化的流程大致如下8):
kvm_set_phys_mem (kvm-all.c) → kvm_set_user_memory_region (kvm-all.c)
static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot) { struct kvm_userspace_memory_region mem; mem.slot = slot->slot; mem.guest_phys_addr = slot->start_addr; mem.memory_size = slot->memory_size; mem.userspace_addr = (unsigned long)slot->ram; mem.flags = slot->flags; if (s->migration_log) { mem.flags |= KVM_MEM_LOG_DIRTY_PAGES; } return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem); }
/* for KVM_CREATE_MEMORY_REGION */ struct kvm_memory_region { __u32 slot; __u32 flags; __u64 guest_phys_addr; __u64 memory_size; /* bytes */ }; /* for KVM_SET_USER_MEMORY_REGION */ struct kvm_userspace_memory_region { __u32 slot; __u32 flags; /* 目前只支援 KVM_MEM_LOG_DIRTY_PAGES 此 flag，KVM 用此 flag 來追蹤客戶機內存是否為 dirty。*/ __u64 guest_phys_addr; __u64 memory_size; /* bytes */ __u64 userspace_addr; /* start of the userspace allocated memory */ }; typedef struct KVMSlot { target_phys_addr_t start_addr; ram_addr_t memory_size; void *ram; int slot; int flags; } KVMSlot; struct KVMState { KVMSlot slots[32]; int fd; int vmfd; int coalesced_mmio; struct kvm_coalesced_mmio_ring *coalesced_mmio_ring; bool coalesced_flush_in_progress; int broken_set_mem_region; int migration_log; int vcpu_events; int robust_singlestep; int debugregs; int pit_state2; int xsave, xcrs; int many_ioeventfds; /* The man page (and posix) say ioctl numbers are signed int, but * they're not. Linux, glibc and *BSD all treat ioctl numbers as * unsigned, and treating them as signed here can break things */ unsigned irqchip_inject_ioctl; };
硬體 TLB 和 CR3 存放與指向的都是影子頁表或其內容。當客戶機欲存取 CR3 或是使用 INVLPG，會陷入 VMM，爾後由 VMM 接手。若是開啟 EPT，因為有另一個 EPTP 可供操作，客戶機可以存取 CR3 或是使用 INVLPG 而不會陷入 VMM。
setup_vmcs_config 會將 VM 設置為存取 CR3 或是執行 INVLPG 就觸發 VMExit。static __init int setup_vmcs_config(struct vmcs_config *vmcs_conf) { min = CPU_BASED_HLT_EXITING | CPU_BASED_CR3_LOAD_EXITING | CPU_BASED_CR3_STORE_EXITING | ... 略 ... CPU_BASED_INVLPG_EXITING | }
vcpu_enter_guest (arch/x86/kvm/x86.c) 切入非根模式。static int vcpu_enter_guest(struct kvm_vcpu *vcpu) { // 檢查 VCPU 是否有待處理的事件。 if (vcpu->requests) { } // 注入中斷至 VCPU。 if (kvm_check_request(KVM_REQ_EVENT, vcpu) || req_int_win) { inject_pending_event(vcpu); { // 載入客戶機頁表。 r = kvm_mmu_reload(vcpu); // 進入非根模式。 kvm_guest_enter(); // 運行客戶機。 kvm_x86_ops->run(vcpu); ... 略 ... // 處理 VMExit。 r = kvm_x86_ops->handle_exit(vcpu); out: return r; }
kvm_mmu_load (arch/x86/kvm/mmu.c)，分配該影子頁表並將 CR3 指向它。int kvm_mmu_load(struct kvm_vcpu *vcpu) { int r; r = mmu_topup_memory_caches(vcpu); if (r) goto out; r = mmu_alloc_roots(vcpu); spin_lock(&vcpu->kvm->mmu_lock); mmu_sync_roots(vcpu); spin_unlock(&vcpu->kvm->mmu_lock); if (r) goto out; /* set_cr3() should ensure TLB has been flushed */ vcpu->arch.mmu.set_cr3(vcpu, vcpu->arch.mmu.root_hpa); out: return r; }
struct kvm_mmu_page。arch.mmu.set_cr3 會被賦值為平台特定的函式。以 VMX 為例，set_cr3 會被賦值為 vmx_set_cr3 (arch/x86/kvm/vmx.c)，vmx_set_cr3 透過寫入 VMCS 特定欄位設定 CR3 和 EPTP。static void vmx_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3) { unsigned long guest_cr3; u64 eptp; guest_cr3 = cr3; if (enable_ept) { eptp = construct_eptp(cr3); // 設定 EPTP，指向 EPT (GPA -> HPA)。 vmcs_write64(EPT_POINTER, eptp); guest_cr3 = is_paging(vcpu) ? kvm_read_cr3(vcpu) : vcpu->kvm->arch.ept_identity_map_addr; ept_load_pdptrs(vcpu); } vmx_flush_tlb(vcpu); // 設定 CR3，指向客戶機頁表 (GVA -> GPA)。 vmcs_writel(GUEST_CR3, guest_cr3); }
page_fault (arch/x86/kvm/paging_tmpl.h)。/* * Page fault handler. There are several causes for a page fault: * - there is no shadow pte for the guest pte * - write access through a shadow pte marked read only so that we can set * the dirty bit * - write access to a shadow pte marked read only so we can update the page * dirty bitmap, when userspace requests it * - mmio access; in this case we will never install a present shadow pte * - normal guest page fault due to the guest pte marked not present, not * writable, or not executable * * Returns: 1 if we need to emulate the instruction, 0 otherwise, or * a negative value on error. */ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr, u32 error_code, bool prefault) { // MMIO。 if (unlikely(error_code & PFERR_RSVD_MASK)) return handle_mmio_page_fault(vcpu, addr, error_code, mmu_is_nested(vcpu)); // 查找客戶機頁表。 r = FNAME(walk_addr)(&walker, vcpu, addr, error_code); // 如果客戶機頁表沒有 GVA -> GPA 映射，向 VCPU 注入頁缺失例外，交由客戶機 OS 處理。 if (!r) { pgprintk("%s: guest page fault\n", __func__); if (!prefault) inject_page_fault(vcpu, &walker.fault); return 0; } // 或是影子頁表沒有 GVA -> HPA 映射，VMM 填充該影子頁表項。 sptep = FNAME(fetch)(vcpu, addr, &walker, user_fault, write_fault, level, &emulate, pfn, map_writable, prefault); }
rmap_write_protect (arch/x86/kvm/mmu.c) 9)。static bool __rmap_write_protect(struct kvm *kvm, unsigned long *rmapp, int level, bool pt_protect) { u64 *sptep; struct rmap_iterator iter; bool flush = false; for (sptep = rmap_get_first(*rmapp, &iter); sptep;) { BUG_ON(!(*sptep & PT_PRESENT_MASK)); if (spte_write_protect(kvm, sptep, &flush, pt_protect)) { sptep = rmap_get_first(*rmapp, &iter); continue; } sptep = rmap_get_next(&iter); } return flush; }
spte_write_protect (arch/x86/kvm/mmu.c)。static bool spte_write_protect(struct kvm *kvm, u64 *sptep, bool *flush, bool pt_protect) { u64 spte = *sptep; if (!is_writable_pte(spte) && !(pt_protect && spte_is_locklessly_modifiable(spte))) return false; rmap_printk("rmap_write_protect: spte %p %llx\n", sptep, *sptep); if (__drop_large_spte(kvm, sptep)) { *flush |= true; return true; } if (pt_protect) spte &= ~SPTE_MMU_WRITEABLE; spte = spte & ~PT_WRITABLE_MASK; *flush |= mmu_spte_update(sptep, spte); return false; }
But making the guest kernel trap on page table accesses might not be as easy: Page tables are normal data in memory, and while switching page tables, i.e. loading the pointer to the root page table is a privileged instruction that will trap when issued in user mode, accessing page table entries just means accessing memory, and it won't trap. The trick to still make these accesses trap is to mark the pages the page table entries reside on as invalid on the shadow page tables.
http://events.ccc.de/congress/2006/Fahrplan/attachments/1132-InsideVMware.pdf
注意! 先不談虛擬化的情況下，以 Linux 為例。每個進程在內核中會有一份對映的資料結構，其中會有該進程的頁表基址。內核會被映射在每一個進程虛擬內存高位址處，所以每一個進程共享一份內核。如果內核要修改當前運行進程的頁表項，會執行以下偽碼，並沖掉當前 TLB 的內容。
*pte = new_value;
這是一個內存存取，所以需要經過頁表作地址轉換。一般內存存取和修改頁表項都會經過同一套頁表。如果內核要修改一個被上下文切換出去進程的頁表項，基本上內核會先讀取該進程在內核中的資料結構，取得該進程的頁表基址，進而取得欲修改的頁表項，再執行上述偽碼。這裡，一般內存存取和修改頁表項還是經過同一套頁表 (內核位於虛擬內存高位址處，且每一個進程都共用一份內核)。10)11)12)13)。簡而言之，要存取頁目錄項或是頁表項，都需要先經過頁表作位址轉換，再行修改。
KVM 实现机制 描述 VTLB (TLB + 影子頁表) 的实现。
我们先来看一下 VTLB 的基本操作。客户真正的访存是通过影子页表进行的，如果影子页表中存在客户线性地址到物理地址的映射，那么访存操作就正常进行了。如果影子页表中不存在客户线性地址到物理地址的映射，那么将引发一次页故障，从而导致一次 VM exit。VMM 获得控制后，将首先根据引发异常的客户线性地址去查找客户页表，如果客户页表本身限制这次访问，如到物理地址的映射不存在、违反页级保护规则等，VMM 将把异常事件回注给客户，由客户操作系统处理该页故障。如果客户页表允许本次访问，那么通常本次页故障是由于影子页表中不存在客户线性地址到物理地址的映射引起的，此时就需要根据客户页表的内容来构建相应的影子页表，或称为对客户页表进行影射（Shadowing）。
如图所示，SPD 是 PD 的影子页表，SPT1/SPT2 是 PT1/PT2 的影子页表。由于客户 PDE 和 PTE 给出的页表基址和页基址并不是真正的物理地址，所以我们采用虚线表示 PDE 到客户页表以及 PTE 到普通客户页的映射关系。
VMM 中用于影子页表的内存是受限的，因此当内存紧张时，VMM 可能回收一部分影子页表。例如，可能回收图中的影子页表 SPT2，以后客户访问 P1 时将导致页故障，VMM 将再次分配影子页表，查询客户页表，并修补客户线性地址到 P1 的映射。
如果完全模拟物理 TLB 的行为，客户机在切换 CR3 时，VMM 需要清空整个 VTLB，使所有影子页表的内容无效。在多进程客户操作系统中，CR3 将被频繁地切换，某些影子页表的内容可能很快就会被再次用到，而重建影子页表是一项十分耗时的工作。因此，采用完全模拟物理 TLB 行为的方法构建 VTLB 在效率上是较差的。
提高效率的主要做法就是缓存影子页表，即客户切换 CR3 时不清空影子页表。例如，假设客户机上有两个进程 A 和 B，参见图 3，在 T1 时刻之前 A 正在运行，此时 CR3 指向进程 A 的影子页表。在 T1 至 T2 时刻进程 B 运行，此时 CR3 指向进程 B 的影子页表，但并不丢弃进程 A 的影子页表。以后在 T3 时刻再次切换到进程 A 时，原来 A 的影子页表还可以重用，这就避免了全部重新构建 A 的影子页表，提高了效率。
为了实现缓存影子页表的做法，必须意识到以下问题的存在：客户可能在不通知 VMM 的情况下 (即不刷新 TLB，導致無法陷入 VMM)，象修改普通内存一样修改影子页表。例如，在进程 B 运行时，客户 OS 可能由于内存的紧张，将属于进程 A 的内存换出，并将相应的页表项的 P 位置 0，而由于 A 不是当前进程，所以客户 OS 不会使用 INVLPG 指令刷新 TLB，VMM 也就无从得知客户修改了进程 A 的页表。以后，当进程 A 恢复运行时，由于影子页表与客户页表不一致，将导致错误。
因此，在采用缓存影子页表的做法时，必须有某种机制保持客户页表与影子页表间的一致性，这可通过为客户页表所在的页 (位於實際機器上的物理內存) 设置写保护来实现。
首先必须区别普通客户内存和客户页表，因为效率上的考虑，不能对所有的客户页面进行写保护。当一个页表没有用于访存时，VMM 是无从知道该页的身份的。例如，客户操作系统在初始化某张页表时，VMM 不能确定该页是普通客户内存还是客户页表，只有以后该表页用于访存时，由于在 VTLB (TLB + 影子頁表) 中没有影射，将导致一次 VTLB Fill，并触发 VMM 搜索客户页表结构，从而得知与引起页面故障的客户线性地址相关的客户页面的真实身份 (即知道該頁是頁表，而非普通客戶內存)。
VTLB Fill 操作实际上在客户页表和影子页表之间进行了一次同步，为了跟踪客户页表的后续变化，应该对客户页表进行写保护。注意，客户页表也是通过影子页表来访问的，为了设置写保护就必须知道影子页表中访问客户页表所使用的 PTE，为了做到这一点，KVM 在影子页表中建立客户线性地址到物理地址的映射关系的同时，还维护了物理地址到末级页表 PTE 间的逆向映射，即给定客户页面，能够方便地得到访问该客户页面的末级页表 PTE。图中，红色箭头表示逆向映射。同时，给定一个客户页面，如果其逆向映射存在，那么正向映射一定存在，即该客户页面可以通过影子页表被访问到。
当 VMM 在 VTLB Fill 操作过程中识别一个客户页表，如 PT1，就会通过逆向映射找到访问其所需的影子页表项，如 SPT2 中的某个 PTE (指向 PT1)，将 PTE 的 WP 位置 1。以后，客户对该客户页表的修改将导致 VM exit，从而使 VMM 有机会与客户页表保持同步。
需注意的问题是，内存的紧张可能导致 SPT2 被回收，因此在识别到一个客户页表时，影子页表中不一定总存在到客户页表的映射，也就不能为其设置写保护。但以后客户要修改该页表时，总要首先在影子页表中建立到它的映射，在建立映射时，VMM 检查该客户页面是否是客户页表且已被影射，如果是，则置页表项的 WP 位为 1。
还有一种可能是，影子页表中存在到 PT1 的映射，但 PT1 并没有被影射 (即不存在 SPT1，但 SPT2 中有頁表項指到 PT1)。这时，VMM 不对 PT1 进行写保护，客户可随意修改 PT1，以后当客户使用 PT1 进行访存时，必然会引起一次 VTLB Fill，从而使 VMM 有机会影射该客户页表 (即建立 SPT1)，与其同步，并在 SPT2 没有被回收的情况下设置对该客户页表的写保护。
由以上分析中我们还可看出，客户页表可以很大，但 VMM 没有必要对它们全部进行影射，VMM 仅需影射那些真正用于访存的客户页表。
注意! 因為影子頁表和 EPT 基本上都是在客戶機頁表之外，再多一套頁表，而該頁表 (影子頁表或 EPT) 都是由 VMM 負責填充其中的內容。因此填充 EPT 的代碼基本上是重用影子頁表的代碼。
setup_vmcs_config 會設置 VM 使其不在存取 CR3 或是 INVLPG 的時候，陷入 VMM。static __init int setup_vmcs_config(struct vmcs_config *vmcs_conf) { if (_cpu_based_2nd_exec_control & SECONDARY_EXEC_ENABLE_EPT) { /* CR3 accesses and invlpg don't need to cause VM Exits when EPT enabled */ _cpu_based_exec_control &= ~(CPU_BASED_CR3_LOAD_EXITING | CPU_BASED_CR3_STORE_EXITING | CPU_BASED_INVLPG_EXITING); rdmsr(MSR_IA32_VMX_EPT_VPID_CAP, vmx_capability.ept, vmx_capability.vpid); } ... 略 ... }
vmx_set_cr3 (arch/x86/kvm/vmx.c) 設定 CR3 和 EPTP。static void vmx_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3) { unsigned long guest_cr3; u64 eptp; guest_cr3 = cr3; if (enable_ept) { eptp = construct_eptp(cr3); // 設定 EPTP，指向 EPT (GPA -> HPA)。 vmcs_write64(EPT_POINTER, eptp); guest_cr3 = is_paging(vcpu) ? kvm_read_cr3(vcpu) : vcpu->kvm->arch.ept_identity_map_addr; ept_load_pdptrs(vcpu); } vmx_flush_tlb(vcpu); // 設定 CR3，指向客戶機頁表 (GVA -> GPA)。 vmcs_writel(GUEST_CR3, guest_cr3); }
tdp_page_fault 在因為發生 EPT 頁缺失，陷入 VMM 時，設置 EPT 項。static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa, u32 error_code, bool prefault) { spin_lock(&vcpu->kvm->mmu_lock); if (mmu_notifier_retry(vcpu, mmu_seq)) goto out_unlock; kvm_mmu_free_some_pages(vcpu); if (likely(!force_pt_level)) transparent_hugepage_adjust(vcpu, &gfn, &pfn, &level); // 設置 EPT 項。 r = __direct_map(vcpu, gpa, write, map_writable, level, gfn, pfn, prefault); spin_unlock(&vcpu->kvm->mmu_lock); return r; }
spte 即 shadow page table entry。static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write, int map_writable, int level, gfn_t gfn, pfn_t pfn, bool prefault) { for_each_shadow_entry(vcpu, (u64)gfn << PAGE_SHIFT, iterator) { if (iterator.level == level) { unsigned pte_access = ACC_ALL; mmu_set_spte(vcpu, iterator.sptep, ACC_ALL, pte_access, 0, write, &emulate, level, gfn, pfn, prefault, map_writable); direct_pte_prefetch(vcpu, iterator.sptep); ++vcpu->stat.pf_fixed; break; } if (!is_shadow_present_pte(*iterator.sptep)) { u64 base_addr = iterator.addr; base_addr &= PT64_LVL_ADDR_MASK(iterator.level); pseudo_gfn = base_addr >> PAGE_SHIFT; sp = kvm_mmu_get_page(vcpu, pseudo_gfn, iterator.addr, iterator.level - 1, 1, ACC_ALL, iterator.sptep); if (!sp) { pgprintk("nonpaging_map: ENOMEM\n"); kvm_release_pfn_clean(pfn); return -ENOMEM; } mmu_spte_set(iterator.sptep, __pa(sp->spt) | PT_PRESENT_MASK | PT_WRITABLE_MASK | shadow_user_mask | shadow_x_mask | shadow_accessed_mask); } } return emulate; }
為了避免在 VMExit 沖掉所有 TLB 內容，x86 支援 VPID，可在 TLB 項加上 VPID。
vmx_create_vcpu 在創建 VCPU 時，會配置其 VPID。static struct kvm_vcpu *vmx_create_vcpu(struct kvm *kvm, unsigned int id) { int err; struct vcpu_vmx *vmx = kmem_cache_zalloc(kvm_vcpu_cache, GFP_KERNEL); int cpu; if (!vmx) return ERR_PTR(-ENOMEM); allocate_vpid(vmx); ... 略 ... }
allocate_vpid 使用 vmx_vpid_bitmap 替每一個客戶機分配一個 VPID，VMM 自己的 VPID 恆為 0。static void allocate_vpid(struct vcpu_vmx *vmx) { int vpid; vmx->vpid = 0; if (!enable_vpid) return; spin_lock(&vmx_vpid_lock); vpid = find_first_zero_bit(vmx_vpid_bitmap, VMX_NR_VPIDS); if (vpid < VMX_NR_VPIDS) { vmx->vpid = vpid; __set_bit(vpid, vmx_vpid_bitmap); } spin_unlock(&vmx_vpid_lock); }
客戶機存取客戶機內存，經由影子頁表或是 EPT 轉換位址時觸發頁缺失。Intel® 64 and IA-32 Architectures Software Developer's Manual Volume 3C 第 27 章，第 127 頁描述發生 EPT violation，Exit Qualification 會帶有什麼資訊。
static int handle_ept_violation(struct kvm_vcpu *vcpu) { unsigned long exit_qualification; gpa_t gpa; u32 error_code; int gla_validity; exit_qualification = vmcs_readl(EXIT_QUALIFICATION); if (exit_qualification & (1 << 6)) { printk(KERN_ERR "EPT: GPA exceeds GAW!\n"); return -EINVAL; } gla_validity = (exit_qualification >> 7) & 0x3; if (gla_validity != 0x3 && gla_validity != 0x1 && gla_validity != 0) { printk(KERN_ERR "EPT: Handling EPT violation failed!\n"); printk(KERN_ERR "EPT: GPA: 0x%lx, GVA: 0x%lx\n", (long unsigned int)vmcs_read64(GUEST_PHYSICAL_ADDRESS), vmcs_readl(GUEST_LINEAR_ADDRESS)); printk(KERN_ERR "EPT: Exit qualification is 0x%lx\n", (long unsigned int)exit_qualification); vcpu->run->exit_reason = KVM_EXIT_UNKNOWN; vcpu->run->hw.hardware_exit_reason = EXIT_REASON_EPT_VIOLATION; return 0; } gpa = vmcs_read64(GUEST_PHYSICAL_ADDRESS); trace_kvm_page_fault(gpa, exit_qualification); /* It is a write fault? */ error_code = exit_qualification & (1U << 1); /* ept page table is present? */ error_code |= (exit_qualification >> 3) & 0x1; return kvm_mmu_page_fault(vcpu, gpa, error_code, NULL, 0); }
int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code, void *insn, int insn_len) { int r, emulation_type = EMULTYPE_RETRY; enum emulation_result er; r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code, false); if (r < 0) goto out; if (!r) { r = 1; goto out; } if (is_mmio_page_fault(vcpu, cr2)) emulation_type = 0; er = x86_emulate_instruction(vcpu, cr2, emulation_type, insn, insn_len); switch (er) { case EMULATE_DONE: return 1; case EMULATE_DO_MMIO: ++vcpu->stat.mmio_exits; /* fall through */ case EMULATE_FAIL: return 0; default: BUG(); } out: return r; }
int x86_emulate_instruction(struct kvm_vcpu *vcpu, unsigned long cr2, int emulation_type, void *insn, int insn_len) { int r; struct x86_emulate_ctxt *ctxt = &vcpu->arch.emulate_ctxt; bool writeback = true; kvm_clear_exception_queue(vcpu); if (!(emulation_type & EMULTYPE_NO_DECODE)) { init_emulate_ctxt(vcpu); ctxt->interruptibility = 0; ctxt->have_exception = false; ctxt->perm_ok = false; ctxt->only_vendor_specific_insn = emulation_type & EMULTYPE_TRAP_UD; r = x86_decode_insn(ctxt, insn, insn_len); trace_kvm_emulate_insn_start(vcpu); ++vcpu->stat.insn_emulation; if (r != EMULATION_OK) { if (emulation_type & EMULTYPE_TRAP_UD) return EMULATE_FAIL; if (reexecute_instruction(vcpu, cr2)) return EMULATE_DONE; if (emulation_type & EMULTYPE_SKIP) return EMULATE_FAIL; return handle_emulation_failure(vcpu); } } if (emulation_type & EMULTYPE_SKIP) { kvm_rip_write(vcpu, ctxt->_eip); return EMULATE_DONE; } if (retry_instruction(ctxt, cr2, emulation_type)) return EMULATE_DONE; /* this is needed for vmware backdoor interface to work since it changes registers values during IO operation */ if (vcpu->arch.emulate_regs_need_sync_from_vcpu) { vcpu->arch.emulate_regs_need_sync_from_vcpu = false; memcpy(ctxt->regs, vcpu->arch.regs, sizeof ctxt->regs); } restart: r = x86_emulate_insn(ctxt); if (r == EMULATION_INTERCEPTED) return EMULATE_DONE; if (r == EMULATION_FAILED) { if (reexecute_instruction(vcpu, cr2)) return EMULATE_DONE; return handle_emulation_failure(vcpu); } if (ctxt->have_exception) { inject_emulated_exception(vcpu); r = EMULATE_DONE; } else if (vcpu->arch.pio.count) { if (!vcpu->arch.pio.in) vcpu->arch.pio.count = 0; else writeback = false; r = EMULATE_DO_MMIO; } else if (vcpu->mmio_needed) { if (!vcpu->mmio_is_write) writeback = false; r = EMULATE_DO_MMIO; } else if (r == EMULATION_RESTART) goto restart; else r = EMULATE_DONE; if (writeback) { toggle_interruptibility(vcpu, ctxt->interruptibility); kvm_set_rflags(vcpu, ctxt->eflags); kvm_make_request(KVM_REQ_EVENT, vcpu); memcpy(vcpu->arch.regs, ctxt->regs, sizeof ctxt->regs); vcpu->arch.emulate_regs_need_sync_to_vcpu = false; kvm_rip_write(vcpu, ctxt->eip); } else vcpu->arch.emulate_regs_need_sync_to_vcpu = true; return r; }
VMM 只需模擬裝置的軟件接口，如: port IO、MMIO、DMA 和中斷。無需模擬裝置物理上的結構。
ioapic_deliver (virt/kvm/ioapic.c)。static int ioapic_deliver(struct kvm_ioapic *ioapic, int irq) { union kvm_ioapic_redirect_entry *entry = &ioapic->redirtbl[irq]; struct kvm_lapic_irq irqe; irqe.dest_id = entry->fields.dest_id; ... 略 ... // 將 IRQ 注入到 KVM 所擁有的 VCPU/LAPIC。 return kvm_irq_delivery_to_apic(ioapic->kvm, NULL, &irqe); }
kvm_irq_delivery_to_apic (virt/kvm/irq_comm.c)。int kvm_irq_delivery_to_apic(struct kvm *kvm, struct kvm_lapic *src, struct kvm_lapic_irq *irq) { int i, r = -1; struct kvm_vcpu *vcpu, *lowest = NULL; kvm_for_each_vcpu(i, vcpu, kvm) { if (!kvm_apic_present(vcpu)) continue; if (!kvm_apic_match_dest(vcpu, src, irq->shorthand, irq->dest_id, irq->dest_mode)) continue; if (!kvm_is_dm_lowest_prio(irq)) { if (r < 0) r = 0; r += kvm_apic_set_irq(vcpu, irq); } else if (kvm_lapic_enabled(vcpu)) { if (!lowest) lowest = vcpu; else if (kvm_apic_compare_prio(vcpu, lowest) < 0) lowest = vcpu; } } if (lowest) r = kvm_apic_set_irq(lowest, irq); return r; }
kvm_apic_set_irq。Postcopy Live migration for QEmu/KVM 是就 kvm: the Linux Virtual Machine Monitor 的方法加以改良。How to Mitigate Latency Problems during KVM/QEMU Live Migration 則是就現行做法進行優化，SRCU (Sleepable RCU) 是 RCU 的變種。
客戶機作業系統其時間是否需要跟實際時間一致端看其如何應用。
KVM 是 Linux 內核的一個模組，它會以裝置 /dev/kvm 向外界提供它的功能。QEMU 透過 ioctl 去讀寫該裝置請求 KVM 完成特定任務。KVM 主要的工作有兩個:
第一，它負責檢視客戶機 VM Exit 的原因並做相對應的處理; 第二，它負責透過 VM Entry 啟動客戶機。客戶機若因為操作 IO 而觸發 VM Exit，KVM 會轉交 QEMU 完成 IO。整個 KVM 流程基本如下14):
/dev/kvm 取得 fd。ioctl 操作 /dev/kvm 取得 VM fd。ioctl 操作 VM fd，針對每一個 VCPU 取得個別的 fd。struct KVMState 和 struct KVMSlot 分別是其重要資料結構。typedef struct KVMSlot { target_phys_addr_t start_addr; ram_addr_t memory_size; void *ram; int slot; int flags; } KVMSlot; struct KVMState { KVMSlot slots[32]; int fd; int vmfd; int coalesced_mmio; struct kvm_coalesced_mmio_ring *coalesced_mmio_ring; bool coalesced_flush_in_progress; int broken_set_mem_region; int migration_log; int vcpu_events; int robust_singlestep; int debugregs; int pit_state2; int xsave, xcrs; int many_ioeventfds; /* The man page (and posix) say ioctl numbers are signed int, but * they're not. Linux, glibc and *BSD all treat ioctl numbers as * unsigned, and treating them as signed here can break things */ unsigned irqchip_inject_ioctl; };
CPU_COMMON (cpu-defs.h) 裡有欄位給 KVM 使用。#define CPU_COMMON struct KVMState *kvm_state; \ struct kvm_run *kvm_run; \ int kvm_fd; \ int kvm_vcpu_dirty;
main (vl.c) 會呼叫 configure_accelerator 檢查使用者是否選用 KVM。int main(int argc, char **argv, char **envp) { ... 略 ... /* init the memory */ if (ram_size == 0) { ram_size = DEFAULT_RAM_SIZE * 1024 * 1024; } configure_accelerator(); qemu_init_cpu_loop(); if (qemu_init_main_loop()) { fprintf(stderr, "qemu_init_main_loop failed\n"); exit(1); } ... 略 ... }
kvm_init (kvm-all.c)。int kvm_init(void) { KVMState *s; // slot 是用來記錄客戶機物理位址與 QEMU 虛擬位址的映射。 for (i = 0; i < ARRAY_SIZE(s->slots); i++) { s->slots[i].slot = i; } // 開啟 ''/dev/kvm'' 取得 fd。 s->fd = qemu_open("/dev/kvm", O_RDWR); // 透過 ''ioctl'' 操作 ''/dev/kvm'' 取得 VM fd。 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0); kvm_state = s; // kvm_state 為一全域變數。 memory_listener_register(&kvm_memory_listener, NULL); }
qemu_kvm_cpu_thread_fn，qemu_kvm_start_vcpu 會喚起一個執行緒執行 qemu_kvm_cpu_thread_fn。若是原本 TCG 的模式，則改由 qemu_tcg_init_vcpu 喚起 qemu_tcg_cpu_thread_fn。static void qemu_kvm_start_vcpu(CPUArchState *env) { env->thread = g_malloc0(sizeof(QemuThread)); env->halt_cond = g_malloc0(sizeof(QemuCond)); qemu_cond_init(env->halt_cond); qemu_thread_create(env->thread, qemu_kvm_cpu_thread_fn, env, QEMU_THREAD_JOINABLE); while (env->created == 0) { qemu_cond_wait(&qemu_cpu_cond, &qemu_global_mutex); } } void qemu_init_vcpu(void *_env) { CPUArchState *env = _env; env->nr_cores = smp_cores; env->nr_threads = smp_threads; env->stopped = 1; if (kvm_enabled()) { qemu_kvm_start_vcpu(env); } else if (tcg_enabled()) { qemu_tcg_init_vcpu(env); } else { qemu_dummy_start_vcpu(env); } }
qemu_kvm_cpu_thread_fn 呼叫 kvm_cpu_exec 此一主要執行迴圈。static void *qemu_kvm_cpu_thread_fn(void *arg) { ... 略 ... r = kvm_init_vcpu(env); if (r < 0) { fprintf(stderr, "kvm_init_vcpu failed: %s\n", strerror(-r)); exit(1); } qemu_kvm_init_cpu_signals(env); /* signal CPU creation */ env->created = 1; qemu_cond_signal(&qemu_cpu_cond); while (1) { if (cpu_can_run(env)) { r = kvm_cpu_exec(env); if (r == EXCP_DEBUG) { cpu_handle_guest_debug(env); } } qemu_kvm_wait_io_event(env); } return NULL; }
kvm_init_vcpu (kvm-all.c)。int kvm_init_vcpu(CPUArchState *env) { KVMState *s = kvm_state; long mmap_size; int ret; ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index); env->kvm_fd = ret; // VCPU fd 而非 KVM fd。http://lists.gnu.org/archive/html/qemu-devel/2012-06/msg02302.html env->kvm_state = s; env->kvm_vcpu_dirty = 1; // QEMU 的 kvm_run 被 mmap 到 VCPU fd。這非常重要，當後續 KVM 將客戶機的 IO 交給 QEMU 執行， // QEMU 就是透過 kvm_run 讀取 IO 相關細節。 env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED, env->kvm_fd, 0); ret = kvm_arch_init_vcpu(env); if (ret == 0) { qemu_register_reset(kvm_reset_vcpu, env); kvm_arch_reset_vcpu(env); } err: return ret; }
kvm_cpu_exec (kvm-all.c)。int kvm_cpu_exec(CPUArchState *env) { struct kvm_run *run = env->kvm_run; do { ... 略 ... run_ret = kvm_vcpu_ioctl(env, KVM_RUN, 0); // 檢視 VMExit 的原因，並做相應的處理。若 VMExit 可由 KVM (內核) 處理，由 KVM 處理。 // 其餘諸如 IO 則交給 QEMU。 switch (run->exit_reason) { // IO 交由 QEMU (用戶態) 處理。 case KVM_EXIT_IO: DPRINTF("handle_io\n"); kvm_handle_io(run->io.port, (uint8_t *)run + run->io.data_offset, run->io.direction, run->io.size, run->io.count); ret = 0; break; case KVM_EXIT_MMIO: DPRINTF("handle_mmio\n"); cpu_physical_memory_rw(run->mmio.phys_addr, run->mmio.data, run->mmio.len, run->mmio.is_write); ret = 0; break; ... 略 ... // 其餘交由平台特定的 handler 處理。 default: DPRINTF("kvm_arch_handle_exit\n"); ret = kvm_arch_handle_exit(env, run); break; } } while (ret == 0); }
kvm_arch_handle_exit。以 x86 為例，kvm_arch_handle_exit (target-i386/kvm.c)。int kvm_arch_handle_exit(CPUX86State *env, struct kvm_run *run) { }
struct kvm_run。/* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ struct kvm_run { /* in */ __u8 request_interrupt_window; __u8 padding1[7]; /* out */ __u32 exit_reason; __u8 ready_for_interrupt_injection; __u8 if_flag; __u8 padding2[2]; /* in (pre_kvm_run), out (post_kvm_run) */ __u64 cr8; __u64 apic_base; ... 略 ... }; struct kvm_vcpu { ... 略 ... struct kvm_run *run; ... 略 ... }; static int emulator_pio_in_out(struct kvm_vcpu *vcpu, int size, unsigned short port, void *val, unsigned int count, bool in) { trace_kvm_pio(!in, port, size, count); vcpu->arch.pio.port = port; vcpu->arch.pio.in = in; vcpu->arch.pio.count = count; vcpu->arch.pio.size = size; if (!kernel_pio(vcpu, vcpu->arch.pio_data)) { vcpu->arch.pio.count = 0; return 1; } // 回到 QEMU 之後，QEMU 會檢視以下來欄位。 vcpu->run->exit_reason = KVM_EXIT_IO; vcpu->run->io.direction = in ? KVM_EXIT_IO_IN : KVM_EXIT_IO_OUT; vcpu->run->io.size = size; vcpu->run->io.data_offset = KVM_PIO_PAGE_OFFSET * PAGE_SIZE; vcpu->run->io.count = count; vcpu->run->io.port = port; return 0; }
molcpu_mainloop。