; $Id: PAEand32Bit.mac 28800 2010-04-27 08:22:32Z vboxsync $
;; @file
; VMM - World Switchers, template for PAE and 32-Bit.
;

;
; Copyright (C) 2006-2007 Oracle Corporation
;
; This file is part of VirtualBox Open Source Edition (OSE), as
; available from http://www.virtualbox.org. This file is free software;
; you can redistribute it and/or modify it under the terms of the GNU
; General Public License (GPL) as published by the Free Software
; Foundation, in version 2 as it comes in the "COPYING" file of the
; VirtualBox OSE distribution. VirtualBox OSE is distributed in the
; hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
;

;%define DEBUG_STUFF 1

;*******************************************************************************
;* Header Files                                                                *
;*******************************************************************************
%include "VBox/asmdefs.mac"
%include "VBox/x86.mac"
%include "VBox/cpum.mac"
%include "VBox/stam.mac"
%include "VBox/vm.mac"
%include "CPUMInternal.mac"
%include "VMMSwitcher/VMMSwitcher.mac"

%undef NEED_ID
%ifdef NEED_PAE_ON_32BIT_HOST
%define NEED_ID
%endif
%ifdef NEED_32BIT_ON_PAE_HOST
%define NEED_ID
%endif



;
; Start the fixup records
;   We collect the fixups in the .data section as we go along
;   It is therefore VITAL that no-one is using the .data section
;   for anything else between 'Start' and 'End'.
;
BEGINDATA
GLOBALNAME Fixups



BEGINCODE
GLOBALNAME Start

;;
; The C interface.
;
BEGINPROC vmmR0HostToGuest

%ifdef DEBUG_STUFF
    COM_S_NEWLINE
    COM_S_CHAR '^'
%endif

%ifdef VBOX_WITH_STATISTICS
    ;
    ; Switcher stats.
    ;
    FIXUP FIX_HC_VM_OFF, 1, VM.StatSwitcherToGC
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_START edx
%endif

    ;
    ; Call worker.
    ;
    FIXUP FIX_HC_CPUM_OFF, 1, 0
    mov     edx, 0ffffffffh
    push    cs                          ; allow for far return and restore cs correctly.
    call    NAME(vmmR0HostToGuestAsm)

%ifdef VBOX_WITH_STATISTICS
    ;
    ; Switcher stats.
    ;
    FIXUP FIX_HC_VM_OFF, 1, VM.StatSwitcherToHC
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_STOP edx
%endif

    ret
ENDPROC vmmR0HostToGuest



; *****************************************************************************
; vmmR0HostToGuestAsm
;
; Phase one of the switch from host to guest context (host MMU context)
;
; INPUT:
;       - edx       virtual address of CPUM structure (valid in host context)
;
; USES/DESTROYS:
;       - eax, ecx, edx
;
; ASSUMPTION:
;       - current CS and DS selectors are wide open
;
; *****************************************************************************
ALIGNCODE(16)
BEGINPROC vmmR0HostToGuestAsm
    ;;
    ;; Save CPU host context
    ;;      Skip eax, edx and ecx as these are not preserved over calls.
    ;;
    CPUMCPU_FROM_CPUM(edx)
    ; general registers.
    mov     [edx + CPUMCPU.Host.ebx], ebx
    mov     [edx + CPUMCPU.Host.edi], edi
    mov     [edx + CPUMCPU.Host.esi], esi
    mov     [edx + CPUMCPU.Host.esp], esp
    mov     [edx + CPUMCPU.Host.ebp], ebp
    ; selectors.
    mov     [edx + CPUMCPU.Host.ds], ds
    mov     [edx + CPUMCPU.Host.es], es
    mov     [edx + CPUMCPU.Host.fs], fs
    mov     [edx + CPUMCPU.Host.gs], gs
    mov     [edx + CPUMCPU.Host.ss], ss
    ; special registers.
    sldt    [edx + CPUMCPU.Host.ldtr]
    sidt    [edx + CPUMCPU.Host.idtr]
    sgdt    [edx + CPUMCPU.Host.gdtr]
    str     [edx + CPUMCPU.Host.tr]
    ; flags
    pushfd
    pop     dword [edx + CPUMCPU.Host.eflags]

    FIXUP FIX_NO_SYSENTER_JMP, 0, htg_no_sysenter - NAME(Start) ; this will insert a jmp htg_no_sysenter if host doesn't use sysenter.
    ; save MSR_IA32_SYSENTER_CS register.
    mov     ecx, MSR_IA32_SYSENTER_CS
    mov     ebx, edx                    ; save edx
    rdmsr                               ; edx:eax <- MSR[ecx]
    mov     [ebx + CPUMCPU.Host.SysEnter.cs], eax
    mov     [ebx + CPUMCPU.Host.SysEnter.cs + 4], edx
    xor     eax, eax                    ; load 0:0 to cause #GP upon sysenter
    xor     edx, edx
    wrmsr
    xchg    ebx, edx                    ; restore edx
    jmp short htg_no_sysenter

ALIGNCODE(16)
htg_no_sysenter:

    FIXUP FIX_NO_SYSCALL_JMP, 0, htg_no_syscall - NAME(Start) ; this will insert a jmp htg_no_syscall if host doesn't use syscall.
    ; clear MSR_K6_EFER_SCE.
    mov     ebx, edx                    ; save edx
    mov     ecx, MSR_K6_EFER
    rdmsr                               ; edx:eax <- MSR[ecx]
    and     eax, ~MSR_K6_EFER_SCE
    wrmsr
    mov     edx, ebx                    ; restore edx
    jmp short htg_no_syscall

ALIGNCODE(16)
htg_no_syscall:

    ;; handle use flags.
    mov     esi, [edx + CPUMCPU.fUseFlags] ; esi == use flags.
    and     esi, ~CPUM_USED_FPU         ; Clear CPUM_USED_* flags. ;;@todo FPU check can be optimized to use cr0 flags!
    mov     [edx + CPUMCPU.fUseFlags], esi

    ; debug registers.
    test    esi, CPUM_USE_DEBUG_REGS | CPUM_USE_DEBUG_REGS_HOST
    jz      htg_debug_regs_no
    jmp     htg_debug_regs_save_dr7and6
htg_debug_regs_no:

    ; control registers.
    mov     eax, cr0
    mov     [edx + CPUMCPU.Host.cr0], eax
    ;mov     eax, cr2                   ; assume host os don't suff things in cr2. (safe)
    ;mov     [edx + CPUMCPU.Host.cr2], eax
    mov     eax, cr3
    mov     [edx + CPUMCPU.Host.cr3], eax
    mov     eax, cr4
    mov     [edx + CPUMCPU.Host.cr4], eax

    ;;
    ;; Start switching to VMM context.
    ;;

    ;
    ; Change CR0 and CR4 so we can correctly emulate FPU/MMX/SSE[23] exceptions
    ; Also disable WP. (eax==cr4 now)
    ; Note! X86_CR4_PSE and X86_CR4_PAE are important if the host thinks so :-)
    ; Note! X86_CR4_VMXE must not be touched in case the CPU is in vmx root mode
    ;
    and     eax, X86_CR4_MCE | X86_CR4_PSE | X86_CR4_PAE | X86_CR4_VMXE
    mov     ecx, [edx + CPUMCPU.Guest.cr4]
    ;; @todo Switcher cleanup: Determin base CR4 during CPUMR0Init / VMMR3SelectSwitcher putting it
    ;                          in CPUMCPU.Hyper.cr4 (which isn't currently being used). That should
    ;                          simplify this operation a bit (and improve locality of the data).

    ;
    ; CR4.AndMask and CR4.OrMask are set in CPUMR3Init based on the presence of
    ; FXSAVE support on the host CPU
    ;
    CPUM_FROM_CPUMCPU(edx)
    and     ecx, [edx + CPUM.CR4.AndMask]
    or      eax, ecx
    or      eax, [edx + CPUM.CR4.OrMask]
    mov     cr4, eax

    CPUMCPU_FROM_CPUM(edx)
    mov     eax, [edx + CPUMCPU.Guest.cr0]
    and     eax, X86_CR0_EM
    or      eax, X86_CR0_PE | X86_CR0_PG | X86_CR0_TS | X86_CR0_ET | X86_CR0_NE | X86_CR0_MP
    mov     cr0, eax

    ; Load new gdt so we can do far jump to guest code after cr3 reload.
    lgdt    [edx + CPUMCPU.Hyper.gdtr]
    DEBUG_CHAR('1')                     ; trashes esi

    ; Store the hypervisor cr3 for later loading
    mov     ebp, [edx + CPUMCPU.Hyper.cr3]

    ;;
    ;; Load Intermediate memory context.
    ;;
    FIXUP SWITCHER_FIX_INTER_CR3_HC, 1
    mov     eax, 0ffffffffh
    mov     cr3, eax
    DEBUG_CHAR('2')                     ; trashes esi

%ifdef NEED_ID
    ;;
    ;; Jump to identity mapped location
    ;;
    FIXUP FIX_HC_2_ID_NEAR_REL, 1, NAME(IDEnterTarget) - NAME(Start)
    jmp near NAME(IDEnterTarget)

    ; We're now on identity mapped pages!
ALIGNCODE(16)
GLOBALNAME IDEnterTarget
    DEBUG_CHAR('3')
    mov     edx, cr4
%ifdef NEED_PAE_ON_32BIT_HOST
    or      edx, X86_CR4_PAE
%else
    and     edx, ~X86_CR4_PAE
%endif
    mov     eax, cr0
    and     eax, (~X86_CR0_PG) & 0xffffffff ; prevent yasm warning
    mov     cr0, eax
    DEBUG_CHAR('4')
    mov     cr4, edx
    FIXUP SWITCHER_FIX_INTER_CR3_GC, 1
    mov     edx, 0ffffffffh
    mov     cr3, edx
    or      eax, X86_CR0_PG
    DEBUG_CHAR('5')
    mov     cr0, eax
    DEBUG_CHAR('6')
%endif

    ;;
    ;; Jump to guest code mapping of the code and load the Hypervisor CS.
    ;;
    FIXUP FIX_GC_FAR32, 1, NAME(FarJmpGCTarget) - NAME(Start)
    jmp     0fff8h:0deadfaceh


    ;;
    ;; When we arrive at this label we're at the
    ;; guest code mapping of the switching code.
    ;;
ALIGNCODE(16)
GLOBALNAME FarJmpGCTarget
    DEBUG_CHAR('-')
    ; load final cr3 and do far jump to load cs.
    mov     cr3, ebp            ; ebp set above
    DEBUG_CHAR('0')

    ;;
    ;; We're in VMM MMU context and VMM CS is loaded.
    ;; Setup the rest of the VMM state.
    ;;
    FIXUP FIX_GC_CPUMCPU_OFF, 1, 0
    mov     edx, 0ffffffffh
    ; Activate guest IDT
    DEBUG_CHAR('1')
    lidt    [edx + CPUMCPU.Hyper.idtr]
    ; Load selectors
    DEBUG_CHAR('2')
    FIXUP FIX_HYPER_DS, 1
    mov     eax, 0ffffh
    mov     ds, eax
    mov     es, eax
    xor     eax, eax
    mov     gs, eax
    mov     fs, eax

    ; Setup stack; use the lss_esp, ss pair for lss
    DEBUG_CHAR('3')
    mov     eax, [edx + CPUMCPU.Hyper.esp]
    mov     [edx + CPUMCPU.Hyper.lss_esp], eax
    lss     esp, [edx + CPUMCPU.Hyper.lss_esp]

    ; Restore TSS selector; must mark it as not busy before using ltr (!)
    DEBUG_CHAR('4')
    FIXUP FIX_GC_TSS_GDTE_DW2, 2
    and     dword [0ffffffffh], ~0200h      ; clear busy flag (2nd type2 bit)
    DEBUG_CHAR('5')
    ltr     word [edx + CPUMCPU.Hyper.tr]
    DEBUG_CHAR('6')

    ; Activate the ldt (now we can safely crash).
    lldt    [edx + CPUMCPU.Hyper.ldtr]
    DEBUG_CHAR('7')

    ;; use flags.
    mov     esi, [edx + CPUMCPU.fUseFlags]

    ; debug registers
    test    esi, CPUM_USE_DEBUG_REGS
    jz      htg_debug_regs_guest_no
    jmp     htg_debug_regs_guest
htg_debug_regs_guest_no:
    DEBUG_CHAR('9')

%ifdef VBOX_WITH_NMI
    ;
    ; Setup K7 NMI.
    ;
    mov     esi, edx
    ; clear all PerfEvtSeln registers
    xor     eax, eax
    xor     edx, edx
    mov     ecx, MSR_K7_PERFCTR0
    wrmsr
    mov     ecx, MSR_K7_PERFCTR1
    wrmsr
    mov     ecx, MSR_K7_PERFCTR2
    wrmsr
    mov     ecx, MSR_K7_PERFCTR3
    wrmsr

    mov     eax, RT_BIT(20) | RT_BIT(17) | RT_BIT(16) | 076h
    mov     ecx, MSR_K7_EVNTSEL0
    wrmsr
    mov     eax, 02329B000h
    mov     edx, 0fffffffeh ; -1.6GHz * 5
    mov     ecx, MSR_K7_PERFCTR0
    wrmsr

    FIXUP FIX_GC_APIC_BASE_32BIT, 1
    mov     eax, 0f0f0f0f0h
    add     eax, 0340h                  ; APIC_LVTPC
    mov     dword [eax], 0400h          ; APIC_DM_NMI

    xor     edx, edx
    mov     eax, RT_BIT(20) | RT_BIT(17) | RT_BIT(16) | 076h | RT_BIT(22) ;+EN
    mov     ecx, MSR_K7_EVNTSEL0
    wrmsr

    mov     edx, esi
%endif

    ; General registers.
    mov     ebx, [edx + CPUMCPU.Hyper.ebx]
    mov     ebp, [edx + CPUMCPU.Hyper.ebp]
    mov     esi, [edx + CPUMCPU.Hyper.esi]
    mov     edi, [edx + CPUMCPU.Hyper.edi]
    push    dword [edx + CPUMCPU.Hyper.eflags]
    popfd
    DEBUG_CHAR('!')

    ;;
    ;; Return to the VMM code which either called the switcher or
    ;; the code set up to run by HC.
    ;;
%ifdef DEBUG_STUFF
    COM_S_PRINT ';eip='
    mov     eax, [edx + CPUMCPU.Hyper.eip]
    COM_S_DWORD_REG eax
    COM_S_CHAR ';'
%endif
    mov     eax, [edx + CPUMCPU.Hyper.eip]
    ; callees expect CPUM ptr
    CPUM_FROM_CPUMCPU(edx)

%ifdef VBOX_WITH_STATISTICS
    FIXUP FIX_GC_VM_OFF, 1, VM.StatSwitcherToGC
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_STOP edx
    FIXUP FIX_GC_CPUM_OFF, 1, 0
    mov     edx, 0ffffffffh
%endif
    jmp     eax

;;
; Detour for saving the host DR7 and DR6.
; esi and edx must be preserved.
htg_debug_regs_save_dr7and6:
DEBUG_S_CHAR('s');
    mov     eax, dr7                    ; not sure, but if I read the docs right this will trap if GD is set. FIXME!!!
    mov     [edx + CPUMCPU.Host.dr7], eax
    xor     eax, eax                    ; clear everything. (bit 12? is read as 1...)
    mov     dr7, eax
    mov     eax, dr6                    ; just in case we save the state register too.
    mov     [edx + CPUMCPU.Host.dr6], eax
    jmp     htg_debug_regs_no

;;
; Detour for saving host DR0-3 and loading hypervisor debug registers.
; esi and edx must be preserved.
htg_debug_regs_guest:
    DEBUG_S_CHAR('D')
    DEBUG_S_CHAR('R')
    DEBUG_S_CHAR('x')
    ; save host DR0-3.
    mov     eax, dr0
    mov     [edx + CPUMCPU.Host.dr0], eax
    mov     ebx, dr1
    mov     [edx + CPUMCPU.Host.dr1], ebx
    mov     ecx, dr2
    mov     [edx + CPUMCPU.Host.dr2], ecx
    mov     eax, dr3
    mov     [edx + CPUMCPU.Host.dr3], eax

    ; load hyper DR0-7
    mov     ebx, [edx + CPUMCPU.Hyper.dr]
    mov     dr0, ebx
    mov     ecx, [edx + CPUMCPU.Hyper.dr + 8*1]
    mov     dr1, ecx
    mov     eax, [edx + CPUMCPU.Hyper.dr + 8*2]
    mov     dr2, eax
    mov     ebx, [edx + CPUMCPU.Hyper.dr + 8*3]
    mov     dr3, ebx
    ;mov     eax, [edx + CPUMCPU.Hyper.dr + 8*6]
    mov     ecx, 0ffff0ff0h
    mov     dr6, ecx
    mov     eax, [edx + CPUMCPU.Hyper.dr + 8*7]
    mov     dr7, eax
    jmp     htg_debug_regs_guest_no

ENDPROC vmmR0HostToGuestAsm


;;
; Trampoline for doing a call when starting the hyper visor execution.
;
; Push any arguments to the routine.
; Push the argument frame size (cArg * 4).
; Push the call target (_cdecl convention).
; Push the address of this routine.
;
;
ALIGNCODE(16)
BEGINPROC vmmGCCallTrampoline
%ifdef DEBUG_STUFF
    COM_S_CHAR 'c'
    COM_S_CHAR 't'
    COM_S_CHAR '!'
%endif

    ; call routine
    pop     eax                         ; call address
    mov     esi, edx                    ; save edx
    pop     edi                         ; argument count.
%ifdef DEBUG_STUFF
    COM_S_PRINT ';eax='
    COM_S_DWORD_REG eax
    COM_S_CHAR ';'
%endif
    call    eax                         ; do call
    add     esp, edi                    ; cleanup stack

    ; return to the host context.
    push    byte 0                      ; eip
    mov     edx, esi                    ; CPUM pointer

%ifdef DEBUG_STUFF
    COM_S_CHAR '`'
%endif
    jmp     NAME(VMMGCGuestToHostAsm)   ; eax = returncode.
ENDPROC vmmGCCallTrampoline



;;
; The C interface.
;
ALIGNCODE(16)
BEGINPROC vmmGCGuestToHost
%ifdef DEBUG_STUFF
    push    esi
    COM_NEWLINE
    DEBUG_CHAR('b')
    DEBUG_CHAR('a')
    DEBUG_CHAR('c')
    DEBUG_CHAR('k')
    DEBUG_CHAR('!')
    COM_NEWLINE
    pop     esi
%endif
    mov     eax, [esp + 4]
    jmp     NAME(VMMGCGuestToHostAsm)
ENDPROC vmmGCGuestToHost


;;
; VMMGCGuestToHostAsmGuestCtx
;
; Switches from Guest Context to Host Context.
; Of course it's only called from within the GC.
;
; @param    eax     Return code.
; @param    esp + 4 Pointer to CPUMCTXCORE.
;
; @remark   ASSUMES interrupts disabled.
;
ALIGNCODE(16)
BEGINPROC VMMGCGuestToHostAsmGuestCtx
    DEBUG_CHAR('~')

%ifdef VBOX_WITH_STATISTICS
    FIXUP FIX_GC_VM_OFF, 1, VM.StatTotalInGC
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_STOP edx

    FIXUP FIX_GC_VM_OFF, 1, VM.StatTotalGCToQemu
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_START edx

    FIXUP FIX_GC_VM_OFF, 1, VM.StatSwitcherToHC
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_START edx
%endif

    ;
    ; Load the CPUMCPU pointer.
    ;
    FIXUP FIX_GC_CPUMCPU_OFF, 1, 0
    mov     edx, 0ffffffffh

    ; Skip return address (assumes called!)
    lea     esp, [esp + 4]

    ;
    ; Guest Context (assumes esp now points to CPUMCTXCORE structure).
    ;
    ; general purpose registers.
    push    eax

    mov     eax, [esp + 4 + CPUMCTXCORE.eax]
    mov     [edx + CPUMCPU.Guest.eax], eax
    mov     eax, [esp + 4 + CPUMCTXCORE.ecx]
    mov     [edx + CPUMCPU.Guest.ecx], eax
    mov     eax, [esp + 4 + CPUMCTXCORE.edx]
    mov     [edx + CPUMCPU.Guest.edx], eax
    mov     eax, [esp + 4 + CPUMCTXCORE.ebx]
    mov     [edx + CPUMCPU.Guest.ebx], eax
    mov     eax, [esp + 4 + CPUMCTXCORE.esp]
    mov     [edx + CPUMCPU.Guest.esp], eax
    mov     eax, [esp + 4 + CPUMCTXCORE.ebp]
    mov     [edx + CPUMCPU.Guest.ebp], eax
    mov     eax, [esp + 4 + CPUMCTXCORE.esi]
    mov     [edx + CPUMCPU.Guest.esi], eax
    mov     eax, [esp + 4 + CPUMCTXCORE.edi]
    mov     [edx + CPUMCPU.Guest.edi], eax
    mov     eax, dword [esp + 4 + CPUMCTXCORE.es]
    mov     dword [edx + CPUMCPU.Guest.es], eax
    mov     eax, dword [esp + 4 + CPUMCTXCORE.cs]
    mov     dword [edx + CPUMCPU.Guest.cs], eax
    mov     eax, dword [esp + 4 + CPUMCTXCORE.ss]
    mov     dword [edx + CPUMCPU.Guest.ss], eax
    mov     eax, dword [esp + 4 + CPUMCTXCORE.ds]
    mov     dword [edx + CPUMCPU.Guest.ds], eax
    mov     eax, dword [esp + 4 + CPUMCTXCORE.fs]
    mov     dword [edx + CPUMCPU.Guest.fs], eax
    mov     eax, dword [esp + 4 + CPUMCTXCORE.gs]
    mov     dword [edx + CPUMCPU.Guest.gs], eax
    mov     eax, [esp + 4 + CPUMCTXCORE.eflags]
    mov     dword [edx + CPUMCPU.Guest.eflags], eax
    mov     eax, [esp + 4 + CPUMCTXCORE.eip]
    mov     dword [edx + CPUMCPU.Guest.eip], eax
    pop     eax

    add     esp, CPUMCTXCORE_size      ; skip CPUMCTXCORE structure

    jmp     vmmGCGuestToHostAsm_EIPDone
ENDPROC VMMGCGuestToHostAsmGuestCtx


;;
; VMMGCGuestToHostAsmHyperCtx
;
; This is an alternative entry point which we'll be using
; when the we have the hypervisor context and need to save
; that before going to the host.
;
; This is typically useful when abandoning the hypervisor
; because of a trap and want the trap state to be saved.
;
; @param    eax     Return code.
; @param    ecx     Points to CPUMCTXCORE.
; @uses     eax,edx,ecx
ALIGNCODE(16)
BEGINPROC VMMGCGuestToHostAsmHyperCtx
    DEBUG_CHAR('#')

%ifdef VBOX_WITH_STATISTICS
    FIXUP FIX_GC_VM_OFF, 1, VM.StatTotalInGC
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_STOP edx

    FIXUP FIX_GC_VM_OFF, 1, VM.StatTotalGCToQemu
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_START edx

    FIXUP FIX_GC_VM_OFF, 1, VM.StatSwitcherToHC
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_START edx
%endif

    ;
    ; Load the CPUM pointer.
    ;
    FIXUP FIX_GC_CPUMCPU_OFF, 1, 0
    mov     edx, 0ffffffffh

    push    eax                         ; save return code.
    ; general purpose registers
    mov     eax, [ecx + CPUMCTXCORE.edi]
    mov     [edx + CPUMCPU.Hyper.edi], eax
    mov     eax, [ecx + CPUMCTXCORE.esi]
    mov     [edx + CPUMCPU.Hyper.esi], eax
    mov     eax, [ecx + CPUMCTXCORE.ebp]
    mov     [edx + CPUMCPU.Hyper.ebp], eax
    mov     eax, [ecx + CPUMCTXCORE.eax]
    mov     [edx + CPUMCPU.Hyper.eax], eax
    mov     eax, [ecx + CPUMCTXCORE.ebx]
    mov     [edx + CPUMCPU.Hyper.ebx], eax
    mov     eax, [ecx + CPUMCTXCORE.edx]
    mov     [edx + CPUMCPU.Hyper.edx], eax
    mov     eax, [ecx + CPUMCTXCORE.ecx]
    mov     [edx + CPUMCPU.Hyper.ecx], eax
    mov     eax, [ecx + CPUMCTXCORE.esp]
    mov     [edx + CPUMCPU.Hyper.esp], eax
    ; selectors
    mov     eax, [ecx + CPUMCTXCORE.ss]
    mov     [edx + CPUMCPU.Hyper.ss], eax
    mov     eax, [ecx + CPUMCTXCORE.gs]
    mov     [edx + CPUMCPU.Hyper.gs], eax
    mov     eax, [ecx + CPUMCTXCORE.fs]
    mov     [edx + CPUMCPU.Hyper.fs], eax
    mov     eax, [ecx + CPUMCTXCORE.es]
    mov     [edx + CPUMCPU.Hyper.es], eax
    mov     eax, [ecx + CPUMCTXCORE.ds]
    mov     [edx + CPUMCPU.Hyper.ds], eax
    mov     eax, [ecx + CPUMCTXCORE.cs]
    mov     [edx + CPUMCPU.Hyper.cs], eax
    ; flags
    mov     eax, [ecx + CPUMCTXCORE.eflags]
    mov     [edx + CPUMCPU.Hyper.eflags], eax
    ; eip
    mov     eax, [ecx + CPUMCTXCORE.eip]
    mov     [edx + CPUMCPU.Hyper.eip], eax
    ; jump to common worker code.
    pop     eax                         ; restore return code.
    jmp     vmmGCGuestToHostAsm_SkipHyperRegs

ENDPROC VMMGCGuestToHostAsmHyperCtx


;;
; VMMGCGuestToHostAsm
;
; This is an alternative entry point which we'll be using
; when the we have saved the guest state already or we haven't
; been messing with the guest at all.
;
; @param    eax     Return code.
; @uses     eax, edx, ecx (or it may use them in the future)
;
ALIGNCODE(16)
BEGINPROC VMMGCGuestToHostAsm
    DEBUG_CHAR('%')

%ifdef VBOX_WITH_STATISTICS
    FIXUP FIX_GC_VM_OFF, 1, VM.StatTotalInGC
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_STOP edx

    FIXUP FIX_GC_VM_OFF, 1, VM.StatTotalGCToQemu
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_START edx

    FIXUP FIX_GC_VM_OFF, 1, VM.StatSwitcherToHC
    mov     edx, 0ffffffffh
    STAM_PROFILE_ADV_START edx
%endif

    ;
    ; Load the CPUMCPU pointer.
    ;
    FIXUP FIX_GC_CPUMCPU_OFF, 1, 0
    mov     edx, 0ffffffffh

    pop     dword [edx + CPUMCPU.Hyper.eip] ; call return from stack
    jmp short vmmGCGuestToHostAsm_EIPDone

ALIGNCODE(16)
vmmGCGuestToHostAsm_EIPDone:
    ; general registers which we care about.
    mov     dword [edx + CPUMCPU.Hyper.ebx], ebx
    mov     dword [edx + CPUMCPU.Hyper.esi], esi
    mov     dword [edx + CPUMCPU.Hyper.edi], edi
    mov     dword [edx + CPUMCPU.Hyper.ebp], ebp
    mov     dword [edx + CPUMCPU.Hyper.esp], esp

    ; special registers which may change.
vmmGCGuestToHostAsm_SkipHyperRegs:
    ; str     [edx + CPUMCPU.Hyper.tr] - double fault only, and it won't be right then either.
    sldt    [edx + CPUMCPU.Hyper.ldtr]

    ; No need to save CRx here. They are set dynamically according to Guest/Host requirements.
    ; FPU context is saved before restore of host saving (another) branch.

%ifdef VBOX_WITH_NMI
    ;
    ; Disarm K7 NMI.
    ;
    mov     esi, edx
    mov     edi, eax

    xor     edx, edx
    xor     eax, eax
    mov     ecx, MSR_K7_EVNTSEL0
    wrmsr

    mov     eax, edi
    mov     edx, esi
%endif


    ;;
    ;; Load Intermediate memory context.
    ;;
    mov     edi, eax                    ; save return code in EDI (careful with COM_DWORD_REG from here on!)
    mov     ecx, [edx + CPUMCPU.Host.cr3]
    FIXUP SWITCHER_FIX_INTER_CR3_GC, 1
    mov     eax, 0ffffffffh
    mov     cr3, eax
    DEBUG_CHAR('?')

    ;; We're now in intermediate memory context!
%ifdef NEED_ID
    ;;
    ;; Jump to identity mapped location
    ;;
    FIXUP FIX_GC_2_ID_NEAR_REL, 1, NAME(IDExitTarget) - NAME(Start)
    jmp near NAME(IDExitTarget)

    ; We're now on identity mapped pages!
ALIGNCODE(16)
GLOBALNAME IDExitTarget
    DEBUG_CHAR('1')
    mov     edx, cr4
%ifdef NEED_PAE_ON_32BIT_HOST
    and     edx, ~X86_CR4_PAE
%else
    or      edx, X86_CR4_PAE
%endif
    mov     eax, cr0
    and     eax, (~X86_CR0_PG) & 0xffffffff ; prevent yasm warning
    mov     cr0, eax
    DEBUG_CHAR('2')
    mov     cr4, edx
    FIXUP SWITCHER_FIX_INTER_CR3_HC, 1
    mov     edx, 0ffffffffh
    mov     cr3, edx
    or      eax, X86_CR0_PG
    DEBUG_CHAR('3')
    mov     cr0, eax
    DEBUG_CHAR('4')

    ;;
    ;; Jump to HC mapping.
    ;;
    FIXUP FIX_ID_2_HC_NEAR_REL, 1, NAME(HCExitTarget) - NAME(Start)
    jmp     near NAME(HCExitTarget)
%else
    ;;
    ;; Jump to HC mapping.
    ;;
    FIXUP FIX_GC_2_HC_NEAR_REL, 1, NAME(HCExitTarget) - NAME(Start)
    jmp     near NAME(HCExitTarget)
%endif


    ;
    ; When we arrive here we're at the host context
    ; mapping of the switcher code.
    ;
ALIGNCODE(16)
GLOBALNAME HCExitTarget
    DEBUG_CHAR('9')
    ; load final cr3
    mov     cr3, ecx
    DEBUG_CHAR('@')


    ;;
    ;; Restore Host context.
    ;;
    ; Load CPUM pointer into edx
    FIXUP FIX_HC_CPUM_OFF, 1, 0
    mov     edx, 0ffffffffh
    CPUMCPU_FROM_CPUM(edx)
    ; activate host gdt and idt
    lgdt    [edx + CPUMCPU.Host.gdtr]
    DEBUG_CHAR('0')
    lidt    [edx + CPUMCPU.Host.idtr]
    DEBUG_CHAR('1')
    ; Restore TSS selector; must mark it as not busy before using ltr (!)
%if 1 ; ASSUME that this is supposed to be 'BUSY'. (saves 20-30 ticks on the T42p)
    movzx   eax, word [edx + CPUMCPU.Host.tr]          ; eax <- TR
    and     al, 0F8h                                ; mask away TI and RPL bits, get descriptor offset.
    add     eax, [edx + CPUMCPU.Host.gdtr + 2]         ; eax <- GDTR.address + descriptor offset.
    and     dword [eax + 4], ~0200h                 ; clear busy flag (2nd type2 bit)
    ltr     word [edx + CPUMCPU.Host.tr]
%else
    movzx   eax, word [edx + CPUMCPU.Host.tr]          ; eax <- TR
    and     al, 0F8h                                ; mask away TI and RPL bits, get descriptor offset.
    add     eax, [edx + CPUMCPU.Host.gdtr + 2]         ; eax <- GDTR.address + descriptor offset.
    mov     ecx, [eax + 4]                          ; ecx <- 2nd descriptor dword
    mov     ebx, ecx                                ; save orginal value
    and     ecx, ~0200h                             ; clear busy flag (2nd type2 bit)
    mov     [eax + 4], ecx                          ; not using xchg here is paranoia..
    ltr     word [edx + CPUMCPU.Host.tr]
    xchg    [eax + 4], ebx                          ; using xchg is paranoia too...
%endif
    ; activate ldt
    DEBUG_CHAR('2')
    lldt    [edx + CPUMCPU.Host.ldtr]
    ; Restore segment registers
    mov     eax, [edx + CPUMCPU.Host.ds]
    mov     ds, eax
    mov     eax, [edx + CPUMCPU.Host.es]
    mov     es, eax
    mov     eax, [edx + CPUMCPU.Host.fs]
    mov     fs, eax
    mov     eax, [edx + CPUMCPU.Host.gs]
    mov     gs, eax
    ; restore stack
    lss     esp, [edx + CPUMCPU.Host.esp]


    FIXUP FIX_NO_SYSENTER_JMP, 0, gth_sysenter_no - NAME(Start) ; this will insert a jmp gth_sysenter_no if host doesn't use sysenter.
    ; restore MSR_IA32_SYSENTER_CS register.
    mov     ecx, MSR_IA32_SYSENTER_CS
    mov     eax, [edx + CPUMCPU.Host.SysEnter.cs]
    mov     ebx, [edx + CPUMCPU.Host.SysEnter.cs + 4]
    xchg    edx, ebx                    ; save/load edx
    wrmsr                               ; MSR[ecx] <- edx:eax
    xchg    edx, ebx                    ; restore edx
    jmp short gth_sysenter_no

ALIGNCODE(16)
gth_sysenter_no:

    FIXUP FIX_NO_SYSCALL_JMP, 0, gth_syscall_no - NAME(Start) ; this will insert a jmp gth_syscall_no if host doesn't use syscall.
    ; set MSR_K6_EFER_SCE.
    mov     ebx, edx                    ; save edx
    mov     ecx, MSR_K6_EFER
    rdmsr
    or      eax, MSR_K6_EFER_SCE
    wrmsr
    mov     edx, ebx                    ; restore edx
    jmp short gth_syscall_no

ALIGNCODE(16)
gth_syscall_no:

    ; Restore FPU if guest has used it.
    ; Using fxrstor should ensure that we're not causing unwanted exception on the host.
    mov     esi, [edx + CPUMCPU.fUseFlags] ; esi == use flags.
    test    esi, CPUM_USED_FPU
    jz near gth_fpu_no
    mov     ecx, cr0
    and     ecx, ~(X86_CR0_TS | X86_CR0_EM)
    mov     cr0, ecx

    FIXUP FIX_NO_FXSAVE_JMP, 0, gth_no_fxsave - NAME(Start) ; this will insert a jmp gth_no_fxsave if fxsave isn't supported.
    fxsave  [edx + CPUMCPU.Guest.fpu]
    fxrstor [edx + CPUMCPU.Host.fpu]
    jmp near gth_fpu_no

gth_no_fxsave:
    fnsave  [edx + CPUMCPU.Guest.fpu]
    mov     eax, [edx + CPUMCPU.Host.fpu]     ; control word
    not     eax                            ; 1 means exception ignored (6 LS bits)
    and     eax, byte 03Fh                 ; 6 LS bits only
    test    eax, [edx + CPUMCPU.Host.fpu + 4] ; status word
    jz      gth_no_exceptions_pending

    ; technically incorrect, but we certainly don't want any exceptions now!!
    and     dword [edx + CPUMCPU.Host.fpu + 4], ~03Fh

gth_no_exceptions_pending:
    frstor  [edx + CPUMCPU.Host.fpu]
    jmp short gth_fpu_no

ALIGNCODE(16)
gth_fpu_no:

    ; Control registers.
    ; Would've liked to have these highere up in case of crashes, but
    ; the fpu stuff must be done before we restore cr0.
    mov     ecx, [edx + CPUMCPU.Host.cr4]
    mov     cr4, ecx
    mov     ecx, [edx + CPUMCPU.Host.cr0]
    mov     cr0, ecx
    ;mov     ecx, [edx + CPUMCPU.Host.cr2] ; assumes this is waste of time.
    ;mov     cr2, ecx

    ; restore debug registers (if modified) (esi must still be fUseFlags!)
    ; (must be done after cr4 reload because of the debug extension.)
    test    esi, CPUM_USE_DEBUG_REGS | CPUM_USE_DEBUG_REGS_HOST
    jz short gth_debug_regs_no
    jmp     gth_debug_regs_restore
gth_debug_regs_no:

    ; restore general registers.
    mov     eax, edi                    ; restore return code. eax = return code !!
    mov     edi, [edx + CPUMCPU.Host.edi]
    mov     esi, [edx + CPUMCPU.Host.esi]
    mov     ebx, [edx + CPUMCPU.Host.ebx]
    mov     ebp, [edx + CPUMCPU.Host.ebp]
    push    dword [edx + CPUMCPU.Host.eflags]
    popfd

%ifdef DEBUG_STUFF
;    COM_S_CHAR '4'
%endif
    retf

;;
; Detour for restoring the host debug registers.
; edx and edi must be preserved.
gth_debug_regs_restore:
    DEBUG_S_CHAR('d')
    xor     eax, eax
    mov     dr7, eax                    ; paranoia or not?
    test    esi, CPUM_USE_DEBUG_REGS
    jz short gth_debug_regs_dr7
    DEBUG_S_CHAR('r')
    mov     eax, [edx + CPUMCPU.Host.dr0]
    mov     dr0, eax
    mov     ebx, [edx + CPUMCPU.Host.dr1]
    mov     dr1, ebx
    mov     ecx, [edx + CPUMCPU.Host.dr2]
    mov     dr2, ecx
    mov     eax, [edx + CPUMCPU.Host.dr3]
    mov     dr3, eax
gth_debug_regs_dr7:
    mov     ebx, [edx + CPUMCPU.Host.dr6]
    mov     dr6, ebx
    mov     ecx, [edx + CPUMCPU.Host.dr7]
    mov     dr7, ecx
    jmp     gth_debug_regs_no

ENDPROC VMMGCGuestToHostAsm


GLOBALNAME End
;
; The description string (in the text section).
;
NAME(Description):
    db SWITCHER_DESCRIPTION
    db 0

extern NAME(Relocate)

;
; End the fixup records.
;
BEGINDATA
    db FIX_THE_END                      ; final entry.
GLOBALNAME FixupsEnd

;;
; The switcher definition structure.
ALIGNDATA(16)
GLOBALNAME Def
    istruc VMMSWITCHERDEF
        at VMMSWITCHERDEF.pvCode,                       RTCCPTR_DEF NAME(Start)
        at VMMSWITCHERDEF.pvFixups,                     RTCCPTR_DEF NAME(Fixups)
        at VMMSWITCHERDEF.pszDesc,                      RTCCPTR_DEF NAME(Description)
        at VMMSWITCHERDEF.pfnRelocate,                  RTCCPTR_DEF NAME(Relocate)
        at VMMSWITCHERDEF.enmType,                      dd SWITCHER_TYPE
        at VMMSWITCHERDEF.cbCode,                       dd NAME(End)                        - NAME(Start)
        at VMMSWITCHERDEF.offR0HostToGuest,             dd NAME(vmmR0HostToGuest)           - NAME(Start)
        at VMMSWITCHERDEF.offGCGuestToHost,             dd NAME(vmmGCGuestToHost)           - NAME(Start)
        at VMMSWITCHERDEF.offGCCallTrampoline,          dd NAME(vmmGCCallTrampoline)        - NAME(Start)
        at VMMSWITCHERDEF.offGCGuestToHostAsm,          dd NAME(VMMGCGuestToHostAsm)        - NAME(Start)
        at VMMSWITCHERDEF.offGCGuestToHostAsmHyperCtx,  dd NAME(VMMGCGuestToHostAsmHyperCtx)- NAME(Start)
        at VMMSWITCHERDEF.offGCGuestToHostAsmGuestCtx,  dd NAME(VMMGCGuestToHostAsmGuestCtx)- NAME(Start)
        ; disasm help
        at VMMSWITCHERDEF.offHCCode0,                   dd 0
%ifdef NEED_ID
        at VMMSWITCHERDEF.cbHCCode0,                    dd NAME(IDEnterTarget)              - NAME(Start)
%else
        at VMMSWITCHERDEF.cbHCCode0,                    dd NAME(FarJmpGCTarget)             - NAME(Start)
%endif
        at VMMSWITCHERDEF.offHCCode1,                   dd NAME(HCExitTarget)               - NAME(Start)
        at VMMSWITCHERDEF.cbHCCode1,                    dd NAME(End)                        - NAME(HCExitTarget)
%ifdef NEED_ID
        at VMMSWITCHERDEF.offIDCode0,                   dd NAME(IDEnterTarget)              - NAME(Start)
        at VMMSWITCHERDEF.cbIDCode0,                    dd NAME(FarJmpGCTarget)             - NAME(IDEnterTarget)
        at VMMSWITCHERDEF.offIDCode1,                   dd NAME(IDExitTarget)               - NAME(Start)
        at VMMSWITCHERDEF.cbIDCode1,                    dd NAME(HCExitTarget)               - NAME(IDExitTarget)
%else
        at VMMSWITCHERDEF.offIDCode0,                   dd 0
        at VMMSWITCHERDEF.cbIDCode0,                    dd 0
        at VMMSWITCHERDEF.offIDCode1,                   dd 0
        at VMMSWITCHERDEF.cbIDCode1,                    dd 0
%endif
        at VMMSWITCHERDEF.offGCCode,                    dd NAME(FarJmpGCTarget)             - NAME(Start)
%ifdef NEED_ID
        at VMMSWITCHERDEF.cbGCCode,                     dd NAME(IDExitTarget)               - NAME(FarJmpGCTarget)
%else
        at VMMSWITCHERDEF.cbGCCode,                     dd NAME(HCExitTarget)               - NAME(FarJmpGCTarget)
%endif

    iend