source: vbox/trunk/src/VBox/VMM/VMMAll/HMAll.cpp@ 57076

/* $Id: HMAll.cpp 57066 2015-07-24 10:11:17Z vboxsync $ */
/** @file
 * HM - All contexts.
 */

/*
 * Copyright (C) 2006-2015 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.
 */


/*******************************************************************************
*   Header Files                                                               *
*******************************************************************************/
#define LOG_GROUP LOG_GROUP_HM
#include <VBox/vmm/hm.h>
#include <VBox/vmm/pgm.h>
#include "HMInternal.h"
#include <VBox/vmm/vm.h>
#include <VBox/vmm/hm_vmx.h>
#include <VBox/vmm/hm_svm.h>
#include <VBox/err.h>
#include <VBox/log.h>
#include <iprt/param.h>
#include <iprt/assert.h>
#include <iprt/asm.h>
#include <iprt/string.h>
#include <iprt/thread.h>
#include <iprt/x86.h>
#include <iprt/asm-amd64-x86.h>


/**
 * Checks whether HM (VT-x/AMD-V) is being used by this VM.
 *
 * @retval  @c true if used.
 * @retval  @c false if software virtualization (raw-mode) is used.
 * @param   pVM         The cross context VM structure.
 * @sa      HMIsEnabled, HMR3IsEnabled
 * @internal
 */
VMMDECL(bool) HMIsEnabledNotMacro(PVM pVM)
{
    Assert(pVM->fHMEnabledFixed);
    return pVM->fHMEnabled;
}


/**
 * Queues a guest page for invalidation.
 *
 * @returns VBox status code.
 * @param   pVCpu       Pointer to the VMCPU.
 * @param   GCVirt      Page to invalidate
 */
static void hmQueueInvlPage(PVMCPU pVCpu, RTGCPTR GCVirt)
{
    /* Nothing to do if a TLB flush is already pending */
    if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_TLB_FLUSH))
        return;
#if 1
    VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
    NOREF(GCVirt);
#else
    /* Be very careful when activating this code! */
    if (iPage == RT_ELEMENTS(pVCpu->hm.s.TlbShootdown.aPages))
        VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
    else
        VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_SHOOTDOWN);
#endif
}


/**
 * Invalidates a guest page.
 *
 * @returns VBox status code.
 * @param   pVCpu       Pointer to the VMCPU.
 * @param   GCVirt      Page to invalidate
 */
VMM_INT_DECL(int) HMInvalidatePage(PVMCPU pVCpu, RTGCPTR GCVirt)
{
    STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushPageManual);
#ifdef IN_RING0
    PVM pVM = pVCpu->CTX_SUFF(pVM);
    if (pVM->hm.s.vmx.fSupported)
        return VMXR0InvalidatePage(pVM, pVCpu, GCVirt);

    Assert(pVM->hm.s.svm.fSupported);
    return SVMR0InvalidatePage(pVM, pVCpu, GCVirt);

#else
    hmQueueInvlPage(pVCpu, GCVirt);
    return VINF_SUCCESS;
#endif
}


/**
 * Flushes the guest TLB.
 *
 * @returns VBox status code.
 * @param   pVCpu       Pointer to the VMCPU.
 */
VMM_INT_DECL(int) HMFlushTLB(PVMCPU pVCpu)
{
    LogFlow(("HMFlushTLB\n"));

    VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
    STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbManual);
    return VINF_SUCCESS;
}


#ifdef IN_RING0
/**
 * Dummy RTMpOnSpecific handler since RTMpPokeCpu couldn't be used.
 *
 */
static DECLCALLBACK(void) hmFlushHandler(RTCPUID idCpu, void *pvUser1, void *pvUser2)
{
    NOREF(idCpu); NOREF(pvUser1); NOREF(pvUser2);
    return;
}

/**
 * Wrapper for RTMpPokeCpu to deal with VERR_NOT_SUPPORTED.
 */
static void hmR0PokeCpu(PVMCPU pVCpu, RTCPUID idHostCpu)
{
    uint32_t cWorldSwitchExits = ASMAtomicUoReadU32(&pVCpu->hm.s.cWorldSwitchExits);

    STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatPoke, x);
    int rc = RTMpPokeCpu(idHostCpu);
    STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatPoke, x);

    /* Not implemented on some platforms (Darwin, Linux kernel < 2.6.19); fall
       back to a less efficient implementation (broadcast). */
    if (rc == VERR_NOT_SUPPORTED)
    {
        STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatSpinPoke, z);
        /* synchronous. */
        RTMpOnSpecific(idHostCpu, hmFlushHandler, 0, 0);
        STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatSpinPoke, z);
    }
    else
    {
        if (rc == VINF_SUCCESS)
            STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatSpinPoke, z);
        else
            STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatSpinPokeFailed, z);

/** @todo If more than one CPU is going to be poked, we could optimize this
 *        operation by poking them first and wait afterwards.  Would require
 *        recording who to poke and their current cWorldSwitchExits values,
 *        that's something not suitable for stack... So, pVCpu->hm.s.something
 *        then. */
        /* Spin until the VCPU has switched back (poking is async). */
        while (   ASMAtomicUoReadBool(&pVCpu->hm.s.fCheckedTLBFlush)
               && cWorldSwitchExits == ASMAtomicUoReadU32(&pVCpu->hm.s.cWorldSwitchExits))
            ASMNopPause();

        if (rc == VINF_SUCCESS)
            STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatSpinPoke, z);
        else
            STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatSpinPokeFailed, z);
    }
}
#endif /* IN_RING0 */


#ifndef IN_RC
/**
 * Poke an EMT so it can perform the appropriate TLB shootdowns.
 *
 * @param   pVCpu               The handle of the virtual CPU to poke.
 * @param   fAccountFlushStat   Whether to account the call to
 *                              StatTlbShootdownFlush or StatTlbShootdown.
 */
static void hmPokeCpuForTlbFlush(PVMCPU pVCpu, bool fAccountFlushStat)
{
    if (ASMAtomicUoReadBool(&pVCpu->hm.s.fCheckedTLBFlush))
    {
        if (fAccountFlushStat)
            STAM_COUNTER_INC(&pVCpu->hm.s.StatTlbShootdownFlush);
        else
            STAM_COUNTER_INC(&pVCpu->hm.s.StatTlbShootdown);
#ifdef IN_RING0
        RTCPUID idHostCpu = pVCpu->hm.s.idEnteredCpu;
        if (idHostCpu != NIL_RTCPUID)
            hmR0PokeCpu(pVCpu, idHostCpu);
#else
        VMR3NotifyCpuFFU(pVCpu->pUVCpu, VMNOTIFYFF_FLAGS_POKE);
#endif
    }
    else
        STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushPageManual);
}


/**
 * Invalidates a guest page on all VCPUs.
 *
 * @returns VBox status code.
 * @param   pVM         Pointer to the VM.
 * @param   GCVirt      Page to invalidate.
 */
VMM_INT_DECL(int) HMInvalidatePageOnAllVCpus(PVM pVM, RTGCPTR GCVirt)
{
    /*
     * The VT-x/AMD-V code will be flushing TLB each time a VCPU migrates to a different
     * host CPU, see hmR0VmxFlushTaggedTlbBoth() and hmR0SvmFlushTaggedTlb().
     *
     * This is the reason why we do not care about thread preemption here and just
     * execute HMInvalidatePage() assuming it might be the 'right' CPU.
     */
    VMCPUID idCurCpu = VMMGetCpuId(pVM);
    STAM_COUNTER_INC(&pVM->aCpus[idCurCpu].hm.s.StatFlushPage);

    for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
    {
        PVMCPU pVCpu = &pVM->aCpus[idCpu];

        /* Nothing to do if a TLB flush is already pending; the VCPU should
           have already been poked if it were active. */
        if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_TLB_FLUSH))
            continue;

        if (pVCpu->idCpu == idCurCpu)
            HMInvalidatePage(pVCpu, GCVirt);
        else
        {
            hmQueueInvlPage(pVCpu, GCVirt);
            hmPokeCpuForTlbFlush(pVCpu, false /* fAccountFlushStat */);
        }
    }

    return VINF_SUCCESS;
}


/**
 * Flush the TLBs of all VCPUs.
 *
 * @returns VBox status code.
 * @param   pVM       Pointer to the VM.
 */
VMM_INT_DECL(int) HMFlushTLBOnAllVCpus(PVM pVM)
{
    if (pVM->cCpus == 1)
        return HMFlushTLB(&pVM->aCpus[0]);

    VMCPUID idThisCpu = VMMGetCpuId(pVM);

    STAM_COUNTER_INC(&pVM->aCpus[idThisCpu].hm.s.StatFlushTlb);

    for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
    {
        PVMCPU pVCpu = &pVM->aCpus[idCpu];

        /* Nothing to do if a TLB flush is already pending; the VCPU should
           have already been poked if it were active. */
        if (!VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_TLB_FLUSH))
        {
            VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
            if (idThisCpu != idCpu)
                hmPokeCpuForTlbFlush(pVCpu, true /* fAccountFlushStat */);
        }
    }

    return VINF_SUCCESS;
}
#endif /* !IN_RC */

/**
 * Checks if nested paging is enabled.
 *
 * @returns true if nested paging is active, false otherwise.
 * @param   pVM         Pointer to the VM.
 *
 * @remarks Works before hmR3InitFinalizeR0.
 */
VMM_INT_DECL(bool) HMIsNestedPagingActive(PVM pVM)
{
    return HMIsEnabled(pVM) && pVM->hm.s.fNestedPaging;
}


/**
 * Checks if both nested paging and unhampered guest execution are enabled.
 *
 * The almost complete guest execution in hardware is only applicable to VT-x.
 *
 * @returns true if we have both enabled, otherwise false.
 * @param   pVM         Pointer to the VM.
 *
 * @remarks Works before hmR3InitFinalizeR0.
 */
VMM_INT_DECL(bool) HMAreNestedPagingAndFullGuestExecEnabled(PVM pVM)
{
    return HMIsEnabled(pVM)
        && pVM->hm.s.fNestedPaging
        && (   pVM->hm.s.vmx.fUnrestrictedGuest
            || pVM->hm.s.svm.fSupported);
}


/**
 * Checks if this VM is long-mode capable.
 *
 * @returns true if long mode is allowed, false otherwise.
 * @param   pUVM        The user mode VM handle.
 */
VMM_INT_DECL(bool) HMIsLongModeAllowed(PVM pVM)
{
    return HMIsEnabled(pVM) && pVM->hm.s.fAllow64BitGuests;
}


/**
 * Checks if MSR bitmaps are available. It is assumed that when it's available
 * it will be used as well.
 *
 * @returns true if MSR bitmaps are available, false otherwise.
 * @param   pVM         Pointer to the VM.
 */
VMM_INT_DECL(bool) HMAreMsrBitmapsAvailable(PVM pVM)
{
    if (HMIsEnabled(pVM))
    {
        if (pVM->hm.s.svm.fSupported)
            return true;

        if (   pVM->hm.s.vmx.fSupported
            && (pVM->hm.s.vmx.Msrs.VmxProcCtls.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_MSR_BITMAPS))
        {
            return true;
        }
    }
    return false;
}


/**
 * Return the shadow paging mode for nested paging/ept
 *
 * @returns shadow paging mode
 * @param   pVM         Pointer to the VM.
 */
VMM_INT_DECL(PGMMODE) HMGetShwPagingMode(PVM pVM)
{
    Assert(HMIsNestedPagingActive(pVM));
    if (pVM->hm.s.svm.fSupported)
        return PGMMODE_NESTED;

    Assert(pVM->hm.s.vmx.fSupported);
    return PGMMODE_EPT;
}

/**
 * Invalidates a guest page by physical address
 *
 * NOTE: Assumes the current instruction references this physical page though a virtual address!!
 *
 * @returns VBox status code.
 * @param   pVM         Pointer to the VM.
 * @param   GCPhys      Page to invalidate
 */
VMM_INT_DECL(int) HMInvalidatePhysPage(PVM pVM, RTGCPHYS GCPhys)
{
    if (!HMIsNestedPagingActive(pVM))
        return VINF_SUCCESS;

#ifdef IN_RING0
    if (pVM->hm.s.vmx.fSupported)
    {
        VMCPUID idThisCpu = VMMGetCpuId(pVM);

        for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
        {
            PVMCPU pVCpu = &pVM->aCpus[idCpu];

            if (idThisCpu == idCpu)
            {
                /** @todo r=ramshankar: Intel does not support flushing by guest physical
                 *        address either. See comment in VMXR0InvalidatePhysPage(). Fix this. */
                VMXR0InvalidatePhysPage(pVM, pVCpu, GCPhys);
            }
            else
            {
                VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
                hmPokeCpuForTlbFlush(pVCpu, true /*fAccountFlushStat*/);
            }
        }
        return VINF_SUCCESS;
    }

    /* AMD-V doesn't support invalidation with guest physical addresses; see
       comment in SVMR0InvalidatePhysPage. */
    Assert(pVM->hm.s.svm.fSupported);
#else
    NOREF(GCPhys);
#endif

    HMFlushTLBOnAllVCpus(pVM);
    return VINF_SUCCESS;
}

/**
 * Checks if an interrupt event is currently pending.
 *
 * @returns Interrupt event pending state.
 * @param   pVM         Pointer to the VM.
 */
VMM_INT_DECL(bool) HMHasPendingIrq(PVM pVM)
{
    PVMCPU pVCpu = VMMGetCpu(pVM);
    return !!pVCpu->hm.s.Event.fPending;
}


/**
 * Return the PAE PDPE entries.
 *
 * @returns Pointer to the PAE PDPE array.
 * @param   pVCpu       Pointer to the VMCPU.
 */
VMM_INT_DECL(PX86PDPE) HMGetPaePdpes(PVMCPU pVCpu)
{
    return &pVCpu->hm.s.aPdpes[0];
}


/**
 * Checks if the current AMD CPU is subject to erratum 170 "In SVM mode,
 * incorrect code bytes may be fetched after a world-switch".
 *
 * @param   pu32Family      Where to store the CPU family (can be NULL).
 * @param   pu32Model       Where to store the CPU model (can be NULL).
 * @param   pu32Stepping    Where to store the CPU stepping (can be NULL).
 * @returns true if the erratum applies, false otherwise.
 */
VMM_INT_DECL(int) HMAmdIsSubjectToErratum170(uint32_t *pu32Family, uint32_t *pu32Model, uint32_t *pu32Stepping)
{
    /*
     * Erratum 170 which requires a forced TLB flush for each world switch:
     * See AMD spec. "Revision Guide for AMD NPT Family 0Fh Processors".
     *
     * All BH-G1/2 and DH-G1/2 models include a fix:
     * Athlon X2:   0x6b 1/2
     *              0x68 1/2
     * Athlon 64:   0x7f 1
     *              0x6f 2
     * Sempron:     0x7f 1/2
     *              0x6f 2
     *              0x6c 2
     *              0x7c 2
     * Turion 64:   0x68 2
     */
    uint32_t u32Dummy;
    uint32_t u32Version, u32Family, u32Model, u32Stepping, u32BaseFamily;
    ASMCpuId(1, &u32Version, &u32Dummy, &u32Dummy, &u32Dummy);
    u32BaseFamily = (u32Version >> 8) & 0xf;
    u32Family     = u32BaseFamily + (u32BaseFamily == 0xf ? ((u32Version >> 20) & 0x7f) : 0);
    u32Model      = ((u32Version >> 4) & 0xf);
    u32Model      = u32Model | ((u32BaseFamily == 0xf ? (u32Version >> 16) & 0x0f : 0) << 4);
    u32Stepping   = u32Version & 0xf;

    bool fErratumApplies = false;
    if (   u32Family == 0xf
        && !((u32Model == 0x68 || u32Model == 0x6b || u32Model == 0x7f) && u32Stepping >= 1)
        && !((u32Model == 0x6f || u32Model == 0x6c || u32Model == 0x7c) && u32Stepping >= 2))
    {
        fErratumApplies = true;
    }

    if (pu32Family)
        *pu32Family   = u32Family;
    if (pu32Model)
        *pu32Model    = u32Model;
    if (pu32Stepping)
        *pu32Stepping = u32Stepping;

    return fErratumApplies;
}


/**
 * Sets or clears the single instruction flag.
 *
 * When set, HM will try its best to return to ring-3 after executing a single
 * instruction.  This can be used for debugging.  See also
 * EMR3HmSingleInstruction.
 *
 * @returns The old flag state.
 * @param   pVCpu               Pointer to the cross context CPU structure of
 *                              the calling EMT.
 * @param   fEnable             The new flag state.
 */
VMM_INT_DECL(bool) HMSetSingleInstruction(PVMCPU pVCpu, bool fEnable)
{
    VMCPU_ASSERT_EMT(pVCpu);
    bool fOld = pVCpu->hm.s.fSingleInstruction;
    pVCpu->hm.s.fSingleInstruction = fEnable;
    return fOld;
}


/**
 * Notifies HM that paravirtualized hypercalls are now enabled.
 *
 * @param   pVCpu   Pointer to the VMCPU.
 */
VMM_INT_DECL(void) HMHypercallsEnable(PVMCPU pVCpu)
{
    pVCpu->hm.s.fHypercallsEnabled = true;
}


/**
 * Notifies HM that paravirtualized hypercalls are now disabled.
 *
 * @param   pVCpu   Pointer to the VMCPU.
 */
VMM_INT_DECL(void) HMHypercallsDisable(PVMCPU pVCpu)
{
    pVCpu->hm.s.fHypercallsEnabled = false;
}


/**
 * Notifies HM that GIM provider wants to trap #UD.
 *
 * @param   pVCpu   Pointer to the VMCPU.
 */
VMM_INT_DECL(void) HMTrapXcptUDForGIMEnable(PVMCPU pVCpu)
{
    pVCpu->hm.s.fGIMTrapXcptUD = true;
    HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_XCPT_INTERCEPTS);
}


/**
 * Notifies HM that GIM provider no longer wants to trap #UD.
 *
 * @param   pVCpu   Pointer to the VMCPU.
 */
VMM_INT_DECL(void) HMTrapXcptUDForGIMDisable(PVMCPU pVCpu)
{
    pVCpu->hm.s.fGIMTrapXcptUD = false;
    HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_XCPT_INTERCEPTS);
}