/* $Id: GIMKvm.cpp 99220 2023-03-30 12:40:46Z vboxsync $ */ /** @file * GIM - Guest Interface Manager, KVM implementation. */ /* * Copyright (C) 2015-2023 Oracle and/or its affiliates. * * This file is part of VirtualBox base platform packages, as * available from https://www.virtualbox.org. * * This program is free software; you can redistribute it and/or * modify it under the terms of the GNU General Public License * as published by the Free Software Foundation, in version 3 of the * License. * * This program is distributed in the hope that it will be useful, but * WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU * General Public License for more details. * * You should have received a copy of the GNU General Public License * along with this program; if not, see . * * SPDX-License-Identifier: GPL-3.0-only */ /********************************************************************************************************************************* * Header Files * *********************************************************************************************************************************/ #define LOG_GROUP LOG_GROUP_GIM #include #include #include #include #include #include #include "GIMInternal.h" #include #include #include #include #include #include #include #include /********************************************************************************************************************************* * Defined Constants And Macros * *********************************************************************************************************************************/ /** * GIM KVM saved-state version. */ #define GIM_KVM_SAVED_STATE_VERSION UINT32_C(1) /** * VBox internal struct. to passback to EMT rendezvous callback while enabling * the KVM wall-clock. */ typedef struct KVMWALLCLOCKINFO { /** Guest physical address of the wall-clock struct. */ RTGCPHYS GCPhysWallClock; } KVMWALLCLOCKINFO; /** Pointer to the wall-clock info. struct. */ typedef KVMWALLCLOCKINFO *PKVMWALLCLOCKINFO; /********************************************************************************************************************************* * Global Variables * *********************************************************************************************************************************/ #ifdef VBOX_WITH_STATISTICS # define GIMKVM_MSRRANGE(a_uFirst, a_uLast, a_szName) \ { (a_uFirst), (a_uLast), kCpumMsrRdFn_Gim, kCpumMsrWrFn_Gim, 0, 0, 0, 0, 0, a_szName, { 0 }, { 0 }, { 0 }, { 0 } } #else # define GIMKVM_MSRRANGE(a_uFirst, a_uLast, a_szName) \ { (a_uFirst), (a_uLast), kCpumMsrRdFn_Gim, kCpumMsrWrFn_Gim, 0, 0, 0, 0, 0, a_szName } #endif /** * Array of MSR ranges supported by KVM. */ static CPUMMSRRANGE const g_aMsrRanges_Kvm[] = { GIMKVM_MSRRANGE(MSR_GIM_KVM_RANGE0_FIRST, MSR_GIM_KVM_RANGE0_LAST, "KVM range 0"), GIMKVM_MSRRANGE(MSR_GIM_KVM_RANGE1_FIRST, MSR_GIM_KVM_RANGE1_LAST, "KVM range 1") }; #undef GIMKVM_MSRRANGE /** * Updates the KVM VCPU system-time structure in guest memory. * * @returns VBox status code. * @param pVM The cross context VM structure. * @param pVCpu The cross context virtual CPU structure. * * @remarks This must be called after the system time MSR value has been updated. */ static int gimR3KvmUpdateSystemTime(PVM pVM, PVMCPU pVCpu) { PGIMKVM pKvm = &pVM->gim.s.u.Kvm; PGIMKVMCPU pKvmCpu = &pVCpu->gim.s.u.KvmCpu; /* * Validate the MSR has the enable bit and the guest's system time struct. address. */ MSR_GIM_KVM_SYSTEM_TIME_IS_ENABLED(pKvmCpu->u64SystemTimeMsr); if (!PGMPhysIsGCPhysNormal(pVM, pKvmCpu->GCPhysSystemTime)) { LogRel(("GIM: KVM: VCPU%3d: Invalid physical addr requested for mapping system-time struct. GCPhysSystemTime=%#RGp\n", pVCpu->idCpu, pKvmCpu->GCPhysSystemTime)); return VERR_GIM_OPERATION_FAILED; } VMSTATE const enmVMState = pVM->enmVMState; bool const fRunning = VMSTATE_IS_RUNNING(enmVMState); Assert(!(pKvmCpu->u32SystemTimeVersion & UINT32_C(1))); /* * Construct a system-time struct. */ GIMKVMSYSTEMTIME SystemTime; RT_ZERO(SystemTime); SystemTime.u32Version = pKvmCpu->u32SystemTimeVersion + !!fRunning; SystemTime.u64NanoTS = pKvmCpu->uVirtNanoTS; SystemTime.u64Tsc = pKvmCpu->uTsc; SystemTime.fFlags = pKvmCpu->fSystemTimeFlags | GIM_KVM_SYSTEM_TIME_FLAGS_TSC_STABLE; /* * How the guest calculates the system time (nanoseconds): * * tsc = rdtsc - SysTime.u64Tsc * if (SysTime.i8TscShift >= 0) * tsc <<= i8TscShift; * else * tsc >>= -i8TscShift; * time = ((tsc * SysTime.u32TscScale) >> 32) + SysTime.u64NanoTS */ uint64_t u64TscFreq = pKvm->cTscTicksPerSecond; SystemTime.i8TscShift = 0; while (u64TscFreq > 2 * RT_NS_1SEC_64) { u64TscFreq >>= 1; SystemTime.i8TscShift--; } uint32_t uTscFreqLo = (uint32_t)u64TscFreq; while (uTscFreqLo <= RT_NS_1SEC) { uTscFreqLo <<= 1; SystemTime.i8TscShift++; } SystemTime.u32TscScale = ASMDivU64ByU32RetU32(RT_NS_1SEC_64 << 32, uTscFreqLo); /* * For informational purposes, back-calculate the exact TSC frequency the guest will see. * Note that the frequency is in kHz, not Hz, since that's what Linux uses. */ uint64_t uTscKHz = (RT_NS_1MS_64 << 32) / SystemTime.u32TscScale; if (SystemTime.i8TscShift < 0) uTscKHz <<= -SystemTime.i8TscShift; else uTscKHz >>= SystemTime.i8TscShift; /* * Update guest memory with the system-time struct. * * We update the struct with an incremented, odd version field to indicate to the guest * that the memory is being updated concurrently by the host and it should discard any * data from this struct when it reads an odd version. * * When the VM is not running, we don't need to do this two step update for obvious * reasons and so we skip it. */ if (fRunning) Assert(SystemTime.u32Version & UINT32_C(1)); else Assert(!(SystemTime.u32Version & UINT32_C(1))); int rc = PGMPhysSimpleWriteGCPhys(pVM, pKvmCpu->GCPhysSystemTime, &SystemTime, sizeof(GIMKVMSYSTEMTIME)); if (RT_SUCCESS(rc)) { LogRel(("GIM: KVM: VCPU%3d: Enabled system-time struct. at %#RGp - u32TscScale=%#RX32 i8TscShift=%d uVersion=%#RU32 " "fFlags=%#x uTsc=%#RX64 uVirtNanoTS=%#RX64 TscKHz=%RU64\n", pVCpu->idCpu, pKvmCpu->GCPhysSystemTime, SystemTime.u32TscScale, SystemTime.i8TscShift, SystemTime.u32Version + !!fRunning, SystemTime.fFlags, pKvmCpu->uTsc, pKvmCpu->uVirtNanoTS, uTscKHz)); TMR3CpuTickParavirtEnable(pVM); } else { LogRel(("GIM: KVM: VCPU%3d: Failed to write system-time struct. at %#RGp. rc=%Rrc\n", pVCpu->idCpu, pKvmCpu->GCPhysSystemTime, rc)); } if (fRunning) { ++SystemTime.u32Version; Assert(!(SystemTime.u32Version & UINT32_C(1))); rc = PGMPhysSimpleWriteGCPhys(pVM, pKvmCpu->GCPhysSystemTime + RT_UOFFSETOF(GIMKVMSYSTEMTIME, u32Version), &SystemTime.u32Version, sizeof(SystemTime.u32Version)); if (RT_FAILURE(rc)) { LogRel(("GIM: KVM: VCPU%3d: Failed to write system-time struct. while updating version field at %#RGp. rc=%Rrc\n", pVCpu->idCpu, pKvmCpu->GCPhysSystemTime, rc)); return rc; } /* Update the version so our next write will start with an even value. */ pKvmCpu->u32SystemTimeVersion += 2; } return rc; } /** * Initializes the KVM GIM provider. * * @returns VBox status code. * @param pVM The cross context VM structure. */ VMMR3_INT_DECL(int) gimR3KvmInit(PVM pVM) { AssertReturn(pVM, VERR_INVALID_PARAMETER); AssertReturn(pVM->gim.s.enmProviderId == GIMPROVIDERID_KVM, VERR_INTERNAL_ERROR_5); int rc; PGIMKVM pKvm = &pVM->gim.s.u.Kvm; /* * Determine interface capabilities based on the version. */ if (!pVM->gim.s.u32Version) { /* Basic features. */ pKvm->uBaseFeat = 0 | GIM_KVM_BASE_FEAT_CLOCK_OLD //| GIM_KVM_BASE_FEAT_NOP_IO_DELAY //| GIM_KVM_BASE_FEAT_MMU_OP | GIM_KVM_BASE_FEAT_CLOCK //| GIM_KVM_BASE_FEAT_ASYNC_PF //| GIM_KVM_BASE_FEAT_STEAL_TIME //| GIM_KVM_BASE_FEAT_PV_EOI | GIM_KVM_BASE_FEAT_PV_UNHALT ; /* Rest of the features are determined in gimR3KvmInitCompleted(). */ } /* * Expose HVP (Hypervisor Present) bit to the guest. */ CPUMR3SetGuestCpuIdFeature(pVM, CPUMCPUIDFEATURE_HVP); /* * Modify the standard hypervisor leaves for KVM. */ CPUMCPUIDLEAF HyperLeaf; RT_ZERO(HyperLeaf); HyperLeaf.uLeaf = UINT32_C(0x40000000); HyperLeaf.uEax = UINT32_C(0x40000001); /* Minimum value for KVM is 0x40000001. */ HyperLeaf.uEbx = 0x4B4D564B; /* 'KVMK' */ HyperLeaf.uEcx = 0x564B4D56; /* 'VMKV' */ HyperLeaf.uEdx = 0x0000004D; /* 'M000' */ rc = CPUMR3CpuIdInsert(pVM, &HyperLeaf); AssertLogRelRCReturn(rc, rc); /* * Add KVM specific leaves. */ HyperLeaf.uLeaf = UINT32_C(0x40000001); HyperLeaf.uEax = pKvm->uBaseFeat; HyperLeaf.uEbx = 0; /* Reserved */ HyperLeaf.uEcx = 0; /* Reserved */ HyperLeaf.uEdx = 0; /* Reserved */ rc = CPUMR3CpuIdInsert(pVM, &HyperLeaf); AssertLogRelRCReturn(rc, rc); /* * Insert all MSR ranges of KVM. */ for (unsigned i = 0; i < RT_ELEMENTS(g_aMsrRanges_Kvm); i++) { rc = CPUMR3MsrRangesInsert(pVM, &g_aMsrRanges_Kvm[i]); AssertLogRelRCReturn(rc, rc); } /* * Setup hypercall and #UD handling. * Note! We always need to trap VMCALL/VMMCALL hypercall using #UDs for raw-mode VMs. */ for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++) EMSetHypercallInstructionsEnabled(pVM->apCpusR3[idCpu], true); size_t cbHypercall = 0; rc = GIMQueryHypercallOpcodeBytes(pVM, pKvm->abOpcodeNative, sizeof(pKvm->abOpcodeNative), &cbHypercall, &pKvm->uOpcodeNative); AssertLogRelRCReturn(rc, rc); AssertLogRelReturn(cbHypercall == sizeof(pKvm->abOpcodeNative), VERR_GIM_IPE_1); pKvm->fTrapXcptUD = pKvm->uOpcodeNative != OP_VMCALL; return VINF_SUCCESS; } /** * Initializes remaining bits of the KVM provider. * * This is called after initializing HM and almost all other VMM components. * * @returns VBox status code. * @param pVM The cross context VM structure. */ VMMR3_INT_DECL(int) gimR3KvmInitCompleted(PVM pVM) { PGIMKVM pKvm = &pVM->gim.s.u.Kvm; pKvm->cTscTicksPerSecond = TMCpuTicksPerSecond(pVM); if (TMR3CpuTickIsFixedRateMonotonic(pVM, true /* fWithParavirtEnabled */)) { /** @todo We might want to consider just enabling this bit *always*. As far * as I can see in the Linux guest, the "TSC_STABLE" bit is only * translated as a "monotonic" bit which even in Async systems we * -should- be reporting a strictly monotonic TSC to the guest. */ pKvm->uBaseFeat |= GIM_KVM_BASE_FEAT_TSC_STABLE; CPUMCPUIDLEAF HyperLeaf; RT_ZERO(HyperLeaf); HyperLeaf.uLeaf = UINT32_C(0x40000001); HyperLeaf.uEax = pKvm->uBaseFeat; HyperLeaf.uEbx = 0; HyperLeaf.uEcx = 0; HyperLeaf.uEdx = 0; int rc = CPUMR3CpuIdInsert(pVM, &HyperLeaf); AssertLogRelRCReturn(rc, rc); } return VINF_SUCCESS; } /** * Terminates the KVM GIM provider. * * @returns VBox status code. * @param pVM The cross context VM structure. */ VMMR3_INT_DECL(int) gimR3KvmTerm(PVM pVM) { gimR3KvmReset(pVM); return VINF_SUCCESS; } /** * This resets KVM provider MSRs and unmaps whatever KVM regions that * the guest may have mapped. * * This is called when the VM is being reset. * * @param pVM The cross context VM structure. * @thread EMT(0) */ VMMR3_INT_DECL(void) gimR3KvmReset(PVM pVM) { VM_ASSERT_EMT0(pVM); LogRel(("GIM: KVM: Resetting MSRs\n")); /* * Reset MSRs. */ PGIMKVM pKvm = &pVM->gim.s.u.Kvm; pKvm->u64WallClockMsr = 0; for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++) { PGIMKVMCPU pKvmCpu = &pVM->apCpusR3[idCpu]->gim.s.u.KvmCpu; pKvmCpu->u64SystemTimeMsr = 0; pKvmCpu->u32SystemTimeVersion = 0; pKvmCpu->fSystemTimeFlags = 0; pKvmCpu->GCPhysSystemTime = 0; pKvmCpu->uTsc = 0; pKvmCpu->uVirtNanoTS = 0; } } /** * KVM state-save operation. * * @returns VBox status code. * @param pVM The cross context VM structure. * @param pSSM The saved state handle. */ VMMR3_INT_DECL(int) gimR3KvmSave(PVM pVM, PSSMHANDLE pSSM) { PCGIMKVM pKvm = &pVM->gim.s.u.Kvm; /* * Save the KVM SSM version. */ SSMR3PutU32(pSSM, GIM_KVM_SAVED_STATE_VERSION); /* * Save per-VCPU data. */ for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++) { PCGIMKVMCPU pKvmCpu = &pVM->apCpusR3[idCpu]->gim.s.u.KvmCpu; SSMR3PutU64(pSSM, pKvmCpu->u64SystemTimeMsr); SSMR3PutU64(pSSM, pKvmCpu->uTsc); SSMR3PutU64(pSSM, pKvmCpu->uVirtNanoTS); SSMR3PutGCPhys(pSSM, pKvmCpu->GCPhysSystemTime); SSMR3PutU32(pSSM, pKvmCpu->u32SystemTimeVersion); SSMR3PutU8(pSSM, pKvmCpu->fSystemTimeFlags); } /* * Save per-VM data. */ SSMR3PutU64(pSSM, pKvm->u64WallClockMsr); return SSMR3PutU32(pSSM, pKvm->uBaseFeat); } /** * KVM state-load operation, final pass. * * @returns VBox status code. * @param pVM The cross context VM structure. * @param pSSM The saved state handle. */ VMMR3_INT_DECL(int) gimR3KvmLoad(PVM pVM, PSSMHANDLE pSSM) { /* * Load the KVM SSM version first. */ uint32_t uKvmSavedStatVersion; int rc = SSMR3GetU32(pSSM, &uKvmSavedStatVersion); AssertRCReturn(rc, rc); if (uKvmSavedStatVersion != GIM_KVM_SAVED_STATE_VERSION) return SSMR3SetLoadError(pSSM, VERR_SSM_UNSUPPORTED_DATA_UNIT_VERSION, RT_SRC_POS, N_("Unsupported KVM saved-state version %u (expected %u)."), uKvmSavedStatVersion, GIM_KVM_SAVED_STATE_VERSION); /* * Update the TSC frequency from TM. */ PGIMKVM pKvm = &pVM->gim.s.u.Kvm; pKvm->cTscTicksPerSecond = TMCpuTicksPerSecond(pVM); /* * Load per-VCPU data. */ for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++) { PVMCPU pVCpu = pVM->apCpusR3[idCpu]; PGIMKVMCPU pKvmCpu = &pVCpu->gim.s.u.KvmCpu; SSMR3GetU64(pSSM, &pKvmCpu->u64SystemTimeMsr); SSMR3GetU64(pSSM, &pKvmCpu->uTsc); SSMR3GetU64(pSSM, &pKvmCpu->uVirtNanoTS); SSMR3GetGCPhys(pSSM, &pKvmCpu->GCPhysSystemTime); SSMR3GetU32(pSSM, &pKvmCpu->u32SystemTimeVersion); rc = SSMR3GetU8(pSSM, &pKvmCpu->fSystemTimeFlags); AssertRCReturn(rc, rc); /* Enable the system-time struct. if necessary. */ /** @todo update guest struct only if cTscTicksPerSecond doesn't match host * anymore. */ if (MSR_GIM_KVM_SYSTEM_TIME_IS_ENABLED(pKvmCpu->u64SystemTimeMsr)) { Assert(!TMVirtualIsTicking(pVM)); /* paranoia. */ Assert(!TMCpuTickIsTicking(pVCpu)); gimR3KvmUpdateSystemTime(pVM, pVCpu); } } /* * Load per-VM data. */ SSMR3GetU64(pSSM, &pKvm->u64WallClockMsr); rc = SSMR3GetU32(pSSM, &pKvm->uBaseFeat); AssertRCReturn(rc, rc); return VINF_SUCCESS; } /** * Disables the KVM system-time struct. * * @returns VBox status code. * @param pVM The cross context VM structure. */ VMMR3_INT_DECL(int) gimR3KvmDisableSystemTime(PVM pVM) { TMR3CpuTickParavirtDisable(pVM); return VINF_SUCCESS; } /** * @callback_method_impl{PFNVMMEMTRENDEZVOUS, * Worker for gimR3KvmEnableWallClock} */ static DECLCALLBACK(VBOXSTRICTRC) gimR3KvmEnableWallClockCallback(PVM pVM, PVMCPU pVCpu, void *pvUser) { PKVMWALLCLOCKINFO pWallClockInfo = (PKVMWALLCLOCKINFO)pvUser; AssertPtr(pWallClockInfo); RTGCPHYS GCPhysWallClock = pWallClockInfo->GCPhysWallClock; RT_NOREF1(pVCpu); /* * Read the wall-clock version (sequence) from the guest. */ uint32_t uVersion; Assert(PGMPhysIsGCPhysNormal(pVM, GCPhysWallClock)); int rc = PGMPhysSimpleReadGCPhys(pVM, &uVersion, GCPhysWallClock, sizeof(uVersion)); if (RT_FAILURE(rc)) { LogRel(("GIM: KVM: Failed to read wall-clock struct. version at %#RGp. rc=%Rrc\n", GCPhysWallClock, rc)); return rc; } /* * Ensure the version is incrementally even. */ /* faster: uVersion = (uVersion | 1) + 1; */ if (!(uVersion & 1)) ++uVersion; ++uVersion; /* * Update wall-clock guest struct. with UTC information. */ RTTIMESPEC TimeSpec; int32_t iSec; int32_t iNano; TMR3UtcNow(pVM, &TimeSpec); RTTimeSpecGetSecondsAndNano(&TimeSpec, &iSec, &iNano); GIMKVMWALLCLOCK WallClock; RT_ZERO(WallClock); AssertCompile(sizeof(uVersion) == sizeof(WallClock.u32Version)); WallClock.u32Version = uVersion; WallClock.u32Sec = iSec; WallClock.u32Nano = iNano; /* * Write out the wall-clock struct. to guest memory. */ Assert(!(WallClock.u32Version & 1)); rc = PGMPhysSimpleWriteGCPhys(pVM, GCPhysWallClock, &WallClock, sizeof(GIMKVMWALLCLOCK)); if (RT_SUCCESS(rc)) LogRel(("GIM: KVM: Enabled wall-clock struct. at %#RGp - u32Sec=%u u32Nano=%u uVersion=%#RU32\n", GCPhysWallClock, WallClock.u32Sec, WallClock.u32Nano, WallClock.u32Version)); else LogRel(("GIM: KVM: Failed to write wall-clock struct. at %#RGp. rc=%Rrc\n", GCPhysWallClock, rc)); return rc; } /** * Enables the KVM wall-clock structure. * * Since the wall-clock can be read by any VCPU but it is a global struct. in * guest-memory, we do an EMT rendezvous here to be on the safe side. The * alternative is to use an MMIO2 region and use the WallClock.u32Version field * for transactional update. However, this MSR is rarely written to (typically * once during bootup) it's currently not a performance issue especially since * we're already in ring-3. If we really wanted better performance in this code * path, we should be doing it in ring-0 with transactional update while make * sure there is only 1 writer as well. * * @returns VBox status code. * @param pVM The cross context VM structure. * @param GCPhysWallClock Where the guest wall-clock structure is located. * * @remarks Don't do any release assertions here, these can be triggered by * guest R0 code. */ VMMR3_INT_DECL(int) gimR3KvmEnableWallClock(PVM pVM, RTGCPHYS GCPhysWallClock) { KVMWALLCLOCKINFO WallClockInfo; WallClockInfo.GCPhysWallClock = GCPhysWallClock; return VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, gimR3KvmEnableWallClockCallback, &WallClockInfo); } /** * Enables the KVM system time structure. * * This can be done concurrently because the guest memory being updated is per-VCPU * and the struct even has a "version" field which needs to be incremented * before/after altering guest memory to allow concurrent updates from the host. * Hence this is not being done in an EMT rendezvous. It -is- done in ring-3 since * we call into ring-3 only TM code in the end. * * @returns VBox status code. * @param pVM The cross context VM structure. * @param pVCpu The cross context virtual CPU structure. * @param uMsrSystemTime The system time MSR value being written. */ VMMR3_INT_DECL(int) gimR3KvmEnableSystemTime(PVMCC pVM, PVMCPUCC pVCpu, uint64_t uMsrSystemTime) { Assert(uMsrSystemTime & MSR_GIM_KVM_SYSTEM_TIME_ENABLE_BIT); PGIMKVM pKvm = &pVM->gim.s.u.Kvm; PGIMKVMCPU pKvmCpu = &pVCpu->gim.s.u.KvmCpu; /* * Update the system-time struct. * The system-time structs are usually placed at a different guest address for each VCPU. */ pKvmCpu->uTsc = TMCpuTickGetNoCheck(pVCpu); pKvmCpu->uVirtNanoTS = ASMMultU64ByU32DivByU32(pKvmCpu->uTsc, RT_NS_1SEC, pKvm->cTscTicksPerSecond); pKvmCpu->u64SystemTimeMsr = uMsrSystemTime; pKvmCpu->GCPhysSystemTime = MSR_GIM_KVM_SYSTEM_TIME_GUEST_GPA(uMsrSystemTime); int rc = gimR3KvmUpdateSystemTime(pVM, pVCpu); if (RT_FAILURE(rc)) { pKvmCpu->u64SystemTimeMsr = 0; /* We shouldn't throw a #GP(0) here for buggy guests (neither does KVM apparently), see @bugref{8627}. */ } return rc; }