diff options
Diffstat (limited to 'sys/amd64/vmm')
-rw-r--r-- | sys/amd64/vmm/amd/amdv.c | 143 | ||||
-rw-r--r-- | sys/amd64/vmm/amd/npt.c | 87 | ||||
-rw-r--r-- | sys/amd64/vmm/amd/npt.h | 36 | ||||
-rw-r--r-- | sys/amd64/vmm/amd/svm.c | 2092 | ||||
-rw-r--r-- | sys/amd64/vmm/amd/svm.h | 54 | ||||
-rw-r--r-- | sys/amd64/vmm/amd/svm_genassym.c | 48 | ||||
-rw-r--r-- | sys/amd64/vmm/amd/svm_msr.c | 136 | ||||
-rw-r--r-- | sys/amd64/vmm/amd/svm_msr.h | 44 | ||||
-rw-r--r-- | sys/amd64/vmm/amd/svm_softc.h | 113 | ||||
-rw-r--r-- | sys/amd64/vmm/amd/svm_support.S | 115 | ||||
-rw-r--r-- | sys/amd64/vmm/amd/vmcb.c | 442 | ||||
-rw-r--r-- | sys/amd64/vmm/amd/vmcb.h | 332 | ||||
-rw-r--r-- | sys/amd64/vmm/intel/vmx.c | 3 | ||||
-rw-r--r-- | sys/amd64/vmm/io/vlapic.c | 8 | ||||
-rw-r--r-- | sys/amd64/vmm/vmm.c | 26 | ||||
-rw-r--r-- | sys/amd64/vmm/vmm_instruction_emul.c | 11 | ||||
-rw-r--r-- | sys/amd64/vmm/x86.c | 75 |
17 files changed, 3604 insertions, 161 deletions
diff --git a/sys/amd64/vmm/amd/amdv.c b/sys/amd64/vmm/amd/amdv.c index ca639df..acb3a3d 100644 --- a/sys/amd64/vmm/amd/amdv.c +++ b/sys/amd64/vmm/amd/amdv.c @@ -38,149 +38,6 @@ __FBSDID("$FreeBSD$"); #include "io/iommu.h" static int -amdv_init(int ipinum) -{ - - printf("amdv_init: not implemented\n"); - return (ENXIO); -} - -static int -amdv_cleanup(void) -{ - - printf("amdv_cleanup: not implemented\n"); - return (ENXIO); -} - -static void -amdv_resume(void) -{ -} - -static void * -amdv_vminit(struct vm *vm, struct pmap *pmap) -{ - - printf("amdv_vminit: not implemented\n"); - return (NULL); -} - -static int -amdv_vmrun(void *arg, int vcpu, register_t rip, struct pmap *pmap, - void *rptr, void *sptr) -{ - - printf("amdv_vmrun: not implemented\n"); - return (ENXIO); -} - -static void -amdv_vmcleanup(void *arg) -{ - - printf("amdv_vmcleanup: not implemented\n"); - return; -} - -static int -amdv_getreg(void *arg, int vcpu, int regnum, uint64_t *retval) -{ - - printf("amdv_getreg: not implemented\n"); - return (EINVAL); -} - -static int -amdv_setreg(void *arg, int vcpu, int regnum, uint64_t val) -{ - - printf("amdv_setreg: not implemented\n"); - return (EINVAL); -} - -static int -amdv_getdesc(void *vmi, int vcpu, int num, struct seg_desc *desc) -{ - - printf("amdv_get_desc: not implemented\n"); - return (EINVAL); -} - -static int -amdv_setdesc(void *vmi, int vcpu, int num, struct seg_desc *desc) -{ - - printf("amdv_get_desc: not implemented\n"); - return (EINVAL); -} - -static int -amdv_getcap(void *arg, int vcpu, int type, int *retval) -{ - - printf("amdv_getcap: not implemented\n"); - return (EINVAL); -} - -static int -amdv_setcap(void *arg, int vcpu, int type, int val) -{ - - printf("amdv_setcap: not implemented\n"); - return (EINVAL); -} - -static struct vmspace * -amdv_vmspace_alloc(vm_offset_t min, vm_offset_t max) -{ - - printf("amdv_vmspace_alloc: not implemented\n"); - return (NULL); -} - -static void -amdv_vmspace_free(struct vmspace *vmspace) -{ - - printf("amdv_vmspace_free: not implemented\n"); - return; -} - -static struct vlapic * -amdv_vlapic_init(void *arg, int vcpuid) -{ - - panic("amdv_vlapic_init: not implmented"); -} - -static void -amdv_vlapic_cleanup(void *arg, struct vlapic *vlapic) -{ - - panic("amdv_vlapic_cleanup: not implemented"); -} - -struct vmm_ops vmm_ops_amd = { - amdv_init, - amdv_cleanup, - amdv_resume, - amdv_vminit, - amdv_vmrun, - amdv_vmcleanup, - amdv_getreg, - amdv_setreg, - amdv_getdesc, - amdv_setdesc, - amdv_getcap, - amdv_setcap, - amdv_vmspace_alloc, - amdv_vmspace_free, - amdv_vlapic_init, - amdv_vlapic_cleanup, -}; - -static int amd_iommu_init(void) { diff --git a/sys/amd64/vmm/amd/npt.c b/sys/amd64/vmm/amd/npt.c new file mode 100644 index 0000000..bebb4d5 --- /dev/null +++ b/sys/amd64/vmm/amd/npt.c @@ -0,0 +1,87 @@ +/*- + * Copyright (c) 2013 Anish Gupta (akgupt3@gmail.com) + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice unmodified, this list of conditions, and the following + * disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +#include <sys/cdefs.h> +__FBSDID("$FreeBSD$"); + +#include <sys/param.h> +#include <sys/kernel.h> +#include <sys/systm.h> +#include <sys/sysctl.h> + +#include <vm/vm.h> +#include <vm/pmap.h> +#include <vm/vm_extern.h> + +#include <machine/pmap.h> + +#include "npt.h" + +SYSCTL_DECL(_hw_vmm); +SYSCTL_NODE(_hw_vmm, OID_AUTO, npt, CTLFLAG_RW, NULL, NULL); + +static int npt_flags; +SYSCTL_INT(_hw_vmm_npt, OID_AUTO, pmap_flags, CTLFLAG_RD, + &npt_flags, 0, NULL); + +#define NPT_IPIMASK 0xFF + +/* + * AMD nested page table init. + */ +int +svm_npt_init(int ipinum) +{ + int enable_superpage = 1; + + npt_flags = ipinum & NPT_IPIMASK; + TUNABLE_INT_FETCH("hw.vmm.npt.enable_superpage", &enable_superpage); + if (enable_superpage) + npt_flags |= PMAP_PDE_SUPERPAGE; + + return (0); +} + +static int +npt_pinit(pmap_t pmap) +{ + + return (pmap_pinit_type(pmap, PT_RVI, npt_flags)); +} + +struct vmspace * +svm_npt_alloc(vm_offset_t min, vm_offset_t max) +{ + + return (vmspace_alloc(min, max, npt_pinit)); +} + +void +svm_npt_free(struct vmspace *vmspace) +{ + + vmspace_free(vmspace); +} diff --git a/sys/amd64/vmm/amd/npt.h b/sys/amd64/vmm/amd/npt.h new file mode 100644 index 0000000..5966474 --- /dev/null +++ b/sys/amd64/vmm/amd/npt.h @@ -0,0 +1,36 @@ +/*- + * Copyright (c) 2013 Anish Gupta (akgupt3@gmail.com) + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice unmodified, this list of conditions, and the following + * disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + * $FreeBSD$ + */ + +#ifndef _SVM_NPT_H_ +#define _SVM_NPT_H_ + +int svm_npt_init(int ipinum); +struct vmspace *svm_npt_alloc(vm_offset_t min, vm_offset_t max); +void svm_npt_free(struct vmspace *vmspace); + +#endif /* _SVM_NPT_H_ */ diff --git a/sys/amd64/vmm/amd/svm.c b/sys/amd64/vmm/amd/svm.c new file mode 100644 index 0000000..e59f723 --- /dev/null +++ b/sys/amd64/vmm/amd/svm.c @@ -0,0 +1,2092 @@ +/*- + * Copyright (c) 2013, Anish Gupta (akgupt3@gmail.com) + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice unmodified, this list of conditions, and the following + * disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +#include <sys/cdefs.h> +__FBSDID("$FreeBSD$"); + +#include <sys/param.h> +#include <sys/systm.h> +#include <sys/smp.h> +#include <sys/kernel.h> +#include <sys/malloc.h> +#include <sys/pcpu.h> +#include <sys/proc.h> +#include <sys/sysctl.h> + +#include <vm/vm.h> +#include <vm/pmap.h> + +#include <machine/cpufunc.h> +#include <machine/psl.h> +#include <machine/pmap.h> +#include <machine/md_var.h> +#include <machine/specialreg.h> +#include <machine/smp.h> +#include <machine/vmm.h> +#include <machine/vmm_instruction_emul.h> + +#include "vmm_lapic.h" +#include "vmm_stat.h" +#include "vmm_ktr.h" +#include "vmm_ioport.h" +#include "vatpic.h" +#include "vlapic.h" +#include "vlapic_priv.h" + +#include "x86.h" +#include "vmcb.h" +#include "svm.h" +#include "svm_softc.h" +#include "svm_msr.h" +#include "npt.h" + +SYSCTL_DECL(_hw_vmm); +SYSCTL_NODE(_hw_vmm, OID_AUTO, svm, CTLFLAG_RW, NULL, NULL); + +/* + * SVM CPUID function 0x8000_000A, edx bit decoding. + */ +#define AMD_CPUID_SVM_NP BIT(0) /* Nested paging or RVI */ +#define AMD_CPUID_SVM_LBR BIT(1) /* Last branch virtualization */ +#define AMD_CPUID_SVM_SVML BIT(2) /* SVM lock */ +#define AMD_CPUID_SVM_NRIP_SAVE BIT(3) /* Next RIP is saved */ +#define AMD_CPUID_SVM_TSC_RATE BIT(4) /* TSC rate control. */ +#define AMD_CPUID_SVM_VMCB_CLEAN BIT(5) /* VMCB state caching */ +#define AMD_CPUID_SVM_FLUSH_BY_ASID BIT(6) /* Flush by ASID */ +#define AMD_CPUID_SVM_DECODE_ASSIST BIT(7) /* Decode assist */ +#define AMD_CPUID_SVM_PAUSE_INC BIT(10) /* Pause intercept filter. */ +#define AMD_CPUID_SVM_PAUSE_FTH BIT(12) /* Pause filter threshold */ + +#define VMCB_CACHE_DEFAULT (VMCB_CACHE_ASID | \ + VMCB_CACHE_IOPM | \ + VMCB_CACHE_I | \ + VMCB_CACHE_TPR | \ + VMCB_CACHE_CR2 | \ + VMCB_CACHE_CR | \ + VMCB_CACHE_DT | \ + VMCB_CACHE_SEG | \ + VMCB_CACHE_NP) + +static uint32_t vmcb_clean = VMCB_CACHE_DEFAULT; +SYSCTL_INT(_hw_vmm_svm, OID_AUTO, vmcb_clean, CTLFLAG_RDTUN, &vmcb_clean, + 0, NULL); + +static MALLOC_DEFINE(M_SVM, "svm", "svm"); +static MALLOC_DEFINE(M_SVM_VLAPIC, "svm-vlapic", "svm-vlapic"); + +/* Per-CPU context area. */ +extern struct pcpu __pcpu[]; + +static uint32_t svm_feature; /* AMD SVM features. */ +SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, features, CTLFLAG_RD, &svm_feature, 0, + "SVM features advertised by CPUID.8000000AH:EDX"); + +static int disable_npf_assist; +SYSCTL_INT(_hw_vmm_svm, OID_AUTO, disable_npf_assist, CTLFLAG_RWTUN, + &disable_npf_assist, 0, NULL); + +/* Maximum ASIDs supported by the processor */ +static uint32_t nasid; +SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, num_asids, CTLFLAG_RD, &nasid, 0, + "Number of ASIDs supported by this processor"); + +/* Current ASID generation for each host cpu */ +static struct asid asid[MAXCPU]; + +/* + * SVM host state saved area of size 4KB for each core. + */ +static uint8_t hsave[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE); + +static VMM_STAT_AMD(VCPU_EXITINTINFO, "VM exits during event delivery"); +static VMM_STAT_AMD(VCPU_INTINFO_INJECTED, "Events pending at VM entry"); +static VMM_STAT_AMD(VMEXIT_VINTR, "VM exits due to interrupt window"); + +static int svm_setreg(void *arg, int vcpu, int ident, uint64_t val); + +static __inline int +flush_by_asid(void) +{ + + return (svm_feature & AMD_CPUID_SVM_FLUSH_BY_ASID); +} + +static __inline int +decode_assist(void) +{ + + return (svm_feature & AMD_CPUID_SVM_DECODE_ASSIST); +} + +static void +svm_disable(void *arg __unused) +{ + uint64_t efer; + + efer = rdmsr(MSR_EFER); + efer &= ~EFER_SVM; + wrmsr(MSR_EFER, efer); +} + +/* + * Disable SVM on all CPUs. + */ +static int +svm_cleanup(void) +{ + + smp_rendezvous(NULL, svm_disable, NULL, NULL); + return (0); +} + +/* + * Verify that all the features required by bhyve are available. + */ +static int +check_svm_features(void) +{ + u_int regs[4]; + + /* CPUID Fn8000_000A is for SVM */ + do_cpuid(0x8000000A, regs); + svm_feature = regs[3]; + + printf("SVM: Revision %d\n", regs[0] & 0xFF); + printf("SVM: NumASID %u\n", regs[1]); + + nasid = regs[1]; + KASSERT(nasid > 1, ("Insufficient ASIDs for guests: %#x", nasid)); + + printf("SVM: Features 0x%b\n", svm_feature, + "\020" + "\001NP" /* Nested paging */ + "\002LbrVirt" /* LBR virtualization */ + "\003SVML" /* SVM lock */ + "\004NRIPS" /* NRIP save */ + "\005TscRateMsr" /* MSR based TSC rate control */ + "\006VmcbClean" /* VMCB clean bits */ + "\007FlushByAsid" /* Flush by ASID */ + "\010DecodeAssist" /* Decode assist */ + "\011<b8>" + "\012<b9>" + "\013PauseFilter" + "\014<b11>" + "\015PauseFilterThreshold" + "\016AVIC" + ); + + /* bhyve requires the Nested Paging feature */ + if (!(svm_feature & AMD_CPUID_SVM_NP)) { + printf("SVM: Nested Paging feature not available.\n"); + return (ENXIO); + } + + /* bhyve requires the NRIP Save feature */ + if (!(svm_feature & AMD_CPUID_SVM_NRIP_SAVE)) { + printf("SVM: NRIP Save feature not available.\n"); + return (ENXIO); + } + + return (0); +} + +static void +svm_enable(void *arg __unused) +{ + uint64_t efer; + + efer = rdmsr(MSR_EFER); + efer |= EFER_SVM; + wrmsr(MSR_EFER, efer); + + wrmsr(MSR_VM_HSAVE_PA, vtophys(hsave[curcpu])); +} + +/* + * Return 1 if SVM is enabled on this processor and 0 otherwise. + */ +static int +svm_available(void) +{ + uint64_t msr; + + /* Section 15.4 Enabling SVM from APM2. */ + if ((amd_feature2 & AMDID2_SVM) == 0) { + printf("SVM: not available.\n"); + return (0); + } + + msr = rdmsr(MSR_VM_CR); + if ((msr & VM_CR_SVMDIS) != 0) { + printf("SVM: disabled by BIOS.\n"); + return (0); + } + + return (1); +} + +static int +svm_init(int ipinum) +{ + int error, cpu; + + if (!svm_available()) + return (ENXIO); + + error = check_svm_features(); + if (error) + return (error); + + vmcb_clean &= VMCB_CACHE_DEFAULT; + + for (cpu = 0; cpu < MAXCPU; cpu++) { + /* + * Initialize the host ASIDs to their "highest" valid values. + * + * The next ASID allocation will rollover both 'gen' and 'num' + * and start off the sequence at {1,1}. + */ + asid[cpu].gen = ~0UL; + asid[cpu].num = nasid - 1; + } + + svm_msr_init(); + svm_npt_init(ipinum); + + /* Enable SVM on all CPUs */ + smp_rendezvous(NULL, svm_enable, NULL, NULL); + + return (0); +} + +static void +svm_restore(void) +{ + + svm_enable(NULL); +} + +/* Pentium compatible MSRs */ +#define MSR_PENTIUM_START 0 +#define MSR_PENTIUM_END 0x1FFF +/* AMD 6th generation and Intel compatible MSRs */ +#define MSR_AMD6TH_START 0xC0000000UL +#define MSR_AMD6TH_END 0xC0001FFFUL +/* AMD 7th and 8th generation compatible MSRs */ +#define MSR_AMD7TH_START 0xC0010000UL +#define MSR_AMD7TH_END 0xC0011FFFUL + +/* + * Get the index and bit position for a MSR in permission bitmap. + * Two bits are used for each MSR: lower bit for read and higher bit for write. + */ +static int +svm_msr_index(uint64_t msr, int *index, int *bit) +{ + uint32_t base, off; + + *index = -1; + *bit = (msr % 4) * 2; + base = 0; + + if (msr >= MSR_PENTIUM_START && msr <= MSR_PENTIUM_END) { + *index = msr / 4; + return (0); + } + + base += (MSR_PENTIUM_END - MSR_PENTIUM_START + 1); + if (msr >= MSR_AMD6TH_START && msr <= MSR_AMD6TH_END) { + off = (msr - MSR_AMD6TH_START); + *index = (off + base) / 4; + return (0); + } + + base += (MSR_AMD6TH_END - MSR_AMD6TH_START + 1); + if (msr >= MSR_AMD7TH_START && msr <= MSR_AMD7TH_END) { + off = (msr - MSR_AMD7TH_START); + *index = (off + base) / 4; + return (0); + } + + return (EINVAL); +} + +/* + * Allow vcpu to read or write the 'msr' without trapping into the hypervisor. + */ +static void +svm_msr_perm(uint8_t *perm_bitmap, uint64_t msr, bool read, bool write) +{ + int index, bit, error; + + error = svm_msr_index(msr, &index, &bit); + KASSERT(error == 0, ("%s: invalid msr %#lx", __func__, msr)); + KASSERT(index >= 0 && index < SVM_MSR_BITMAP_SIZE, + ("%s: invalid index %d for msr %#lx", __func__, index, msr)); + KASSERT(bit >= 0 && bit <= 6, ("%s: invalid bit position %d " + "msr %#lx", __func__, bit, msr)); + + if (read) + perm_bitmap[index] &= ~(1UL << bit); + + if (write) + perm_bitmap[index] &= ~(2UL << bit); +} + +static void +svm_msr_rw_ok(uint8_t *perm_bitmap, uint64_t msr) +{ + + svm_msr_perm(perm_bitmap, msr, true, true); +} + +static void +svm_msr_rd_ok(uint8_t *perm_bitmap, uint64_t msr) +{ + + svm_msr_perm(perm_bitmap, msr, true, false); +} + +static __inline int +svm_get_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask) +{ + struct vmcb_ctrl *ctrl; + + KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx)); + + ctrl = svm_get_vmcb_ctrl(sc, vcpu); + return (ctrl->intercept[idx] & bitmask ? 1 : 0); +} + +static __inline void +svm_set_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask, + int enabled) +{ + struct vmcb_ctrl *ctrl; + uint32_t oldval; + + KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx)); + + ctrl = svm_get_vmcb_ctrl(sc, vcpu); + oldval = ctrl->intercept[idx]; + + if (enabled) + ctrl->intercept[idx] |= bitmask; + else + ctrl->intercept[idx] &= ~bitmask; + + if (ctrl->intercept[idx] != oldval) { + svm_set_dirty(sc, vcpu, VMCB_CACHE_I); + VCPU_CTR3(sc->vm, vcpu, "intercept[%d] modified " + "from %#x to %#x", idx, oldval, ctrl->intercept[idx]); + } +} + +static __inline void +svm_disable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask) +{ + + svm_set_intercept(sc, vcpu, off, bitmask, 0); +} + +static __inline void +svm_enable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask) +{ + + svm_set_intercept(sc, vcpu, off, bitmask, 1); +} + +static void +vmcb_init(struct svm_softc *sc, int vcpu, uint64_t iopm_base_pa, + uint64_t msrpm_base_pa, uint64_t np_pml4) +{ + struct vmcb_ctrl *ctrl; + struct vmcb_state *state; + uint32_t mask; + int n; + + ctrl = svm_get_vmcb_ctrl(sc, vcpu); + state = svm_get_vmcb_state(sc, vcpu); + + ctrl->iopm_base_pa = iopm_base_pa; + ctrl->msrpm_base_pa = msrpm_base_pa; + + /* Enable nested paging */ + ctrl->np_enable = 1; + ctrl->n_cr3 = np_pml4; + + /* + * Intercept accesses to the control registers that are not shadowed + * in the VMCB - i.e. all except cr0, cr2, cr3, cr4 and cr8. + */ + for (n = 0; n < 16; n++) { + mask = (BIT(n) << 16) | BIT(n); + if (n == 0 || n == 2 || n == 3 || n == 4 || n == 8) + svm_disable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask); + else + svm_enable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask); + } + + /* Intercept Machine Check exceptions. */ + svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(IDT_MC)); + + /* Intercept various events (for e.g. I/O, MSR and CPUID accesses) */ + svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IO); + svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_MSR); + svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_CPUID); + svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INTR); + svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INIT); + svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_NMI); + svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SMI); + svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SHUTDOWN); + svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, + VMCB_INTCPT_FERR_FREEZE); + + /* + * From section "Canonicalization and Consistency Checks" in APMv2 + * the VMRUN intercept bit must be set to pass the consistency check. + */ + svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMRUN); + + /* + * The ASID will be set to a non-zero value just before VMRUN. + */ + ctrl->asid = 0; + + /* + * Section 15.21.1, Interrupt Masking in EFLAGS + * Section 15.21.2, Virtualizing APIC.TPR + * + * This must be set for %rflag and %cr8 isolation of guest and host. + */ + ctrl->v_intr_masking = 1; + + /* Enable Last Branch Record aka LBR for debugging */ + ctrl->lbr_virt_en = 1; + state->dbgctl = BIT(0); + + /* EFER_SVM must always be set when the guest is executing */ + state->efer = EFER_SVM; + + /* Set up the PAT to power-on state */ + state->g_pat = PAT_VALUE(0, PAT_WRITE_BACK) | + PAT_VALUE(1, PAT_WRITE_THROUGH) | + PAT_VALUE(2, PAT_UNCACHED) | + PAT_VALUE(3, PAT_UNCACHEABLE) | + PAT_VALUE(4, PAT_WRITE_BACK) | + PAT_VALUE(5, PAT_WRITE_THROUGH) | + PAT_VALUE(6, PAT_UNCACHED) | + PAT_VALUE(7, PAT_UNCACHEABLE); +} + +/* + * Initialize a virtual machine. + */ +static void * +svm_vminit(struct vm *vm, pmap_t pmap) +{ + struct svm_softc *svm_sc; + struct svm_vcpu *vcpu; + vm_paddr_t msrpm_pa, iopm_pa, pml4_pa; + int i; + + svm_sc = malloc(sizeof (struct svm_softc), M_SVM, M_WAITOK | M_ZERO); + svm_sc->vm = vm; + svm_sc->nptp = (vm_offset_t)vtophys(pmap->pm_pml4); + + /* + * Intercept read and write accesses to all MSRs. + */ + memset(svm_sc->msr_bitmap, 0xFF, sizeof(svm_sc->msr_bitmap)); + + /* + * Access to the following MSRs is redirected to the VMCB when the + * guest is executing. Therefore it is safe to allow the guest to + * read/write these MSRs directly without hypervisor involvement. + */ + svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE); + svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE); + svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE); + + svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR); + svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR); + svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR); + svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK); + svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR); + svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR); + svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR); + svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT); + + svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC); + + /* + * Intercept writes to make sure that the EFER_SVM bit is not cleared. + */ + svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER); + + /* Intercept access to all I/O ports. */ + memset(svm_sc->iopm_bitmap, 0xFF, sizeof(svm_sc->iopm_bitmap)); + + iopm_pa = vtophys(svm_sc->iopm_bitmap); + msrpm_pa = vtophys(svm_sc->msr_bitmap); + pml4_pa = svm_sc->nptp; + for (i = 0; i < VM_MAXCPU; i++) { + vcpu = svm_get_vcpu(svm_sc, i); + vcpu->lastcpu = NOCPU; + vcpu->vmcb_pa = vtophys(&vcpu->vmcb); + vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa); + svm_msr_guest_init(svm_sc, i); + } + return (svm_sc); +} + +static int +svm_cpl(struct vmcb_state *state) +{ + + /* + * From APMv2: + * "Retrieve the CPL from the CPL field in the VMCB, not + * from any segment DPL" + */ + return (state->cpl); +} + +static enum vm_cpu_mode +svm_vcpu_mode(struct vmcb *vmcb) +{ + struct vmcb_segment seg; + struct vmcb_state *state; + int error; + + state = &vmcb->state; + + if (state->efer & EFER_LMA) { + error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg); + KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__, + error)); + + /* + * Section 4.8.1 for APM2, check if Code Segment has + * Long attribute set in descriptor. + */ + if (seg.attrib & VMCB_CS_ATTRIB_L) + return (CPU_MODE_64BIT); + else + return (CPU_MODE_COMPATIBILITY); + } else if (state->cr0 & CR0_PE) { + return (CPU_MODE_PROTECTED); + } else { + return (CPU_MODE_REAL); + } +} + +static enum vm_paging_mode +svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer) +{ + + if ((cr0 & CR0_PG) == 0) + return (PAGING_MODE_FLAT); + if ((cr4 & CR4_PAE) == 0) + return (PAGING_MODE_32); + if (efer & EFER_LME) + return (PAGING_MODE_64); + else + return (PAGING_MODE_PAE); +} + +/* + * ins/outs utility routines + */ +static uint64_t +svm_inout_str_index(struct svm_regctx *regs, int in) +{ + uint64_t val; + + val = in ? regs->sctx_rdi : regs->sctx_rsi; + + return (val); +} + +static uint64_t +svm_inout_str_count(struct svm_regctx *regs, int rep) +{ + uint64_t val; + + val = rep ? regs->sctx_rcx : 1; + + return (val); +} + +static void +svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1, + int in, struct vm_inout_str *vis) +{ + int error, s; + + if (in) { + vis->seg_name = VM_REG_GUEST_ES; + } else { + /* The segment field has standard encoding */ + s = (info1 >> 10) & 0x7; + vis->seg_name = vm_segment_name(s); + } + + error = vmcb_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc); + KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error)); +} + +static int +svm_inout_str_addrsize(uint64_t info1) +{ + uint32_t size; + + size = (info1 >> 7) & 0x7; + switch (size) { + case 1: + return (2); /* 16 bit */ + case 2: + return (4); /* 32 bit */ + case 4: + return (8); /* 64 bit */ + default: + panic("%s: invalid size encoding %d", __func__, size); + } +} + +static void +svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging) +{ + struct vmcb_state *state; + + state = &vmcb->state; + paging->cr3 = state->cr3; + paging->cpl = svm_cpl(state); + paging->cpu_mode = svm_vcpu_mode(vmcb); + paging->paging_mode = svm_paging_mode(state->cr0, state->cr4, + state->efer); +} + +#define UNHANDLED 0 + +/* + * Handle guest I/O intercept. + */ +static int +svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit) +{ + struct vmcb_ctrl *ctrl; + struct vmcb_state *state; + struct svm_regctx *regs; + struct vm_inout_str *vis; + uint64_t info1; + int inout_string; + + state = svm_get_vmcb_state(svm_sc, vcpu); + ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu); + regs = svm_get_guest_regctx(svm_sc, vcpu); + + info1 = ctrl->exitinfo1; + inout_string = info1 & BIT(2) ? 1 : 0; + + /* + * The effective segment number in EXITINFO1[12:10] is populated + * only if the processor has the DecodeAssist capability. + * + * XXX this is not specified explicitly in APMv2 but can be verified + * empirically. + */ + if (inout_string && !decode_assist()) + return (UNHANDLED); + + vmexit->exitcode = VM_EXITCODE_INOUT; + vmexit->u.inout.in = (info1 & BIT(0)) ? 1 : 0; + vmexit->u.inout.string = inout_string; + vmexit->u.inout.rep = (info1 & BIT(3)) ? 1 : 0; + vmexit->u.inout.bytes = (info1 >> 4) & 0x7; + vmexit->u.inout.port = (uint16_t)(info1 >> 16); + vmexit->u.inout.eax = (uint32_t)(state->rax); + + if (inout_string) { + vmexit->exitcode = VM_EXITCODE_INOUT_STR; + vis = &vmexit->u.inout_str; + svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging); + vis->rflags = state->rflags; + vis->cr0 = state->cr0; + vis->index = svm_inout_str_index(regs, vmexit->u.inout.in); + vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep); + vis->addrsize = svm_inout_str_addrsize(info1); + svm_inout_str_seginfo(svm_sc, vcpu, info1, + vmexit->u.inout.in, vis); + } + + return (UNHANDLED); +} + +static int +npf_fault_type(uint64_t exitinfo1) +{ + + if (exitinfo1 & VMCB_NPF_INFO1_W) + return (VM_PROT_WRITE); + else if (exitinfo1 & VMCB_NPF_INFO1_ID) + return (VM_PROT_EXECUTE); + else + return (VM_PROT_READ); +} + +static bool +svm_npf_emul_fault(uint64_t exitinfo1) +{ + + if (exitinfo1 & VMCB_NPF_INFO1_ID) { + return (false); + } + + if (exitinfo1 & VMCB_NPF_INFO1_GPT) { + return (false); + } + + if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) { + return (false); + } + + return (true); +} + +static void +svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit) +{ + struct vm_guest_paging *paging; + struct vmcb_segment seg; + struct vmcb_ctrl *ctrl; + char *inst_bytes; + int error, inst_len; + + ctrl = &vmcb->ctrl; + paging = &vmexit->u.inst_emul.paging; + + vmexit->exitcode = VM_EXITCODE_INST_EMUL; + vmexit->u.inst_emul.gpa = gpa; + vmexit->u.inst_emul.gla = VIE_INVALID_GLA; + svm_paging_info(vmcb, paging); + + error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg); + KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error)); + + switch(paging->cpu_mode) { + case CPU_MODE_PROTECTED: + case CPU_MODE_COMPATIBILITY: + /* + * Section 4.8.1 of APM2, Default Operand Size or D bit. + */ + vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ? + 1 : 0; + break; + default: + vmexit->u.inst_emul.cs_d = 0; + break; + } + + /* + * Copy the instruction bytes into 'vie' if available. + */ + if (decode_assist() && !disable_npf_assist) { + inst_len = ctrl->inst_len; + inst_bytes = ctrl->inst_bytes; + } else { + inst_len = 0; + inst_bytes = NULL; + } + vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len); +} + +#ifdef KTR +static const char * +intrtype_to_str(int intr_type) +{ + switch (intr_type) { + case VMCB_EVENTINJ_TYPE_INTR: + return ("hwintr"); + case VMCB_EVENTINJ_TYPE_NMI: + return ("nmi"); + case VMCB_EVENTINJ_TYPE_INTn: + return ("swintr"); + case VMCB_EVENTINJ_TYPE_EXCEPTION: + return ("exception"); + default: + panic("%s: unknown intr_type %d", __func__, intr_type); + } +} +#endif + +/* + * Inject an event to vcpu as described in section 15.20, "Event injection". + */ +static void +svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector, + uint32_t error, bool ec_valid) +{ + struct vmcb_ctrl *ctrl; + + ctrl = svm_get_vmcb_ctrl(sc, vcpu); + + KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, + ("%s: event already pending %#lx", __func__, ctrl->eventinj)); + + KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d", + __func__, vector)); + + switch (intr_type) { + case VMCB_EVENTINJ_TYPE_INTR: + case VMCB_EVENTINJ_TYPE_NMI: + case VMCB_EVENTINJ_TYPE_INTn: + break; + case VMCB_EVENTINJ_TYPE_EXCEPTION: + if (vector >= 0 && vector <= 31 && vector != 2) + break; + /* FALLTHROUGH */ + default: + panic("%s: invalid intr_type/vector: %d/%d", __func__, + intr_type, vector); + } + ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID; + if (ec_valid) { + ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID; + ctrl->eventinj |= (uint64_t)error << 32; + VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x", + intrtype_to_str(intr_type), vector, error); + } else { + VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d", + intrtype_to_str(intr_type), vector); + } +} + +static void +svm_update_virqinfo(struct svm_softc *sc, int vcpu) +{ + struct vm *vm; + struct vlapic *vlapic; + struct vmcb_ctrl *ctrl; + int pending; + + vm = sc->vm; + vlapic = vm_lapic(vm, vcpu); + ctrl = svm_get_vmcb_ctrl(sc, vcpu); + + /* Update %cr8 in the emulated vlapic */ + vlapic_set_cr8(vlapic, ctrl->v_tpr); + + /* + * If V_IRQ indicates that the interrupt injection attempted on then + * last VMRUN was successful then update the vlapic accordingly. + */ + if (ctrl->v_intr_vector != 0) { + pending = ctrl->v_irq; + KASSERT(ctrl->v_intr_vector >= 16, ("%s: invalid " + "v_intr_vector %d", __func__, ctrl->v_intr_vector)); + KASSERT(!ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__)); + VCPU_CTR2(vm, vcpu, "v_intr_vector %d %s", ctrl->v_intr_vector, + pending ? "pending" : "accepted"); + if (!pending) + vlapic_intr_accepted(vlapic, ctrl->v_intr_vector); + } +} + +static void +svm_save_intinfo(struct svm_softc *svm_sc, int vcpu) +{ + struct vmcb_ctrl *ctrl; + uint64_t intinfo; + + ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu); + intinfo = ctrl->exitintinfo; + if (!VMCB_EXITINTINFO_VALID(intinfo)) + return; + + /* + * From APMv2, Section "Intercepts during IDT interrupt delivery" + * + * If a #VMEXIT happened during event delivery then record the event + * that was being delivered. + */ + VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n", + intinfo, VMCB_EXITINTINFO_VECTOR(intinfo)); + vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1); + vm_exit_intinfo(svm_sc->vm, vcpu, intinfo); +} + +static __inline int +vintr_intercept_enabled(struct svm_softc *sc, int vcpu) +{ + + return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, + VMCB_INTCPT_VINTR)); +} + +static __inline void +enable_intr_window_exiting(struct svm_softc *sc, int vcpu) +{ + struct vmcb_ctrl *ctrl; + + ctrl = svm_get_vmcb_ctrl(sc, vcpu); + + if (ctrl->v_irq && ctrl->v_intr_vector == 0) { + KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__)); + KASSERT(vintr_intercept_enabled(sc, vcpu), + ("%s: vintr intercept should be enabled", __func__)); + return; + } + + VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting"); + ctrl->v_irq = 1; + ctrl->v_ign_tpr = 1; + ctrl->v_intr_vector = 0; + svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR); + svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR); +} + +static __inline void +disable_intr_window_exiting(struct svm_softc *sc, int vcpu) +{ + struct vmcb_ctrl *ctrl; + + ctrl = svm_get_vmcb_ctrl(sc, vcpu); + + if (!ctrl->v_irq && ctrl->v_intr_vector == 0) { + KASSERT(!vintr_intercept_enabled(sc, vcpu), + ("%s: vintr intercept should be disabled", __func__)); + return; + } + +#ifdef KTR + if (ctrl->v_intr_vector == 0) + VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting"); + else + VCPU_CTR0(sc->vm, vcpu, "Clearing V_IRQ interrupt injection"); +#endif + ctrl->v_irq = 0; + ctrl->v_intr_vector = 0; + svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR); + svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR); +} + +static int +svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val) +{ + struct vmcb_ctrl *ctrl; + int oldval, newval; + + ctrl = svm_get_vmcb_ctrl(sc, vcpu); + oldval = ctrl->intr_shadow; + newval = val ? 1 : 0; + if (newval != oldval) { + ctrl->intr_shadow = newval; + VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval); + } + return (0); +} + +static int +svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val) +{ + struct vmcb_ctrl *ctrl; + + ctrl = svm_get_vmcb_ctrl(sc, vcpu); + *val = ctrl->intr_shadow; + return (0); +} + +/* + * Once an NMI is injected it blocks delivery of further NMIs until the handler + * executes an IRET. The IRET intercept is enabled when an NMI is injected to + * to track when the vcpu is done handling the NMI. + */ +static int +nmi_blocked(struct svm_softc *sc, int vcpu) +{ + int blocked; + + blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, + VMCB_INTCPT_IRET); + return (blocked); +} + +static void +enable_nmi_blocking(struct svm_softc *sc, int vcpu) +{ + + KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked")); + VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled"); + svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET); +} + +static void +clear_nmi_blocking(struct svm_softc *sc, int vcpu) +{ + int error; + + KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked")); + VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared"); + /* + * When the IRET intercept is cleared the vcpu will attempt to execute + * the "iret" when it runs next. However, it is possible to inject + * another NMI into the vcpu before the "iret" has actually executed. + * + * For e.g. if the "iret" encounters a #NPF when accessing the stack + * it will trap back into the hypervisor. If an NMI is pending for + * the vcpu it will be injected into the guest. + * + * XXX this needs to be fixed + */ + svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET); + + /* + * Set 'intr_shadow' to prevent an NMI from being injected on the + * immediate VMRUN. + */ + error = svm_modify_intr_shadow(sc, vcpu, 1); + KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error)); +} + +static int +emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val, + bool *retu) +{ + int error; + + if (lapic_msr(num)) + error = lapic_wrmsr(sc->vm, vcpu, num, val, retu); + else if (num == MSR_EFER) + error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, val); + else + error = svm_wrmsr(sc, vcpu, num, val, retu); + + return (error); +} + +static int +emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu) +{ + struct vmcb_state *state; + struct svm_regctx *ctx; + uint64_t result; + int error; + + if (lapic_msr(num)) + error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu); + else + error = svm_rdmsr(sc, vcpu, num, &result, retu); + + if (error == 0) { + state = svm_get_vmcb_state(sc, vcpu); + ctx = svm_get_guest_regctx(sc, vcpu); + state->rax = result & 0xffffffff; + ctx->sctx_rdx = result >> 32; + } + + return (error); +} + +#ifdef KTR +static const char * +exit_reason_to_str(uint64_t reason) +{ + static char reasonbuf[32]; + + switch (reason) { + case VMCB_EXIT_INVALID: + return ("invalvmcb"); + case VMCB_EXIT_SHUTDOWN: + return ("shutdown"); + case VMCB_EXIT_NPF: + return ("nptfault"); + case VMCB_EXIT_PAUSE: + return ("pause"); + case VMCB_EXIT_HLT: + return ("hlt"); + case VMCB_EXIT_CPUID: + return ("cpuid"); + case VMCB_EXIT_IO: + return ("inout"); + case VMCB_EXIT_MC: + return ("mchk"); + case VMCB_EXIT_INTR: + return ("extintr"); + case VMCB_EXIT_NMI: + return ("nmi"); + case VMCB_EXIT_VINTR: + return ("vintr"); + case VMCB_EXIT_MSR: + return ("msr"); + case VMCB_EXIT_IRET: + return ("iret"); + default: + snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason); + return (reasonbuf); + } +} +#endif /* KTR */ + +/* + * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs + * that are due to instruction intercepts as well as MSR and IOIO intercepts + * and exceptions caused by INT3, INTO and BOUND instructions. + * + * Return 1 if the nRIP is valid and 0 otherwise. + */ +static int +nrip_valid(uint64_t exitcode) +{ + switch (exitcode) { + case 0x00 ... 0x0F: /* read of CR0 through CR15 */ + case 0x10 ... 0x1F: /* write of CR0 through CR15 */ + case 0x20 ... 0x2F: /* read of DR0 through DR15 */ + case 0x30 ... 0x3F: /* write of DR0 through DR15 */ + case 0x43: /* INT3 */ + case 0x44: /* INTO */ + case 0x45: /* BOUND */ + case 0x65 ... 0x7C: /* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */ + case 0x80 ... 0x8D: /* VMEXIT_VMRUN ... VMEXIT_XSETBV */ + return (1); + default: + return (0); + } +} + +/* + * Collateral for a generic SVM VM-exit. + */ +static void +vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2) +{ + + vme->exitcode = VM_EXITCODE_SVM; + vme->u.svm.exitcode = code; + vme->u.svm.exitinfo1 = info1; + vme->u.svm.exitinfo2 = info2; +} + +static int +svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit) +{ + struct vmcb *vmcb; + struct vmcb_state *state; + struct vmcb_ctrl *ctrl; + struct svm_regctx *ctx; + uint64_t code, info1, info2, val; + uint32_t eax, ecx, edx; + int handled; + bool retu; + + ctx = svm_get_guest_regctx(svm_sc, vcpu); + vmcb = svm_get_vmcb(svm_sc, vcpu); + state = &vmcb->state; + ctrl = &vmcb->ctrl; + + handled = 0; + code = ctrl->exitcode; + info1 = ctrl->exitinfo1; + info2 = ctrl->exitinfo2; + + vmexit->exitcode = VM_EXITCODE_BOGUS; + vmexit->rip = state->rip; + vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0; + + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1); + + /* + * #VMEXIT(INVALID) needs to be handled early because the VMCB is + * in an inconsistent state and can trigger assertions that would + * never happen otherwise. + */ + if (code == VMCB_EXIT_INVALID) { + vm_exit_svm(vmexit, code, info1, info2); + return (0); + } + + KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event " + "injection valid bit is set %#lx", __func__, ctrl->eventinj)); + + KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15, + ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)", + vmexit->inst_length, code, info1, info2)); + + svm_update_virqinfo(svm_sc, vcpu); + svm_save_intinfo(svm_sc, vcpu); + + switch (code) { + case VMCB_EXIT_IRET: + /* + * Restart execution at "iret" but with the intercept cleared. + */ + vmexit->inst_length = 0; + clear_nmi_blocking(svm_sc, vcpu); + handled = 1; + break; + case VMCB_EXIT_VINTR: /* interrupt window exiting */ + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1); + handled = 1; + break; + case VMCB_EXIT_INTR: /* external interrupt */ + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1); + handled = 1; + break; + case VMCB_EXIT_NMI: /* external NMI */ + handled = 1; + break; + case VMCB_EXIT_MC: /* machine check */ + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1); + break; + case VMCB_EXIT_MSR: /* MSR access. */ + eax = state->rax; + ecx = ctx->sctx_rcx; + edx = ctx->sctx_rdx; + retu = false; + + if (info1) { + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1); + val = (uint64_t)edx << 32 | eax; + VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx", + ecx, val); + if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) { + vmexit->exitcode = VM_EXITCODE_WRMSR; + vmexit->u.msr.code = ecx; + vmexit->u.msr.wval = val; + } else if (!retu) { + handled = 1; + } else { + KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, + ("emulate_wrmsr retu with bogus exitcode")); + } + } else { + VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx); + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1); + if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) { + vmexit->exitcode = VM_EXITCODE_RDMSR; + vmexit->u.msr.code = ecx; + } else if (!retu) { + handled = 1; + } else { + KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, + ("emulate_rdmsr retu with bogus exitcode")); + } + } + break; + case VMCB_EXIT_IO: + handled = svm_handle_io(svm_sc, vcpu, vmexit); + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1); + break; + case VMCB_EXIT_CPUID: + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1); + handled = x86_emulate_cpuid(svm_sc->vm, vcpu, + (uint32_t *)&state->rax, + (uint32_t *)&ctx->sctx_rbx, + (uint32_t *)&ctx->sctx_rcx, + (uint32_t *)&ctx->sctx_rdx); + break; + case VMCB_EXIT_HLT: + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1); + vmexit->exitcode = VM_EXITCODE_HLT; + vmexit->u.hlt.rflags = state->rflags; + break; + case VMCB_EXIT_PAUSE: + vmexit->exitcode = VM_EXITCODE_PAUSE; + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1); + break; + case VMCB_EXIT_NPF: + /* EXITINFO2 contains the faulting guest physical address */ + if (info1 & VMCB_NPF_INFO1_RSV) { + VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with " + "reserved bits set: info1(%#lx) info2(%#lx)", + info1, info2); + } else if (vm_mem_allocated(svm_sc->vm, info2)) { + vmexit->exitcode = VM_EXITCODE_PAGING; + vmexit->u.paging.gpa = info2; + vmexit->u.paging.fault_type = npf_fault_type(info1); + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1); + VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault " + "on gpa %#lx/%#lx at rip %#lx", + info2, info1, state->rip); + } else if (svm_npf_emul_fault(info1)) { + svm_handle_inst_emul(vmcb, info2, vmexit); + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1); + VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault " + "for gpa %#lx/%#lx at rip %#lx", + info2, info1, state->rip); + } + break; + default: + vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1); + break; + } + + VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d", + handled ? "handled" : "unhandled", exit_reason_to_str(code), + vmexit->rip, vmexit->inst_length); + + if (handled) { + vmexit->rip += vmexit->inst_length; + vmexit->inst_length = 0; + state->rip = vmexit->rip; + } else { + if (vmexit->exitcode == VM_EXITCODE_BOGUS) { + /* + * If this VM exit was not claimed by anybody then + * treat it as a generic SVM exit. + */ + vm_exit_svm(vmexit, code, info1, info2); + } else { + /* + * The exitcode and collateral have been populated. + * The VM exit will be processed further in userland. + */ + } + } + return (handled); +} + +static void +svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu) +{ + uint64_t intinfo; + + if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo)) + return; + + KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not " + "valid: %#lx", __func__, intinfo)); + + svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo), + VMCB_EXITINTINFO_VECTOR(intinfo), + VMCB_EXITINTINFO_EC(intinfo), + VMCB_EXITINTINFO_EC_VALID(intinfo)); + vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1); + VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo); +} + +/* + * Inject event to virtual cpu. + */ +static void +svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic) +{ + struct vmcb_ctrl *ctrl; + struct vmcb_state *state; + uint8_t v_tpr; + int vector, need_intr_window, pending_apic_vector; + + state = svm_get_vmcb_state(sc, vcpu); + ctrl = svm_get_vmcb_ctrl(sc, vcpu); + + need_intr_window = 0; + pending_apic_vector = 0; + + /* + * Inject pending events or exceptions for this vcpu. + * + * An event might be pending because the previous #VMEXIT happened + * during event delivery (i.e. ctrl->exitintinfo). + * + * An event might also be pending because an exception was injected + * by the hypervisor (e.g. #PF during instruction emulation). + */ + svm_inj_intinfo(sc, vcpu); + + /* NMI event has priority over interrupts. */ + if (vm_nmi_pending(sc->vm, vcpu)) { + if (nmi_blocked(sc, vcpu)) { + /* + * Can't inject another NMI if the guest has not + * yet executed an "iret" after the last NMI. + */ + VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due " + "to NMI-blocking"); + } else if (ctrl->intr_shadow) { + /* + * Can't inject an NMI if the vcpu is in an intr_shadow. + */ + VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to " + "interrupt shadow"); + need_intr_window = 1; + goto done; + } else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) { + /* + * If there is already an exception/interrupt pending + * then defer the NMI until after that. + */ + VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to " + "eventinj %#lx", ctrl->eventinj); + + /* + * Use self-IPI to trigger a VM-exit as soon as + * possible after the event injection is completed. + * + * This works only if the external interrupt exiting + * is at a lower priority than the event injection. + * + * Although not explicitly specified in APMv2 the + * relative priorities were verified empirically. + */ + ipi_cpu(curcpu, IPI_AST); /* XXX vmm_ipinum? */ + } else { + vm_nmi_clear(sc->vm, vcpu); + + /* Inject NMI, vector number is not used */ + svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI, + IDT_NMI, 0, false); + + /* virtual NMI blocking is now in effect */ + enable_nmi_blocking(sc, vcpu); + + VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI"); + } + } + + if (!vm_extint_pending(sc->vm, vcpu)) { + /* + * APIC interrupts are delivered using the V_IRQ offload. + * + * The primary benefit is that the hypervisor doesn't need to + * deal with the various conditions that inhibit interrupts. + * It also means that TPR changes via CR8 will be handled + * without any hypervisor involvement. + * + * Note that the APIC vector must remain pending in the vIRR + * until it is confirmed that it was delivered to the guest. + * This can be confirmed based on the value of V_IRQ at the + * next #VMEXIT (1 = pending, 0 = delivered). + * + * Also note that it is possible that another higher priority + * vector can become pending before this vector is delivered + * to the guest. This is alright because vcpu_notify_event() + * will send an IPI and force the vcpu to trap back into the + * hypervisor. The higher priority vector will be injected on + * the next VMRUN. + */ + if (vlapic_pending_intr(vlapic, &vector)) { + KASSERT(vector >= 16 && vector <= 255, + ("invalid vector %d from local APIC", vector)); + pending_apic_vector = vector; + } + goto done; + } + + /* Ask the legacy pic for a vector to inject */ + vatpic_pending_intr(sc->vm, &vector); + KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR", + vector)); + + /* + * If the guest has disabled interrupts or is in an interrupt shadow + * then we cannot inject the pending interrupt. + */ + if ((state->rflags & PSL_I) == 0) { + VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to " + "rflags %#lx", vector, state->rflags); + need_intr_window = 1; + goto done; + } + + if (ctrl->intr_shadow) { + VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to " + "interrupt shadow", vector); + need_intr_window = 1; + goto done; + } + + if (ctrl->eventinj & VMCB_EVENTINJ_VALID) { + VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to " + "eventinj %#lx", vector, ctrl->eventinj); + need_intr_window = 1; + goto done; + } + + /* + * Legacy PIC interrupts are delivered via the event injection + * mechanism. + */ + svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false); + + vm_extint_clear(sc->vm, vcpu); + vatpic_intr_accepted(sc->vm, vector); + + /* + * Force a VM-exit as soon as the vcpu is ready to accept another + * interrupt. This is done because the PIC might have another vector + * that it wants to inject. Also, if the APIC has a pending interrupt + * that was preempted by the ExtInt then it allows us to inject the + * APIC vector as soon as possible. + */ + need_intr_window = 1; +done: + /* + * The guest can modify the TPR by writing to %CR8. In guest mode + * the processor reflects this write to V_TPR without hypervisor + * intervention. + * + * The guest can also modify the TPR by writing to it via the memory + * mapped APIC page. In this case, the write will be emulated by the + * hypervisor. For this reason V_TPR must be updated before every + * VMRUN. + */ + v_tpr = vlapic_get_cr8(vlapic); + KASSERT(v_tpr >= 0 && v_tpr <= 15, ("invalid v_tpr %#x", v_tpr)); + if (ctrl->v_tpr != v_tpr) { + VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x", + ctrl->v_tpr, v_tpr); + ctrl->v_tpr = v_tpr; + svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR); + } + + if (pending_apic_vector) { + /* + * If an APIC vector is being injected then interrupt window + * exiting is not possible on this VMRUN. + */ + KASSERT(!need_intr_window, ("intr_window exiting impossible")); + VCPU_CTR1(sc->vm, vcpu, "Injecting vector %d using V_IRQ", + pending_apic_vector); + + ctrl->v_irq = 1; + ctrl->v_ign_tpr = 0; + ctrl->v_intr_vector = pending_apic_vector; + ctrl->v_intr_prio = pending_apic_vector >> 4; + svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR); + } else if (need_intr_window) { + /* + * We use V_IRQ in conjunction with the VINTR intercept to + * trap into the hypervisor as soon as a virtual interrupt + * can be delivered. + * + * Since injected events are not subject to intercept checks + * we need to ensure that the V_IRQ is not actually going to + * be delivered on VM entry. The KASSERT below enforces this. + */ + KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 || + (state->rflags & PSL_I) == 0 || ctrl->intr_shadow, + ("Bogus intr_window_exiting: eventinj (%#lx), " + "intr_shadow (%u), rflags (%#lx)", + ctrl->eventinj, ctrl->intr_shadow, state->rflags)); + enable_intr_window_exiting(sc, vcpu); + } else { + disable_intr_window_exiting(sc, vcpu); + } +} + +static __inline void +restore_host_tss(void) +{ + struct system_segment_descriptor *tss_sd; + + /* + * The TSS descriptor was in use prior to launching the guest so it + * has been marked busy. + * + * 'ltr' requires the descriptor to be marked available so change the + * type to "64-bit available TSS". + */ + tss_sd = PCPU_GET(tss); + tss_sd->sd_type = SDT_SYSTSS; + ltr(GSEL(GPROC0_SEL, SEL_KPL)); +} + +static void +check_asid(struct svm_softc *sc, int vcpuid, pmap_t pmap, u_int thiscpu) +{ + struct svm_vcpu *vcpustate; + struct vmcb_ctrl *ctrl; + long eptgen; + bool alloc_asid; + + KASSERT(CPU_ISSET(thiscpu, &pmap->pm_active), ("%s: nested pmap not " + "active on cpu %u", __func__, thiscpu)); + + vcpustate = svm_get_vcpu(sc, vcpuid); + ctrl = svm_get_vmcb_ctrl(sc, vcpuid); + + /* + * The TLB entries associated with the vcpu's ASID are not valid + * if either of the following conditions is true: + * + * 1. The vcpu's ASID generation is different than the host cpu's + * ASID generation. This happens when the vcpu migrates to a new + * host cpu. It can also happen when the number of vcpus executing + * on a host cpu is greater than the number of ASIDs available. + * + * 2. The pmap generation number is different than the value cached in + * the 'vcpustate'. This happens when the host invalidates pages + * belonging to the guest. + * + * asidgen eptgen Action + * mismatch mismatch + * 0 0 (a) + * 0 1 (b1) or (b2) + * 1 0 (c) + * 1 1 (d) + * + * (a) There is no mismatch in eptgen or ASID generation and therefore + * no further action is needed. + * + * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is + * retained and the TLB entries associated with this ASID + * are flushed by VMRUN. + * + * (b2) If the cpu does not support FlushByAsid then a new ASID is + * allocated. + * + * (c) A new ASID is allocated. + * + * (d) A new ASID is allocated. + */ + + alloc_asid = false; + eptgen = pmap->pm_eptgen; + ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING; + + if (vcpustate->asid.gen != asid[thiscpu].gen) { + alloc_asid = true; /* (c) and (d) */ + } else if (vcpustate->eptgen != eptgen) { + if (flush_by_asid()) + ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST; /* (b1) */ + else + alloc_asid = true; /* (b2) */ + } else { + /* + * This is the common case (a). + */ + KASSERT(!alloc_asid, ("ASID allocation not necessary")); + KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING, + ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl)); + } + + if (alloc_asid) { + if (++asid[thiscpu].num >= nasid) { + asid[thiscpu].num = 1; + if (++asid[thiscpu].gen == 0) + asid[thiscpu].gen = 1; + /* + * If this cpu does not support "flush-by-asid" + * then flush the entire TLB on a generation + * bump. Subsequent ASID allocation in this + * generation can be done without a TLB flush. + */ + if (!flush_by_asid()) + ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL; + } + vcpustate->asid.gen = asid[thiscpu].gen; + vcpustate->asid.num = asid[thiscpu].num; + + ctrl->asid = vcpustate->asid.num; + svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID); + /* + * If this cpu supports "flush-by-asid" then the TLB + * was not flushed after the generation bump. The TLB + * is flushed selectively after every new ASID allocation. + */ + if (flush_by_asid()) + ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST; + } + vcpustate->eptgen = eptgen; + + KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero")); + KASSERT(ctrl->asid == vcpustate->asid.num, + ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num)); +} + +static __inline void +disable_gintr(void) +{ + + __asm __volatile("clgi" : : :); +} + +static __inline void +enable_gintr(void) +{ + + __asm __volatile("stgi" : : :); +} + +/* + * Start vcpu with specified RIP. + */ +static int +svm_vmrun(void *arg, int vcpu, register_t rip, pmap_t pmap, + void *rend_cookie, void *suspended_cookie) +{ + struct svm_regctx *gctx; + struct svm_softc *svm_sc; + struct svm_vcpu *vcpustate; + struct vmcb_state *state; + struct vmcb_ctrl *ctrl; + struct vm_exit *vmexit; + struct vlapic *vlapic; + struct vm *vm; + uint64_t vmcb_pa; + u_int thiscpu; + int handled; + + svm_sc = arg; + vm = svm_sc->vm; + + vcpustate = svm_get_vcpu(svm_sc, vcpu); + state = svm_get_vmcb_state(svm_sc, vcpu); + ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu); + vmexit = vm_exitinfo(vm, vcpu); + vlapic = vm_lapic(vm, vcpu); + + /* + * Stash 'curcpu' on the stack as 'thiscpu'. + * + * The per-cpu data area is not accessible until MSR_GSBASE is restored + * after the #VMEXIT. Since VMRUN is executed inside a critical section + * 'curcpu' and 'thiscpu' are guaranteed to identical. + */ + thiscpu = curcpu; + + gctx = svm_get_guest_regctx(svm_sc, vcpu); + vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa; + + if (vcpustate->lastcpu != thiscpu) { + /* + * Force new ASID allocation by invalidating the generation. + */ + vcpustate->asid.gen = 0; + + /* + * Invalidate the VMCB state cache by marking all fields dirty. + */ + svm_set_dirty(svm_sc, vcpu, 0xffffffff); + + /* + * XXX + * Setting 'vcpustate->lastcpu' here is bit premature because + * we may return from this function without actually executing + * the VMRUN instruction. This could happen if a rendezvous + * or an AST is pending on the first time through the loop. + * + * This works for now but any new side-effects of vcpu + * migration should take this case into account. + */ + vcpustate->lastcpu = thiscpu; + vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1); + } + + svm_msr_guest_enter(svm_sc, vcpu); + + /* Update Guest RIP */ + state->rip = rip; + + do { + /* + * Disable global interrupts to guarantee atomicity during + * loading of guest state. This includes not only the state + * loaded by the "vmrun" instruction but also software state + * maintained by the hypervisor: suspended and rendezvous + * state, NPT generation number, vlapic interrupts etc. + */ + disable_gintr(); + + if (vcpu_suspended(suspended_cookie)) { + enable_gintr(); + vm_exit_suspended(vm, vcpu, state->rip); + break; + } + + if (vcpu_rendezvous_pending(rend_cookie)) { + enable_gintr(); + vm_exit_rendezvous(vm, vcpu, state->rip); + break; + } + + /* We are asked to give the cpu by scheduler. */ + if (curthread->td_flags & (TDF_ASTPENDING | TDF_NEEDRESCHED)) { + enable_gintr(); + vm_exit_astpending(vm, vcpu, state->rip); + break; + } + + svm_inj_interrupts(svm_sc, vcpu, vlapic); + + /* Activate the nested pmap on 'thiscpu' */ + CPU_SET_ATOMIC_ACQ(thiscpu, &pmap->pm_active); + + /* + * Check the pmap generation and the ASID generation to + * ensure that the vcpu does not use stale TLB mappings. + */ + check_asid(svm_sc, vcpu, pmap, thiscpu); + + ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty; + vcpustate->dirty = 0; + VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean); + + /* Launch Virtual Machine. */ + VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip); + svm_launch(vmcb_pa, gctx); + + CPU_CLR_ATOMIC(thiscpu, &pmap->pm_active); + + /* + * Restore MSR_GSBASE to point to the pcpu data area. + * + * Note that accesses done via PCPU_GET/PCPU_SET will work + * only after MSR_GSBASE is restored. + * + * Also note that we don't bother restoring MSR_KGSBASE + * since it is not used in the kernel and will be restored + * when the VMRUN ioctl returns to userspace. + */ + wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[thiscpu]); + KASSERT(curcpu == thiscpu, ("thiscpu/curcpu (%u/%u) mismatch", + thiscpu, curcpu)); + + /* + * The host GDTR and IDTR is saved by VMRUN and restored + * automatically on #VMEXIT. However, the host TSS needs + * to be restored explicitly. + */ + restore_host_tss(); + + /* #VMEXIT disables interrupts so re-enable them here. */ + enable_gintr(); + + /* Handle #VMEXIT and if required return to user space. */ + handled = svm_vmexit(svm_sc, vcpu, vmexit); + } while (handled); + + svm_msr_guest_exit(svm_sc, vcpu); + + return (0); +} + +static void +svm_vmcleanup(void *arg) +{ + struct svm_softc *sc = arg; + + free(sc, M_SVM); +} + +static register_t * +swctx_regptr(struct svm_regctx *regctx, int reg) +{ + + switch (reg) { + case VM_REG_GUEST_RBX: + return (®ctx->sctx_rbx); + case VM_REG_GUEST_RCX: + return (®ctx->sctx_rcx); + case VM_REG_GUEST_RDX: + return (®ctx->sctx_rdx); + case VM_REG_GUEST_RDI: + return (®ctx->sctx_rdi); + case VM_REG_GUEST_RSI: + return (®ctx->sctx_rsi); + case VM_REG_GUEST_RBP: + return (®ctx->sctx_rbp); + case VM_REG_GUEST_R8: + return (®ctx->sctx_r8); + case VM_REG_GUEST_R9: + return (®ctx->sctx_r9); + case VM_REG_GUEST_R10: + return (®ctx->sctx_r10); + case VM_REG_GUEST_R11: + return (®ctx->sctx_r11); + case VM_REG_GUEST_R12: + return (®ctx->sctx_r12); + case VM_REG_GUEST_R13: + return (®ctx->sctx_r13); + case VM_REG_GUEST_R14: + return (®ctx->sctx_r14); + case VM_REG_GUEST_R15: + return (®ctx->sctx_r15); + default: + return (NULL); + } +} + +static int +svm_getreg(void *arg, int vcpu, int ident, uint64_t *val) +{ + struct svm_softc *svm_sc; + register_t *reg; + + svm_sc = arg; + + if (ident == VM_REG_GUEST_INTR_SHADOW) { + return (svm_get_intr_shadow(svm_sc, vcpu, val)); + } + + if (vmcb_read(svm_sc, vcpu, ident, val) == 0) { + return (0); + } + + reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident); + + if (reg != NULL) { + *val = *reg; + return (0); + } + + VCPU_CTR1(svm_sc->vm, vcpu, "svm_getreg: unknown register %#x", ident); + return (EINVAL); +} + +static int +svm_setreg(void *arg, int vcpu, int ident, uint64_t val) +{ + struct svm_softc *svm_sc; + register_t *reg; + + svm_sc = arg; + + if (ident == VM_REG_GUEST_INTR_SHADOW) { + return (svm_modify_intr_shadow(svm_sc, vcpu, val)); + } + + if (vmcb_write(svm_sc, vcpu, ident, val) == 0) { + return (0); + } + + reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident); + + if (reg != NULL) { + *reg = val; + return (0); + } + + /* + * XXX deal with CR3 and invalidate TLB entries tagged with the + * vcpu's ASID. This needs to be treated differently depending on + * whether 'running' is true/false. + */ + + VCPU_CTR1(svm_sc->vm, vcpu, "svm_setreg: unknown register %#x", ident); + return (EINVAL); +} + +static int +svm_setcap(void *arg, int vcpu, int type, int val) +{ + struct svm_softc *sc; + int error; + + sc = arg; + error = 0; + switch (type) { + case VM_CAP_HALT_EXIT: + svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, + VMCB_INTCPT_HLT, val); + break; + case VM_CAP_PAUSE_EXIT: + svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, + VMCB_INTCPT_PAUSE, val); + break; + case VM_CAP_UNRESTRICTED_GUEST: + /* Unrestricted guest execution cannot be disabled in SVM */ + if (val == 0) + error = EINVAL; + break; + default: + error = ENOENT; + break; + } + return (error); +} + +static int +svm_getcap(void *arg, int vcpu, int type, int *retval) +{ + struct svm_softc *sc; + int error; + + sc = arg; + error = 0; + + switch (type) { + case VM_CAP_HALT_EXIT: + *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, + VMCB_INTCPT_HLT); + break; + case VM_CAP_PAUSE_EXIT: + *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, + VMCB_INTCPT_PAUSE); + break; + case VM_CAP_UNRESTRICTED_GUEST: + *retval = 1; /* unrestricted guest is always enabled */ + break; + default: + error = ENOENT; + break; + } + return (error); +} + +static struct vlapic * +svm_vlapic_init(void *arg, int vcpuid) +{ + struct svm_softc *svm_sc; + struct vlapic *vlapic; + + svm_sc = arg; + vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO); + vlapic->vm = svm_sc->vm; + vlapic->vcpuid = vcpuid; + vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid]; + + vlapic_init(vlapic); + + return (vlapic); +} + +static void +svm_vlapic_cleanup(void *arg, struct vlapic *vlapic) +{ + + vlapic_cleanup(vlapic); + free(vlapic, M_SVM_VLAPIC); +} + +struct vmm_ops vmm_ops_amd = { + svm_init, + svm_cleanup, + svm_restore, + svm_vminit, + svm_vmrun, + svm_vmcleanup, + svm_getreg, + svm_setreg, + vmcb_getdesc, + vmcb_setdesc, + svm_getcap, + svm_setcap, + svm_npt_alloc, + svm_npt_free, + svm_vlapic_init, + svm_vlapic_cleanup +}; diff --git a/sys/amd64/vmm/amd/svm.h b/sys/amd64/vmm/amd/svm.h new file mode 100644 index 0000000..86bd638 --- /dev/null +++ b/sys/amd64/vmm/amd/svm.h @@ -0,0 +1,54 @@ +/*- + * Copyright (c) 2013 Anish Gupta (akgupt3@gmail.com) + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice unmodified, this list of conditions, and the following + * disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + * $FreeBSD$ + */ + +#ifndef _SVM_H_ +#define _SVM_H_ + +/* + * Guest register state that is saved outside the VMCB. + */ +struct svm_regctx { + register_t sctx_rbp; + register_t sctx_rbx; + register_t sctx_rcx; + register_t sctx_rdx; + register_t sctx_rdi; + register_t sctx_rsi; + register_t sctx_r8; + register_t sctx_r9; + register_t sctx_r10; + register_t sctx_r11; + register_t sctx_r12; + register_t sctx_r13; + register_t sctx_r14; + register_t sctx_r15; +}; + +void svm_launch(uint64_t pa, struct svm_regctx *); + +#endif /* _SVM_H_ */ diff --git a/sys/amd64/vmm/amd/svm_genassym.c b/sys/amd64/vmm/amd/svm_genassym.c new file mode 100644 index 0000000..b7831eb --- /dev/null +++ b/sys/amd64/vmm/amd/svm_genassym.c @@ -0,0 +1,48 @@ +/*- + * Copyright (c) 2013 Anish Gupta (akgupt3@gmail.com) + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice unmodified, this list of conditions, and the following + * disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +#include <sys/cdefs.h> +__FBSDID("$FreeBSD$"); + +#include <sys/param.h> +#include <sys/assym.h> + +#include "svm.h" + +ASSYM(SCTX_RBX, offsetof(struct svm_regctx, sctx_rbx)); +ASSYM(SCTX_RCX, offsetof(struct svm_regctx, sctx_rcx)); +ASSYM(SCTX_RBP, offsetof(struct svm_regctx, sctx_rbp)); +ASSYM(SCTX_RDX, offsetof(struct svm_regctx, sctx_rdx)); +ASSYM(SCTX_RDI, offsetof(struct svm_regctx, sctx_rdi)); +ASSYM(SCTX_RSI, offsetof(struct svm_regctx, sctx_rsi)); +ASSYM(SCTX_R8, offsetof(struct svm_regctx, sctx_r8)); +ASSYM(SCTX_R9, offsetof(struct svm_regctx, sctx_r9)); +ASSYM(SCTX_R10, offsetof(struct svm_regctx, sctx_r10)); +ASSYM(SCTX_R11, offsetof(struct svm_regctx, sctx_r11)); +ASSYM(SCTX_R12, offsetof(struct svm_regctx, sctx_r12)); +ASSYM(SCTX_R13, offsetof(struct svm_regctx, sctx_r13)); +ASSYM(SCTX_R14, offsetof(struct svm_regctx, sctx_r14)); +ASSYM(SCTX_R15, offsetof(struct svm_regctx, sctx_r15)); diff --git a/sys/amd64/vmm/amd/svm_msr.c b/sys/amd64/vmm/amd/svm_msr.c new file mode 100644 index 0000000..100af4b --- /dev/null +++ b/sys/amd64/vmm/amd/svm_msr.c @@ -0,0 +1,136 @@ +/*- + * Copyright (c) 2014, Neel Natu (neel@freebsd.org) + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice unmodified, this list of conditions, and the following + * disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +#include <sys/cdefs.h> +__FBSDID("$FreeBSD$"); + +#include <sys/types.h> +#include <sys/errno.h> + +#include <machine/cpufunc.h> +#include <machine/specialreg.h> + +#include "svm_msr.h" + +#ifndef MSR_AMDK8_IPM +#define MSR_AMDK8_IPM 0xc0010055 +#endif + +enum { + IDX_MSR_LSTAR, + IDX_MSR_CSTAR, + IDX_MSR_STAR, + IDX_MSR_SF_MASK, + HOST_MSR_NUM /* must be the last enumeration */ +}; + +static uint64_t host_msrs[HOST_MSR_NUM]; + +void +svm_msr_init(void) +{ + /* + * It is safe to cache the values of the following MSRs because they + * don't change based on curcpu, curproc or curthread. + */ + host_msrs[IDX_MSR_LSTAR] = rdmsr(MSR_LSTAR); + host_msrs[IDX_MSR_CSTAR] = rdmsr(MSR_CSTAR); + host_msrs[IDX_MSR_STAR] = rdmsr(MSR_STAR); + host_msrs[IDX_MSR_SF_MASK] = rdmsr(MSR_SF_MASK); +} + +void +svm_msr_guest_init(struct svm_softc *sc, int vcpu) +{ + /* + * All the MSRs accessible to the guest are either saved/restored by + * hardware on every #VMEXIT/VMRUN (e.g., G_PAT) or are saved/restored + * by VMSAVE/VMLOAD (e.g., MSR_GSBASE). + * + * There are no guest MSRs that are saved/restored "by hand" so nothing + * more to do here. + */ + return; +} + +void +svm_msr_guest_enter(struct svm_softc *sc, int vcpu) +{ + /* + * Save host MSRs (if any) and restore guest MSRs (if any). + */ +} + +void +svm_msr_guest_exit(struct svm_softc *sc, int vcpu) +{ + /* + * Save guest MSRs (if any) and restore host MSRs. + */ + wrmsr(MSR_LSTAR, host_msrs[IDX_MSR_LSTAR]); + wrmsr(MSR_CSTAR, host_msrs[IDX_MSR_CSTAR]); + wrmsr(MSR_STAR, host_msrs[IDX_MSR_STAR]); + wrmsr(MSR_SF_MASK, host_msrs[IDX_MSR_SF_MASK]); + + /* MSR_KGSBASE will be restored on the way back to userspace */ +} + +int +svm_rdmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t *result, + bool *retu) +{ + int error = 0; + + switch (num) { + case MSR_AMDK8_IPM: + *result = 0; + break; + default: + error = EINVAL; + break; + } + + return (error); +} + +int +svm_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val, bool *retu) +{ + int error = 0; + + switch (num) { + case MSR_AMDK8_IPM: + /* + * Ignore writes to the "Interrupt Pending Message" MSR. + */ + break; + default: + error = EINVAL; + break; + } + + return (error); +} diff --git a/sys/amd64/vmm/amd/svm_msr.h b/sys/amd64/vmm/amd/svm_msr.h new file mode 100644 index 0000000..07716c8 --- /dev/null +++ b/sys/amd64/vmm/amd/svm_msr.h @@ -0,0 +1,44 @@ +/*- + * Copyright (c) 2014 Neel Natu (neel@freebsd.org) + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice unmodified, this list of conditions, and the following + * disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + * $FreeBSD$ + */ + +#ifndef _SVM_MSR_H_ +#define _SVM_MSR_H_ + +struct svm_softc; + +void svm_msr_init(void); +void svm_msr_guest_init(struct svm_softc *sc, int vcpu); +void svm_msr_guest_enter(struct svm_softc *sc, int vcpu); +void svm_msr_guest_exit(struct svm_softc *sc, int vcpu); + +int svm_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val, + bool *retu); +int svm_rdmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t *result, + bool *retu); + +#endif /* _SVM_MSR_H_ */ diff --git a/sys/amd64/vmm/amd/svm_softc.h b/sys/amd64/vmm/amd/svm_softc.h new file mode 100644 index 0000000..a5bb57c --- /dev/null +++ b/sys/amd64/vmm/amd/svm_softc.h @@ -0,0 +1,113 @@ +/*- + * Copyright (c) 2013 Anish Gupta (akgupt3@gmail.com) + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice unmodified, this list of conditions, and the following + * disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + * $FreeBSD$ + */ + +#ifndef _SVM_SOFTC_H_ +#define _SVM_SOFTC_H_ + +#define SVM_IO_BITMAP_SIZE (3 * PAGE_SIZE) +#define SVM_MSR_BITMAP_SIZE (2 * PAGE_SIZE) + +struct asid { + uint64_t gen; /* range is [1, ~0UL] */ + uint32_t num; /* range is [1, nasid - 1] */ +}; + +/* + * XXX separate out 'struct vmcb' from 'svm_vcpu' to avoid wasting space + * due to VMCB alignment requirements. + */ +struct svm_vcpu { + struct vmcb vmcb; /* hardware saved vcpu context */ + struct svm_regctx swctx; /* software saved vcpu context */ + uint64_t vmcb_pa; /* VMCB physical address */ + int lastcpu; /* host cpu that the vcpu last ran on */ + uint32_t dirty; /* state cache bits that must be cleared */ + long eptgen; /* pmap->pm_eptgen when the vcpu last ran */ + struct asid asid; +} __aligned(PAGE_SIZE); + +/* + * SVM softc, one per virtual machine. + */ +struct svm_softc { + uint8_t iopm_bitmap[SVM_IO_BITMAP_SIZE]; /* shared by all vcpus */ + uint8_t msr_bitmap[SVM_MSR_BITMAP_SIZE]; /* shared by all vcpus */ + uint8_t apic_page[VM_MAXCPU][PAGE_SIZE]; + struct svm_vcpu vcpu[VM_MAXCPU]; + vm_offset_t nptp; /* nested page table */ + struct vm *vm; +} __aligned(PAGE_SIZE); + +CTASSERT((offsetof(struct svm_softc, nptp) & PAGE_MASK) == 0); + +static __inline struct svm_vcpu * +svm_get_vcpu(struct svm_softc *sc, int vcpu) +{ + + return (&(sc->vcpu[vcpu])); +} + +static __inline struct vmcb * +svm_get_vmcb(struct svm_softc *sc, int vcpu) +{ + + return (&(sc->vcpu[vcpu].vmcb)); +} + +static __inline struct vmcb_state * +svm_get_vmcb_state(struct svm_softc *sc, int vcpu) +{ + + return (&(sc->vcpu[vcpu].vmcb.state)); +} + +static __inline struct vmcb_ctrl * +svm_get_vmcb_ctrl(struct svm_softc *sc, int vcpu) +{ + + return (&(sc->vcpu[vcpu].vmcb.ctrl)); +} + +static __inline struct svm_regctx * +svm_get_guest_regctx(struct svm_softc *sc, int vcpu) +{ + + return (&(sc->vcpu[vcpu].swctx)); +} + +static __inline void +svm_set_dirty(struct svm_softc *sc, int vcpu, uint32_t dirtybits) +{ + struct svm_vcpu *vcpustate; + + vcpustate = svm_get_vcpu(sc, vcpu); + + vcpustate->dirty |= dirtybits; +} + +#endif /* _SVM_SOFTC_H_ */ diff --git a/sys/amd64/vmm/amd/svm_support.S b/sys/amd64/vmm/amd/svm_support.S new file mode 100644 index 0000000..11da759 --- /dev/null +++ b/sys/amd64/vmm/amd/svm_support.S @@ -0,0 +1,115 @@ +/*- + * Copyright (c) 2013, Anish Gupta (akgupt3@gmail.com) + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice unmodified, this list of conditions, and the following + * disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ +#include <machine/asmacros.h> + +#include "svm_assym.s" + +/* + * Be friendly to DTrace FBT's prologue/epilogue pattern matching. + * + * They are also responsible for saving/restoring the host %rbp across VMRUN. + */ +#define VENTER push %rbp ; mov %rsp,%rbp +#define VLEAVE pop %rbp + +/* + * svm_launch(uint64_t vmcb, struct svm_regctx *gctx) + * %rdi: physical address of VMCB + * %rsi: pointer to guest context + */ +ENTRY(svm_launch) + VENTER + + /* + * Host register state saved across a VMRUN. + * + * All "callee saved registers" except: + * %rsp: because it is preserved by the processor across VMRUN. + * %rbp: because it is saved/restored by the function prologue/epilogue. + */ + push %rbx + push %r12 + push %r13 + push %r14 + push %r15 + + /* Save the physical address of the VMCB in %rax */ + movq %rdi, %rax + + push %rsi /* push guest context pointer on the stack */ + + /* + * Restore guest state. + */ + movq SCTX_R8(%rsi), %r8 + movq SCTX_R9(%rsi), %r9 + movq SCTX_R10(%rsi), %r10 + movq SCTX_R11(%rsi), %r11 + movq SCTX_R12(%rsi), %r12 + movq SCTX_R13(%rsi), %r13 + movq SCTX_R14(%rsi), %r14 + movq SCTX_R15(%rsi), %r15 + movq SCTX_RBP(%rsi), %rbp + movq SCTX_RBX(%rsi), %rbx + movq SCTX_RCX(%rsi), %rcx + movq SCTX_RDX(%rsi), %rdx + movq SCTX_RDI(%rsi), %rdi + movq SCTX_RSI(%rsi), %rsi /* %rsi must be restored last */ + + vmload %rax + vmrun %rax + vmsave %rax + + pop %rax /* pop guest context pointer from the stack */ + + /* + * Save guest state. + */ + movq %r8, SCTX_R8(%rax) + movq %r9, SCTX_R9(%rax) + movq %r10, SCTX_R10(%rax) + movq %r11, SCTX_R11(%rax) + movq %r12, SCTX_R12(%rax) + movq %r13, SCTX_R13(%rax) + movq %r14, SCTX_R14(%rax) + movq %r15, SCTX_R15(%rax) + movq %rbp, SCTX_RBP(%rax) + movq %rbx, SCTX_RBX(%rax) + movq %rcx, SCTX_RCX(%rax) + movq %rdx, SCTX_RDX(%rax) + movq %rdi, SCTX_RDI(%rax) + movq %rsi, SCTX_RSI(%rax) + + /* Restore host state */ + pop %r15 + pop %r14 + pop %r13 + pop %r12 + pop %rbx + + VLEAVE + ret +END(svm_launch) diff --git a/sys/amd64/vmm/amd/vmcb.c b/sys/amd64/vmm/amd/vmcb.c new file mode 100644 index 0000000..8be2308 --- /dev/null +++ b/sys/amd64/vmm/amd/vmcb.c @@ -0,0 +1,442 @@ +/*- + * Copyright (c) 2013 Anish Gupta (akgupt3@gmail.com) + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice unmodified, this list of conditions, and the following + * disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +#include <sys/cdefs.h> +__FBSDID("$FreeBSD$"); + +#include <sys/param.h> +#include <sys/systm.h> +#include <sys/cpuset.h> + +#include <machine/segments.h> +#include <machine/specialreg.h> +#include <machine/vmm.h> + +#include "vmm_ktr.h" + +#include "vmcb.h" +#include "svm.h" +#include "svm_softc.h" + +/* + * The VMCB aka Virtual Machine Control Block is a 4KB aligned page + * in memory that describes the virtual machine. + * + * The VMCB contains: + * - instructions or events in the guest to intercept + * - control bits that modify execution environment of the guest + * - guest processor state (e.g. general purpose registers) + */ + +/* + * Return VMCB segment area. + */ +static struct vmcb_segment * +vmcb_segptr(struct vmcb *vmcb, int type) +{ + struct vmcb_state *state; + struct vmcb_segment *seg; + + state = &vmcb->state; + + switch (type) { + case VM_REG_GUEST_CS: + seg = &state->cs; + break; + + case VM_REG_GUEST_DS: + seg = &state->ds; + break; + + case VM_REG_GUEST_ES: + seg = &state->es; + break; + + case VM_REG_GUEST_FS: + seg = &state->fs; + break; + + case VM_REG_GUEST_GS: + seg = &state->gs; + break; + + case VM_REG_GUEST_SS: + seg = &state->ss; + break; + + case VM_REG_GUEST_GDTR: + seg = &state->gdt; + break; + + case VM_REG_GUEST_IDTR: + seg = &state->idt; + break; + + case VM_REG_GUEST_LDTR: + seg = &state->ldt; + break; + + case VM_REG_GUEST_TR: + seg = &state->tr; + break; + + default: + seg = NULL; + break; + } + + return (seg); +} + +static int +vmcb_access(struct svm_softc *softc, int vcpu, int write, int ident, + uint64_t *val) +{ + struct vmcb *vmcb; + int off, bytes; + char *ptr; + + vmcb = svm_get_vmcb(softc, vcpu); + off = VMCB_ACCESS_OFFSET(ident); + bytes = VMCB_ACCESS_BYTES(ident); + + if ((off + bytes) >= sizeof (struct vmcb)) + return (EINVAL); + + ptr = (char *)vmcb; + + if (!write) + *val = 0; + + switch (bytes) { + case 8: + case 4: + case 2: + if (write) + memcpy(ptr + off, val, bytes); + else + memcpy(val, ptr + off, bytes); + break; + default: + VCPU_CTR1(softc->vm, vcpu, + "Invalid size %d for VMCB access: %d", bytes); + return (EINVAL); + } + + /* Invalidate all VMCB state cached by h/w. */ + if (write) + svm_set_dirty(softc, vcpu, 0xffffffff); + + return (0); +} + +/* + * Read from segment selector, control and general purpose register of VMCB. + */ +int +vmcb_read(struct svm_softc *sc, int vcpu, int ident, uint64_t *retval) +{ + struct vmcb *vmcb; + struct vmcb_state *state; + struct vmcb_segment *seg; + int err; + + vmcb = svm_get_vmcb(sc, vcpu); + state = &vmcb->state; + err = 0; + + if (VMCB_ACCESS_OK(ident)) + return (vmcb_access(sc, vcpu, 0, ident, retval)); + + switch (ident) { + case VM_REG_GUEST_CR0: + *retval = state->cr0; + break; + + case VM_REG_GUEST_CR2: + *retval = state->cr2; + break; + + case VM_REG_GUEST_CR3: + *retval = state->cr3; + break; + + case VM_REG_GUEST_CR4: + *retval = state->cr4; + break; + + case VM_REG_GUEST_DR7: + *retval = state->dr7; + break; + + case VM_REG_GUEST_EFER: + *retval = state->efer; + break; + + case VM_REG_GUEST_RAX: + *retval = state->rax; + break; + + case VM_REG_GUEST_RFLAGS: + *retval = state->rflags; + break; + + case VM_REG_GUEST_RIP: + *retval = state->rip; + break; + + case VM_REG_GUEST_RSP: + *retval = state->rsp; + break; + + case VM_REG_GUEST_CS: + case VM_REG_GUEST_DS: + case VM_REG_GUEST_ES: + case VM_REG_GUEST_FS: + case VM_REG_GUEST_GS: + case VM_REG_GUEST_SS: + case VM_REG_GUEST_LDTR: + case VM_REG_GUEST_TR: + seg = vmcb_segptr(vmcb, ident); + KASSERT(seg != NULL, ("%s: unable to get segment %d from VMCB", + __func__, ident)); + *retval = seg->selector; + break; + + case VM_REG_GUEST_GDTR: + case VM_REG_GUEST_IDTR: + /* GDTR and IDTR don't have segment selectors */ + err = EINVAL; + break; + default: + err = EINVAL; + break; + } + + return (err); +} + +/* + * Write to segment selector, control and general purpose register of VMCB. + */ +int +vmcb_write(struct svm_softc *sc, int vcpu, int ident, uint64_t val) +{ + struct vmcb *vmcb; + struct vmcb_state *state; + struct vmcb_segment *seg; + int err, dirtyseg; + + vmcb = svm_get_vmcb(sc, vcpu); + state = &vmcb->state; + dirtyseg = 0; + err = 0; + + if (VMCB_ACCESS_OK(ident)) + return (vmcb_access(sc, vcpu, 1, ident, &val)); + + switch (ident) { + case VM_REG_GUEST_CR0: + state->cr0 = val; + svm_set_dirty(sc, vcpu, VMCB_CACHE_CR); + break; + + case VM_REG_GUEST_CR2: + state->cr2 = val; + svm_set_dirty(sc, vcpu, VMCB_CACHE_CR2); + break; + + case VM_REG_GUEST_CR3: + state->cr3 = val; + svm_set_dirty(sc, vcpu, VMCB_CACHE_CR); + break; + + case VM_REG_GUEST_CR4: + state->cr4 = val; + svm_set_dirty(sc, vcpu, VMCB_CACHE_CR); + break; + + case VM_REG_GUEST_DR7: + state->dr7 = val; + break; + + case VM_REG_GUEST_EFER: + /* EFER_SVM must always be set when the guest is executing */ + state->efer = val | EFER_SVM; + svm_set_dirty(sc, vcpu, VMCB_CACHE_CR); + break; + + case VM_REG_GUEST_RAX: + state->rax = val; + break; + + case VM_REG_GUEST_RFLAGS: + state->rflags = val; + break; + + case VM_REG_GUEST_RIP: + state->rip = val; + break; + + case VM_REG_GUEST_RSP: + state->rsp = val; + break; + + case VM_REG_GUEST_CS: + case VM_REG_GUEST_DS: + case VM_REG_GUEST_ES: + case VM_REG_GUEST_SS: + dirtyseg = 1; /* FALLTHROUGH */ + case VM_REG_GUEST_FS: + case VM_REG_GUEST_GS: + case VM_REG_GUEST_LDTR: + case VM_REG_GUEST_TR: + seg = vmcb_segptr(vmcb, ident); + KASSERT(seg != NULL, ("%s: unable to get segment %d from VMCB", + __func__, ident)); + seg->selector = val; + if (dirtyseg) + svm_set_dirty(sc, vcpu, VMCB_CACHE_SEG); + break; + + case VM_REG_GUEST_GDTR: + case VM_REG_GUEST_IDTR: + /* GDTR and IDTR don't have segment selectors */ + err = EINVAL; + break; + default: + err = EINVAL; + break; + } + + return (err); +} + +int +vmcb_seg(struct vmcb *vmcb, int ident, struct vmcb_segment *seg2) +{ + struct vmcb_segment *seg; + + seg = vmcb_segptr(vmcb, ident); + if (seg != NULL) { + bcopy(seg, seg2, sizeof(struct vmcb_segment)); + return (0); + } else { + return (EINVAL); + } +} + +int +vmcb_setdesc(void *arg, int vcpu, int reg, struct seg_desc *desc) +{ + struct vmcb *vmcb; + struct svm_softc *sc; + struct vmcb_segment *seg; + uint16_t attrib; + + sc = arg; + vmcb = svm_get_vmcb(sc, vcpu); + + seg = vmcb_segptr(vmcb, reg); + KASSERT(seg != NULL, ("%s: invalid segment descriptor %d", + __func__, reg)); + + seg->base = desc->base; + seg->limit = desc->limit; + if (reg != VM_REG_GUEST_GDTR && reg != VM_REG_GUEST_IDTR) { + /* + * Map seg_desc access to VMCB attribute format. + * + * SVM uses the 'P' bit in the segment attributes to indicate a + * NULL segment so clear it if the segment is marked unusable. + */ + attrib = ((desc->access & 0xF000) >> 4) | (desc->access & 0xFF); + if (SEG_DESC_UNUSABLE(desc->access)) { + attrib &= ~0x80; + } + seg->attrib = attrib; + } + + VCPU_CTR4(sc->vm, vcpu, "Setting desc %d: base (%#lx), limit (%#x), " + "attrib (%#x)", reg, seg->base, seg->limit, seg->attrib); + + switch (reg) { + case VM_REG_GUEST_CS: + case VM_REG_GUEST_DS: + case VM_REG_GUEST_ES: + case VM_REG_GUEST_SS: + svm_set_dirty(sc, vcpu, VMCB_CACHE_SEG); + case VM_REG_GUEST_GDTR: + case VM_REG_GUEST_IDTR: + svm_set_dirty(sc, vcpu, VMCB_CACHE_DT); + break; + default: + break; + } + + return (0); +} + +int +vmcb_getdesc(void *arg, int vcpu, int reg, struct seg_desc *desc) +{ + struct vmcb *vmcb; + struct svm_softc *sc; + struct vmcb_segment *seg; + + sc = arg; + vmcb = svm_get_vmcb(sc, vcpu); + seg = vmcb_segptr(vmcb, reg); + KASSERT(seg != NULL, ("%s: invalid segment descriptor %d", + __func__, reg)); + + desc->base = seg->base; + desc->limit = seg->limit; + desc->access = 0; + + if (reg != VM_REG_GUEST_GDTR && reg != VM_REG_GUEST_IDTR) { + /* Map seg_desc access to VMCB attribute format */ + desc->access = ((seg->attrib & 0xF00) << 4) | + (seg->attrib & 0xFF); + + /* + * VT-x uses bit 16 to indicate a segment that has been loaded + * with a NULL selector (aka unusable). The 'desc->access' + * field is interpreted in the VT-x format by the + * processor-independent code. + * + * SVM uses the 'P' bit to convey the same information so + * convert it into the VT-x format. For more details refer to + * section "Segment State in the VMCB" in APMv2. + */ + if (reg != VM_REG_GUEST_CS && reg != VM_REG_GUEST_TR) { + if ((desc->access & 0x80) == 0) + desc->access |= 0x10000; /* Unusable segment */ + } + } + + return (0); +} diff --git a/sys/amd64/vmm/amd/vmcb.h b/sys/amd64/vmm/amd/vmcb.h new file mode 100644 index 0000000..d3e07ac --- /dev/null +++ b/sys/amd64/vmm/amd/vmcb.h @@ -0,0 +1,332 @@ +/*- + * Copyright (c) 2013 Anish Gupta (akgupt3@gmail.com) + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice unmodified, this list of conditions, and the following + * disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + * $FreeBSD$ + */ + +#ifndef _VMCB_H_ +#define _VMCB_H_ + +struct svm_softc; + +#define BIT(n) (1ULL << n) + +/* + * Secure Virtual Machine: AMD64 Programmer's Manual Vol2, Chapter 15 + * Layout of VMCB: AMD64 Programmer's Manual Vol2, Appendix B + */ + +/* vmcb_ctrl->intercept[] array indices */ +#define VMCB_CR_INTCPT 0 +#define VMCB_DR_INTCPT 1 +#define VMCB_EXC_INTCPT 2 +#define VMCB_CTRL1_INTCPT 3 +#define VMCB_CTRL2_INTCPT 4 + +/* intercept[VMCB_CTRL1_INTCPT] fields */ +#define VMCB_INTCPT_INTR BIT(0) +#define VMCB_INTCPT_NMI BIT(1) +#define VMCB_INTCPT_SMI BIT(2) +#define VMCB_INTCPT_INIT BIT(3) +#define VMCB_INTCPT_VINTR BIT(4) +#define VMCB_INTCPT_CR0_WRITE BIT(5) +#define VMCB_INTCPT_IDTR_READ BIT(6) +#define VMCB_INTCPT_GDTR_READ BIT(7) +#define VMCB_INTCPT_LDTR_READ BIT(8) +#define VMCB_INTCPT_TR_READ BIT(9) +#define VMCB_INTCPT_IDTR_WRITE BIT(10) +#define VMCB_INTCPT_GDTR_WRITE BIT(11) +#define VMCB_INTCPT_LDTR_WRITE BIT(12) +#define VMCB_INTCPT_TR_WRITE BIT(13) +#define VMCB_INTCPT_RDTSC BIT(14) +#define VMCB_INTCPT_RDPMC BIT(15) +#define VMCB_INTCPT_PUSHF BIT(16) +#define VMCB_INTCPT_POPF BIT(17) +#define VMCB_INTCPT_CPUID BIT(18) +#define VMCB_INTCPT_RSM BIT(19) +#define VMCB_INTCPT_IRET BIT(20) +#define VMCB_INTCPT_INTn BIT(21) +#define VMCB_INTCPT_INVD BIT(22) +#define VMCB_INTCPT_PAUSE BIT(23) +#define VMCB_INTCPT_HLT BIT(24) +#define VMCB_INTCPT_INVPG BIT(25) +#define VMCB_INTCPT_INVPGA BIT(26) +#define VMCB_INTCPT_IO BIT(27) +#define VMCB_INTCPT_MSR BIT(28) +#define VMCB_INTCPT_TASK_SWITCH BIT(29) +#define VMCB_INTCPT_FERR_FREEZE BIT(30) +#define VMCB_INTCPT_SHUTDOWN BIT(31) + +/* intercept[VMCB_CTRL2_INTCPT] fields */ +#define VMCB_INTCPT_VMRUN BIT(0) +#define VMCB_INTCPT_VMMCALL BIT(1) +#define VMCB_INTCPT_VMLOAD BIT(2) +#define VMCB_INTCPT_VMSAVE BIT(3) +#define VMCB_INTCPT_STGI BIT(4) +#define VMCB_INTCPT_CLGI BIT(5) +#define VMCB_INTCPT_SKINIT BIT(6) +#define VMCB_INTCPT_RDTSCP BIT(7) +#define VMCB_INTCPT_ICEBP BIT(8) +#define VMCB_INTCPT_WBINVD BIT(9) +#define VMCB_INTCPT_MONITOR BIT(10) +#define VMCB_INTCPT_MWAIT BIT(11) +#define VMCB_INTCPT_MWAIT_ARMED BIT(12) +#define VMCB_INTCPT_XSETBV BIT(13) + +/* VMCB TLB control */ +#define VMCB_TLB_FLUSH_NOTHING 0 /* Flush nothing */ +#define VMCB_TLB_FLUSH_ALL 1 /* Flush entire TLB */ +#define VMCB_TLB_FLUSH_GUEST 3 /* Flush all guest entries */ +#define VMCB_TLB_FLUSH_GUEST_NONGLOBAL 7 /* Flush guest non-PG entries */ + +/* VMCB state caching */ +#define VMCB_CACHE_NONE 0 /* No caching */ +#define VMCB_CACHE_I BIT(0) /* Intercept, TSC off, Pause filter */ +#define VMCB_CACHE_IOPM BIT(1) /* I/O and MSR permission */ +#define VMCB_CACHE_ASID BIT(2) /* ASID */ +#define VMCB_CACHE_TPR BIT(3) /* V_TPR to V_INTR_VECTOR */ +#define VMCB_CACHE_NP BIT(4) /* Nested Paging */ +#define VMCB_CACHE_CR BIT(5) /* CR0, CR3, CR4 & EFER */ +#define VMCB_CACHE_DR BIT(6) /* Debug registers */ +#define VMCB_CACHE_DT BIT(7) /* GDT/IDT */ +#define VMCB_CACHE_SEG BIT(8) /* User segments, CPL */ +#define VMCB_CACHE_CR2 BIT(9) /* page fault address */ +#define VMCB_CACHE_LBR BIT(10) /* Last branch */ + +/* VMCB control event injection */ +#define VMCB_EVENTINJ_EC_VALID BIT(11) /* Error Code valid */ +#define VMCB_EVENTINJ_VALID BIT(31) /* Event valid */ + +/* Event types that can be injected */ +#define VMCB_EVENTINJ_TYPE_INTR 0 +#define VMCB_EVENTINJ_TYPE_NMI 2 +#define VMCB_EVENTINJ_TYPE_EXCEPTION 3 +#define VMCB_EVENTINJ_TYPE_INTn 4 + +/* VMCB exit code, APM vol2 Appendix C */ +#define VMCB_EXIT_MC 0x52 +#define VMCB_EXIT_INTR 0x60 +#define VMCB_EXIT_NMI 0x61 +#define VMCB_EXIT_VINTR 0x64 +#define VMCB_EXIT_PUSHF 0x70 +#define VMCB_EXIT_POPF 0x71 +#define VMCB_EXIT_CPUID 0x72 +#define VMCB_EXIT_IRET 0x74 +#define VMCB_EXIT_PAUSE 0x77 +#define VMCB_EXIT_HLT 0x78 +#define VMCB_EXIT_IO 0x7B +#define VMCB_EXIT_MSR 0x7C +#define VMCB_EXIT_SHUTDOWN 0x7F +#define VMCB_EXIT_VMSAVE 0x83 +#define VMCB_EXIT_NPF 0x400 +#define VMCB_EXIT_INVALID -1 + +/* + * Nested page fault. + * Bit definitions to decode EXITINFO1. + */ +#define VMCB_NPF_INFO1_P BIT(0) /* Nested page present. */ +#define VMCB_NPF_INFO1_W BIT(1) /* Access was write. */ +#define VMCB_NPF_INFO1_U BIT(2) /* Access was user access. */ +#define VMCB_NPF_INFO1_RSV BIT(3) /* Reserved bits present. */ +#define VMCB_NPF_INFO1_ID BIT(4) /* Code read. */ + +#define VMCB_NPF_INFO1_GPA BIT(32) /* Guest physical address. */ +#define VMCB_NPF_INFO1_GPT BIT(33) /* Guest page table. */ + +/* + * EXITINTINFO, Interrupt exit info for all intrecepts. + * Section 15.7.2, Intercepts during IDT Interrupt Delivery. + */ +#define VMCB_EXITINTINFO_VECTOR(x) ((x) & 0xFF) +#define VMCB_EXITINTINFO_TYPE(x) (((x) >> 8) & 0x7) +#define VMCB_EXITINTINFO_EC_VALID(x) (((x) & BIT(11)) ? 1 : 0) +#define VMCB_EXITINTINFO_VALID(x) (((x) & BIT(31)) ? 1 : 0) +#define VMCB_EXITINTINFO_EC(x) (((x) >> 32) & 0xFFFFFFFF) + +/* Offset of various VMCB fields. */ +#define VMCB_OFF_CTRL(x) (x) +#define VMCB_OFF_STATE(x) ((x) + 0x400) + +#define VMCB_OFF_CR_INTERCEPT VMCB_OFF_CTRL(0x0) +#define VMCB_OFF_DR_INTERCEPT VMCB_OFF_CTRL(0x4) +#define VMCB_OFF_EXC_INTERCEPT VMCB_OFF_CTRL(0x8) +#define VMCB_OFF_INST1_INTERCEPT VMCB_OFF_CTRL(0xC) +#define VMCB_OFF_INST2_INTERCEPT VMCB_OFF_CTRL(0x10) +#define VMCB_OFF_IO_PERM VMCB_OFF_CTRL(0x40) +#define VMCB_OFF_MSR_PERM VMCB_OFF_CTRL(0x48) +#define VMCB_OFF_TSC_OFFSET VMCB_OFF_CTRL(0x50) +#define VMCB_OFF_ASID VMCB_OFF_CTRL(0x58) +#define VMCB_OFF_TLB_CTRL VMCB_OFF_CTRL(0x5C) +#define VMCB_OFF_VIRQ VMCB_OFF_CTRL(0x60) +#define VMCB_OFF_EXIT_REASON VMCB_OFF_CTRL(0x70) +#define VMCB_OFF_EXITINFO1 VMCB_OFF_CTRL(0x78) +#define VMCB_OFF_EXITINFO2 VMCB_OFF_CTRL(0x80) +#define VMCB_OFF_EXITINTINFO VMCB_OFF_CTRL(0x88) +#define VMCB_OFF_AVIC_BAR VMCB_OFF_CTRL(0x98) +#define VMCB_OFF_NPT_BASE VMCB_OFF_CTRL(0xB0) +#define VMCB_OFF_AVIC_PAGE VMCB_OFF_CTRL(0xE0) +#define VMCB_OFF_AVIC_LT VMCB_OFF_CTRL(0xF0) +#define VMCB_OFF_AVIC_PT VMCB_OFF_CTRL(0xF8) +#define VMCB_OFF_SYSENTER_CS VMCB_OFF_STATE(0x228) +#define VMCB_OFF_SYSENTER_ESP VMCB_OFF_STATE(0x230) +#define VMCB_OFF_SYSENTER_EIP VMCB_OFF_STATE(0x238) +#define VMCB_OFF_GUEST_PAT VMCB_OFF_STATE(0x268) + +/* + * Encode the VMCB offset and bytes that we want to read from VMCB. + */ +#define VMCB_ACCESS(o, w) (0x80000000 | (((w) & 0xF) << 16) | \ + ((o) & 0xFFF)) +#define VMCB_ACCESS_OK(v) ((v) & 0x80000000 ) +#define VMCB_ACCESS_BYTES(v) (((v) >> 16) & 0xF) +#define VMCB_ACCESS_OFFSET(v) ((v) & 0xFFF) + +#ifdef _KERNEL +/* VMCB save state area segment format */ +struct vmcb_segment { + uint16_t selector; + uint16_t attrib; + uint32_t limit; + uint64_t base; +} __attribute__ ((__packed__)); +CTASSERT(sizeof(struct vmcb_segment) == 16); + +/* Code segment descriptor attribute in 12 bit format as saved by VMCB. */ +#define VMCB_CS_ATTRIB_L BIT(9) /* Long mode. */ +#define VMCB_CS_ATTRIB_D BIT(10) /* OPerand size bit. */ + +/* + * The VMCB is divided into two areas - the first one contains various + * control bits including the intercept vector and the second one contains + * the guest state. + */ + +/* VMCB control area - padded up to 1024 bytes */ +struct vmcb_ctrl { + uint32_t intercept[5]; /* all intercepts */ + uint8_t pad1[0x28]; /* Offsets 0x14-0x3B are reserved. */ + uint16_t pause_filthresh; /* Offset 0x3C, PAUSE filter threshold */ + uint16_t pause_filcnt; /* Offset 0x3E, PAUSE filter count */ + uint64_t iopm_base_pa; /* 0x40: IOPM_BASE_PA */ + uint64_t msrpm_base_pa; /* 0x48: MSRPM_BASE_PA */ + uint64_t tsc_offset; /* 0x50: TSC_OFFSET */ + uint32_t asid; /* 0x58: Guest ASID */ + uint8_t tlb_ctrl; /* 0x5C: TLB_CONTROL */ + uint8_t pad2[3]; /* 0x5D-0x5F: Reserved. */ + uint8_t v_tpr; /* 0x60: V_TPR, guest CR8 */ + uint8_t v_irq:1; /* Is virtual interrupt pending? */ + uint8_t :7; /* Padding */ + uint8_t v_intr_prio:4; /* 0x62: Priority for virtual interrupt. */ + uint8_t v_ign_tpr:1; + uint8_t :3; + uint8_t v_intr_masking:1; /* Guest and host sharing of RFLAGS. */ + uint8_t :7; + uint8_t v_intr_vector; /* 0x65: Vector for virtual interrupt. */ + uint8_t pad3[3]; /* Bit64-40 Reserved. */ + uint64_t intr_shadow:1; /* 0x68: Interrupt shadow, section15.2.1 APM2 */ + uint64_t :63; + uint64_t exitcode; /* 0x70, Exitcode */ + uint64_t exitinfo1; /* 0x78, EXITINFO1 */ + uint64_t exitinfo2; /* 0x80, EXITINFO2 */ + uint64_t exitintinfo; /* 0x88, Interrupt exit value. */ + uint64_t np_enable:1; /* 0x90, Nested paging enable. */ + uint64_t :63; + uint8_t pad4[0x10]; /* 0x98-0xA7 reserved. */ + uint64_t eventinj; /* 0xA8, Event injection. */ + uint64_t n_cr3; /* B0, Nested page table. */ + uint64_t lbr_virt_en:1; /* Enable LBR virtualization. */ + uint64_t :63; + uint32_t vmcb_clean; /* 0xC0: VMCB clean bits for caching */ + uint32_t :32; /* 0xC4: Reserved */ + uint64_t nrip; /* 0xC8: Guest next nRIP. */ + uint8_t inst_len; /* 0xD0: #NPF decode assist */ + uint8_t inst_bytes[15]; + uint8_t padd6[0x320]; +} __attribute__ ((__packed__)); +CTASSERT(sizeof(struct vmcb_ctrl) == 1024); + +struct vmcb_state { + struct vmcb_segment es; + struct vmcb_segment cs; + struct vmcb_segment ss; + struct vmcb_segment ds; + struct vmcb_segment fs; + struct vmcb_segment gs; + struct vmcb_segment gdt; + struct vmcb_segment ldt; + struct vmcb_segment idt; + struct vmcb_segment tr; + uint8_t pad1[0x2b]; /* Reserved: 0xA0-0xCA */ + uint8_t cpl; + uint8_t pad2[4]; + uint64_t efer; + uint8_t pad3[0x70]; /* Reserved: 0xd8-0x147 */ + uint64_t cr4; + uint64_t cr3; /* Guest CR3 */ + uint64_t cr0; + uint64_t dr7; + uint64_t dr6; + uint64_t rflags; + uint64_t rip; + uint8_t pad4[0x58]; /* Reserved: 0x180-0x1D7 */ + uint64_t rsp; + uint8_t pad5[0x18]; /* Reserved 0x1E0-0x1F7 */ + uint64_t rax; + uint64_t star; + uint64_t lstar; + uint64_t cstar; + uint64_t sfmask; + uint64_t kernelgsbase; + uint64_t sysenter_cs; + uint64_t sysenter_esp; + uint64_t sysenter_eip; + uint64_t cr2; + uint8_t pad6[0x20]; + uint64_t g_pat; + uint64_t dbgctl; + uint64_t br_from; + uint64_t br_to; + uint64_t int_from; + uint64_t int_to; + uint8_t pad7[0x968]; /* Reserved upto end of VMCB */ +} __attribute__ ((__packed__)); +CTASSERT(sizeof(struct vmcb_state) == 0xC00); + +struct vmcb { + struct vmcb_ctrl ctrl; + struct vmcb_state state; +} __attribute__ ((__packed__)); +CTASSERT(sizeof(struct vmcb) == PAGE_SIZE); +CTASSERT(offsetof(struct vmcb, state) == 0x400); + +int vmcb_read(struct svm_softc *sc, int vcpu, int ident, uint64_t *retval); +int vmcb_write(struct svm_softc *sc, int vcpu, int ident, uint64_t val); +int vmcb_setdesc(void *arg, int vcpu, int ident, struct seg_desc *desc); +int vmcb_getdesc(void *arg, int vcpu, int ident, struct seg_desc *desc); +int vmcb_seg(struct vmcb *vmcb, int ident, struct vmcb_segment *seg); + +#endif /* _KERNEL */ +#endif /* _VMCB_H_ */ diff --git a/sys/amd64/vmm/intel/vmx.c b/sys/amd64/vmm/intel/vmx.c index 2fe5a27..c855697 100644 --- a/sys/amd64/vmm/intel/vmx.c +++ b/sys/amd64/vmm/intel/vmx.c @@ -1746,8 +1746,6 @@ inout_str_seginfo(struct vmx *vmx, int vcpuid, uint32_t inst_info, int in, error = vmx_getdesc(vmx, vcpuid, vis->seg_name, &vis->seg_desc); KASSERT(error == 0, ("%s: vmx_getdesc error %d", __func__, error)); - - /* XXX modify svm.c to update bit 16 of seg_desc.access (unusable) */ } static void @@ -1781,6 +1779,7 @@ vmexit_inst_emul(struct vm_exit *vmexit, uint64_t gpa, uint64_t gla) vmexit->u.inst_emul.cs_d = 0; break; } + vie_init(&vmexit->u.inst_emul.vie, NULL, 0); } static int diff --git a/sys/amd64/vmm/io/vlapic.c b/sys/amd64/vmm/io/vlapic.c index d684dba..b192735 100644 --- a/sys/amd64/vmm/io/vlapic.c +++ b/sys/amd64/vmm/io/vlapic.c @@ -912,8 +912,12 @@ vlapic_set_tpr(struct vlapic *vlapic, uint8_t val) { struct LAPIC *lapic = vlapic->apic_page; - lapic->tpr = val; - vlapic_update_ppr(vlapic); + if (lapic->tpr != val) { + VCPU_CTR2(vlapic->vm, vlapic->vcpuid, "vlapic TPR changed " + "from %#x to %#x", lapic->tpr, val); + lapic->tpr = val; + vlapic_update_ppr(vlapic); + } } static uint8_t diff --git a/sys/amd64/vmm/vmm.c b/sys/amd64/vmm/vmm.c index dd0829c..7d06749 100644 --- a/sys/amd64/vmm/vmm.c +++ b/sys/amd64/vmm/vmm.c @@ -1231,7 +1231,7 @@ vm_handle_inst_emul(struct vm *vm, int vcpuid, bool *retu) mem_region_read_t mread; mem_region_write_t mwrite; enum vm_cpu_mode cpu_mode; - int cs_d, error; + int cs_d, error, length; vcpu = &vm->vcpu[vcpuid]; vme = &vcpu->exitinfo; @@ -1245,11 +1245,21 @@ vm_handle_inst_emul(struct vm *vm, int vcpuid, bool *retu) VCPU_CTR1(vm, vcpuid, "inst_emul fault accessing gpa %#lx", gpa); - vie_init(vie); - /* Fetch, decode and emulate the faulting instruction */ - error = vmm_fetch_instruction(vm, vcpuid, paging, vme->rip, - vme->inst_length, vie); + if (vie->num_valid == 0) { + /* + * If the instruction length is not known then assume a + * maximum size instruction. + */ + length = vme->inst_length ? vme->inst_length : VIE_INST_SIZE; + error = vmm_fetch_instruction(vm, vcpuid, paging, vme->rip, + length, vie); + } else { + /* + * The instruction bytes have already been copied into 'vie' + */ + error = 0; + } if (error == 1) return (0); /* Resume guest to handle page fault */ else if (error == -1) @@ -1260,6 +1270,12 @@ vm_handle_inst_emul(struct vm *vm, int vcpuid, bool *retu) if (vmm_decode_instruction(vm, vcpuid, gla, cpu_mode, cs_d, vie) != 0) return (EFAULT); + /* + * If the instruction length is not specified the update it now. + */ + if (vme->inst_length == 0) + vme->inst_length = vie->num_processed; + /* return to userland unless this is an in-kernel emulated device */ if (gpa >= DEFAULT_APIC_BASE && gpa < DEFAULT_APIC_BASE + PAGE_SIZE) { mread = lapic_mmio_read; diff --git a/sys/amd64/vmm/vmm_instruction_emul.c b/sys/amd64/vmm/vmm_instruction_emul.c index c6ba01e..d1d7173 100644 --- a/sys/amd64/vmm/vmm_instruction_emul.c +++ b/sys/amd64/vmm/vmm_instruction_emul.c @@ -1184,13 +1184,20 @@ vie_calculate_gla(enum vm_cpu_mode cpu_mode, enum vm_reg_name seg, #ifdef _KERNEL void -vie_init(struct vie *vie) +vie_init(struct vie *vie, const char *inst_bytes, int inst_length) { + KASSERT(inst_length >= 0 && inst_length <= VIE_INST_SIZE, + ("%s: invalid instruction length (%d)", __func__, inst_length)); bzero(vie, sizeof(struct vie)); vie->base_register = VM_REG_LAST; vie->index_register = VM_REG_LAST; + + if (inst_length) { + bcopy(inst_bytes, vie->inst, inst_length); + vie->num_valid = inst_length; + } } static int @@ -1826,7 +1833,7 @@ static int verify_inst_length(struct vie *vie) { - if (vie->num_processed == vie->num_valid) + if (vie->num_processed) return (0); else return (-1); diff --git a/sys/amd64/vmm/x86.c b/sys/amd64/vmm/x86.c index c7515cf..c0c3e9c 100644 --- a/sys/amd64/vmm/x86.c +++ b/sys/amd64/vmm/x86.c @@ -44,6 +44,8 @@ __FBSDID("$FreeBSD$"); #include <machine/vmm.h> #include "vmm_host.h" +#include "vmm_ktr.h" +#include "vmm_util.h" #include "x86.h" SYSCTL_DECL(_hw_vmm); @@ -54,6 +56,8 @@ static SYSCTL_NODE(_hw_vmm, OID_AUTO, topology, CTLFLAG_RD, 0, NULL); static const char bhyve_id[12] = "bhyve bhyve "; static uint64_t bhyve_xcpuids; +SYSCTL_ULONG(_hw_vmm, OID_AUTO, bhyve_xcpuids, CTLFLAG_RW, &bhyve_xcpuids, 0, + "Number of times an unknown cpuid leaf was accessed"); /* * The default CPU topology is a single thread per package. @@ -91,6 +95,8 @@ x86_emulate_cpuid(struct vm *vm, int vcpu_id, unsigned int func, regs[4], logical_cpus; enum x2apic_state x2apic_state; + VCPU_CTR2(vm, vcpu_id, "cpuid %#x,%#x", *eax, *ecx); + /* * Requests for invalid CPUID levels should map to the highest * available level instead. @@ -124,25 +130,80 @@ x86_emulate_cpuid(struct vm *vm, int vcpu_id, case CPUID_8000_0003: case CPUID_8000_0004: case CPUID_8000_0006: + cpuid_count(*eax, *ecx, regs); + break; case CPUID_8000_0008: cpuid_count(*eax, *ecx, regs); + if (vmm_is_amd()) { + /* + * XXX this might appear silly because AMD + * cpus don't have threads. + * + * However this matches the logical cpus as + * advertised by leaf 0x1 and will work even + * if the 'threads_per_core' tunable is set + * incorrectly on an AMD host. + */ + logical_cpus = threads_per_core * + cores_per_package; + regs[2] = logical_cpus - 1; + } break; case CPUID_8000_0001: + cpuid_count(*eax, *ecx, regs); + + /* + * Hide SVM and Topology Extension features from guest. + */ + regs[2] &= ~(AMDID2_SVM | AMDID2_TOPOLOGY); + + /* + * Don't advertise extended performance counter MSRs + * to the guest. + */ + regs[2] &= ~AMDID2_PCXC; + regs[2] &= ~AMDID2_PNXC; + regs[2] &= ~AMDID2_PTSCEL2I; + + /* + * Don't advertise Instruction Based Sampling feature. + */ + regs[2] &= ~AMDID2_IBS; + + /* NodeID MSR not available */ + regs[2] &= ~AMDID2_NODE_ID; + + /* Don't advertise the OS visible workaround feature */ + regs[2] &= ~AMDID2_OSVW; + /* * Hide rdtscp/ia32_tsc_aux until we know how * to deal with them. */ - cpuid_count(*eax, *ecx, regs); regs[3] &= ~AMDID_RDTSCP; break; case CPUID_8000_0007: - cpuid_count(*eax, *ecx, regs); /* - * If the host TSCs are not synchronized across - * physical cpus then we cannot advertise an - * invariant tsc to a vcpu. + * AMD uses this leaf to advertise the processor's + * power monitoring and RAS capabilities. These + * features are hardware-specific and exposing + * them to a guest doesn't make a lot of sense. + * + * Intel uses this leaf only to advertise the + * "Invariant TSC" feature with all other bits + * being reserved (set to zero). + */ + regs[0] = 0; + regs[1] = 0; + regs[2] = 0; + regs[3] = 0; + + /* + * "Invariant TSC" can be advertised to the guest if: + * - host TSC frequency is invariant + * - host TSCs are synchronized across physical cpus * * XXX This still falls short because the vcpu * can observe the TSC moving backwards as it @@ -150,8 +211,8 @@ x86_emulate_cpuid(struct vm *vm, int vcpu_id, * it should discourage the guest from using the * TSC to keep track of time. */ - if (!smp_tsc) - regs[3] &= ~AMDPM_TSC_INVARIANT; + if (tsc_is_invariant && smp_tsc) + regs[3] |= AMDPM_TSC_INVARIANT; break; case CPUID_0000_0001: |