Index: head/sys/amd64/include/vmm.h =================================================================== --- head/sys/amd64/include/vmm.h (revision 283656) +++ head/sys/amd64/include/vmm.h (revision 283657) @@ -1,648 +1,662 @@ /*- * Copyright (c) 2011 NetApp, Inc. * 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, 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 NETAPP, INC ``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 NETAPP, INC OR CONTRIBUTORS 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 _VMM_H_ #define _VMM_H_ #include enum vm_suspend_how { VM_SUSPEND_NONE, VM_SUSPEND_RESET, VM_SUSPEND_POWEROFF, VM_SUSPEND_HALT, VM_SUSPEND_TRIPLEFAULT, VM_SUSPEND_LAST }; /* * Identifiers for architecturally defined registers. */ enum vm_reg_name { VM_REG_GUEST_RAX, VM_REG_GUEST_RBX, VM_REG_GUEST_RCX, VM_REG_GUEST_RDX, VM_REG_GUEST_RSI, VM_REG_GUEST_RDI, VM_REG_GUEST_RBP, VM_REG_GUEST_R8, VM_REG_GUEST_R9, VM_REG_GUEST_R10, VM_REG_GUEST_R11, VM_REG_GUEST_R12, VM_REG_GUEST_R13, VM_REG_GUEST_R14, VM_REG_GUEST_R15, VM_REG_GUEST_CR0, VM_REG_GUEST_CR3, VM_REG_GUEST_CR4, VM_REG_GUEST_DR7, VM_REG_GUEST_RSP, VM_REG_GUEST_RIP, VM_REG_GUEST_RFLAGS, VM_REG_GUEST_ES, VM_REG_GUEST_CS, VM_REG_GUEST_SS, VM_REG_GUEST_DS, VM_REG_GUEST_FS, VM_REG_GUEST_GS, VM_REG_GUEST_LDTR, VM_REG_GUEST_TR, VM_REG_GUEST_IDTR, VM_REG_GUEST_GDTR, VM_REG_GUEST_EFER, VM_REG_GUEST_CR2, VM_REG_GUEST_PDPTE0, VM_REG_GUEST_PDPTE1, VM_REG_GUEST_PDPTE2, VM_REG_GUEST_PDPTE3, VM_REG_GUEST_INTR_SHADOW, VM_REG_LAST }; enum x2apic_state { X2APIC_DISABLED, X2APIC_ENABLED, X2APIC_STATE_LAST }; #define VM_INTINFO_VECTOR(info) ((info) & 0xff) #define VM_INTINFO_DEL_ERRCODE 0x800 #define VM_INTINFO_RSVD 0x7ffff000 #define VM_INTINFO_VALID 0x80000000 #define VM_INTINFO_TYPE 0x700 #define VM_INTINFO_HWINTR (0 << 8) #define VM_INTINFO_NMI (2 << 8) #define VM_INTINFO_HWEXCEPTION (3 << 8) #define VM_INTINFO_SWINTR (4 << 8) #ifdef _KERNEL #define VM_MAX_NAMELEN 32 struct vm; struct vm_exception; struct vm_memory_segment; struct seg_desc; struct vm_exit; struct vm_run; struct vhpet; struct vioapic; struct vlapic; struct vmspace; struct vm_object; struct vm_guest_paging; struct pmap; +struct vm_eventinfo { + void *rptr; /* rendezvous cookie */ + int *sptr; /* suspend cookie */ + int *iptr; /* reqidle cookie */ +}; + typedef int (*vmm_init_func_t)(int ipinum); typedef int (*vmm_cleanup_func_t)(void); typedef void (*vmm_resume_func_t)(void); typedef void * (*vmi_init_func_t)(struct vm *vm, struct pmap *pmap); typedef int (*vmi_run_func_t)(void *vmi, int vcpu, register_t rip, - struct pmap *pmap, void *rendezvous_cookie, - void *suspend_cookie); + struct pmap *pmap, struct vm_eventinfo *info); typedef void (*vmi_cleanup_func_t)(void *vmi); typedef int (*vmi_get_register_t)(void *vmi, int vcpu, int num, uint64_t *retval); typedef int (*vmi_set_register_t)(void *vmi, int vcpu, int num, uint64_t val); typedef int (*vmi_get_desc_t)(void *vmi, int vcpu, int num, struct seg_desc *desc); typedef int (*vmi_set_desc_t)(void *vmi, int vcpu, int num, struct seg_desc *desc); typedef int (*vmi_get_cap_t)(void *vmi, int vcpu, int num, int *retval); typedef int (*vmi_set_cap_t)(void *vmi, int vcpu, int num, int val); typedef struct vmspace * (*vmi_vmspace_alloc)(vm_offset_t min, vm_offset_t max); typedef void (*vmi_vmspace_free)(struct vmspace *vmspace); typedef struct vlapic * (*vmi_vlapic_init)(void *vmi, int vcpu); typedef void (*vmi_vlapic_cleanup)(void *vmi, struct vlapic *vlapic); struct vmm_ops { vmm_init_func_t init; /* module wide initialization */ vmm_cleanup_func_t cleanup; vmm_resume_func_t resume; vmi_init_func_t vminit; /* vm-specific initialization */ vmi_run_func_t vmrun; vmi_cleanup_func_t vmcleanup; vmi_get_register_t vmgetreg; vmi_set_register_t vmsetreg; vmi_get_desc_t vmgetdesc; vmi_set_desc_t vmsetdesc; vmi_get_cap_t vmgetcap; vmi_set_cap_t vmsetcap; vmi_vmspace_alloc vmspace_alloc; vmi_vmspace_free vmspace_free; vmi_vlapic_init vlapic_init; vmi_vlapic_cleanup vlapic_cleanup; }; extern struct vmm_ops vmm_ops_intel; extern struct vmm_ops vmm_ops_amd; int vm_create(const char *name, struct vm **retvm); void vm_destroy(struct vm *vm); int vm_reinit(struct vm *vm); const char *vm_name(struct vm *vm); int vm_malloc(struct vm *vm, vm_paddr_t gpa, size_t len); int vm_map_mmio(struct vm *vm, vm_paddr_t gpa, size_t len, vm_paddr_t hpa); int vm_unmap_mmio(struct vm *vm, vm_paddr_t gpa, size_t len); void *vm_gpa_hold(struct vm *, vm_paddr_t gpa, size_t len, int prot, void **cookie); void vm_gpa_release(void *cookie); int vm_gpabase2memseg(struct vm *vm, vm_paddr_t gpabase, struct vm_memory_segment *seg); int vm_get_memobj(struct vm *vm, vm_paddr_t gpa, size_t len, vm_offset_t *offset, struct vm_object **object); boolean_t vm_mem_allocated(struct vm *vm, vm_paddr_t gpa); int vm_get_register(struct vm *vm, int vcpu, int reg, uint64_t *retval); int vm_set_register(struct vm *vm, int vcpu, int reg, uint64_t val); int vm_get_seg_desc(struct vm *vm, int vcpu, int reg, struct seg_desc *ret_desc); int vm_set_seg_desc(struct vm *vm, int vcpu, int reg, struct seg_desc *desc); int vm_run(struct vm *vm, struct vm_run *vmrun); int vm_suspend(struct vm *vm, enum vm_suspend_how how); int vm_inject_nmi(struct vm *vm, int vcpu); int vm_nmi_pending(struct vm *vm, int vcpuid); void vm_nmi_clear(struct vm *vm, int vcpuid); int vm_inject_extint(struct vm *vm, int vcpu); int vm_extint_pending(struct vm *vm, int vcpuid); void vm_extint_clear(struct vm *vm, int vcpuid); struct vlapic *vm_lapic(struct vm *vm, int cpu); struct vioapic *vm_ioapic(struct vm *vm); struct vhpet *vm_hpet(struct vm *vm); int vm_get_capability(struct vm *vm, int vcpu, int type, int *val); int vm_set_capability(struct vm *vm, int vcpu, int type, int val); int vm_get_x2apic_state(struct vm *vm, int vcpu, enum x2apic_state *state); int vm_set_x2apic_state(struct vm *vm, int vcpu, enum x2apic_state state); int vm_apicid2vcpuid(struct vm *vm, int apicid); int vm_activate_cpu(struct vm *vm, int vcpu); struct vm_exit *vm_exitinfo(struct vm *vm, int vcpuid); void vm_exit_suspended(struct vm *vm, int vcpuid, uint64_t rip); void vm_exit_rendezvous(struct vm *vm, int vcpuid, uint64_t rip); void vm_exit_astpending(struct vm *vm, int vcpuid, uint64_t rip); +void vm_exit_reqidle(struct vm *vm, int vcpuid, uint64_t rip); #ifdef _SYS__CPUSET_H_ /* * Rendezvous all vcpus specified in 'dest' and execute 'func(arg)'. * The rendezvous 'func(arg)' is not allowed to do anything that will * cause the thread to be put to sleep. * * If the rendezvous is being initiated from a vcpu context then the * 'vcpuid' must refer to that vcpu, otherwise it should be set to -1. * * The caller cannot hold any locks when initiating the rendezvous. * * The implementation of this API may cause vcpus other than those specified * by 'dest' to be stalled. The caller should not rely on any vcpus making * forward progress when the rendezvous is in progress. */ typedef void (*vm_rendezvous_func_t)(struct vm *vm, int vcpuid, void *arg); void vm_smp_rendezvous(struct vm *vm, int vcpuid, cpuset_t dest, vm_rendezvous_func_t func, void *arg); cpuset_t vm_active_cpus(struct vm *vm); cpuset_t vm_suspended_cpus(struct vm *vm); #endif /* _SYS__CPUSET_H_ */ static __inline int -vcpu_rendezvous_pending(void *rendezvous_cookie) +vcpu_rendezvous_pending(struct vm_eventinfo *info) { - return (*(uintptr_t *)rendezvous_cookie != 0); + return (*((uintptr_t *)(info->rptr)) != 0); } static __inline int -vcpu_suspended(void *suspend_cookie) +vcpu_suspended(struct vm_eventinfo *info) { - return (*(int *)suspend_cookie); + return (*info->sptr); } +static __inline int +vcpu_reqidle(struct vm_eventinfo *info) +{ + + return (*info->iptr); +} + /* * Return 1 if device indicated by bus/slot/func is supposed to be a * pci passthrough device. * * Return 0 otherwise. */ int vmm_is_pptdev(int bus, int slot, int func); void *vm_iommu_domain(struct vm *vm); enum vcpu_state { VCPU_IDLE, VCPU_FROZEN, VCPU_RUNNING, VCPU_SLEEPING, }; int vcpu_set_state(struct vm *vm, int vcpu, enum vcpu_state state, bool from_idle); enum vcpu_state vcpu_get_state(struct vm *vm, int vcpu, int *hostcpu); static int __inline vcpu_is_running(struct vm *vm, int vcpu, int *hostcpu) { return (vcpu_get_state(vm, vcpu, hostcpu) == VCPU_RUNNING); } #ifdef _SYS_PROC_H_ static int __inline vcpu_should_yield(struct vm *vm, int vcpu) { if (curthread->td_flags & (TDF_ASTPENDING | TDF_NEEDRESCHED)) return (1); else if (curthread->td_owepreempt) return (1); else return (0); } #endif void *vcpu_stats(struct vm *vm, int vcpu); void vcpu_notify_event(struct vm *vm, int vcpuid, bool lapic_intr); struct vmspace *vm_get_vmspace(struct vm *vm); int vm_assign_pptdev(struct vm *vm, int bus, int slot, int func); int vm_unassign_pptdev(struct vm *vm, int bus, int slot, int func); struct vatpic *vm_atpic(struct vm *vm); struct vatpit *vm_atpit(struct vm *vm); struct vpmtmr *vm_pmtmr(struct vm *vm); struct vrtc *vm_rtc(struct vm *vm); /* * Inject exception 'vector' into the guest vcpu. This function returns 0 on * success and non-zero on failure. * * Wrapper functions like 'vm_inject_gp()' should be preferred to calling * this function directly because they enforce the trap-like or fault-like * behavior of an exception. * * This function should only be called in the context of the thread that is * executing this vcpu. */ int vm_inject_exception(struct vm *vm, int vcpuid, int vector, int err_valid, uint32_t errcode, int restart_instruction); /* * This function is called after a VM-exit that occurred during exception or * interrupt delivery through the IDT. The format of 'intinfo' is described * in Figure 15-1, "EXITINTINFO for All Intercepts", APM, Vol 2. * * If a VM-exit handler completes the event delivery successfully then it * should call vm_exit_intinfo() to extinguish the pending event. For e.g., * if the task switch emulation is triggered via a task gate then it should * call this function with 'intinfo=0' to indicate that the external event * is not pending anymore. * * Return value is 0 on success and non-zero on failure. */ int vm_exit_intinfo(struct vm *vm, int vcpuid, uint64_t intinfo); /* * This function is called before every VM-entry to retrieve a pending * event that should be injected into the guest. This function combines * nested events into a double or triple fault. * * Returns 0 if there are no events that need to be injected into the guest * and non-zero otherwise. */ int vm_entry_intinfo(struct vm *vm, int vcpuid, uint64_t *info); int vm_get_intinfo(struct vm *vm, int vcpuid, uint64_t *info1, uint64_t *info2); enum vm_reg_name vm_segment_name(int seg_encoding); struct vm_copyinfo { uint64_t gpa; size_t len; void *hva; void *cookie; }; /* * Set up 'copyinfo[]' to copy to/from guest linear address space starting * at 'gla' and 'len' bytes long. The 'prot' should be set to PROT_READ for * a copyin or PROT_WRITE for a copyout. * * retval is_fault Intepretation * 0 0 Success * 0 1 An exception was injected into the guest * EFAULT N/A Unrecoverable error * * The 'copyinfo[]' can be passed to 'vm_copyin()' or 'vm_copyout()' only if * the return value is 0. The 'copyinfo[]' resources should be freed by calling * 'vm_copy_teardown()' after the copy is done. */ int vm_copy_setup(struct vm *vm, int vcpuid, struct vm_guest_paging *paging, uint64_t gla, size_t len, int prot, struct vm_copyinfo *copyinfo, int num_copyinfo, int *is_fault); void vm_copy_teardown(struct vm *vm, int vcpuid, struct vm_copyinfo *copyinfo, int num_copyinfo); void vm_copyin(struct vm *vm, int vcpuid, struct vm_copyinfo *copyinfo, void *kaddr, size_t len); void vm_copyout(struct vm *vm, int vcpuid, const void *kaddr, struct vm_copyinfo *copyinfo, size_t len); int vcpu_trace_exceptions(struct vm *vm, int vcpuid); #endif /* KERNEL */ #define VM_MAXCPU 16 /* maximum virtual cpus */ /* * Identifiers for optional vmm capabilities */ enum vm_cap_type { VM_CAP_HALT_EXIT, VM_CAP_MTRAP_EXIT, VM_CAP_PAUSE_EXIT, VM_CAP_UNRESTRICTED_GUEST, VM_CAP_ENABLE_INVPCID, VM_CAP_MAX }; enum vm_intr_trigger { EDGE_TRIGGER, LEVEL_TRIGGER }; /* * The 'access' field has the format specified in Table 21-2 of the Intel * Architecture Manual vol 3b. * * XXX The contents of the 'access' field are architecturally defined except * bit 16 - Segment Unusable. */ struct seg_desc { uint64_t base; uint32_t limit; uint32_t access; }; #define SEG_DESC_TYPE(access) ((access) & 0x001f) #define SEG_DESC_DPL(access) (((access) >> 5) & 0x3) #define SEG_DESC_PRESENT(access) (((access) & 0x0080) ? 1 : 0) #define SEG_DESC_DEF32(access) (((access) & 0x4000) ? 1 : 0) #define SEG_DESC_GRANULARITY(access) (((access) & 0x8000) ? 1 : 0) #define SEG_DESC_UNUSABLE(access) (((access) & 0x10000) ? 1 : 0) enum vm_cpu_mode { CPU_MODE_REAL, CPU_MODE_PROTECTED, CPU_MODE_COMPATIBILITY, /* IA-32E mode (CS.L = 0) */ CPU_MODE_64BIT, /* IA-32E mode (CS.L = 1) */ }; enum vm_paging_mode { PAGING_MODE_FLAT, PAGING_MODE_32, PAGING_MODE_PAE, PAGING_MODE_64, }; struct vm_guest_paging { uint64_t cr3; int cpl; enum vm_cpu_mode cpu_mode; enum vm_paging_mode paging_mode; }; /* * The data structures 'vie' and 'vie_op' are meant to be opaque to the * consumers of instruction decoding. The only reason why their contents * need to be exposed is because they are part of the 'vm_exit' structure. */ struct vie_op { uint8_t op_byte; /* actual opcode byte */ uint8_t op_type; /* type of operation (e.g. MOV) */ uint16_t op_flags; }; #define VIE_INST_SIZE 15 struct vie { uint8_t inst[VIE_INST_SIZE]; /* instruction bytes */ uint8_t num_valid; /* size of the instruction */ uint8_t num_processed; uint8_t addrsize:4, opsize:4; /* address and operand sizes */ uint8_t rex_w:1, /* REX prefix */ rex_r:1, rex_x:1, rex_b:1, rex_present:1, repz_present:1, /* REP/REPE/REPZ prefix */ repnz_present:1, /* REPNE/REPNZ prefix */ opsize_override:1, /* Operand size override */ addrsize_override:1, /* Address size override */ segment_override:1; /* Segment override */ uint8_t mod:2, /* ModRM byte */ reg:4, rm:4; uint8_t ss:2, /* SIB byte */ index:4, base:4; uint8_t disp_bytes; uint8_t imm_bytes; uint8_t scale; int base_register; /* VM_REG_GUEST_xyz */ int index_register; /* VM_REG_GUEST_xyz */ int segment_register; /* VM_REG_GUEST_xyz */ int64_t displacement; /* optional addr displacement */ int64_t immediate; /* optional immediate operand */ uint8_t decoded; /* set to 1 if successfully decoded */ struct vie_op op; /* opcode description */ }; enum vm_exitcode { VM_EXITCODE_INOUT, VM_EXITCODE_VMX, VM_EXITCODE_BOGUS, VM_EXITCODE_RDMSR, VM_EXITCODE_WRMSR, VM_EXITCODE_HLT, VM_EXITCODE_MTRAP, VM_EXITCODE_PAUSE, VM_EXITCODE_PAGING, VM_EXITCODE_INST_EMUL, VM_EXITCODE_SPINUP_AP, VM_EXITCODE_DEPRECATED1, /* used to be SPINDOWN_CPU */ VM_EXITCODE_RENDEZVOUS, VM_EXITCODE_IOAPIC_EOI, VM_EXITCODE_SUSPENDED, VM_EXITCODE_INOUT_STR, VM_EXITCODE_TASK_SWITCH, VM_EXITCODE_MONITOR, VM_EXITCODE_MWAIT, VM_EXITCODE_SVM, + VM_EXITCODE_REQIDLE, VM_EXITCODE_MAX }; struct vm_inout { uint16_t bytes:3; /* 1 or 2 or 4 */ uint16_t in:1; uint16_t string:1; uint16_t rep:1; uint16_t port; uint32_t eax; /* valid for out */ }; struct vm_inout_str { struct vm_inout inout; /* must be the first element */ struct vm_guest_paging paging; uint64_t rflags; uint64_t cr0; uint64_t index; uint64_t count; /* rep=1 (%rcx), rep=0 (1) */ int addrsize; enum vm_reg_name seg_name; struct seg_desc seg_desc; }; enum task_switch_reason { TSR_CALL, TSR_IRET, TSR_JMP, TSR_IDT_GATE, /* task gate in IDT */ }; struct vm_task_switch { uint16_t tsssel; /* new TSS selector */ int ext; /* task switch due to external event */ uint32_t errcode; int errcode_valid; /* push 'errcode' on the new stack */ enum task_switch_reason reason; struct vm_guest_paging paging; }; struct vm_exit { enum vm_exitcode exitcode; int inst_length; /* 0 means unknown */ uint64_t rip; union { struct vm_inout inout; struct vm_inout_str inout_str; struct { uint64_t gpa; int fault_type; } paging; struct { uint64_t gpa; uint64_t gla; uint64_t cs_base; int cs_d; /* CS.D */ struct vm_guest_paging paging; struct vie vie; } inst_emul; /* * VMX specific payload. Used when there is no "better" * exitcode to represent the VM-exit. */ struct { int status; /* vmx inst status */ /* * 'exit_reason' and 'exit_qualification' are valid * only if 'status' is zero. */ uint32_t exit_reason; uint64_t exit_qualification; /* * 'inst_error' and 'inst_type' are valid * only if 'status' is non-zero. */ int inst_type; int inst_error; } vmx; /* * SVM specific payload. */ struct { uint64_t exitcode; uint64_t exitinfo1; uint64_t exitinfo2; } svm; struct { uint32_t code; /* ecx value */ uint64_t wval; } msr; struct { int vcpu; uint64_t rip; } spinup_ap; struct { uint64_t rflags; } hlt; struct { int vector; } ioapic_eoi; struct { enum vm_suspend_how how; } suspended; struct vm_task_switch task_switch; } u; }; /* APIs to inject faults into the guest */ void vm_inject_fault(void *vm, int vcpuid, int vector, int errcode_valid, int errcode); static __inline void vm_inject_ud(void *vm, int vcpuid) { vm_inject_fault(vm, vcpuid, IDT_UD, 0, 0); } static __inline void vm_inject_gp(void *vm, int vcpuid) { vm_inject_fault(vm, vcpuid, IDT_GP, 1, 0); } static __inline void vm_inject_ac(void *vm, int vcpuid, int errcode) { vm_inject_fault(vm, vcpuid, IDT_AC, 1, errcode); } static __inline void vm_inject_ss(void *vm, int vcpuid, int errcode) { vm_inject_fault(vm, vcpuid, IDT_SS, 1, errcode); } void vm_inject_pf(void *vm, int vcpuid, int error_code, uint64_t cr2); int vm_restart_instruction(void *vm, int vcpuid); #endif /* _VMM_H_ */ Index: head/sys/amd64/vmm/amd/svm.c =================================================================== --- head/sys/amd64/vmm/amd/svm.c (revision 283656) +++ head/sys/amd64/vmm/amd/svm.c (revision 283657) @@ -1,2259 +1,2265 @@ /*- * 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 __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #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 AMD_CPUID_SVM_AVIC BIT(13) /* AVIC present */ #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]; nasid = regs[1]; KASSERT(nasid > 1, ("Insufficient ASIDs for guests: %#x", nasid)); /* 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 everything when tracing guest exceptions otherwise * just intercept machine check exception. */ if (vcpu_trace_exceptions(sc->vm, vcpu)) { for (n = 0; n < 32; n++) { /* * Skip unimplemented vectors in the exception bitmap. */ if (n == 2 || n == 9) { continue; } svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(n)); } } else { 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); svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MONITOR); svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MWAIT); /* * 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->nextrip = ~0; 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); } /* * 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_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_REAL: vmexit->u.inst_emul.cs_base = seg.base; vmexit->u.inst_emul.cs_d = 0; break; case CPU_MODE_PROTECTED: case CPU_MODE_COMPATIBILITY: vmexit->u.inst_emul.cs_base = seg.base; /* * 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_base = 0; 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)); } #define EFER_MBZ_BITS 0xFFFFFFFFFFFF0200UL static int svm_write_efer(struct svm_softc *sc, int vcpu, uint64_t newval, bool *retu) { struct vm_exit *vme; struct vmcb_state *state; uint64_t changed, lma, oldval; int error; state = svm_get_vmcb_state(sc, vcpu); oldval = state->efer; VCPU_CTR2(sc->vm, vcpu, "wrmsr(efer) %#lx/%#lx", oldval, newval); newval &= ~0xFE; /* clear the Read-As-Zero (RAZ) bits */ changed = oldval ^ newval; if (newval & EFER_MBZ_BITS) goto gpf; /* APMv2 Table 14-5 "Long-Mode Consistency Checks" */ if (changed & EFER_LME) { if (state->cr0 & CR0_PG) goto gpf; } /* EFER.LMA = EFER.LME & CR0.PG */ if ((newval & EFER_LME) != 0 && (state->cr0 & CR0_PG) != 0) lma = EFER_LMA; else lma = 0; if ((newval & EFER_LMA) != lma) goto gpf; if (newval & EFER_NXE) { if (!vm_cpuid_capability(sc->vm, vcpu, VCC_NO_EXECUTE)) goto gpf; } /* * XXX bhyve does not enforce segment limits in 64-bit mode. Until * this is fixed flag guest attempt to set EFER_LMSLE as an error. */ if (newval & EFER_LMSLE) { vme = vm_exitinfo(sc->vm, vcpu); vm_exit_svm(vme, VMCB_EXIT_MSR, 1, 0); *retu = true; return (0); } if (newval & EFER_FFXSR) { if (!vm_cpuid_capability(sc->vm, vcpu, VCC_FFXSR)) goto gpf; } if (newval & EFER_TCE) { if (!vm_cpuid_capability(sc->vm, vcpu, VCC_TCE)) goto gpf; } error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, newval); KASSERT(error == 0, ("%s: error %d updating efer", __func__, error)); return (0); gpf: vm_inject_gp(sc->vm, vcpu); return (0); } 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_write_efer(sc, vcpu, val, retu); 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"); case VMCB_EXIT_MONITOR: return ("monitor"); case VMCB_EXIT_MWAIT: return ("mwait"); 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); } } 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 error, errcode_valid, handled, idtvec, reflect; 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 0x40 ... 0x5F: vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1); reflect = 1; idtvec = code - 0x40; switch (idtvec) { case IDT_MC: /* * Call the machine check handler by hand. Also don't * reflect the machine check back into the guest. */ reflect = 0; VCPU_CTR0(svm_sc->vm, vcpu, "Vectoring to MCE handler"); __asm __volatile("int $18"); break; case IDT_PF: error = svm_setreg(svm_sc, vcpu, VM_REG_GUEST_CR2, info2); KASSERT(error == 0, ("%s: error %d updating cr2", __func__, error)); /* fallthru */ case IDT_NP: case IDT_SS: case IDT_GP: case IDT_AC: case IDT_TS: errcode_valid = 1; break; case IDT_DF: errcode_valid = 1; info1 = 0; break; case IDT_BP: case IDT_OF: case IDT_BR: /* * The 'nrip' field is populated for INT3, INTO and * BOUND exceptions and this also implies that * 'inst_length' is non-zero. * * Reset 'inst_length' to zero so the guest %rip at * event injection is identical to what it was when * the exception originally happened. */ VCPU_CTR2(svm_sc->vm, vcpu, "Reset inst_length from %d " "to zero before injecting exception %d", vmexit->inst_length, idtvec); vmexit->inst_length = 0; /* fallthru */ default: errcode_valid = 0; info1 = 0; break; } KASSERT(vmexit->inst_length == 0, ("invalid inst_length (%d) " "when reflecting exception %d into guest", vmexit->inst_length, idtvec)); if (reflect) { /* Reflect the exception back into the guest */ VCPU_CTR2(svm_sc->vm, vcpu, "Reflecting exception " "%d/%#x into the guest", idtvec, (int)info1); error = vm_inject_exception(svm_sc->vm, vcpu, idtvec, errcode_valid, info1, 0); KASSERT(error == 0, ("%s: vm_inject_exception error %d", __func__, error)); } handled = 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; case VMCB_EXIT_MONITOR: vmexit->exitcode = VM_EXITCODE_MONITOR; break; case VMCB_EXIT_MWAIT: vmexit->exitcode = VM_EXITCODE_MWAIT; 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; struct svm_vcpu *vcpustate; 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); vcpustate = svm_get_vcpu(sc, vcpu); need_intr_window = 0; pending_apic_vector = 0; if (vcpustate->nextrip != state->rip) { ctrl->intr_shadow = 0; VCPU_CTR2(sc->vm, vcpu, "Guest interrupt blocking " "cleared due to rip change: %#lx/%#lx", vcpustate->nextrip, state->rip); } /* * 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 <= 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 vm_eventinfo *evinfo) { 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)) { + if (vcpu_suspended(evinfo)) { enable_gintr(); vm_exit_suspended(vm, vcpu, state->rip); break; } - if (vcpu_rendezvous_pending(rend_cookie)) { + if (vcpu_rendezvous_pending(evinfo)) { enable_gintr(); vm_exit_rendezvous(vm, vcpu, state->rip); + break; + } + + if (vcpu_reqidle(evinfo)) { + enable_gintr(); + vm_exit_reqidle(vm, vcpu, state->rip); break; } /* We are asked to give the cpu by scheduler. */ if (vcpu_should_yield(vm, vcpu)) { 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(); /* Update 'nextrip' */ vcpustate->nextrip = state->rip; /* 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 }; Index: head/sys/amd64/vmm/intel/vmx.c =================================================================== --- head/sys/amd64/vmm/intel/vmx.c (revision 283656) +++ head/sys/amd64/vmm/intel/vmx.c (revision 283657) @@ -1,3417 +1,3423 @@ /*- * Copyright (c) 2011 NetApp, Inc. * 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, 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 NETAPP, INC ``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 NETAPP, INC OR CONTRIBUTORS 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. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "vmm_lapic.h" #include "vmm_host.h" #include "vmm_ioport.h" #include "vmm_ktr.h" #include "vmm_stat.h" #include "vatpic.h" #include "vlapic.h" #include "vlapic_priv.h" #include "ept.h" #include "vmx_cpufunc.h" #include "vmx.h" #include "vmx_msr.h" #include "x86.h" #include "vmx_controls.h" #define PINBASED_CTLS_ONE_SETTING \ (PINBASED_EXTINT_EXITING | \ PINBASED_NMI_EXITING | \ PINBASED_VIRTUAL_NMI) #define PINBASED_CTLS_ZERO_SETTING 0 #define PROCBASED_CTLS_WINDOW_SETTING \ (PROCBASED_INT_WINDOW_EXITING | \ PROCBASED_NMI_WINDOW_EXITING) #define PROCBASED_CTLS_ONE_SETTING \ (PROCBASED_SECONDARY_CONTROLS | \ PROCBASED_MWAIT_EXITING | \ PROCBASED_MONITOR_EXITING | \ PROCBASED_IO_EXITING | \ PROCBASED_MSR_BITMAPS | \ PROCBASED_CTLS_WINDOW_SETTING | \ PROCBASED_CR8_LOAD_EXITING | \ PROCBASED_CR8_STORE_EXITING) #define PROCBASED_CTLS_ZERO_SETTING \ (PROCBASED_CR3_LOAD_EXITING | \ PROCBASED_CR3_STORE_EXITING | \ PROCBASED_IO_BITMAPS) #define PROCBASED_CTLS2_ONE_SETTING PROCBASED2_ENABLE_EPT #define PROCBASED_CTLS2_ZERO_SETTING 0 #define VM_EXIT_CTLS_ONE_SETTING \ (VM_EXIT_HOST_LMA | \ VM_EXIT_SAVE_EFER | \ VM_EXIT_LOAD_EFER | \ VM_EXIT_ACKNOWLEDGE_INTERRUPT) #define VM_EXIT_CTLS_ZERO_SETTING VM_EXIT_SAVE_DEBUG_CONTROLS #define VM_ENTRY_CTLS_ONE_SETTING (VM_ENTRY_LOAD_EFER) #define VM_ENTRY_CTLS_ZERO_SETTING \ (VM_ENTRY_LOAD_DEBUG_CONTROLS | \ VM_ENTRY_INTO_SMM | \ VM_ENTRY_DEACTIVATE_DUAL_MONITOR) #define HANDLED 1 #define UNHANDLED 0 static MALLOC_DEFINE(M_VMX, "vmx", "vmx"); static MALLOC_DEFINE(M_VLAPIC, "vlapic", "vlapic"); SYSCTL_DECL(_hw_vmm); SYSCTL_NODE(_hw_vmm, OID_AUTO, vmx, CTLFLAG_RW, NULL, NULL); int vmxon_enabled[MAXCPU]; static char vmxon_region[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE); static uint32_t pinbased_ctls, procbased_ctls, procbased_ctls2; static uint32_t exit_ctls, entry_ctls; static uint64_t cr0_ones_mask, cr0_zeros_mask; SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr0_ones_mask, CTLFLAG_RD, &cr0_ones_mask, 0, NULL); SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr0_zeros_mask, CTLFLAG_RD, &cr0_zeros_mask, 0, NULL); static uint64_t cr4_ones_mask, cr4_zeros_mask; SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr4_ones_mask, CTLFLAG_RD, &cr4_ones_mask, 0, NULL); SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr4_zeros_mask, CTLFLAG_RD, &cr4_zeros_mask, 0, NULL); static int vmx_initialized; SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, initialized, CTLFLAG_RD, &vmx_initialized, 0, "Intel VMX initialized"); /* * Optional capabilities */ static SYSCTL_NODE(_hw_vmm_vmx, OID_AUTO, cap, CTLFLAG_RW, NULL, NULL); static int cap_halt_exit; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, halt_exit, CTLFLAG_RD, &cap_halt_exit, 0, "HLT triggers a VM-exit"); static int cap_pause_exit; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, pause_exit, CTLFLAG_RD, &cap_pause_exit, 0, "PAUSE triggers a VM-exit"); static int cap_unrestricted_guest; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, unrestricted_guest, CTLFLAG_RD, &cap_unrestricted_guest, 0, "Unrestricted guests"); static int cap_monitor_trap; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, monitor_trap, CTLFLAG_RD, &cap_monitor_trap, 0, "Monitor trap flag"); static int cap_invpcid; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, invpcid, CTLFLAG_RD, &cap_invpcid, 0, "Guests are allowed to use INVPCID"); static int virtual_interrupt_delivery; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, virtual_interrupt_delivery, CTLFLAG_RD, &virtual_interrupt_delivery, 0, "APICv virtual interrupt delivery support"); static int posted_interrupts; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, posted_interrupts, CTLFLAG_RD, &posted_interrupts, 0, "APICv posted interrupt support"); static int pirvec = -1; SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, posted_interrupt_vector, CTLFLAG_RD, &pirvec, 0, "APICv posted interrupt vector"); static struct unrhdr *vpid_unr; static u_int vpid_alloc_failed; SYSCTL_UINT(_hw_vmm_vmx, OID_AUTO, vpid_alloc_failed, CTLFLAG_RD, &vpid_alloc_failed, 0, NULL); /* * Use the last page below 4GB as the APIC access address. This address is * occupied by the boot firmware so it is guaranteed that it will not conflict * with a page in system memory. */ #define APIC_ACCESS_ADDRESS 0xFFFFF000 static int vmx_getdesc(void *arg, int vcpu, int reg, struct seg_desc *desc); static int vmx_getreg(void *arg, int vcpu, int reg, uint64_t *retval); static int vmxctx_setreg(struct vmxctx *vmxctx, int reg, uint64_t val); static void vmx_inject_pir(struct vlapic *vlapic); #ifdef KTR static const char * exit_reason_to_str(int reason) { static char reasonbuf[32]; switch (reason) { case EXIT_REASON_EXCEPTION: return "exception"; case EXIT_REASON_EXT_INTR: return "extint"; case EXIT_REASON_TRIPLE_FAULT: return "triplefault"; case EXIT_REASON_INIT: return "init"; case EXIT_REASON_SIPI: return "sipi"; case EXIT_REASON_IO_SMI: return "iosmi"; case EXIT_REASON_SMI: return "smi"; case EXIT_REASON_INTR_WINDOW: return "intrwindow"; case EXIT_REASON_NMI_WINDOW: return "nmiwindow"; case EXIT_REASON_TASK_SWITCH: return "taskswitch"; case EXIT_REASON_CPUID: return "cpuid"; case EXIT_REASON_GETSEC: return "getsec"; case EXIT_REASON_HLT: return "hlt"; case EXIT_REASON_INVD: return "invd"; case EXIT_REASON_INVLPG: return "invlpg"; case EXIT_REASON_RDPMC: return "rdpmc"; case EXIT_REASON_RDTSC: return "rdtsc"; case EXIT_REASON_RSM: return "rsm"; case EXIT_REASON_VMCALL: return "vmcall"; case EXIT_REASON_VMCLEAR: return "vmclear"; case EXIT_REASON_VMLAUNCH: return "vmlaunch"; case EXIT_REASON_VMPTRLD: return "vmptrld"; case EXIT_REASON_VMPTRST: return "vmptrst"; case EXIT_REASON_VMREAD: return "vmread"; case EXIT_REASON_VMRESUME: return "vmresume"; case EXIT_REASON_VMWRITE: return "vmwrite"; case EXIT_REASON_VMXOFF: return "vmxoff"; case EXIT_REASON_VMXON: return "vmxon"; case EXIT_REASON_CR_ACCESS: return "craccess"; case EXIT_REASON_DR_ACCESS: return "draccess"; case EXIT_REASON_INOUT: return "inout"; case EXIT_REASON_RDMSR: return "rdmsr"; case EXIT_REASON_WRMSR: return "wrmsr"; case EXIT_REASON_INVAL_VMCS: return "invalvmcs"; case EXIT_REASON_INVAL_MSR: return "invalmsr"; case EXIT_REASON_MWAIT: return "mwait"; case EXIT_REASON_MTF: return "mtf"; case EXIT_REASON_MONITOR: return "monitor"; case EXIT_REASON_PAUSE: return "pause"; case EXIT_REASON_MCE_DURING_ENTRY: return "mce-during-entry"; case EXIT_REASON_TPR: return "tpr"; case EXIT_REASON_APIC_ACCESS: return "apic-access"; case EXIT_REASON_GDTR_IDTR: return "gdtridtr"; case EXIT_REASON_LDTR_TR: return "ldtrtr"; case EXIT_REASON_EPT_FAULT: return "eptfault"; case EXIT_REASON_EPT_MISCONFIG: return "eptmisconfig"; case EXIT_REASON_INVEPT: return "invept"; case EXIT_REASON_RDTSCP: return "rdtscp"; case EXIT_REASON_VMX_PREEMPT: return "vmxpreempt"; case EXIT_REASON_INVVPID: return "invvpid"; case EXIT_REASON_WBINVD: return "wbinvd"; case EXIT_REASON_XSETBV: return "xsetbv"; case EXIT_REASON_APIC_WRITE: return "apic-write"; default: snprintf(reasonbuf, sizeof(reasonbuf), "%d", reason); return (reasonbuf); } } #endif /* KTR */ static int vmx_allow_x2apic_msrs(struct vmx *vmx) { int i, error; error = 0; /* * Allow readonly access to the following x2APIC MSRs from the guest. */ error += guest_msr_ro(vmx, MSR_APIC_ID); error += guest_msr_ro(vmx, MSR_APIC_VERSION); error += guest_msr_ro(vmx, MSR_APIC_LDR); error += guest_msr_ro(vmx, MSR_APIC_SVR); for (i = 0; i < 8; i++) error += guest_msr_ro(vmx, MSR_APIC_ISR0 + i); for (i = 0; i < 8; i++) error += guest_msr_ro(vmx, MSR_APIC_TMR0 + i); for (i = 0; i < 8; i++) error += guest_msr_ro(vmx, MSR_APIC_IRR0 + i); error += guest_msr_ro(vmx, MSR_APIC_ESR); error += guest_msr_ro(vmx, MSR_APIC_LVT_TIMER); error += guest_msr_ro(vmx, MSR_APIC_LVT_THERMAL); error += guest_msr_ro(vmx, MSR_APIC_LVT_PCINT); error += guest_msr_ro(vmx, MSR_APIC_LVT_LINT0); error += guest_msr_ro(vmx, MSR_APIC_LVT_LINT1); error += guest_msr_ro(vmx, MSR_APIC_LVT_ERROR); error += guest_msr_ro(vmx, MSR_APIC_ICR_TIMER); error += guest_msr_ro(vmx, MSR_APIC_DCR_TIMER); error += guest_msr_ro(vmx, MSR_APIC_ICR); /* * Allow TPR, EOI and SELF_IPI MSRs to be read and written by the guest. * * These registers get special treatment described in the section * "Virtualizing MSR-Based APIC Accesses". */ error += guest_msr_rw(vmx, MSR_APIC_TPR); error += guest_msr_rw(vmx, MSR_APIC_EOI); error += guest_msr_rw(vmx, MSR_APIC_SELF_IPI); return (error); } u_long vmx_fix_cr0(u_long cr0) { return ((cr0 | cr0_ones_mask) & ~cr0_zeros_mask); } u_long vmx_fix_cr4(u_long cr4) { return ((cr4 | cr4_ones_mask) & ~cr4_zeros_mask); } static void vpid_free(int vpid) { if (vpid < 0 || vpid > 0xffff) panic("vpid_free: invalid vpid %d", vpid); /* * VPIDs [0,VM_MAXCPU] are special and are not allocated from * the unit number allocator. */ if (vpid > VM_MAXCPU) free_unr(vpid_unr, vpid); } static void vpid_alloc(uint16_t *vpid, int num) { int i, x; if (num <= 0 || num > VM_MAXCPU) panic("invalid number of vpids requested: %d", num); /* * If the "enable vpid" execution control is not enabled then the * VPID is required to be 0 for all vcpus. */ if ((procbased_ctls2 & PROCBASED2_ENABLE_VPID) == 0) { for (i = 0; i < num; i++) vpid[i] = 0; return; } /* * Allocate a unique VPID for each vcpu from the unit number allocator. */ for (i = 0; i < num; i++) { x = alloc_unr(vpid_unr); if (x == -1) break; else vpid[i] = x; } if (i < num) { atomic_add_int(&vpid_alloc_failed, 1); /* * If the unit number allocator does not have enough unique * VPIDs then we need to allocate from the [1,VM_MAXCPU] range. * * These VPIDs are not be unique across VMs but this does not * affect correctness because the combined mappings are also * tagged with the EP4TA which is unique for each VM. * * It is still sub-optimal because the invvpid will invalidate * combined mappings for a particular VPID across all EP4TAs. */ while (i-- > 0) vpid_free(vpid[i]); for (i = 0; i < num; i++) vpid[i] = i + 1; } } static void vpid_init(void) { /* * VPID 0 is required when the "enable VPID" execution control is * disabled. * * VPIDs [1,VM_MAXCPU] are used as the "overflow namespace" when the * unit number allocator does not have sufficient unique VPIDs to * satisfy the allocation. * * The remaining VPIDs are managed by the unit number allocator. */ vpid_unr = new_unrhdr(VM_MAXCPU + 1, 0xffff, NULL); } static void vmx_disable(void *arg __unused) { struct invvpid_desc invvpid_desc = { 0 }; struct invept_desc invept_desc = { 0 }; if (vmxon_enabled[curcpu]) { /* * See sections 25.3.3.3 and 25.3.3.4 in Intel Vol 3b. * * VMXON or VMXOFF are not required to invalidate any TLB * caching structures. This prevents potential retention of * cached information in the TLB between distinct VMX episodes. */ invvpid(INVVPID_TYPE_ALL_CONTEXTS, invvpid_desc); invept(INVEPT_TYPE_ALL_CONTEXTS, invept_desc); vmxoff(); } load_cr4(rcr4() & ~CR4_VMXE); } static int vmx_cleanup(void) { if (pirvec >= 0) lapic_ipi_free(pirvec); if (vpid_unr != NULL) { delete_unrhdr(vpid_unr); vpid_unr = NULL; } smp_rendezvous(NULL, vmx_disable, NULL, NULL); return (0); } static void vmx_enable(void *arg __unused) { int error; uint64_t feature_control; feature_control = rdmsr(MSR_IA32_FEATURE_CONTROL); if ((feature_control & IA32_FEATURE_CONTROL_LOCK) == 0 || (feature_control & IA32_FEATURE_CONTROL_VMX_EN) == 0) { wrmsr(MSR_IA32_FEATURE_CONTROL, feature_control | IA32_FEATURE_CONTROL_VMX_EN | IA32_FEATURE_CONTROL_LOCK); } load_cr4(rcr4() | CR4_VMXE); *(uint32_t *)vmxon_region[curcpu] = vmx_revision(); error = vmxon(vmxon_region[curcpu]); if (error == 0) vmxon_enabled[curcpu] = 1; } static void vmx_restore(void) { if (vmxon_enabled[curcpu]) vmxon(vmxon_region[curcpu]); } static int vmx_init(int ipinum) { int error, use_tpr_shadow; uint64_t basic, fixed0, fixed1, feature_control; uint32_t tmp, procbased2_vid_bits; /* CPUID.1:ECX[bit 5] must be 1 for processor to support VMX */ if (!(cpu_feature2 & CPUID2_VMX)) { printf("vmx_init: processor does not support VMX operation\n"); return (ENXIO); } /* * Verify that MSR_IA32_FEATURE_CONTROL lock and VMXON enable bits * are set (bits 0 and 2 respectively). */ feature_control = rdmsr(MSR_IA32_FEATURE_CONTROL); if ((feature_control & IA32_FEATURE_CONTROL_LOCK) == 1 && (feature_control & IA32_FEATURE_CONTROL_VMX_EN) == 0) { printf("vmx_init: VMX operation disabled by BIOS\n"); return (ENXIO); } /* * Verify capabilities MSR_VMX_BASIC: * - bit 54 indicates support for INS/OUTS decoding */ basic = rdmsr(MSR_VMX_BASIC); if ((basic & (1UL << 54)) == 0) { printf("vmx_init: processor does not support desired basic " "capabilities\n"); return (EINVAL); } /* Check support for primary processor-based VM-execution controls */ error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_CTLS_ONE_SETTING, PROCBASED_CTLS_ZERO_SETTING, &procbased_ctls); if (error) { printf("vmx_init: processor does not support desired primary " "processor-based controls\n"); return (error); } /* Clear the processor-based ctl bits that are set on demand */ procbased_ctls &= ~PROCBASED_CTLS_WINDOW_SETTING; /* Check support for secondary processor-based VM-execution controls */ error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED_CTLS2_ONE_SETTING, PROCBASED_CTLS2_ZERO_SETTING, &procbased_ctls2); if (error) { printf("vmx_init: processor does not support desired secondary " "processor-based controls\n"); return (error); } /* Check support for VPID */ error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED2_ENABLE_VPID, 0, &tmp); if (error == 0) procbased_ctls2 |= PROCBASED2_ENABLE_VPID; /* Check support for pin-based VM-execution controls */ error = vmx_set_ctlreg(MSR_VMX_PINBASED_CTLS, MSR_VMX_TRUE_PINBASED_CTLS, PINBASED_CTLS_ONE_SETTING, PINBASED_CTLS_ZERO_SETTING, &pinbased_ctls); if (error) { printf("vmx_init: processor does not support desired " "pin-based controls\n"); return (error); } /* Check support for VM-exit controls */ error = vmx_set_ctlreg(MSR_VMX_EXIT_CTLS, MSR_VMX_TRUE_EXIT_CTLS, VM_EXIT_CTLS_ONE_SETTING, VM_EXIT_CTLS_ZERO_SETTING, &exit_ctls); if (error) { printf("vmx_init: processor does not support desired " "exit controls\n"); return (error); } /* Check support for VM-entry controls */ error = vmx_set_ctlreg(MSR_VMX_ENTRY_CTLS, MSR_VMX_TRUE_ENTRY_CTLS, VM_ENTRY_CTLS_ONE_SETTING, VM_ENTRY_CTLS_ZERO_SETTING, &entry_ctls); if (error) { printf("vmx_init: processor does not support desired " "entry controls\n"); return (error); } /* * Check support for optional features by testing them * as individual bits */ cap_halt_exit = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_HLT_EXITING, 0, &tmp) == 0); cap_monitor_trap = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_PROCBASED_CTLS, PROCBASED_MTF, 0, &tmp) == 0); cap_pause_exit = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_PAUSE_EXITING, 0, &tmp) == 0); cap_unrestricted_guest = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED2_UNRESTRICTED_GUEST, 0, &tmp) == 0); cap_invpcid = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED2_ENABLE_INVPCID, 0, &tmp) == 0); /* * Check support for virtual interrupt delivery. */ procbased2_vid_bits = (PROCBASED2_VIRTUALIZE_APIC_ACCESSES | PROCBASED2_VIRTUALIZE_X2APIC_MODE | PROCBASED2_APIC_REGISTER_VIRTUALIZATION | PROCBASED2_VIRTUAL_INTERRUPT_DELIVERY); use_tpr_shadow = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_USE_TPR_SHADOW, 0, &tmp) == 0); error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, procbased2_vid_bits, 0, &tmp); if (error == 0 && use_tpr_shadow) { virtual_interrupt_delivery = 1; TUNABLE_INT_FETCH("hw.vmm.vmx.use_apic_vid", &virtual_interrupt_delivery); } if (virtual_interrupt_delivery) { procbased_ctls |= PROCBASED_USE_TPR_SHADOW; procbased_ctls2 |= procbased2_vid_bits; procbased_ctls2 &= ~PROCBASED2_VIRTUALIZE_X2APIC_MODE; /* * No need to emulate accesses to %CR8 if virtual * interrupt delivery is enabled. */ procbased_ctls &= ~PROCBASED_CR8_LOAD_EXITING; procbased_ctls &= ~PROCBASED_CR8_STORE_EXITING; /* * Check for Posted Interrupts only if Virtual Interrupt * Delivery is enabled. */ error = vmx_set_ctlreg(MSR_VMX_PINBASED_CTLS, MSR_VMX_TRUE_PINBASED_CTLS, PINBASED_POSTED_INTERRUPT, 0, &tmp); if (error == 0) { pirvec = lapic_ipi_alloc(&IDTVEC(justreturn)); if (pirvec < 0) { if (bootverbose) { printf("vmx_init: unable to allocate " "posted interrupt vector\n"); } } else { posted_interrupts = 1; TUNABLE_INT_FETCH("hw.vmm.vmx.use_apic_pir", &posted_interrupts); } } } if (posted_interrupts) pinbased_ctls |= PINBASED_POSTED_INTERRUPT; /* Initialize EPT */ error = ept_init(ipinum); if (error) { printf("vmx_init: ept initialization failed (%d)\n", error); return (error); } /* * Stash the cr0 and cr4 bits that must be fixed to 0 or 1 */ fixed0 = rdmsr(MSR_VMX_CR0_FIXED0); fixed1 = rdmsr(MSR_VMX_CR0_FIXED1); cr0_ones_mask = fixed0 & fixed1; cr0_zeros_mask = ~fixed0 & ~fixed1; /* * CR0_PE and CR0_PG can be set to zero in VMX non-root operation * if unrestricted guest execution is allowed. */ if (cap_unrestricted_guest) cr0_ones_mask &= ~(CR0_PG | CR0_PE); /* * Do not allow the guest to set CR0_NW or CR0_CD. */ cr0_zeros_mask |= (CR0_NW | CR0_CD); fixed0 = rdmsr(MSR_VMX_CR4_FIXED0); fixed1 = rdmsr(MSR_VMX_CR4_FIXED1); cr4_ones_mask = fixed0 & fixed1; cr4_zeros_mask = ~fixed0 & ~fixed1; vpid_init(); vmx_msr_init(); /* enable VMX operation */ smp_rendezvous(NULL, vmx_enable, NULL, NULL); vmx_initialized = 1; return (0); } static void vmx_trigger_hostintr(int vector) { uintptr_t func; struct gate_descriptor *gd; gd = &idt[vector]; KASSERT(vector >= 32 && vector <= 255, ("vmx_trigger_hostintr: " "invalid vector %d", vector)); KASSERT(gd->gd_p == 1, ("gate descriptor for vector %d not present", vector)); KASSERT(gd->gd_type == SDT_SYSIGT, ("gate descriptor for vector %d " "has invalid type %d", vector, gd->gd_type)); KASSERT(gd->gd_dpl == SEL_KPL, ("gate descriptor for vector %d " "has invalid dpl %d", vector, gd->gd_dpl)); KASSERT(gd->gd_selector == GSEL(GCODE_SEL, SEL_KPL), ("gate descriptor " "for vector %d has invalid selector %d", vector, gd->gd_selector)); KASSERT(gd->gd_ist == 0, ("gate descriptor for vector %d has invalid " "IST %d", vector, gd->gd_ist)); func = ((long)gd->gd_hioffset << 16 | gd->gd_looffset); vmx_call_isr(func); } static int vmx_setup_cr_shadow(int which, struct vmcs *vmcs, uint32_t initial) { int error, mask_ident, shadow_ident; uint64_t mask_value; if (which != 0 && which != 4) panic("vmx_setup_cr_shadow: unknown cr%d", which); if (which == 0) { mask_ident = VMCS_CR0_MASK; mask_value = cr0_ones_mask | cr0_zeros_mask; shadow_ident = VMCS_CR0_SHADOW; } else { mask_ident = VMCS_CR4_MASK; mask_value = cr4_ones_mask | cr4_zeros_mask; shadow_ident = VMCS_CR4_SHADOW; } error = vmcs_setreg(vmcs, 0, VMCS_IDENT(mask_ident), mask_value); if (error) return (error); error = vmcs_setreg(vmcs, 0, VMCS_IDENT(shadow_ident), initial); if (error) return (error); return (0); } #define vmx_setup_cr0_shadow(vmcs,init) vmx_setup_cr_shadow(0, (vmcs), (init)) #define vmx_setup_cr4_shadow(vmcs,init) vmx_setup_cr_shadow(4, (vmcs), (init)) static void * vmx_vminit(struct vm *vm, pmap_t pmap) { uint16_t vpid[VM_MAXCPU]; int i, error; struct vmx *vmx; struct vmcs *vmcs; uint32_t exc_bitmap; vmx = malloc(sizeof(struct vmx), M_VMX, M_WAITOK | M_ZERO); if ((uintptr_t)vmx & PAGE_MASK) { panic("malloc of struct vmx not aligned on %d byte boundary", PAGE_SIZE); } vmx->vm = vm; vmx->eptp = eptp(vtophys((vm_offset_t)pmap->pm_pml4)); /* * Clean up EPTP-tagged guest physical and combined mappings * * VMX transitions are not required to invalidate any guest physical * mappings. So, it may be possible for stale guest physical mappings * to be present in the processor TLBs. * * Combined mappings for this EP4TA are also invalidated for all VPIDs. */ ept_invalidate_mappings(vmx->eptp); msr_bitmap_initialize(vmx->msr_bitmap); /* * It is safe to allow direct access to MSR_GSBASE and MSR_FSBASE. * The guest FSBASE and GSBASE are saved and restored during * vm-exit and vm-entry respectively. The host FSBASE and GSBASE are * always restored from the vmcs host state area on vm-exit. * * The SYSENTER_CS/ESP/EIP MSRs are identical to FS/GSBASE in * how they are saved/restored so can be directly accessed by the * guest. * * MSR_EFER is saved and restored in the guest VMCS area on a * VM exit and entry respectively. It is also restored from the * host VMCS area on a VM exit. * * The TSC MSR is exposed read-only. Writes are disallowed as that * will impact the host TSC. * XXX Writes would be implemented with a wrmsr trap, and * then modifying the TSC offset in the VMCS. */ if (guest_msr_rw(vmx, MSR_GSBASE) || guest_msr_rw(vmx, MSR_FSBASE) || guest_msr_rw(vmx, MSR_SYSENTER_CS_MSR) || guest_msr_rw(vmx, MSR_SYSENTER_ESP_MSR) || guest_msr_rw(vmx, MSR_SYSENTER_EIP_MSR) || guest_msr_rw(vmx, MSR_EFER) || guest_msr_ro(vmx, MSR_TSC)) panic("vmx_vminit: error setting guest msr access"); vpid_alloc(vpid, VM_MAXCPU); if (virtual_interrupt_delivery) { error = vm_map_mmio(vm, DEFAULT_APIC_BASE, PAGE_SIZE, APIC_ACCESS_ADDRESS); /* XXX this should really return an error to the caller */ KASSERT(error == 0, ("vm_map_mmio(apicbase) error %d", error)); } for (i = 0; i < VM_MAXCPU; i++) { vmcs = &vmx->vmcs[i]; vmcs->identifier = vmx_revision(); error = vmclear(vmcs); if (error != 0) { panic("vmx_vminit: vmclear error %d on vcpu %d\n", error, i); } vmx_msr_guest_init(vmx, i); error = vmcs_init(vmcs); KASSERT(error == 0, ("vmcs_init error %d", error)); VMPTRLD(vmcs); error = 0; error += vmwrite(VMCS_HOST_RSP, (u_long)&vmx->ctx[i]); error += vmwrite(VMCS_EPTP, vmx->eptp); error += vmwrite(VMCS_PIN_BASED_CTLS, pinbased_ctls); error += vmwrite(VMCS_PRI_PROC_BASED_CTLS, procbased_ctls); error += vmwrite(VMCS_SEC_PROC_BASED_CTLS, procbased_ctls2); error += vmwrite(VMCS_EXIT_CTLS, exit_ctls); error += vmwrite(VMCS_ENTRY_CTLS, entry_ctls); error += vmwrite(VMCS_MSR_BITMAP, vtophys(vmx->msr_bitmap)); error += vmwrite(VMCS_VPID, vpid[i]); /* exception bitmap */ if (vcpu_trace_exceptions(vm, i)) exc_bitmap = 0xffffffff; else exc_bitmap = 1 << IDT_MC; error += vmwrite(VMCS_EXCEPTION_BITMAP, exc_bitmap); if (virtual_interrupt_delivery) { error += vmwrite(VMCS_APIC_ACCESS, APIC_ACCESS_ADDRESS); error += vmwrite(VMCS_VIRTUAL_APIC, vtophys(&vmx->apic_page[i])); error += vmwrite(VMCS_EOI_EXIT0, 0); error += vmwrite(VMCS_EOI_EXIT1, 0); error += vmwrite(VMCS_EOI_EXIT2, 0); error += vmwrite(VMCS_EOI_EXIT3, 0); } if (posted_interrupts) { error += vmwrite(VMCS_PIR_VECTOR, pirvec); error += vmwrite(VMCS_PIR_DESC, vtophys(&vmx->pir_desc[i])); } VMCLEAR(vmcs); KASSERT(error == 0, ("vmx_vminit: error customizing the vmcs")); vmx->cap[i].set = 0; vmx->cap[i].proc_ctls = procbased_ctls; vmx->cap[i].proc_ctls2 = procbased_ctls2; vmx->state[i].nextrip = ~0; vmx->state[i].lastcpu = NOCPU; vmx->state[i].vpid = vpid[i]; /* * Set up the CR0/4 shadows, and init the read shadow * to the power-on register value from the Intel Sys Arch. * CR0 - 0x60000010 * CR4 - 0 */ error = vmx_setup_cr0_shadow(vmcs, 0x60000010); if (error != 0) panic("vmx_setup_cr0_shadow %d", error); error = vmx_setup_cr4_shadow(vmcs, 0); if (error != 0) panic("vmx_setup_cr4_shadow %d", error); vmx->ctx[i].pmap = pmap; } return (vmx); } static int vmx_handle_cpuid(struct vm *vm, int vcpu, struct vmxctx *vmxctx) { int handled, func; func = vmxctx->guest_rax; handled = x86_emulate_cpuid(vm, vcpu, (uint32_t*)(&vmxctx->guest_rax), (uint32_t*)(&vmxctx->guest_rbx), (uint32_t*)(&vmxctx->guest_rcx), (uint32_t*)(&vmxctx->guest_rdx)); return (handled); } static __inline void vmx_run_trace(struct vmx *vmx, int vcpu) { #ifdef KTR VCPU_CTR1(vmx->vm, vcpu, "Resume execution at %#lx", vmcs_guest_rip()); #endif } static __inline void vmx_exit_trace(struct vmx *vmx, int vcpu, uint64_t rip, uint32_t exit_reason, int handled) { #ifdef KTR VCPU_CTR3(vmx->vm, vcpu, "%s %s vmexit at 0x%0lx", handled ? "handled" : "unhandled", exit_reason_to_str(exit_reason), rip); #endif } static __inline void vmx_astpending_trace(struct vmx *vmx, int vcpu, uint64_t rip) { #ifdef KTR VCPU_CTR1(vmx->vm, vcpu, "astpending vmexit at 0x%0lx", rip); #endif } static VMM_STAT_INTEL(VCPU_INVVPID_SAVED, "Number of vpid invalidations saved"); static VMM_STAT_INTEL(VCPU_INVVPID_DONE, "Number of vpid invalidations done"); /* * Invalidate guest mappings identified by its vpid from the TLB. */ static __inline void vmx_invvpid(struct vmx *vmx, int vcpu, pmap_t pmap, int running) { struct vmxstate *vmxstate; struct invvpid_desc invvpid_desc; vmxstate = &vmx->state[vcpu]; if (vmxstate->vpid == 0) return; if (!running) { /* * Set the 'lastcpu' to an invalid host cpu. * * This will invalidate TLB entries tagged with the vcpu's * vpid the next time it runs via vmx_set_pcpu_defaults(). */ vmxstate->lastcpu = NOCPU; return; } KASSERT(curthread->td_critnest > 0, ("%s: vcpu %d running outside " "critical section", __func__, vcpu)); /* * Invalidate all mappings tagged with 'vpid' * * We do this because this vcpu was executing on a different host * cpu when it last ran. We do not track whether it invalidated * mappings associated with its 'vpid' during that run. So we must * assume that the mappings associated with 'vpid' on 'curcpu' are * stale and invalidate them. * * Note that we incur this penalty only when the scheduler chooses to * move the thread associated with this vcpu between host cpus. * * Note also that this will invalidate mappings tagged with 'vpid' * for "all" EP4TAs. */ if (pmap->pm_eptgen == vmx->eptgen[curcpu]) { invvpid_desc._res1 = 0; invvpid_desc._res2 = 0; invvpid_desc.vpid = vmxstate->vpid; invvpid_desc.linear_addr = 0; invvpid(INVVPID_TYPE_SINGLE_CONTEXT, invvpid_desc); vmm_stat_incr(vmx->vm, vcpu, VCPU_INVVPID_DONE, 1); } else { /* * The invvpid can be skipped if an invept is going to * be performed before entering the guest. The invept * will invalidate combined mappings tagged with * 'vmx->eptp' for all vpids. */ vmm_stat_incr(vmx->vm, vcpu, VCPU_INVVPID_SAVED, 1); } } static void vmx_set_pcpu_defaults(struct vmx *vmx, int vcpu, pmap_t pmap) { struct vmxstate *vmxstate; vmxstate = &vmx->state[vcpu]; if (vmxstate->lastcpu == curcpu) return; vmxstate->lastcpu = curcpu; vmm_stat_incr(vmx->vm, vcpu, VCPU_MIGRATIONS, 1); vmcs_write(VMCS_HOST_TR_BASE, vmm_get_host_trbase()); vmcs_write(VMCS_HOST_GDTR_BASE, vmm_get_host_gdtrbase()); vmcs_write(VMCS_HOST_GS_BASE, vmm_get_host_gsbase()); vmx_invvpid(vmx, vcpu, pmap, 1); } /* * We depend on 'procbased_ctls' to have the Interrupt Window Exiting bit set. */ CTASSERT((PROCBASED_CTLS_ONE_SETTING & PROCBASED_INT_WINDOW_EXITING) != 0); static void __inline vmx_set_int_window_exiting(struct vmx *vmx, int vcpu) { if ((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) == 0) { vmx->cap[vcpu].proc_ctls |= PROCBASED_INT_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Enabling interrupt window exiting"); } } static void __inline vmx_clear_int_window_exiting(struct vmx *vmx, int vcpu) { KASSERT((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) != 0, ("intr_window_exiting not set: %#x", vmx->cap[vcpu].proc_ctls)); vmx->cap[vcpu].proc_ctls &= ~PROCBASED_INT_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Disabling interrupt window exiting"); } static void __inline vmx_set_nmi_window_exiting(struct vmx *vmx, int vcpu) { if ((vmx->cap[vcpu].proc_ctls & PROCBASED_NMI_WINDOW_EXITING) == 0) { vmx->cap[vcpu].proc_ctls |= PROCBASED_NMI_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Enabling NMI window exiting"); } } static void __inline vmx_clear_nmi_window_exiting(struct vmx *vmx, int vcpu) { KASSERT((vmx->cap[vcpu].proc_ctls & PROCBASED_NMI_WINDOW_EXITING) != 0, ("nmi_window_exiting not set %#x", vmx->cap[vcpu].proc_ctls)); vmx->cap[vcpu].proc_ctls &= ~PROCBASED_NMI_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Disabling NMI window exiting"); } #define NMI_BLOCKING (VMCS_INTERRUPTIBILITY_NMI_BLOCKING | \ VMCS_INTERRUPTIBILITY_MOVSS_BLOCKING) #define HWINTR_BLOCKING (VMCS_INTERRUPTIBILITY_STI_BLOCKING | \ VMCS_INTERRUPTIBILITY_MOVSS_BLOCKING) static void vmx_inject_nmi(struct vmx *vmx, int vcpu) { uint32_t gi, info; gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); KASSERT((gi & NMI_BLOCKING) == 0, ("vmx_inject_nmi: invalid guest " "interruptibility-state %#x", gi)); info = vmcs_read(VMCS_ENTRY_INTR_INFO); KASSERT((info & VMCS_INTR_VALID) == 0, ("vmx_inject_nmi: invalid " "VM-entry interruption information %#x", info)); /* * Inject the virtual NMI. The vector must be the NMI IDT entry * or the VMCS entry check will fail. */ info = IDT_NMI | VMCS_INTR_T_NMI | VMCS_INTR_VALID; vmcs_write(VMCS_ENTRY_INTR_INFO, info); VCPU_CTR0(vmx->vm, vcpu, "Injecting vNMI"); /* Clear the request */ vm_nmi_clear(vmx->vm, vcpu); } static void vmx_inject_interrupts(struct vmx *vmx, int vcpu, struct vlapic *vlapic, uint64_t guestrip) { int vector, need_nmi_exiting, extint_pending; uint64_t rflags, entryinfo; uint32_t gi, info; if (vmx->state[vcpu].nextrip != guestrip) { gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); if (gi & HWINTR_BLOCKING) { VCPU_CTR2(vmx->vm, vcpu, "Guest interrupt blocking " "cleared due to rip change: %#lx/%#lx", vmx->state[vcpu].nextrip, guestrip); gi &= ~HWINTR_BLOCKING; vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi); } } if (vm_entry_intinfo(vmx->vm, vcpu, &entryinfo)) { KASSERT((entryinfo & VMCS_INTR_VALID) != 0, ("%s: entry " "intinfo is not valid: %#lx", __func__, entryinfo)); info = vmcs_read(VMCS_ENTRY_INTR_INFO); KASSERT((info & VMCS_INTR_VALID) == 0, ("%s: cannot inject " "pending exception: %#lx/%#x", __func__, entryinfo, info)); info = entryinfo; vector = info & 0xff; if (vector == IDT_BP || vector == IDT_OF) { /* * VT-x requires #BP and #OF to be injected as software * exceptions. */ info &= ~VMCS_INTR_T_MASK; info |= VMCS_INTR_T_SWEXCEPTION; } if (info & VMCS_INTR_DEL_ERRCODE) vmcs_write(VMCS_ENTRY_EXCEPTION_ERROR, entryinfo >> 32); vmcs_write(VMCS_ENTRY_INTR_INFO, info); } if (vm_nmi_pending(vmx->vm, vcpu)) { /* * If there are no conditions blocking NMI injection then * inject it directly here otherwise enable "NMI window * exiting" to inject it as soon as we can. * * We also check for STI_BLOCKING because some implementations * don't allow NMI injection in this case. If we are running * on a processor that doesn't have this restriction it will * immediately exit and the NMI will be injected in the * "NMI window exiting" handler. */ need_nmi_exiting = 1; gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); if ((gi & (HWINTR_BLOCKING | NMI_BLOCKING)) == 0) { info = vmcs_read(VMCS_ENTRY_INTR_INFO); if ((info & VMCS_INTR_VALID) == 0) { vmx_inject_nmi(vmx, vcpu); need_nmi_exiting = 0; } else { VCPU_CTR1(vmx->vm, vcpu, "Cannot inject NMI " "due to VM-entry intr info %#x", info); } } else { VCPU_CTR1(vmx->vm, vcpu, "Cannot inject NMI due to " "Guest Interruptibility-state %#x", gi); } if (need_nmi_exiting) vmx_set_nmi_window_exiting(vmx, vcpu); } extint_pending = vm_extint_pending(vmx->vm, vcpu); if (!extint_pending && virtual_interrupt_delivery) { vmx_inject_pir(vlapic); return; } /* * If interrupt-window exiting is already in effect then don't bother * checking for pending interrupts. This is just an optimization and * not needed for correctness. */ if ((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) != 0) { VCPU_CTR0(vmx->vm, vcpu, "Skip interrupt injection due to " "pending int_window_exiting"); return; } if (!extint_pending) { /* Ask the local apic for a vector to inject */ if (!vlapic_pending_intr(vlapic, &vector)) return; /* * From the Intel SDM, Volume 3, Section "Maskable * Hardware Interrupts": * - maskable interrupt vectors [16,255] can be delivered * through the local APIC. */ KASSERT(vector >= 16 && vector <= 255, ("invalid vector %d from local APIC", vector)); } else { /* Ask the legacy pic for a vector to inject */ vatpic_pending_intr(vmx->vm, &vector); /* * From the Intel SDM, Volume 3, Section "Maskable * Hardware Interrupts": * - maskable interrupt vectors [0,255] can be delivered * through the INTR pin. */ KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR", vector)); } /* Check RFLAGS.IF and the interruptibility state of the guest */ rflags = vmcs_read(VMCS_GUEST_RFLAGS); if ((rflags & PSL_I) == 0) { VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to " "rflags %#lx", vector, rflags); goto cantinject; } gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); if (gi & HWINTR_BLOCKING) { VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to " "Guest Interruptibility-state %#x", vector, gi); goto cantinject; } info = vmcs_read(VMCS_ENTRY_INTR_INFO); if (info & VMCS_INTR_VALID) { /* * This is expected and could happen for multiple reasons: * - A vectoring VM-entry was aborted due to astpending * - A VM-exit happened during event injection. * - An exception was injected above. * - An NMI was injected above or after "NMI window exiting" */ VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to " "VM-entry intr info %#x", vector, info); goto cantinject; } /* Inject the interrupt */ info = VMCS_INTR_T_HWINTR | VMCS_INTR_VALID; info |= vector; vmcs_write(VMCS_ENTRY_INTR_INFO, info); if (!extint_pending) { /* Update the Local APIC ISR */ vlapic_intr_accepted(vlapic, vector); } else { vm_extint_clear(vmx->vm, vcpu); vatpic_intr_accepted(vmx->vm, vector); /* * After we accepted the current ExtINT the PIC may * have posted another one. If that is the case, set * the Interrupt Window Exiting execution control so * we can inject that one too. * * Also, interrupt window exiting allows us to inject any * pending APIC vector that was preempted by the ExtINT * as soon as possible. This applies both for the software * emulated vlapic and the hardware assisted virtual APIC. */ vmx_set_int_window_exiting(vmx, vcpu); } VCPU_CTR1(vmx->vm, vcpu, "Injecting hwintr at vector %d", vector); return; cantinject: /* * Set the Interrupt Window Exiting execution control so we can inject * the interrupt as soon as blocking condition goes away. */ vmx_set_int_window_exiting(vmx, vcpu); } /* * If the Virtual NMIs execution control is '1' then the logical processor * tracks virtual-NMI blocking in the Guest Interruptibility-state field of * the VMCS. An IRET instruction in VMX non-root operation will remove any * virtual-NMI blocking. * * This unblocking occurs even if the IRET causes a fault. In this case the * hypervisor needs to restore virtual-NMI blocking before resuming the guest. */ static void vmx_restore_nmi_blocking(struct vmx *vmx, int vcpuid) { uint32_t gi; VCPU_CTR0(vmx->vm, vcpuid, "Restore Virtual-NMI blocking"); gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); gi |= VMCS_INTERRUPTIBILITY_NMI_BLOCKING; vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi); } static void vmx_clear_nmi_blocking(struct vmx *vmx, int vcpuid) { uint32_t gi; VCPU_CTR0(vmx->vm, vcpuid, "Clear Virtual-NMI blocking"); gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); gi &= ~VMCS_INTERRUPTIBILITY_NMI_BLOCKING; vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi); } static void vmx_assert_nmi_blocking(struct vmx *vmx, int vcpuid) { uint32_t gi; gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); KASSERT(gi & VMCS_INTERRUPTIBILITY_NMI_BLOCKING, ("NMI blocking is not in effect %#x", gi)); } static int vmx_emulate_xsetbv(struct vmx *vmx, int vcpu, struct vm_exit *vmexit) { struct vmxctx *vmxctx; uint64_t xcrval; const struct xsave_limits *limits; vmxctx = &vmx->ctx[vcpu]; limits = vmm_get_xsave_limits(); /* * Note that the processor raises a GP# fault on its own if * xsetbv is executed for CPL != 0, so we do not have to * emulate that fault here. */ /* Only xcr0 is supported. */ if (vmxctx->guest_rcx != 0) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* We only handle xcr0 if both the host and guest have XSAVE enabled. */ if (!limits->xsave_enabled || !(vmcs_read(VMCS_GUEST_CR4) & CR4_XSAVE)) { vm_inject_ud(vmx->vm, vcpu); return (HANDLED); } xcrval = vmxctx->guest_rdx << 32 | (vmxctx->guest_rax & 0xffffffff); if ((xcrval & ~limits->xcr0_allowed) != 0) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } if (!(xcrval & XFEATURE_ENABLED_X87)) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* AVX (YMM_Hi128) requires SSE. */ if (xcrval & XFEATURE_ENABLED_AVX && (xcrval & XFEATURE_AVX) != XFEATURE_AVX) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* * AVX512 requires base AVX (YMM_Hi128) as well as OpMask, * ZMM_Hi256, and Hi16_ZMM. */ if (xcrval & XFEATURE_AVX512 && (xcrval & (XFEATURE_AVX512 | XFEATURE_AVX)) != (XFEATURE_AVX512 | XFEATURE_AVX)) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* * Intel MPX requires both bound register state flags to be * set. */ if (((xcrval & XFEATURE_ENABLED_BNDREGS) != 0) != ((xcrval & XFEATURE_ENABLED_BNDCSR) != 0)) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* * This runs "inside" vmrun() with the guest's FPU state, so * modifying xcr0 directly modifies the guest's xcr0, not the * host's. */ load_xcr(0, xcrval); return (HANDLED); } static uint64_t vmx_get_guest_reg(struct vmx *vmx, int vcpu, int ident) { const struct vmxctx *vmxctx; vmxctx = &vmx->ctx[vcpu]; switch (ident) { case 0: return (vmxctx->guest_rax); case 1: return (vmxctx->guest_rcx); case 2: return (vmxctx->guest_rdx); case 3: return (vmxctx->guest_rbx); case 4: return (vmcs_read(VMCS_GUEST_RSP)); case 5: return (vmxctx->guest_rbp); case 6: return (vmxctx->guest_rsi); case 7: return (vmxctx->guest_rdi); case 8: return (vmxctx->guest_r8); case 9: return (vmxctx->guest_r9); case 10: return (vmxctx->guest_r10); case 11: return (vmxctx->guest_r11); case 12: return (vmxctx->guest_r12); case 13: return (vmxctx->guest_r13); case 14: return (vmxctx->guest_r14); case 15: return (vmxctx->guest_r15); default: panic("invalid vmx register %d", ident); } } static void vmx_set_guest_reg(struct vmx *vmx, int vcpu, int ident, uint64_t regval) { struct vmxctx *vmxctx; vmxctx = &vmx->ctx[vcpu]; switch (ident) { case 0: vmxctx->guest_rax = regval; break; case 1: vmxctx->guest_rcx = regval; break; case 2: vmxctx->guest_rdx = regval; break; case 3: vmxctx->guest_rbx = regval; break; case 4: vmcs_write(VMCS_GUEST_RSP, regval); break; case 5: vmxctx->guest_rbp = regval; break; case 6: vmxctx->guest_rsi = regval; break; case 7: vmxctx->guest_rdi = regval; break; case 8: vmxctx->guest_r8 = regval; break; case 9: vmxctx->guest_r9 = regval; break; case 10: vmxctx->guest_r10 = regval; break; case 11: vmxctx->guest_r11 = regval; break; case 12: vmxctx->guest_r12 = regval; break; case 13: vmxctx->guest_r13 = regval; break; case 14: vmxctx->guest_r14 = regval; break; case 15: vmxctx->guest_r15 = regval; break; default: panic("invalid vmx register %d", ident); } } static int vmx_emulate_cr0_access(struct vmx *vmx, int vcpu, uint64_t exitqual) { uint64_t crval, regval; /* We only handle mov to %cr0 at this time */ if ((exitqual & 0xf0) != 0x00) return (UNHANDLED); regval = vmx_get_guest_reg(vmx, vcpu, (exitqual >> 8) & 0xf); vmcs_write(VMCS_CR0_SHADOW, regval); crval = regval | cr0_ones_mask; crval &= ~cr0_zeros_mask; vmcs_write(VMCS_GUEST_CR0, crval); if (regval & CR0_PG) { uint64_t efer, entry_ctls; /* * If CR0.PG is 1 and EFER.LME is 1 then EFER.LMA and * the "IA-32e mode guest" bit in VM-entry control must be * equal. */ efer = vmcs_read(VMCS_GUEST_IA32_EFER); if (efer & EFER_LME) { efer |= EFER_LMA; vmcs_write(VMCS_GUEST_IA32_EFER, efer); entry_ctls = vmcs_read(VMCS_ENTRY_CTLS); entry_ctls |= VM_ENTRY_GUEST_LMA; vmcs_write(VMCS_ENTRY_CTLS, entry_ctls); } } return (HANDLED); } static int vmx_emulate_cr4_access(struct vmx *vmx, int vcpu, uint64_t exitqual) { uint64_t crval, regval; /* We only handle mov to %cr4 at this time */ if ((exitqual & 0xf0) != 0x00) return (UNHANDLED); regval = vmx_get_guest_reg(vmx, vcpu, (exitqual >> 8) & 0xf); vmcs_write(VMCS_CR4_SHADOW, regval); crval = regval | cr4_ones_mask; crval &= ~cr4_zeros_mask; vmcs_write(VMCS_GUEST_CR4, crval); return (HANDLED); } static int vmx_emulate_cr8_access(struct vmx *vmx, int vcpu, uint64_t exitqual) { struct vlapic *vlapic; uint64_t cr8; int regnum; /* We only handle mov %cr8 to/from a register at this time. */ if ((exitqual & 0xe0) != 0x00) { return (UNHANDLED); } vlapic = vm_lapic(vmx->vm, vcpu); regnum = (exitqual >> 8) & 0xf; if (exitqual & 0x10) { cr8 = vlapic_get_cr8(vlapic); vmx_set_guest_reg(vmx, vcpu, regnum, cr8); } else { cr8 = vmx_get_guest_reg(vmx, vcpu, regnum); vlapic_set_cr8(vlapic, cr8); } return (HANDLED); } /* * From section "Guest Register State" in the Intel SDM: CPL = SS.DPL */ static int vmx_cpl(void) { uint32_t ssar; ssar = vmcs_read(VMCS_GUEST_SS_ACCESS_RIGHTS); return ((ssar >> 5) & 0x3); } static enum vm_cpu_mode vmx_cpu_mode(void) { uint32_t csar; if (vmcs_read(VMCS_GUEST_IA32_EFER) & EFER_LMA) { csar = vmcs_read(VMCS_GUEST_CS_ACCESS_RIGHTS); if (csar & 0x2000) return (CPU_MODE_64BIT); /* CS.L = 1 */ else return (CPU_MODE_COMPATIBILITY); } else if (vmcs_read(VMCS_GUEST_CR0) & CR0_PE) { return (CPU_MODE_PROTECTED); } else { return (CPU_MODE_REAL); } } static enum vm_paging_mode vmx_paging_mode(void) { if (!(vmcs_read(VMCS_GUEST_CR0) & CR0_PG)) return (PAGING_MODE_FLAT); if (!(vmcs_read(VMCS_GUEST_CR4) & CR4_PAE)) return (PAGING_MODE_32); if (vmcs_read(VMCS_GUEST_IA32_EFER) & EFER_LME) return (PAGING_MODE_64); else return (PAGING_MODE_PAE); } static uint64_t inout_str_index(struct vmx *vmx, int vcpuid, int in) { uint64_t val; int error; enum vm_reg_name reg; reg = in ? VM_REG_GUEST_RDI : VM_REG_GUEST_RSI; error = vmx_getreg(vmx, vcpuid, reg, &val); KASSERT(error == 0, ("%s: vmx_getreg error %d", __func__, error)); return (val); } static uint64_t inout_str_count(struct vmx *vmx, int vcpuid, int rep) { uint64_t val; int error; if (rep) { error = vmx_getreg(vmx, vcpuid, VM_REG_GUEST_RCX, &val); KASSERT(!error, ("%s: vmx_getreg error %d", __func__, error)); } else { val = 1; } return (val); } static int inout_str_addrsize(uint32_t inst_info) { uint32_t size; size = (inst_info >> 7) & 0x7; switch (size) { case 0: return (2); /* 16 bit */ case 1: return (4); /* 32 bit */ case 2: return (8); /* 64 bit */ default: panic("%s: invalid size encoding %d", __func__, size); } } static void inout_str_seginfo(struct vmx *vmx, int vcpuid, uint32_t inst_info, int in, struct vm_inout_str *vis) { int error, s; if (in) { vis->seg_name = VM_REG_GUEST_ES; } else { s = (inst_info >> 15) & 0x7; vis->seg_name = vm_segment_name(s); } error = vmx_getdesc(vmx, vcpuid, vis->seg_name, &vis->seg_desc); KASSERT(error == 0, ("%s: vmx_getdesc error %d", __func__, error)); } static void vmx_paging_info(struct vm_guest_paging *paging) { paging->cr3 = vmcs_guest_cr3(); paging->cpl = vmx_cpl(); paging->cpu_mode = vmx_cpu_mode(); paging->paging_mode = vmx_paging_mode(); } static void vmexit_inst_emul(struct vm_exit *vmexit, uint64_t gpa, uint64_t gla) { struct vm_guest_paging *paging; uint32_t csar; paging = &vmexit->u.inst_emul.paging; vmexit->exitcode = VM_EXITCODE_INST_EMUL; vmexit->inst_length = 0; vmexit->u.inst_emul.gpa = gpa; vmexit->u.inst_emul.gla = gla; vmx_paging_info(paging); switch (paging->cpu_mode) { case CPU_MODE_REAL: vmexit->u.inst_emul.cs_base = vmcs_read(VMCS_GUEST_CS_BASE); vmexit->u.inst_emul.cs_d = 0; break; case CPU_MODE_PROTECTED: case CPU_MODE_COMPATIBILITY: vmexit->u.inst_emul.cs_base = vmcs_read(VMCS_GUEST_CS_BASE); csar = vmcs_read(VMCS_GUEST_CS_ACCESS_RIGHTS); vmexit->u.inst_emul.cs_d = SEG_DESC_DEF32(csar); break; default: vmexit->u.inst_emul.cs_base = 0; vmexit->u.inst_emul.cs_d = 0; break; } vie_init(&vmexit->u.inst_emul.vie, NULL, 0); } static int ept_fault_type(uint64_t ept_qual) { int fault_type; if (ept_qual & EPT_VIOLATION_DATA_WRITE) fault_type = VM_PROT_WRITE; else if (ept_qual & EPT_VIOLATION_INST_FETCH) fault_type = VM_PROT_EXECUTE; else fault_type= VM_PROT_READ; return (fault_type); } static boolean_t ept_emulation_fault(uint64_t ept_qual) { int read, write; /* EPT fault on an instruction fetch doesn't make sense here */ if (ept_qual & EPT_VIOLATION_INST_FETCH) return (FALSE); /* EPT fault must be a read fault or a write fault */ read = ept_qual & EPT_VIOLATION_DATA_READ ? 1 : 0; write = ept_qual & EPT_VIOLATION_DATA_WRITE ? 1 : 0; if ((read | write) == 0) return (FALSE); /* * The EPT violation must have been caused by accessing a * guest-physical address that is a translation of a guest-linear * address. */ if ((ept_qual & EPT_VIOLATION_GLA_VALID) == 0 || (ept_qual & EPT_VIOLATION_XLAT_VALID) == 0) { return (FALSE); } return (TRUE); } static __inline int apic_access_virtualization(struct vmx *vmx, int vcpuid) { uint32_t proc_ctls2; proc_ctls2 = vmx->cap[vcpuid].proc_ctls2; return ((proc_ctls2 & PROCBASED2_VIRTUALIZE_APIC_ACCESSES) ? 1 : 0); } static __inline int x2apic_virtualization(struct vmx *vmx, int vcpuid) { uint32_t proc_ctls2; proc_ctls2 = vmx->cap[vcpuid].proc_ctls2; return ((proc_ctls2 & PROCBASED2_VIRTUALIZE_X2APIC_MODE) ? 1 : 0); } static int vmx_handle_apic_write(struct vmx *vmx, int vcpuid, struct vlapic *vlapic, uint64_t qual) { int error, handled, offset; uint32_t *apic_regs, vector; bool retu; handled = HANDLED; offset = APIC_WRITE_OFFSET(qual); if (!apic_access_virtualization(vmx, vcpuid)) { /* * In general there should not be any APIC write VM-exits * unless APIC-access virtualization is enabled. * * However self-IPI virtualization can legitimately trigger * an APIC-write VM-exit so treat it specially. */ if (x2apic_virtualization(vmx, vcpuid) && offset == APIC_OFFSET_SELF_IPI) { apic_regs = (uint32_t *)(vlapic->apic_page); vector = apic_regs[APIC_OFFSET_SELF_IPI / 4]; vlapic_self_ipi_handler(vlapic, vector); return (HANDLED); } else return (UNHANDLED); } switch (offset) { case APIC_OFFSET_ID: vlapic_id_write_handler(vlapic); break; case APIC_OFFSET_LDR: vlapic_ldr_write_handler(vlapic); break; case APIC_OFFSET_DFR: vlapic_dfr_write_handler(vlapic); break; case APIC_OFFSET_SVR: vlapic_svr_write_handler(vlapic); break; case APIC_OFFSET_ESR: vlapic_esr_write_handler(vlapic); break; case APIC_OFFSET_ICR_LOW: retu = false; error = vlapic_icrlo_write_handler(vlapic, &retu); if (error != 0 || retu) handled = UNHANDLED; break; case APIC_OFFSET_CMCI_LVT: case APIC_OFFSET_TIMER_LVT ... APIC_OFFSET_ERROR_LVT: vlapic_lvt_write_handler(vlapic, offset); break; case APIC_OFFSET_TIMER_ICR: vlapic_icrtmr_write_handler(vlapic); break; case APIC_OFFSET_TIMER_DCR: vlapic_dcr_write_handler(vlapic); break; default: handled = UNHANDLED; break; } return (handled); } static bool apic_access_fault(struct vmx *vmx, int vcpuid, uint64_t gpa) { if (apic_access_virtualization(vmx, vcpuid) && (gpa >= DEFAULT_APIC_BASE && gpa < DEFAULT_APIC_BASE + PAGE_SIZE)) return (true); else return (false); } static int vmx_handle_apic_access(struct vmx *vmx, int vcpuid, struct vm_exit *vmexit) { uint64_t qual; int access_type, offset, allowed; if (!apic_access_virtualization(vmx, vcpuid)) return (UNHANDLED); qual = vmexit->u.vmx.exit_qualification; access_type = APIC_ACCESS_TYPE(qual); offset = APIC_ACCESS_OFFSET(qual); allowed = 0; if (access_type == 0) { /* * Read data access to the following registers is expected. */ switch (offset) { case APIC_OFFSET_APR: case APIC_OFFSET_PPR: case APIC_OFFSET_RRR: case APIC_OFFSET_CMCI_LVT: case APIC_OFFSET_TIMER_CCR: allowed = 1; break; default: break; } } else if (access_type == 1) { /* * Write data access to the following registers is expected. */ switch (offset) { case APIC_OFFSET_VER: case APIC_OFFSET_APR: case APIC_OFFSET_PPR: case APIC_OFFSET_RRR: case APIC_OFFSET_ISR0 ... APIC_OFFSET_ISR7: case APIC_OFFSET_TMR0 ... APIC_OFFSET_TMR7: case APIC_OFFSET_IRR0 ... APIC_OFFSET_IRR7: case APIC_OFFSET_CMCI_LVT: case APIC_OFFSET_TIMER_CCR: allowed = 1; break; default: break; } } if (allowed) { vmexit_inst_emul(vmexit, DEFAULT_APIC_BASE + offset, VIE_INVALID_GLA); } /* * Regardless of whether the APIC-access is allowed this handler * always returns UNHANDLED: * - if the access is allowed then it is handled by emulating the * instruction that caused the VM-exit (outside the critical section) * - if the access is not allowed then it will be converted to an * exitcode of VM_EXITCODE_VMX and will be dealt with in userland. */ return (UNHANDLED); } static enum task_switch_reason vmx_task_switch_reason(uint64_t qual) { int reason; reason = (qual >> 30) & 0x3; switch (reason) { case 0: return (TSR_CALL); case 1: return (TSR_IRET); case 2: return (TSR_JMP); case 3: return (TSR_IDT_GATE); default: panic("%s: invalid reason %d", __func__, reason); } } static int emulate_wrmsr(struct vmx *vmx, int vcpuid, u_int num, uint64_t val, bool *retu) { int error; if (lapic_msr(num)) error = lapic_wrmsr(vmx->vm, vcpuid, num, val, retu); else error = vmx_wrmsr(vmx, vcpuid, num, val, retu); return (error); } static int emulate_rdmsr(struct vmx *vmx, int vcpuid, u_int num, bool *retu) { struct vmxctx *vmxctx; uint64_t result; uint32_t eax, edx; int error; if (lapic_msr(num)) error = lapic_rdmsr(vmx->vm, vcpuid, num, &result, retu); else error = vmx_rdmsr(vmx, vcpuid, num, &result, retu); if (error == 0) { eax = result; vmxctx = &vmx->ctx[vcpuid]; error = vmxctx_setreg(vmxctx, VM_REG_GUEST_RAX, eax); KASSERT(error == 0, ("vmxctx_setreg(rax) error %d", error)); edx = result >> 32; error = vmxctx_setreg(vmxctx, VM_REG_GUEST_RDX, edx); KASSERT(error == 0, ("vmxctx_setreg(rdx) error %d", error)); } return (error); } static int vmx_exit_process(struct vmx *vmx, int vcpu, struct vm_exit *vmexit) { int error, errcode, errcode_valid, handled, in; struct vmxctx *vmxctx; struct vlapic *vlapic; struct vm_inout_str *vis; struct vm_task_switch *ts; uint32_t eax, ecx, edx, idtvec_info, idtvec_err, intr_info, inst_info; uint32_t intr_type, intr_vec, reason; uint64_t exitintinfo, qual, gpa; bool retu; CTASSERT((PINBASED_CTLS_ONE_SETTING & PINBASED_VIRTUAL_NMI) != 0); CTASSERT((PINBASED_CTLS_ONE_SETTING & PINBASED_NMI_EXITING) != 0); handled = UNHANDLED; vmxctx = &vmx->ctx[vcpu]; qual = vmexit->u.vmx.exit_qualification; reason = vmexit->u.vmx.exit_reason; vmexit->exitcode = VM_EXITCODE_BOGUS; vmm_stat_incr(vmx->vm, vcpu, VMEXIT_COUNT, 1); /* * VM-entry failures during or after loading guest state. * * These VM-exits are uncommon but must be handled specially * as most VM-exit fields are not populated as usual. */ if (__predict_false(reason == EXIT_REASON_MCE_DURING_ENTRY)) { VCPU_CTR0(vmx->vm, vcpu, "Handling MCE during VM-entry"); __asm __volatile("int $18"); return (1); } /* * VM exits that can be triggered during event delivery need to * be handled specially by re-injecting the event if the IDT * vectoring information field's valid bit is set. * * See "Information for VM Exits During Event Delivery" in Intel SDM * for details. */ idtvec_info = vmcs_idt_vectoring_info(); if (idtvec_info & VMCS_IDT_VEC_VALID) { idtvec_info &= ~(1 << 12); /* clear undefined bit */ exitintinfo = idtvec_info; if (idtvec_info & VMCS_IDT_VEC_ERRCODE_VALID) { idtvec_err = vmcs_idt_vectoring_err(); exitintinfo |= (uint64_t)idtvec_err << 32; } error = vm_exit_intinfo(vmx->vm, vcpu, exitintinfo); KASSERT(error == 0, ("%s: vm_set_intinfo error %d", __func__, error)); /* * If 'virtual NMIs' are being used and the VM-exit * happened while injecting an NMI during the previous * VM-entry, then clear "blocking by NMI" in the * Guest Interruptibility-State so the NMI can be * reinjected on the subsequent VM-entry. * * However, if the NMI was being delivered through a task * gate, then the new task must start execution with NMIs * blocked so don't clear NMI blocking in this case. */ intr_type = idtvec_info & VMCS_INTR_T_MASK; if (intr_type == VMCS_INTR_T_NMI) { if (reason != EXIT_REASON_TASK_SWITCH) vmx_clear_nmi_blocking(vmx, vcpu); else vmx_assert_nmi_blocking(vmx, vcpu); } /* * Update VM-entry instruction length if the event being * delivered was a software interrupt or software exception. */ if (intr_type == VMCS_INTR_T_SWINTR || intr_type == VMCS_INTR_T_PRIV_SWEXCEPTION || intr_type == VMCS_INTR_T_SWEXCEPTION) { vmcs_write(VMCS_ENTRY_INST_LENGTH, vmexit->inst_length); } } switch (reason) { case EXIT_REASON_TASK_SWITCH: ts = &vmexit->u.task_switch; ts->tsssel = qual & 0xffff; ts->reason = vmx_task_switch_reason(qual); ts->ext = 0; ts->errcode_valid = 0; vmx_paging_info(&ts->paging); /* * If the task switch was due to a CALL, JMP, IRET, software * interrupt (INT n) or software exception (INT3, INTO), * then the saved %rip references the instruction that caused * the task switch. The instruction length field in the VMCS * is valid in this case. * * In all other cases (e.g., NMI, hardware exception) the * saved %rip is one that would have been saved in the old TSS * had the task switch completed normally so the instruction * length field is not needed in this case and is explicitly * set to 0. */ if (ts->reason == TSR_IDT_GATE) { KASSERT(idtvec_info & VMCS_IDT_VEC_VALID, ("invalid idtvec_info %#x for IDT task switch", idtvec_info)); intr_type = idtvec_info & VMCS_INTR_T_MASK; if (intr_type != VMCS_INTR_T_SWINTR && intr_type != VMCS_INTR_T_SWEXCEPTION && intr_type != VMCS_INTR_T_PRIV_SWEXCEPTION) { /* Task switch triggered by external event */ ts->ext = 1; vmexit->inst_length = 0; if (idtvec_info & VMCS_IDT_VEC_ERRCODE_VALID) { ts->errcode_valid = 1; ts->errcode = vmcs_idt_vectoring_err(); } } } vmexit->exitcode = VM_EXITCODE_TASK_SWITCH; VCPU_CTR4(vmx->vm, vcpu, "task switch reason %d, tss 0x%04x, " "%s errcode 0x%016lx", ts->reason, ts->tsssel, ts->ext ? "external" : "internal", ((uint64_t)ts->errcode << 32) | ts->errcode_valid); break; case EXIT_REASON_CR_ACCESS: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_CR_ACCESS, 1); switch (qual & 0xf) { case 0: handled = vmx_emulate_cr0_access(vmx, vcpu, qual); break; case 4: handled = vmx_emulate_cr4_access(vmx, vcpu, qual); break; case 8: handled = vmx_emulate_cr8_access(vmx, vcpu, qual); break; } break; case EXIT_REASON_RDMSR: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_RDMSR, 1); retu = false; ecx = vmxctx->guest_rcx; VCPU_CTR1(vmx->vm, vcpu, "rdmsr 0x%08x", ecx); error = emulate_rdmsr(vmx, vcpu, ecx, &retu); if (error) { vmexit->exitcode = VM_EXITCODE_RDMSR; vmexit->u.msr.code = ecx; } else if (!retu) { handled = HANDLED; } else { /* Return to userspace with a valid exitcode */ KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, ("emulate_rdmsr retu with bogus exitcode")); } break; case EXIT_REASON_WRMSR: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_WRMSR, 1); retu = false; eax = vmxctx->guest_rax; ecx = vmxctx->guest_rcx; edx = vmxctx->guest_rdx; VCPU_CTR2(vmx->vm, vcpu, "wrmsr 0x%08x value 0x%016lx", ecx, (uint64_t)edx << 32 | eax); error = emulate_wrmsr(vmx, vcpu, ecx, (uint64_t)edx << 32 | eax, &retu); if (error) { vmexit->exitcode = VM_EXITCODE_WRMSR; vmexit->u.msr.code = ecx; vmexit->u.msr.wval = (uint64_t)edx << 32 | eax; } else if (!retu) { handled = HANDLED; } else { /* Return to userspace with a valid exitcode */ KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, ("emulate_wrmsr retu with bogus exitcode")); } break; case EXIT_REASON_HLT: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_HLT, 1); vmexit->exitcode = VM_EXITCODE_HLT; vmexit->u.hlt.rflags = vmcs_read(VMCS_GUEST_RFLAGS); break; case EXIT_REASON_MTF: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_MTRAP, 1); vmexit->exitcode = VM_EXITCODE_MTRAP; vmexit->inst_length = 0; break; case EXIT_REASON_PAUSE: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_PAUSE, 1); vmexit->exitcode = VM_EXITCODE_PAUSE; break; case EXIT_REASON_INTR_WINDOW: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INTR_WINDOW, 1); vmx_clear_int_window_exiting(vmx, vcpu); return (1); case EXIT_REASON_EXT_INTR: /* * External interrupts serve only to cause VM exits and allow * the host interrupt handler to run. * * If this external interrupt triggers a virtual interrupt * to a VM, then that state will be recorded by the * host interrupt handler in the VM's softc. We will inject * this virtual interrupt during the subsequent VM enter. */ intr_info = vmcs_read(VMCS_EXIT_INTR_INFO); /* * XXX: Ignore this exit if VMCS_INTR_VALID is not set. * This appears to be a bug in VMware Fusion? */ if (!(intr_info & VMCS_INTR_VALID)) return (1); KASSERT((intr_info & VMCS_INTR_VALID) != 0 && (intr_info & VMCS_INTR_T_MASK) == VMCS_INTR_T_HWINTR, ("VM exit interruption info invalid: %#x", intr_info)); vmx_trigger_hostintr(intr_info & 0xff); /* * This is special. We want to treat this as an 'handled' * VM-exit but not increment the instruction pointer. */ vmm_stat_incr(vmx->vm, vcpu, VMEXIT_EXTINT, 1); return (1); case EXIT_REASON_NMI_WINDOW: /* Exit to allow the pending virtual NMI to be injected */ if (vm_nmi_pending(vmx->vm, vcpu)) vmx_inject_nmi(vmx, vcpu); vmx_clear_nmi_window_exiting(vmx, vcpu); vmm_stat_incr(vmx->vm, vcpu, VMEXIT_NMI_WINDOW, 1); return (1); case EXIT_REASON_INOUT: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INOUT, 1); vmexit->exitcode = VM_EXITCODE_INOUT; vmexit->u.inout.bytes = (qual & 0x7) + 1; vmexit->u.inout.in = in = (qual & 0x8) ? 1 : 0; vmexit->u.inout.string = (qual & 0x10) ? 1 : 0; vmexit->u.inout.rep = (qual & 0x20) ? 1 : 0; vmexit->u.inout.port = (uint16_t)(qual >> 16); vmexit->u.inout.eax = (uint32_t)(vmxctx->guest_rax); if (vmexit->u.inout.string) { inst_info = vmcs_read(VMCS_EXIT_INSTRUCTION_INFO); vmexit->exitcode = VM_EXITCODE_INOUT_STR; vis = &vmexit->u.inout_str; vmx_paging_info(&vis->paging); vis->rflags = vmcs_read(VMCS_GUEST_RFLAGS); vis->cr0 = vmcs_read(VMCS_GUEST_CR0); vis->index = inout_str_index(vmx, vcpu, in); vis->count = inout_str_count(vmx, vcpu, vis->inout.rep); vis->addrsize = inout_str_addrsize(inst_info); inout_str_seginfo(vmx, vcpu, inst_info, in, vis); } break; case EXIT_REASON_CPUID: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_CPUID, 1); handled = vmx_handle_cpuid(vmx->vm, vcpu, vmxctx); break; case EXIT_REASON_EXCEPTION: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_EXCEPTION, 1); intr_info = vmcs_read(VMCS_EXIT_INTR_INFO); KASSERT((intr_info & VMCS_INTR_VALID) != 0, ("VM exit interruption info invalid: %#x", intr_info)); intr_vec = intr_info & 0xff; intr_type = intr_info & VMCS_INTR_T_MASK; /* * If Virtual NMIs control is 1 and the VM-exit is due to a * fault encountered during the execution of IRET then we must * restore the state of "virtual-NMI blocking" before resuming * the guest. * * See "Resuming Guest Software after Handling an Exception". * See "Information for VM Exits Due to Vectored Events". */ if ((idtvec_info & VMCS_IDT_VEC_VALID) == 0 && (intr_vec != IDT_DF) && (intr_info & EXIT_QUAL_NMIUDTI) != 0) vmx_restore_nmi_blocking(vmx, vcpu); /* * The NMI has already been handled in vmx_exit_handle_nmi(). */ if (intr_type == VMCS_INTR_T_NMI) return (1); /* * Call the machine check handler by hand. Also don't reflect * the machine check back into the guest. */ if (intr_vec == IDT_MC) { VCPU_CTR0(vmx->vm, vcpu, "Vectoring to MCE handler"); __asm __volatile("int $18"); return (1); } if (intr_vec == IDT_PF) { error = vmxctx_setreg(vmxctx, VM_REG_GUEST_CR2, qual); KASSERT(error == 0, ("%s: vmxctx_setreg(cr2) error %d", __func__, error)); } /* * Software exceptions exhibit trap-like behavior. This in * turn requires populating the VM-entry instruction length * so that the %rip in the trap frame is past the INT3/INTO * instruction. */ if (intr_type == VMCS_INTR_T_SWEXCEPTION) vmcs_write(VMCS_ENTRY_INST_LENGTH, vmexit->inst_length); /* Reflect all other exceptions back into the guest */ errcode_valid = errcode = 0; if (intr_info & VMCS_INTR_DEL_ERRCODE) { errcode_valid = 1; errcode = vmcs_read(VMCS_EXIT_INTR_ERRCODE); } VCPU_CTR2(vmx->vm, vcpu, "Reflecting exception %d/%#x into " "the guest", intr_vec, errcode); error = vm_inject_exception(vmx->vm, vcpu, intr_vec, errcode_valid, errcode, 0); KASSERT(error == 0, ("%s: vm_inject_exception error %d", __func__, error)); return (1); case EXIT_REASON_EPT_FAULT: /* * If 'gpa' lies within the address space allocated to * memory then this must be a nested page fault otherwise * this must be an instruction that accesses MMIO space. */ gpa = vmcs_gpa(); if (vm_mem_allocated(vmx->vm, gpa) || apic_access_fault(vmx, vcpu, gpa)) { vmexit->exitcode = VM_EXITCODE_PAGING; vmexit->inst_length = 0; vmexit->u.paging.gpa = gpa; vmexit->u.paging.fault_type = ept_fault_type(qual); vmm_stat_incr(vmx->vm, vcpu, VMEXIT_NESTED_FAULT, 1); } else if (ept_emulation_fault(qual)) { vmexit_inst_emul(vmexit, gpa, vmcs_gla()); vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INST_EMUL, 1); } /* * If Virtual NMIs control is 1 and the VM-exit is due to an * EPT fault during the execution of IRET then we must restore * the state of "virtual-NMI blocking" before resuming. * * See description of "NMI unblocking due to IRET" in * "Exit Qualification for EPT Violations". */ if ((idtvec_info & VMCS_IDT_VEC_VALID) == 0 && (qual & EXIT_QUAL_NMIUDTI) != 0) vmx_restore_nmi_blocking(vmx, vcpu); break; case EXIT_REASON_VIRTUALIZED_EOI: vmexit->exitcode = VM_EXITCODE_IOAPIC_EOI; vmexit->u.ioapic_eoi.vector = qual & 0xFF; vmexit->inst_length = 0; /* trap-like */ break; case EXIT_REASON_APIC_ACCESS: handled = vmx_handle_apic_access(vmx, vcpu, vmexit); break; case EXIT_REASON_APIC_WRITE: /* * APIC-write VM exit is trap-like so the %rip is already * pointing to the next instruction. */ vmexit->inst_length = 0; vlapic = vm_lapic(vmx->vm, vcpu); handled = vmx_handle_apic_write(vmx, vcpu, vlapic, qual); break; case EXIT_REASON_XSETBV: handled = vmx_emulate_xsetbv(vmx, vcpu, vmexit); break; case EXIT_REASON_MONITOR: vmexit->exitcode = VM_EXITCODE_MONITOR; break; case EXIT_REASON_MWAIT: vmexit->exitcode = VM_EXITCODE_MWAIT; break; default: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_UNKNOWN, 1); break; } if (handled) { /* * It is possible that control is returned to userland * even though we were able to handle the VM exit in the * kernel. * * In such a case we want to make sure that the userland * restarts guest execution at the instruction *after* * the one we just processed. Therefore we update the * guest rip in the VMCS and in 'vmexit'. */ vmexit->rip += vmexit->inst_length; vmexit->inst_length = 0; vmcs_write(VMCS_GUEST_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 VMX exit. */ vmexit->exitcode = VM_EXITCODE_VMX; vmexit->u.vmx.status = VM_SUCCESS; vmexit->u.vmx.inst_type = 0; vmexit->u.vmx.inst_error = 0; } else { /* * The exitcode and collateral have been populated. * The VM exit will be processed further in userland. */ } } return (handled); } static __inline void vmx_exit_inst_error(struct vmxctx *vmxctx, int rc, struct vm_exit *vmexit) { KASSERT(vmxctx->inst_fail_status != VM_SUCCESS, ("vmx_exit_inst_error: invalid inst_fail_status %d", vmxctx->inst_fail_status)); vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_VMX; vmexit->u.vmx.status = vmxctx->inst_fail_status; vmexit->u.vmx.inst_error = vmcs_instruction_error(); vmexit->u.vmx.exit_reason = ~0; vmexit->u.vmx.exit_qualification = ~0; switch (rc) { case VMX_VMRESUME_ERROR: case VMX_VMLAUNCH_ERROR: case VMX_INVEPT_ERROR: vmexit->u.vmx.inst_type = rc; break; default: panic("vm_exit_inst_error: vmx_enter_guest returned %d", rc); } } /* * If the NMI-exiting VM execution control is set to '1' then an NMI in * non-root operation causes a VM-exit. NMI blocking is in effect so it is * sufficient to simply vector to the NMI handler via a software interrupt. * However, this must be done before maskable interrupts are enabled * otherwise the "iret" issued by an interrupt handler will incorrectly * clear NMI blocking. */ static __inline void vmx_exit_handle_nmi(struct vmx *vmx, int vcpuid, struct vm_exit *vmexit) { uint32_t intr_info; KASSERT((read_rflags() & PSL_I) == 0, ("interrupts enabled")); if (vmexit->u.vmx.exit_reason != EXIT_REASON_EXCEPTION) return; intr_info = vmcs_read(VMCS_EXIT_INTR_INFO); KASSERT((intr_info & VMCS_INTR_VALID) != 0, ("VM exit interruption info invalid: %#x", intr_info)); if ((intr_info & VMCS_INTR_T_MASK) == VMCS_INTR_T_NMI) { KASSERT((intr_info & 0xff) == IDT_NMI, ("VM exit due " "to NMI has invalid vector: %#x", intr_info)); VCPU_CTR0(vmx->vm, vcpuid, "Vectoring to NMI handler"); __asm __volatile("int $2"); } } static int vmx_run(void *arg, int vcpu, register_t rip, pmap_t pmap, - void *rendezvous_cookie, void *suspend_cookie) + struct vm_eventinfo *evinfo) { int rc, handled, launched; struct vmx *vmx; struct vm *vm; struct vmxctx *vmxctx; struct vmcs *vmcs; struct vm_exit *vmexit; struct vlapic *vlapic; uint32_t exit_reason; vmx = arg; vm = vmx->vm; vmcs = &vmx->vmcs[vcpu]; vmxctx = &vmx->ctx[vcpu]; vlapic = vm_lapic(vm, vcpu); vmexit = vm_exitinfo(vm, vcpu); launched = 0; KASSERT(vmxctx->pmap == pmap, ("pmap %p different than ctx pmap %p", pmap, vmxctx->pmap)); vmx_msr_guest_enter(vmx, vcpu); VMPTRLD(vmcs); /* * XXX * We do this every time because we may setup the virtual machine * from a different process than the one that actually runs it. * * If the life of a virtual machine was spent entirely in the context * of a single process we could do this once in vmx_vminit(). */ vmcs_write(VMCS_HOST_CR3, rcr3()); vmcs_write(VMCS_GUEST_RIP, rip); vmx_set_pcpu_defaults(vmx, vcpu, pmap); do { KASSERT(vmcs_guest_rip() == rip, ("%s: vmcs guest rip mismatch " "%#lx/%#lx", __func__, vmcs_guest_rip(), rip)); handled = UNHANDLED; /* * Interrupts are disabled from this point on until the * guest starts executing. This is done for the following * reasons: * * If an AST is asserted on this thread after the check below, * then the IPI_AST notification will not be lost, because it * will cause a VM exit due to external interrupt as soon as * the guest state is loaded. * * A posted interrupt after 'vmx_inject_interrupts()' will * not be "lost" because it will be held pending in the host * APIC because interrupts are disabled. The pending interrupt * will be recognized as soon as the guest state is loaded. * * The same reasoning applies to the IPI generated by * pmap_invalidate_ept(). */ disable_intr(); vmx_inject_interrupts(vmx, vcpu, vlapic, rip); /* * Check for vcpu suspension after injecting events because * vmx_inject_interrupts() can suspend the vcpu due to a * triple fault. */ - if (vcpu_suspended(suspend_cookie)) { + if (vcpu_suspended(evinfo)) { enable_intr(); vm_exit_suspended(vmx->vm, vcpu, rip); break; } - if (vcpu_rendezvous_pending(rendezvous_cookie)) { + if (vcpu_rendezvous_pending(evinfo)) { enable_intr(); vm_exit_rendezvous(vmx->vm, vcpu, rip); + break; + } + + if (vcpu_reqidle(evinfo)) { + enable_intr(); + vm_exit_reqidle(vmx->vm, vcpu, rip); break; } if (vcpu_should_yield(vm, vcpu)) { enable_intr(); vm_exit_astpending(vmx->vm, vcpu, rip); vmx_astpending_trace(vmx, vcpu, rip); handled = HANDLED; break; } vmx_run_trace(vmx, vcpu); rc = vmx_enter_guest(vmxctx, vmx, launched); /* Collect some information for VM exit processing */ vmexit->rip = rip = vmcs_guest_rip(); vmexit->inst_length = vmexit_instruction_length(); vmexit->u.vmx.exit_reason = exit_reason = vmcs_exit_reason(); vmexit->u.vmx.exit_qualification = vmcs_exit_qualification(); /* Update 'nextrip' */ vmx->state[vcpu].nextrip = rip; if (rc == VMX_GUEST_VMEXIT) { vmx_exit_handle_nmi(vmx, vcpu, vmexit); enable_intr(); handled = vmx_exit_process(vmx, vcpu, vmexit); } else { enable_intr(); vmx_exit_inst_error(vmxctx, rc, vmexit); } launched = 1; vmx_exit_trace(vmx, vcpu, rip, exit_reason, handled); rip = vmexit->rip; } while (handled); /* * If a VM exit has been handled then the exitcode must be BOGUS * If a VM exit is not handled then the exitcode must not be BOGUS */ if ((handled && vmexit->exitcode != VM_EXITCODE_BOGUS) || (!handled && vmexit->exitcode == VM_EXITCODE_BOGUS)) { panic("Mismatch between handled (%d) and exitcode (%d)", handled, vmexit->exitcode); } if (!handled) vmm_stat_incr(vm, vcpu, VMEXIT_USERSPACE, 1); VCPU_CTR1(vm, vcpu, "returning from vmx_run: exitcode %d", vmexit->exitcode); VMCLEAR(vmcs); vmx_msr_guest_exit(vmx, vcpu); return (0); } static void vmx_vmcleanup(void *arg) { int i; struct vmx *vmx = arg; if (apic_access_virtualization(vmx, 0)) vm_unmap_mmio(vmx->vm, DEFAULT_APIC_BASE, PAGE_SIZE); for (i = 0; i < VM_MAXCPU; i++) vpid_free(vmx->state[i].vpid); free(vmx, M_VMX); return; } static register_t * vmxctx_regptr(struct vmxctx *vmxctx, int reg) { switch (reg) { case VM_REG_GUEST_RAX: return (&vmxctx->guest_rax); case VM_REG_GUEST_RBX: return (&vmxctx->guest_rbx); case VM_REG_GUEST_RCX: return (&vmxctx->guest_rcx); case VM_REG_GUEST_RDX: return (&vmxctx->guest_rdx); case VM_REG_GUEST_RSI: return (&vmxctx->guest_rsi); case VM_REG_GUEST_RDI: return (&vmxctx->guest_rdi); case VM_REG_GUEST_RBP: return (&vmxctx->guest_rbp); case VM_REG_GUEST_R8: return (&vmxctx->guest_r8); case VM_REG_GUEST_R9: return (&vmxctx->guest_r9); case VM_REG_GUEST_R10: return (&vmxctx->guest_r10); case VM_REG_GUEST_R11: return (&vmxctx->guest_r11); case VM_REG_GUEST_R12: return (&vmxctx->guest_r12); case VM_REG_GUEST_R13: return (&vmxctx->guest_r13); case VM_REG_GUEST_R14: return (&vmxctx->guest_r14); case VM_REG_GUEST_R15: return (&vmxctx->guest_r15); case VM_REG_GUEST_CR2: return (&vmxctx->guest_cr2); default: break; } return (NULL); } static int vmxctx_getreg(struct vmxctx *vmxctx, int reg, uint64_t *retval) { register_t *regp; if ((regp = vmxctx_regptr(vmxctx, reg)) != NULL) { *retval = *regp; return (0); } else return (EINVAL); } static int vmxctx_setreg(struct vmxctx *vmxctx, int reg, uint64_t val) { register_t *regp; if ((regp = vmxctx_regptr(vmxctx, reg)) != NULL) { *regp = val; return (0); } else return (EINVAL); } static int vmx_get_intr_shadow(struct vmx *vmx, int vcpu, int running, uint64_t *retval) { uint64_t gi; int error; error = vmcs_getreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(VMCS_GUEST_INTERRUPTIBILITY), &gi); *retval = (gi & HWINTR_BLOCKING) ? 1 : 0; return (error); } static int vmx_modify_intr_shadow(struct vmx *vmx, int vcpu, int running, uint64_t val) { struct vmcs *vmcs; uint64_t gi; int error, ident; /* * Forcing the vcpu into an interrupt shadow is not supported. */ if (val) { error = EINVAL; goto done; } vmcs = &vmx->vmcs[vcpu]; ident = VMCS_IDENT(VMCS_GUEST_INTERRUPTIBILITY); error = vmcs_getreg(vmcs, running, ident, &gi); if (error == 0) { gi &= ~HWINTR_BLOCKING; error = vmcs_setreg(vmcs, running, ident, gi); } done: VCPU_CTR2(vmx->vm, vcpu, "Setting intr_shadow to %#lx %s", val, error ? "failed" : "succeeded"); return (error); } static int vmx_shadow_reg(int reg) { int shreg; shreg = -1; switch (reg) { case VM_REG_GUEST_CR0: shreg = VMCS_CR0_SHADOW; break; case VM_REG_GUEST_CR4: shreg = VMCS_CR4_SHADOW; break; default: break; } return (shreg); } static int vmx_getreg(void *arg, int vcpu, int reg, uint64_t *retval) { int running, hostcpu; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_getreg: %s%d is running", vm_name(vmx->vm), vcpu); if (reg == VM_REG_GUEST_INTR_SHADOW) return (vmx_get_intr_shadow(vmx, vcpu, running, retval)); if (vmxctx_getreg(&vmx->ctx[vcpu], reg, retval) == 0) return (0); return (vmcs_getreg(&vmx->vmcs[vcpu], running, reg, retval)); } static int vmx_setreg(void *arg, int vcpu, int reg, uint64_t val) { int error, hostcpu, running, shadow; uint64_t ctls; pmap_t pmap; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_setreg: %s%d is running", vm_name(vmx->vm), vcpu); if (reg == VM_REG_GUEST_INTR_SHADOW) return (vmx_modify_intr_shadow(vmx, vcpu, running, val)); if (vmxctx_setreg(&vmx->ctx[vcpu], reg, val) == 0) return (0); error = vmcs_setreg(&vmx->vmcs[vcpu], running, reg, val); if (error == 0) { /* * If the "load EFER" VM-entry control is 1 then the * value of EFER.LMA must be identical to "IA-32e mode guest" * bit in the VM-entry control. */ if ((entry_ctls & VM_ENTRY_LOAD_EFER) != 0 && (reg == VM_REG_GUEST_EFER)) { vmcs_getreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(VMCS_ENTRY_CTLS), &ctls); if (val & EFER_LMA) ctls |= VM_ENTRY_GUEST_LMA; else ctls &= ~VM_ENTRY_GUEST_LMA; vmcs_setreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(VMCS_ENTRY_CTLS), ctls); } shadow = vmx_shadow_reg(reg); if (shadow > 0) { /* * Store the unmodified value in the shadow */ error = vmcs_setreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(shadow), val); } if (reg == VM_REG_GUEST_CR3) { /* * Invalidate the guest vcpu's TLB mappings to emulate * the behavior of updating %cr3. * * XXX the processor retains global mappings when %cr3 * is updated but vmx_invvpid() does not. */ pmap = vmx->ctx[vcpu].pmap; vmx_invvpid(vmx, vcpu, pmap, running); } } return (error); } static int vmx_getdesc(void *arg, int vcpu, int reg, struct seg_desc *desc) { int hostcpu, running; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_getdesc: %s%d is running", vm_name(vmx->vm), vcpu); return (vmcs_getdesc(&vmx->vmcs[vcpu], running, reg, desc)); } static int vmx_setdesc(void *arg, int vcpu, int reg, struct seg_desc *desc) { int hostcpu, running; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_setdesc: %s%d is running", vm_name(vmx->vm), vcpu); return (vmcs_setdesc(&vmx->vmcs[vcpu], running, reg, desc)); } static int vmx_getcap(void *arg, int vcpu, int type, int *retval) { struct vmx *vmx = arg; int vcap; int ret; ret = ENOENT; vcap = vmx->cap[vcpu].set; switch (type) { case VM_CAP_HALT_EXIT: if (cap_halt_exit) ret = 0; break; case VM_CAP_PAUSE_EXIT: if (cap_pause_exit) ret = 0; break; case VM_CAP_MTRAP_EXIT: if (cap_monitor_trap) ret = 0; break; case VM_CAP_UNRESTRICTED_GUEST: if (cap_unrestricted_guest) ret = 0; break; case VM_CAP_ENABLE_INVPCID: if (cap_invpcid) ret = 0; break; default: break; } if (ret == 0) *retval = (vcap & (1 << type)) ? 1 : 0; return (ret); } static int vmx_setcap(void *arg, int vcpu, int type, int val) { struct vmx *vmx = arg; struct vmcs *vmcs = &vmx->vmcs[vcpu]; uint32_t baseval; uint32_t *pptr; int error; int flag; int reg; int retval; retval = ENOENT; pptr = NULL; switch (type) { case VM_CAP_HALT_EXIT: if (cap_halt_exit) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls; baseval = *pptr; flag = PROCBASED_HLT_EXITING; reg = VMCS_PRI_PROC_BASED_CTLS; } break; case VM_CAP_MTRAP_EXIT: if (cap_monitor_trap) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls; baseval = *pptr; flag = PROCBASED_MTF; reg = VMCS_PRI_PROC_BASED_CTLS; } break; case VM_CAP_PAUSE_EXIT: if (cap_pause_exit) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls; baseval = *pptr; flag = PROCBASED_PAUSE_EXITING; reg = VMCS_PRI_PROC_BASED_CTLS; } break; case VM_CAP_UNRESTRICTED_GUEST: if (cap_unrestricted_guest) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls2; baseval = *pptr; flag = PROCBASED2_UNRESTRICTED_GUEST; reg = VMCS_SEC_PROC_BASED_CTLS; } break; case VM_CAP_ENABLE_INVPCID: if (cap_invpcid) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls2; baseval = *pptr; flag = PROCBASED2_ENABLE_INVPCID; reg = VMCS_SEC_PROC_BASED_CTLS; } break; default: break; } if (retval == 0) { if (val) { baseval |= flag; } else { baseval &= ~flag; } VMPTRLD(vmcs); error = vmwrite(reg, baseval); VMCLEAR(vmcs); if (error) { retval = error; } else { /* * Update optional stored flags, and record * setting */ if (pptr != NULL) { *pptr = baseval; } if (val) { vmx->cap[vcpu].set |= (1 << type); } else { vmx->cap[vcpu].set &= ~(1 << type); } } } return (retval); } struct vlapic_vtx { struct vlapic vlapic; struct pir_desc *pir_desc; struct vmx *vmx; }; #define VMX_CTR_PIR(vm, vcpuid, pir_desc, notify, vector, level, msg) \ do { \ VCPU_CTR2(vm, vcpuid, msg " assert %s-triggered vector %d", \ level ? "level" : "edge", vector); \ VCPU_CTR1(vm, vcpuid, msg " pir0 0x%016lx", pir_desc->pir[0]); \ VCPU_CTR1(vm, vcpuid, msg " pir1 0x%016lx", pir_desc->pir[1]); \ VCPU_CTR1(vm, vcpuid, msg " pir2 0x%016lx", pir_desc->pir[2]); \ VCPU_CTR1(vm, vcpuid, msg " pir3 0x%016lx", pir_desc->pir[3]); \ VCPU_CTR1(vm, vcpuid, msg " notify: %s", notify ? "yes" : "no");\ } while (0) /* * vlapic->ops handlers that utilize the APICv hardware assist described in * Chapter 29 of the Intel SDM. */ static int vmx_set_intr_ready(struct vlapic *vlapic, int vector, bool level) { struct vlapic_vtx *vlapic_vtx; struct pir_desc *pir_desc; uint64_t mask; int idx, notify; vlapic_vtx = (struct vlapic_vtx *)vlapic; pir_desc = vlapic_vtx->pir_desc; /* * Keep track of interrupt requests in the PIR descriptor. This is * because the virtual APIC page pointed to by the VMCS cannot be * modified if the vcpu is running. */ idx = vector / 64; mask = 1UL << (vector % 64); atomic_set_long(&pir_desc->pir[idx], mask); notify = atomic_cmpset_long(&pir_desc->pending, 0, 1); VMX_CTR_PIR(vlapic->vm, vlapic->vcpuid, pir_desc, notify, vector, level, "vmx_set_intr_ready"); return (notify); } static int vmx_pending_intr(struct vlapic *vlapic, int *vecptr) { struct vlapic_vtx *vlapic_vtx; struct pir_desc *pir_desc; struct LAPIC *lapic; uint64_t pending, pirval; uint32_t ppr, vpr; int i; /* * This function is only expected to be called from the 'HLT' exit * handler which does not care about the vector that is pending. */ KASSERT(vecptr == NULL, ("vmx_pending_intr: vecptr must be NULL")); vlapic_vtx = (struct vlapic_vtx *)vlapic; pir_desc = vlapic_vtx->pir_desc; pending = atomic_load_acq_long(&pir_desc->pending); if (!pending) return (0); /* common case */ /* * If there is an interrupt pending then it will be recognized only * if its priority is greater than the processor priority. * * Special case: if the processor priority is zero then any pending * interrupt will be recognized. */ lapic = vlapic->apic_page; ppr = lapic->ppr & 0xf0; if (ppr == 0) return (1); VCPU_CTR1(vlapic->vm, vlapic->vcpuid, "HLT with non-zero PPR %d", lapic->ppr); for (i = 3; i >= 0; i--) { pirval = pir_desc->pir[i]; if (pirval != 0) { vpr = (i * 64 + flsl(pirval) - 1) & 0xf0; return (vpr > ppr); } } return (0); } static void vmx_intr_accepted(struct vlapic *vlapic, int vector) { panic("vmx_intr_accepted: not expected to be called"); } static void vmx_set_tmr(struct vlapic *vlapic, int vector, bool level) { struct vlapic_vtx *vlapic_vtx; struct vmx *vmx; struct vmcs *vmcs; uint64_t mask, val; KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d", vector)); KASSERT(!vcpu_is_running(vlapic->vm, vlapic->vcpuid, NULL), ("vmx_set_tmr: vcpu cannot be running")); vlapic_vtx = (struct vlapic_vtx *)vlapic; vmx = vlapic_vtx->vmx; vmcs = &vmx->vmcs[vlapic->vcpuid]; mask = 1UL << (vector % 64); VMPTRLD(vmcs); val = vmcs_read(VMCS_EOI_EXIT(vector)); if (level) val |= mask; else val &= ~mask; vmcs_write(VMCS_EOI_EXIT(vector), val); VMCLEAR(vmcs); } static void vmx_enable_x2apic_mode(struct vlapic *vlapic) { struct vmx *vmx; struct vmcs *vmcs; uint32_t proc_ctls2; int vcpuid, error; vcpuid = vlapic->vcpuid; vmx = ((struct vlapic_vtx *)vlapic)->vmx; vmcs = &vmx->vmcs[vcpuid]; proc_ctls2 = vmx->cap[vcpuid].proc_ctls2; KASSERT((proc_ctls2 & PROCBASED2_VIRTUALIZE_APIC_ACCESSES) != 0, ("%s: invalid proc_ctls2 %#x", __func__, proc_ctls2)); proc_ctls2 &= ~PROCBASED2_VIRTUALIZE_APIC_ACCESSES; proc_ctls2 |= PROCBASED2_VIRTUALIZE_X2APIC_MODE; vmx->cap[vcpuid].proc_ctls2 = proc_ctls2; VMPTRLD(vmcs); vmcs_write(VMCS_SEC_PROC_BASED_CTLS, proc_ctls2); VMCLEAR(vmcs); if (vlapic->vcpuid == 0) { /* * The nested page table mappings are shared by all vcpus * so unmap the APIC access page just once. */ error = vm_unmap_mmio(vmx->vm, DEFAULT_APIC_BASE, PAGE_SIZE); KASSERT(error == 0, ("%s: vm_unmap_mmio error %d", __func__, error)); /* * The MSR bitmap is shared by all vcpus so modify it only * once in the context of vcpu 0. */ error = vmx_allow_x2apic_msrs(vmx); KASSERT(error == 0, ("%s: vmx_allow_x2apic_msrs error %d", __func__, error)); } } static void vmx_post_intr(struct vlapic *vlapic, int hostcpu) { ipi_cpu(hostcpu, pirvec); } /* * Transfer the pending interrupts in the PIR descriptor to the IRR * in the virtual APIC page. */ static void vmx_inject_pir(struct vlapic *vlapic) { struct vlapic_vtx *vlapic_vtx; struct pir_desc *pir_desc; struct LAPIC *lapic; uint64_t val, pirval; int rvi, pirbase = -1; uint16_t intr_status_old, intr_status_new; vlapic_vtx = (struct vlapic_vtx *)vlapic; pir_desc = vlapic_vtx->pir_desc; if (atomic_cmpset_long(&pir_desc->pending, 1, 0) == 0) { VCPU_CTR0(vlapic->vm, vlapic->vcpuid, "vmx_inject_pir: " "no posted interrupt pending"); return; } pirval = 0; pirbase = -1; lapic = vlapic->apic_page; val = atomic_readandclear_long(&pir_desc->pir[0]); if (val != 0) { lapic->irr0 |= val; lapic->irr1 |= val >> 32; pirbase = 0; pirval = val; } val = atomic_readandclear_long(&pir_desc->pir[1]); if (val != 0) { lapic->irr2 |= val; lapic->irr3 |= val >> 32; pirbase = 64; pirval = val; } val = atomic_readandclear_long(&pir_desc->pir[2]); if (val != 0) { lapic->irr4 |= val; lapic->irr5 |= val >> 32; pirbase = 128; pirval = val; } val = atomic_readandclear_long(&pir_desc->pir[3]); if (val != 0) { lapic->irr6 |= val; lapic->irr7 |= val >> 32; pirbase = 192; pirval = val; } VLAPIC_CTR_IRR(vlapic, "vmx_inject_pir"); /* * Update RVI so the processor can evaluate pending virtual * interrupts on VM-entry. * * It is possible for pirval to be 0 here, even though the * pending bit has been set. The scenario is: * CPU-Y is sending a posted interrupt to CPU-X, which * is running a guest and processing posted interrupts in h/w. * CPU-X will eventually exit and the state seen in s/w is * the pending bit set, but no PIR bits set. * * CPU-X CPU-Y * (vm running) (host running) * rx posted interrupt * CLEAR pending bit * SET PIR bit * READ/CLEAR PIR bits * SET pending bit * (vm exit) * pending bit set, PIR 0 */ if (pirval != 0) { rvi = pirbase + flsl(pirval) - 1; intr_status_old = vmcs_read(VMCS_GUEST_INTR_STATUS); intr_status_new = (intr_status_old & 0xFF00) | rvi; if (intr_status_new > intr_status_old) { vmcs_write(VMCS_GUEST_INTR_STATUS, intr_status_new); VCPU_CTR2(vlapic->vm, vlapic->vcpuid, "vmx_inject_pir: " "guest_intr_status changed from 0x%04x to 0x%04x", intr_status_old, intr_status_new); } } } static struct vlapic * vmx_vlapic_init(void *arg, int vcpuid) { struct vmx *vmx; struct vlapic *vlapic; struct vlapic_vtx *vlapic_vtx; vmx = arg; vlapic = malloc(sizeof(struct vlapic_vtx), M_VLAPIC, M_WAITOK | M_ZERO); vlapic->vm = vmx->vm; vlapic->vcpuid = vcpuid; vlapic->apic_page = (struct LAPIC *)&vmx->apic_page[vcpuid]; vlapic_vtx = (struct vlapic_vtx *)vlapic; vlapic_vtx->pir_desc = &vmx->pir_desc[vcpuid]; vlapic_vtx->vmx = vmx; if (virtual_interrupt_delivery) { vlapic->ops.set_intr_ready = vmx_set_intr_ready; vlapic->ops.pending_intr = vmx_pending_intr; vlapic->ops.intr_accepted = vmx_intr_accepted; vlapic->ops.set_tmr = vmx_set_tmr; vlapic->ops.enable_x2apic_mode = vmx_enable_x2apic_mode; } if (posted_interrupts) vlapic->ops.post_intr = vmx_post_intr; vlapic_init(vlapic); return (vlapic); } static void vmx_vlapic_cleanup(void *arg, struct vlapic *vlapic) { vlapic_cleanup(vlapic); free(vlapic, M_VLAPIC); } struct vmm_ops vmm_ops_intel = { vmx_init, vmx_cleanup, vmx_restore, vmx_vminit, vmx_run, vmx_vmcleanup, vmx_getreg, vmx_setreg, vmx_getdesc, vmx_setdesc, vmx_getcap, vmx_setcap, ept_vmspace_alloc, ept_vmspace_free, vmx_vlapic_init, vmx_vlapic_cleanup, }; Index: head/sys/amd64/vmm/vmm.c =================================================================== --- head/sys/amd64/vmm/vmm.c (revision 283656) +++ head/sys/amd64/vmm/vmm.c (revision 283657) @@ -1,2435 +1,2505 @@ /*- * Copyright (c) 2011 NetApp, Inc. * 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, 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 NETAPP, INC ``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 NETAPP, INC OR CONTRIBUTORS 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. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "vmm_ioport.h" #include "vmm_ktr.h" #include "vmm_host.h" #include "vmm_mem.h" #include "vmm_util.h" #include "vatpic.h" #include "vatpit.h" #include "vhpet.h" #include "vioapic.h" #include "vlapic.h" #include "vpmtmr.h" #include "vrtc.h" #include "vmm_stat.h" #include "vmm_lapic.h" #include "io/ppt.h" #include "io/iommu.h" struct vlapic; /* * Initialization: * (a) allocated when vcpu is created * (i) initialized when vcpu is created and when it is reinitialized * (o) initialized the first time the vcpu is created * (x) initialized before use */ struct vcpu { struct mtx mtx; /* (o) protects 'state' and 'hostcpu' */ enum vcpu_state state; /* (o) vcpu state */ int hostcpu; /* (o) vcpu's host cpu */ + int reqidle; /* (i) request vcpu to idle */ struct vlapic *vlapic; /* (i) APIC device model */ enum x2apic_state x2apic_state; /* (i) APIC mode */ uint64_t exitintinfo; /* (i) events pending at VM exit */ int nmi_pending; /* (i) NMI pending */ int extint_pending; /* (i) INTR pending */ int exception_pending; /* (i) exception pending */ int exc_vector; /* (x) exception collateral */ int exc_errcode_valid; uint32_t exc_errcode; struct savefpu *guestfpu; /* (a,i) guest fpu state */ uint64_t guest_xcr0; /* (i) guest %xcr0 register */ void *stats; /* (a,i) statistics */ struct vm_exit exitinfo; /* (x) exit reason and collateral */ uint64_t nextrip; /* (x) next instruction to execute */ }; #define vcpu_lock_initialized(v) mtx_initialized(&((v)->mtx)) #define vcpu_lock_init(v) mtx_init(&((v)->mtx), "vcpu lock", 0, MTX_SPIN) #define vcpu_lock(v) mtx_lock_spin(&((v)->mtx)) #define vcpu_unlock(v) mtx_unlock_spin(&((v)->mtx)) #define vcpu_assert_locked(v) mtx_assert(&((v)->mtx), MA_OWNED) struct mem_seg { vm_paddr_t gpa; size_t len; boolean_t wired; vm_object_t object; }; #define VM_MAX_MEMORY_SEGMENTS 2 /* * Initialization: * (o) initialized the first time the VM is created * (i) initialized when VM is created and when it is reinitialized * (x) initialized before use */ struct vm { void *cookie; /* (i) cpu-specific data */ void *iommu; /* (x) iommu-specific data */ struct vhpet *vhpet; /* (i) virtual HPET */ struct vioapic *vioapic; /* (i) virtual ioapic */ struct vatpic *vatpic; /* (i) virtual atpic */ struct vatpit *vatpit; /* (i) virtual atpit */ struct vpmtmr *vpmtmr; /* (i) virtual ACPI PM timer */ struct vrtc *vrtc; /* (o) virtual RTC */ volatile cpuset_t active_cpus; /* (i) active vcpus */ int suspend; /* (i) stop VM execution */ volatile cpuset_t suspended_cpus; /* (i) suspended vcpus */ volatile cpuset_t halted_cpus; /* (x) cpus in a hard halt */ cpuset_t rendezvous_req_cpus; /* (x) rendezvous requested */ cpuset_t rendezvous_done_cpus; /* (x) rendezvous finished */ void *rendezvous_arg; /* (x) rendezvous func/arg */ vm_rendezvous_func_t rendezvous_func; struct mtx rendezvous_mtx; /* (o) rendezvous lock */ int num_mem_segs; /* (o) guest memory segments */ struct mem_seg mem_segs[VM_MAX_MEMORY_SEGMENTS]; struct vmspace *vmspace; /* (o) guest's address space */ char name[VM_MAX_NAMELEN]; /* (o) virtual machine name */ struct vcpu vcpu[VM_MAXCPU]; /* (i) guest vcpus */ }; static int vmm_initialized; static struct vmm_ops *ops; #define VMM_INIT(num) (ops != NULL ? (*ops->init)(num) : 0) #define VMM_CLEANUP() (ops != NULL ? (*ops->cleanup)() : 0) #define VMM_RESUME() (ops != NULL ? (*ops->resume)() : 0) #define VMINIT(vm, pmap) (ops != NULL ? (*ops->vminit)(vm, pmap): NULL) -#define VMRUN(vmi, vcpu, rip, pmap, rptr, sptr) \ - (ops != NULL ? (*ops->vmrun)(vmi, vcpu, rip, pmap, rptr, sptr) : ENXIO) +#define VMRUN(vmi, vcpu, rip, pmap, evinfo) \ + (ops != NULL ? (*ops->vmrun)(vmi, vcpu, rip, pmap, evinfo) : ENXIO) #define VMCLEANUP(vmi) (ops != NULL ? (*ops->vmcleanup)(vmi) : NULL) #define VMSPACE_ALLOC(min, max) \ (ops != NULL ? (*ops->vmspace_alloc)(min, max) : NULL) #define VMSPACE_FREE(vmspace) \ (ops != NULL ? (*ops->vmspace_free)(vmspace) : ENXIO) #define VMGETREG(vmi, vcpu, num, retval) \ (ops != NULL ? (*ops->vmgetreg)(vmi, vcpu, num, retval) : ENXIO) #define VMSETREG(vmi, vcpu, num, val) \ (ops != NULL ? (*ops->vmsetreg)(vmi, vcpu, num, val) : ENXIO) #define VMGETDESC(vmi, vcpu, num, desc) \ (ops != NULL ? (*ops->vmgetdesc)(vmi, vcpu, num, desc) : ENXIO) #define VMSETDESC(vmi, vcpu, num, desc) \ (ops != NULL ? (*ops->vmsetdesc)(vmi, vcpu, num, desc) : ENXIO) #define VMGETCAP(vmi, vcpu, num, retval) \ (ops != NULL ? (*ops->vmgetcap)(vmi, vcpu, num, retval) : ENXIO) #define VMSETCAP(vmi, vcpu, num, val) \ (ops != NULL ? (*ops->vmsetcap)(vmi, vcpu, num, val) : ENXIO) #define VLAPIC_INIT(vmi, vcpu) \ (ops != NULL ? (*ops->vlapic_init)(vmi, vcpu) : NULL) #define VLAPIC_CLEANUP(vmi, vlapic) \ (ops != NULL ? (*ops->vlapic_cleanup)(vmi, vlapic) : NULL) #define fpu_start_emulating() load_cr0(rcr0() | CR0_TS) #define fpu_stop_emulating() clts() static MALLOC_DEFINE(M_VM, "vm", "vm"); /* statistics */ static VMM_STAT(VCPU_TOTAL_RUNTIME, "vcpu total runtime"); SYSCTL_NODE(_hw, OID_AUTO, vmm, CTLFLAG_RW, NULL, NULL); /* * Halt the guest if all vcpus are executing a HLT instruction with * interrupts disabled. */ static int halt_detection_enabled = 1; SYSCTL_INT(_hw_vmm, OID_AUTO, halt_detection, CTLFLAG_RDTUN, &halt_detection_enabled, 0, "Halt VM if all vcpus execute HLT with interrupts disabled"); static int vmm_ipinum; SYSCTL_INT(_hw_vmm, OID_AUTO, ipinum, CTLFLAG_RD, &vmm_ipinum, 0, "IPI vector used for vcpu notifications"); static int trace_guest_exceptions; SYSCTL_INT(_hw_vmm, OID_AUTO, trace_guest_exceptions, CTLFLAG_RDTUN, &trace_guest_exceptions, 0, "Trap into hypervisor on all guest exceptions and reflect them back"); static int vmm_force_iommu = 0; TUNABLE_INT("hw.vmm.force_iommu", &vmm_force_iommu); SYSCTL_INT(_hw_vmm, OID_AUTO, force_iommu, CTLFLAG_RDTUN, &vmm_force_iommu, 0, "Force use of I/O MMU even if no passthrough devices were found."); +static void vcpu_notify_event_locked(struct vcpu *vcpu, bool lapic_intr); + +#ifdef KTR +static const char * +vcpu_state2str(enum vcpu_state state) +{ + + switch (state) { + case VCPU_IDLE: + return ("idle"); + case VCPU_FROZEN: + return ("frozen"); + case VCPU_RUNNING: + return ("running"); + case VCPU_SLEEPING: + return ("sleeping"); + default: + return ("unknown"); + } +} +#endif + static void vcpu_cleanup(struct vm *vm, int i, bool destroy) { struct vcpu *vcpu = &vm->vcpu[i]; VLAPIC_CLEANUP(vm->cookie, vcpu->vlapic); if (destroy) { vmm_stat_free(vcpu->stats); fpu_save_area_free(vcpu->guestfpu); } } static void vcpu_init(struct vm *vm, int vcpu_id, bool create) { struct vcpu *vcpu; KASSERT(vcpu_id >= 0 && vcpu_id < VM_MAXCPU, ("vcpu_init: invalid vcpu %d", vcpu_id)); vcpu = &vm->vcpu[vcpu_id]; if (create) { KASSERT(!vcpu_lock_initialized(vcpu), ("vcpu %d already " "initialized", vcpu_id)); vcpu_lock_init(vcpu); vcpu->state = VCPU_IDLE; vcpu->hostcpu = NOCPU; vcpu->guestfpu = fpu_save_area_alloc(); vcpu->stats = vmm_stat_alloc(); } vcpu->vlapic = VLAPIC_INIT(vm->cookie, vcpu_id); vm_set_x2apic_state(vm, vcpu_id, X2APIC_DISABLED); + vcpu->reqidle = 0; vcpu->exitintinfo = 0; vcpu->nmi_pending = 0; vcpu->extint_pending = 0; vcpu->exception_pending = 0; vcpu->guest_xcr0 = XFEATURE_ENABLED_X87; fpu_save_area_reset(vcpu->guestfpu); vmm_stat_init(vcpu->stats); } int vcpu_trace_exceptions(struct vm *vm, int vcpuid) { return (trace_guest_exceptions); } struct vm_exit * vm_exitinfo(struct vm *vm, int cpuid) { struct vcpu *vcpu; if (cpuid < 0 || cpuid >= VM_MAXCPU) panic("vm_exitinfo: invalid cpuid %d", cpuid); vcpu = &vm->vcpu[cpuid]; return (&vcpu->exitinfo); } static void vmm_resume(void) { VMM_RESUME(); } static int vmm_init(void) { int error; vmm_host_state_init(); vmm_ipinum = lapic_ipi_alloc(&IDTVEC(justreturn)); if (vmm_ipinum < 0) vmm_ipinum = IPI_AST; error = vmm_mem_init(); if (error) return (error); if (vmm_is_intel()) ops = &vmm_ops_intel; else if (vmm_is_amd()) ops = &vmm_ops_amd; else return (ENXIO); vmm_resume_p = vmm_resume; return (VMM_INIT(vmm_ipinum)); } static int vmm_handler(module_t mod, int what, void *arg) { int error; switch (what) { case MOD_LOAD: vmmdev_init(); if (vmm_force_iommu || ppt_avail_devices() > 0) iommu_init(); error = vmm_init(); if (error == 0) vmm_initialized = 1; break; case MOD_UNLOAD: error = vmmdev_cleanup(); if (error == 0) { vmm_resume_p = NULL; iommu_cleanup(); if (vmm_ipinum != IPI_AST) lapic_ipi_free(vmm_ipinum); error = VMM_CLEANUP(); /* * Something bad happened - prevent new * VMs from being created */ if (error) vmm_initialized = 0; } break; default: error = 0; break; } return (error); } static moduledata_t vmm_kmod = { "vmm", vmm_handler, NULL }; /* * vmm initialization has the following dependencies: * * - iommu initialization must happen after the pci passthru driver has had * a chance to attach to any passthru devices (after SI_SUB_CONFIGURE). * * - VT-x initialization requires smp_rendezvous() and therefore must happen * after SMP is fully functional (after SI_SUB_SMP). */ DECLARE_MODULE(vmm, vmm_kmod, SI_SUB_SMP + 1, SI_ORDER_ANY); MODULE_VERSION(vmm, 1); static void vm_init(struct vm *vm, bool create) { int i; vm->cookie = VMINIT(vm, vmspace_pmap(vm->vmspace)); vm->iommu = NULL; vm->vioapic = vioapic_init(vm); vm->vhpet = vhpet_init(vm); vm->vatpic = vatpic_init(vm); vm->vatpit = vatpit_init(vm); vm->vpmtmr = vpmtmr_init(vm); if (create) vm->vrtc = vrtc_init(vm); CPU_ZERO(&vm->active_cpus); vm->suspend = 0; CPU_ZERO(&vm->suspended_cpus); for (i = 0; i < VM_MAXCPU; i++) vcpu_init(vm, i, create); } int vm_create(const char *name, struct vm **retvm) { struct vm *vm; struct vmspace *vmspace; /* * If vmm.ko could not be successfully initialized then don't attempt * to create the virtual machine. */ if (!vmm_initialized) return (ENXIO); if (name == NULL || strlen(name) >= VM_MAX_NAMELEN) return (EINVAL); vmspace = VMSPACE_ALLOC(0, VM_MAXUSER_ADDRESS); if (vmspace == NULL) return (ENOMEM); vm = malloc(sizeof(struct vm), M_VM, M_WAITOK | M_ZERO); strcpy(vm->name, name); vm->num_mem_segs = 0; vm->vmspace = vmspace; mtx_init(&vm->rendezvous_mtx, "vm rendezvous lock", 0, MTX_DEF); vm_init(vm, true); *retvm = vm; return (0); } static void vm_free_mem_seg(struct vm *vm, struct mem_seg *seg) { if (seg->object != NULL) vmm_mem_free(vm->vmspace, seg->gpa, seg->len); bzero(seg, sizeof(*seg)); } static void vm_cleanup(struct vm *vm, bool destroy) { int i; ppt_unassign_all(vm); if (vm->iommu != NULL) iommu_destroy_domain(vm->iommu); if (destroy) vrtc_cleanup(vm->vrtc); else vrtc_reset(vm->vrtc); vpmtmr_cleanup(vm->vpmtmr); vatpit_cleanup(vm->vatpit); vhpet_cleanup(vm->vhpet); vatpic_cleanup(vm->vatpic); vioapic_cleanup(vm->vioapic); for (i = 0; i < VM_MAXCPU; i++) vcpu_cleanup(vm, i, destroy); VMCLEANUP(vm->cookie); if (destroy) { for (i = 0; i < vm->num_mem_segs; i++) vm_free_mem_seg(vm, &vm->mem_segs[i]); vm->num_mem_segs = 0; VMSPACE_FREE(vm->vmspace); vm->vmspace = NULL; } } void vm_destroy(struct vm *vm) { vm_cleanup(vm, true); free(vm, M_VM); } int vm_reinit(struct vm *vm) { int error; /* * A virtual machine can be reset only if all vcpus are suspended. */ if (CPU_CMP(&vm->suspended_cpus, &vm->active_cpus) == 0) { vm_cleanup(vm, false); vm_init(vm, false); error = 0; } else { error = EBUSY; } return (error); } const char * vm_name(struct vm *vm) { return (vm->name); } int vm_map_mmio(struct vm *vm, vm_paddr_t gpa, size_t len, vm_paddr_t hpa) { vm_object_t obj; if ((obj = vmm_mmio_alloc(vm->vmspace, gpa, len, hpa)) == NULL) return (ENOMEM); else return (0); } int vm_unmap_mmio(struct vm *vm, vm_paddr_t gpa, size_t len) { vmm_mmio_free(vm->vmspace, gpa, len); return (0); } boolean_t vm_mem_allocated(struct vm *vm, vm_paddr_t gpa) { int i; vm_paddr_t gpabase, gpalimit; for (i = 0; i < vm->num_mem_segs; i++) { gpabase = vm->mem_segs[i].gpa; gpalimit = gpabase + vm->mem_segs[i].len; if (gpa >= gpabase && gpa < gpalimit) return (TRUE); /* 'gpa' is regular memory */ } if (ppt_is_mmio(vm, gpa)) return (TRUE); /* 'gpa' is pci passthru mmio */ return (FALSE); } int vm_malloc(struct vm *vm, vm_paddr_t gpa, size_t len) { int available, allocated; struct mem_seg *seg; vm_object_t object; vm_paddr_t g; if ((gpa & PAGE_MASK) || (len & PAGE_MASK) || len == 0) return (EINVAL); available = allocated = 0; g = gpa; while (g < gpa + len) { if (vm_mem_allocated(vm, g)) allocated++; else available++; g += PAGE_SIZE; } /* * If there are some allocated and some available pages in the address * range then it is an error. */ if (allocated && available) return (EINVAL); /* * If the entire address range being requested has already been * allocated then there isn't anything more to do. */ if (allocated && available == 0) return (0); if (vm->num_mem_segs >= VM_MAX_MEMORY_SEGMENTS) return (E2BIG); seg = &vm->mem_segs[vm->num_mem_segs]; if ((object = vmm_mem_alloc(vm->vmspace, gpa, len)) == NULL) return (ENOMEM); seg->gpa = gpa; seg->len = len; seg->object = object; seg->wired = FALSE; vm->num_mem_segs++; return (0); } static vm_paddr_t vm_maxmem(struct vm *vm) { int i; vm_paddr_t gpa, maxmem; maxmem = 0; for (i = 0; i < vm->num_mem_segs; i++) { gpa = vm->mem_segs[i].gpa + vm->mem_segs[i].len; if (gpa > maxmem) maxmem = gpa; } return (maxmem); } static void vm_gpa_unwire(struct vm *vm) { int i, rv; struct mem_seg *seg; for (i = 0; i < vm->num_mem_segs; i++) { seg = &vm->mem_segs[i]; if (!seg->wired) continue; rv = vm_map_unwire(&vm->vmspace->vm_map, seg->gpa, seg->gpa + seg->len, VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); KASSERT(rv == KERN_SUCCESS, ("vm(%s) memory segment " "%#lx/%ld could not be unwired: %d", vm_name(vm), seg->gpa, seg->len, rv)); seg->wired = FALSE; } } static int vm_gpa_wire(struct vm *vm) { int i, rv; struct mem_seg *seg; for (i = 0; i < vm->num_mem_segs; i++) { seg = &vm->mem_segs[i]; if (seg->wired) continue; /* XXX rlimits? */ rv = vm_map_wire(&vm->vmspace->vm_map, seg->gpa, seg->gpa + seg->len, VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); if (rv != KERN_SUCCESS) break; seg->wired = TRUE; } if (i < vm->num_mem_segs) { /* * Undo the wiring before returning an error. */ vm_gpa_unwire(vm); return (EAGAIN); } return (0); } static void vm_iommu_modify(struct vm *vm, boolean_t map) { int i, sz; vm_paddr_t gpa, hpa; struct mem_seg *seg; void *vp, *cookie, *host_domain; sz = PAGE_SIZE; host_domain = iommu_host_domain(); for (i = 0; i < vm->num_mem_segs; i++) { seg = &vm->mem_segs[i]; KASSERT(seg->wired, ("vm(%s) memory segment %#lx/%ld not wired", vm_name(vm), seg->gpa, seg->len)); gpa = seg->gpa; while (gpa < seg->gpa + seg->len) { vp = vm_gpa_hold(vm, gpa, PAGE_SIZE, VM_PROT_WRITE, &cookie); KASSERT(vp != NULL, ("vm(%s) could not map gpa %#lx", vm_name(vm), gpa)); vm_gpa_release(cookie); hpa = DMAP_TO_PHYS((uintptr_t)vp); if (map) { iommu_create_mapping(vm->iommu, gpa, hpa, sz); iommu_remove_mapping(host_domain, hpa, sz); } else { iommu_remove_mapping(vm->iommu, gpa, sz); iommu_create_mapping(host_domain, hpa, hpa, sz); } gpa += PAGE_SIZE; } } /* * Invalidate the cached translations associated with the domain * from which pages were removed. */ if (map) iommu_invalidate_tlb(host_domain); else iommu_invalidate_tlb(vm->iommu); } #define vm_iommu_unmap(vm) vm_iommu_modify((vm), FALSE) #define vm_iommu_map(vm) vm_iommu_modify((vm), TRUE) int vm_unassign_pptdev(struct vm *vm, int bus, int slot, int func) { int error; error = ppt_unassign_device(vm, bus, slot, func); if (error) return (error); if (ppt_assigned_devices(vm) == 0) { vm_iommu_unmap(vm); vm_gpa_unwire(vm); } return (0); } int vm_assign_pptdev(struct vm *vm, int bus, int slot, int func) { int error; vm_paddr_t maxaddr; /* * Virtual machines with pci passthru devices get special treatment: * - the guest physical memory is wired * - the iommu is programmed to do the 'gpa' to 'hpa' translation * * We need to do this before the first pci passthru device is attached. */ if (ppt_assigned_devices(vm) == 0) { KASSERT(vm->iommu == NULL, ("vm_assign_pptdev: iommu must be NULL")); maxaddr = vm_maxmem(vm); vm->iommu = iommu_create_domain(maxaddr); error = vm_gpa_wire(vm); if (error) return (error); vm_iommu_map(vm); } error = ppt_assign_device(vm, bus, slot, func); return (error); } void * vm_gpa_hold(struct vm *vm, vm_paddr_t gpa, size_t len, int reqprot, void **cookie) { int count, pageoff; vm_page_t m; pageoff = gpa & PAGE_MASK; if (len > PAGE_SIZE - pageoff) panic("vm_gpa_hold: invalid gpa/len: 0x%016lx/%lu", gpa, len); count = vm_fault_quick_hold_pages(&vm->vmspace->vm_map, trunc_page(gpa), PAGE_SIZE, reqprot, &m, 1); if (count == 1) { *cookie = m; return ((void *)(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)) + pageoff)); } else { *cookie = NULL; return (NULL); } } void vm_gpa_release(void *cookie) { vm_page_t m = cookie; vm_page_lock(m); vm_page_unhold(m); vm_page_unlock(m); } int vm_gpabase2memseg(struct vm *vm, vm_paddr_t gpabase, struct vm_memory_segment *seg) { int i; for (i = 0; i < vm->num_mem_segs; i++) { if (gpabase == vm->mem_segs[i].gpa) { seg->gpa = vm->mem_segs[i].gpa; seg->len = vm->mem_segs[i].len; seg->wired = vm->mem_segs[i].wired; return (0); } } return (-1); } int vm_get_memobj(struct vm *vm, vm_paddr_t gpa, size_t len, vm_offset_t *offset, struct vm_object **object) { int i; size_t seg_len; vm_paddr_t seg_gpa; vm_object_t seg_obj; for (i = 0; i < vm->num_mem_segs; i++) { if ((seg_obj = vm->mem_segs[i].object) == NULL) continue; seg_gpa = vm->mem_segs[i].gpa; seg_len = vm->mem_segs[i].len; if (gpa >= seg_gpa && gpa < seg_gpa + seg_len) { *offset = gpa - seg_gpa; *object = seg_obj; vm_object_reference(seg_obj); return (0); } } return (EINVAL); } int vm_get_register(struct vm *vm, int vcpu, int reg, uint64_t *retval) { if (vcpu < 0 || vcpu >= VM_MAXCPU) return (EINVAL); if (reg >= VM_REG_LAST) return (EINVAL); return (VMGETREG(vm->cookie, vcpu, reg, retval)); } int vm_set_register(struct vm *vm, int vcpuid, int reg, uint64_t val) { struct vcpu *vcpu; int error; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); if (reg >= VM_REG_LAST) return (EINVAL); error = VMSETREG(vm->cookie, vcpuid, reg, val); if (error || reg != VM_REG_GUEST_RIP) return (error); /* Set 'nextrip' to match the value of %rip */ VCPU_CTR1(vm, vcpuid, "Setting nextrip to %#lx", val); vcpu = &vm->vcpu[vcpuid]; vcpu->nextrip = val; return (0); } static boolean_t is_descriptor_table(int reg) { switch (reg) { case VM_REG_GUEST_IDTR: case VM_REG_GUEST_GDTR: return (TRUE); default: return (FALSE); } } static boolean_t is_segment_register(int reg) { switch (reg) { case VM_REG_GUEST_ES: case VM_REG_GUEST_CS: case VM_REG_GUEST_SS: case VM_REG_GUEST_DS: case VM_REG_GUEST_FS: case VM_REG_GUEST_GS: case VM_REG_GUEST_TR: case VM_REG_GUEST_LDTR: return (TRUE); default: return (FALSE); } } int vm_get_seg_desc(struct vm *vm, int vcpu, int reg, struct seg_desc *desc) { if (vcpu < 0 || vcpu >= VM_MAXCPU) return (EINVAL); if (!is_segment_register(reg) && !is_descriptor_table(reg)) return (EINVAL); return (VMGETDESC(vm->cookie, vcpu, reg, desc)); } int vm_set_seg_desc(struct vm *vm, int vcpu, int reg, struct seg_desc *desc) { if (vcpu < 0 || vcpu >= VM_MAXCPU) return (EINVAL); if (!is_segment_register(reg) && !is_descriptor_table(reg)) return (EINVAL); return (VMSETDESC(vm->cookie, vcpu, reg, desc)); } static void restore_guest_fpustate(struct vcpu *vcpu) { /* flush host state to the pcb */ fpuexit(curthread); /* restore guest FPU state */ fpu_stop_emulating(); fpurestore(vcpu->guestfpu); /* restore guest XCR0 if XSAVE is enabled in the host */ if (rcr4() & CR4_XSAVE) load_xcr(0, vcpu->guest_xcr0); /* * The FPU is now "dirty" with the guest's state so turn on emulation * to trap any access to the FPU by the host. */ fpu_start_emulating(); } static void save_guest_fpustate(struct vcpu *vcpu) { if ((rcr0() & CR0_TS) == 0) panic("fpu emulation not enabled in host!"); /* save guest XCR0 and restore host XCR0 */ if (rcr4() & CR4_XSAVE) { vcpu->guest_xcr0 = rxcr(0); load_xcr(0, vmm_get_host_xcr0()); } /* save guest FPU state */ fpu_stop_emulating(); fpusave(vcpu->guestfpu); fpu_start_emulating(); } static VMM_STAT(VCPU_IDLE_TICKS, "number of ticks vcpu was idle"); static int -vcpu_set_state_locked(struct vcpu *vcpu, enum vcpu_state newstate, +vcpu_set_state_locked(struct vm *vm, int vcpuid, enum vcpu_state newstate, bool from_idle) { + struct vcpu *vcpu; int error; + vcpu = &vm->vcpu[vcpuid]; vcpu_assert_locked(vcpu); /* * State transitions from the vmmdev_ioctl() must always begin from * the VCPU_IDLE state. This guarantees that there is only a single * ioctl() operating on a vcpu at any point. */ if (from_idle) { - while (vcpu->state != VCPU_IDLE) + while (vcpu->state != VCPU_IDLE) { + vcpu->reqidle = 1; + vcpu_notify_event_locked(vcpu, false); + VCPU_CTR1(vm, vcpuid, "vcpu state change from %s to " + "idle requested", vcpu_state2str(vcpu->state)); msleep_spin(&vcpu->state, &vcpu->mtx, "vmstat", hz); + } } else { KASSERT(vcpu->state != VCPU_IDLE, ("invalid transition from " "vcpu idle state")); } if (vcpu->state == VCPU_RUNNING) { KASSERT(vcpu->hostcpu == curcpu, ("curcpu %d and hostcpu %d " "mismatch for running vcpu", curcpu, vcpu->hostcpu)); } else { KASSERT(vcpu->hostcpu == NOCPU, ("Invalid hostcpu %d for a " "vcpu that is not running", vcpu->hostcpu)); } /* * The following state transitions are allowed: * IDLE -> FROZEN -> IDLE * FROZEN -> RUNNING -> FROZEN * FROZEN -> SLEEPING -> FROZEN */ switch (vcpu->state) { case VCPU_IDLE: case VCPU_RUNNING: case VCPU_SLEEPING: error = (newstate != VCPU_FROZEN); break; case VCPU_FROZEN: error = (newstate == VCPU_FROZEN); break; default: error = 1; break; } if (error) return (EBUSY); + VCPU_CTR2(vm, vcpuid, "vcpu state changed from %s to %s", + vcpu_state2str(vcpu->state), vcpu_state2str(newstate)); + vcpu->state = newstate; if (newstate == VCPU_RUNNING) vcpu->hostcpu = curcpu; else vcpu->hostcpu = NOCPU; if (newstate == VCPU_IDLE) wakeup(&vcpu->state); return (0); } static void vcpu_require_state(struct vm *vm, int vcpuid, enum vcpu_state newstate) { int error; if ((error = vcpu_set_state(vm, vcpuid, newstate, false)) != 0) panic("Error %d setting state to %d\n", error, newstate); } static void -vcpu_require_state_locked(struct vcpu *vcpu, enum vcpu_state newstate) +vcpu_require_state_locked(struct vm *vm, int vcpuid, enum vcpu_state newstate) { int error; - if ((error = vcpu_set_state_locked(vcpu, newstate, false)) != 0) + if ((error = vcpu_set_state_locked(vm, vcpuid, newstate, false)) != 0) panic("Error %d setting state to %d", error, newstate); } static void vm_set_rendezvous_func(struct vm *vm, vm_rendezvous_func_t func) { KASSERT(mtx_owned(&vm->rendezvous_mtx), ("rendezvous_mtx not locked")); /* * Update 'rendezvous_func' and execute a write memory barrier to * ensure that it is visible across all host cpus. This is not needed * for correctness but it does ensure that all the vcpus will notice * that the rendezvous is requested immediately. */ vm->rendezvous_func = func; wmb(); } #define RENDEZVOUS_CTR0(vm, vcpuid, fmt) \ do { \ if (vcpuid >= 0) \ VCPU_CTR0(vm, vcpuid, fmt); \ else \ VM_CTR0(vm, fmt); \ } while (0) static void vm_handle_rendezvous(struct vm *vm, int vcpuid) { KASSERT(vcpuid == -1 || (vcpuid >= 0 && vcpuid < VM_MAXCPU), ("vm_handle_rendezvous: invalid vcpuid %d", vcpuid)); mtx_lock(&vm->rendezvous_mtx); while (vm->rendezvous_func != NULL) { /* 'rendezvous_req_cpus' must be a subset of 'active_cpus' */ CPU_AND(&vm->rendezvous_req_cpus, &vm->active_cpus); if (vcpuid != -1 && CPU_ISSET(vcpuid, &vm->rendezvous_req_cpus) && !CPU_ISSET(vcpuid, &vm->rendezvous_done_cpus)) { VCPU_CTR0(vm, vcpuid, "Calling rendezvous func"); (*vm->rendezvous_func)(vm, vcpuid, vm->rendezvous_arg); CPU_SET(vcpuid, &vm->rendezvous_done_cpus); } if (CPU_CMP(&vm->rendezvous_req_cpus, &vm->rendezvous_done_cpus) == 0) { VCPU_CTR0(vm, vcpuid, "Rendezvous completed"); vm_set_rendezvous_func(vm, NULL); wakeup(&vm->rendezvous_func); break; } RENDEZVOUS_CTR0(vm, vcpuid, "Wait for rendezvous completion"); mtx_sleep(&vm->rendezvous_func, &vm->rendezvous_mtx, 0, "vmrndv", 0); } mtx_unlock(&vm->rendezvous_mtx); } /* * Emulate a guest 'hlt' by sleeping until the vcpu is ready to run. */ static int vm_handle_hlt(struct vm *vm, int vcpuid, bool intr_disabled, bool *retu) { struct vcpu *vcpu; const char *wmesg; int t, vcpu_halted, vm_halted; KASSERT(!CPU_ISSET(vcpuid, &vm->halted_cpus), ("vcpu already halted")); vcpu = &vm->vcpu[vcpuid]; vcpu_halted = 0; vm_halted = 0; vcpu_lock(vcpu); while (1) { /* * Do a final check for pending NMI or interrupts before * really putting this thread to sleep. Also check for * software events that would cause this vcpu to wakeup. * * These interrupts/events could have happened after the * vcpu returned from VMRUN() and before it acquired the * vcpu lock above. */ - if (vm->rendezvous_func != NULL || vm->suspend) + if (vm->rendezvous_func != NULL || vm->suspend || vcpu->reqidle) break; if (vm_nmi_pending(vm, vcpuid)) break; if (!intr_disabled) { if (vm_extint_pending(vm, vcpuid) || vlapic_pending_intr(vcpu->vlapic, NULL)) { break; } } /* Don't go to sleep if the vcpu thread needs to yield */ if (vcpu_should_yield(vm, vcpuid)) break; /* * Some Linux guests implement "halt" by having all vcpus * execute HLT with interrupts disabled. 'halted_cpus' keeps * track of the vcpus that have entered this state. When all * vcpus enter the halted state the virtual machine is halted. */ if (intr_disabled) { wmesg = "vmhalt"; VCPU_CTR0(vm, vcpuid, "Halted"); if (!vcpu_halted && halt_detection_enabled) { vcpu_halted = 1; CPU_SET_ATOMIC(vcpuid, &vm->halted_cpus); } if (CPU_CMP(&vm->halted_cpus, &vm->active_cpus) == 0) { vm_halted = 1; break; } } else { wmesg = "vmidle"; } t = ticks; - vcpu_require_state_locked(vcpu, VCPU_SLEEPING); + vcpu_require_state_locked(vm, vcpuid, VCPU_SLEEPING); /* * XXX msleep_spin() cannot be interrupted by signals so * wake up periodically to check pending signals. */ msleep_spin(vcpu, &vcpu->mtx, wmesg, hz); - vcpu_require_state_locked(vcpu, VCPU_FROZEN); + vcpu_require_state_locked(vm, vcpuid, VCPU_FROZEN); vmm_stat_incr(vm, vcpuid, VCPU_IDLE_TICKS, ticks - t); } if (vcpu_halted) CPU_CLR_ATOMIC(vcpuid, &vm->halted_cpus); vcpu_unlock(vcpu); if (vm_halted) vm_suspend(vm, VM_SUSPEND_HALT); return (0); } static int vm_handle_paging(struct vm *vm, int vcpuid, bool *retu) { int rv, ftype; struct vm_map *map; struct vcpu *vcpu; struct vm_exit *vme; vcpu = &vm->vcpu[vcpuid]; vme = &vcpu->exitinfo; KASSERT(vme->inst_length == 0, ("%s: invalid inst_length %d", __func__, vme->inst_length)); ftype = vme->u.paging.fault_type; KASSERT(ftype == VM_PROT_READ || ftype == VM_PROT_WRITE || ftype == VM_PROT_EXECUTE, ("vm_handle_paging: invalid fault_type %d", ftype)); if (ftype == VM_PROT_READ || ftype == VM_PROT_WRITE) { rv = pmap_emulate_accessed_dirty(vmspace_pmap(vm->vmspace), vme->u.paging.gpa, ftype); if (rv == 0) { VCPU_CTR2(vm, vcpuid, "%s bit emulation for gpa %#lx", ftype == VM_PROT_READ ? "accessed" : "dirty", vme->u.paging.gpa); goto done; } } map = &vm->vmspace->vm_map; rv = vm_fault(map, vme->u.paging.gpa, ftype, VM_FAULT_NORMAL); VCPU_CTR3(vm, vcpuid, "vm_handle_paging rv = %d, gpa = %#lx, " "ftype = %d", rv, vme->u.paging.gpa, ftype); if (rv != KERN_SUCCESS) return (EFAULT); done: return (0); } static int vm_handle_inst_emul(struct vm *vm, int vcpuid, bool *retu) { struct vie *vie; struct vcpu *vcpu; struct vm_exit *vme; uint64_t gla, gpa, cs_base; struct vm_guest_paging *paging; mem_region_read_t mread; mem_region_write_t mwrite; enum vm_cpu_mode cpu_mode; int cs_d, error, fault; vcpu = &vm->vcpu[vcpuid]; vme = &vcpu->exitinfo; KASSERT(vme->inst_length == 0, ("%s: invalid inst_length %d", __func__, vme->inst_length)); gla = vme->u.inst_emul.gla; gpa = vme->u.inst_emul.gpa; cs_base = vme->u.inst_emul.cs_base; cs_d = vme->u.inst_emul.cs_d; vie = &vme->u.inst_emul.vie; paging = &vme->u.inst_emul.paging; cpu_mode = paging->cpu_mode; VCPU_CTR1(vm, vcpuid, "inst_emul fault accessing gpa %#lx", gpa); /* Fetch, decode and emulate the faulting instruction */ if (vie->num_valid == 0) { error = vmm_fetch_instruction(vm, vcpuid, paging, vme->rip + cs_base, VIE_INST_SIZE, vie, &fault); } else { /* * The instruction bytes have already been copied into 'vie' */ error = fault = 0; } if (error || fault) return (error); if (vmm_decode_instruction(vm, vcpuid, gla, cpu_mode, cs_d, vie) != 0) { VCPU_CTR1(vm, vcpuid, "Error decoding instruction at %#lx", vme->rip + cs_base); *retu = true; /* dump instruction bytes in userspace */ return (0); } /* * Update 'nextrip' based on the length of the emulated instruction. */ vme->inst_length = vie->num_processed; vcpu->nextrip += vie->num_processed; VCPU_CTR1(vm, vcpuid, "nextrip updated to %#lx after instruction " "decoding", vcpu->nextrip); /* 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; mwrite = lapic_mmio_write; } else if (gpa >= VIOAPIC_BASE && gpa < VIOAPIC_BASE + VIOAPIC_SIZE) { mread = vioapic_mmio_read; mwrite = vioapic_mmio_write; } else if (gpa >= VHPET_BASE && gpa < VHPET_BASE + VHPET_SIZE) { mread = vhpet_mmio_read; mwrite = vhpet_mmio_write; } else { *retu = true; return (0); } error = vmm_emulate_instruction(vm, vcpuid, gpa, vie, paging, mread, mwrite, retu); return (error); } static int vm_handle_suspend(struct vm *vm, int vcpuid, bool *retu) { int i, done; struct vcpu *vcpu; done = 0; vcpu = &vm->vcpu[vcpuid]; CPU_SET_ATOMIC(vcpuid, &vm->suspended_cpus); /* * Wait until all 'active_cpus' have suspended themselves. * * Since a VM may be suspended at any time including when one or * more vcpus are doing a rendezvous we need to call the rendezvous * handler while we are waiting to prevent a deadlock. */ vcpu_lock(vcpu); while (1) { if (CPU_CMP(&vm->suspended_cpus, &vm->active_cpus) == 0) { VCPU_CTR0(vm, vcpuid, "All vcpus suspended"); break; } if (vm->rendezvous_func == NULL) { VCPU_CTR0(vm, vcpuid, "Sleeping during suspend"); - vcpu_require_state_locked(vcpu, VCPU_SLEEPING); + vcpu_require_state_locked(vm, vcpuid, VCPU_SLEEPING); msleep_spin(vcpu, &vcpu->mtx, "vmsusp", hz); - vcpu_require_state_locked(vcpu, VCPU_FROZEN); + vcpu_require_state_locked(vm, vcpuid, VCPU_FROZEN); } else { VCPU_CTR0(vm, vcpuid, "Rendezvous during suspend"); vcpu_unlock(vcpu); vm_handle_rendezvous(vm, vcpuid); vcpu_lock(vcpu); } } vcpu_unlock(vcpu); /* * Wakeup the other sleeping vcpus and return to userspace. */ for (i = 0; i < VM_MAXCPU; i++) { if (CPU_ISSET(i, &vm->suspended_cpus)) { vcpu_notify_event(vm, i, false); } } *retu = true; return (0); } +static int +vm_handle_reqidle(struct vm *vm, int vcpuid, bool *retu) +{ + struct vcpu *vcpu = &vm->vcpu[vcpuid]; + + vcpu_lock(vcpu); + KASSERT(vcpu->reqidle, ("invalid vcpu reqidle %d", vcpu->reqidle)); + vcpu->reqidle = 0; + vcpu_unlock(vcpu); + *retu = true; + return (0); +} + int vm_suspend(struct vm *vm, enum vm_suspend_how how) { int i; if (how <= VM_SUSPEND_NONE || how >= VM_SUSPEND_LAST) return (EINVAL); if (atomic_cmpset_int(&vm->suspend, 0, how) == 0) { VM_CTR2(vm, "virtual machine already suspended %d/%d", vm->suspend, how); return (EALREADY); } VM_CTR1(vm, "virtual machine successfully suspended %d", how); /* * Notify all active vcpus that they are now suspended. */ for (i = 0; i < VM_MAXCPU; i++) { if (CPU_ISSET(i, &vm->active_cpus)) vcpu_notify_event(vm, i, false); } return (0); } void vm_exit_suspended(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; KASSERT(vm->suspend > VM_SUSPEND_NONE && vm->suspend < VM_SUSPEND_LAST, ("vm_exit_suspended: invalid suspend type %d", vm->suspend)); vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_SUSPENDED; vmexit->u.suspended.how = vm->suspend; } void vm_exit_rendezvous(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; KASSERT(vm->rendezvous_func != NULL, ("rendezvous not in progress")); vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_RENDEZVOUS; vmm_stat_incr(vm, vcpuid, VMEXIT_RENDEZVOUS, 1); } void +vm_exit_reqidle(struct vm *vm, int vcpuid, uint64_t rip) +{ + struct vm_exit *vmexit; + + vmexit = vm_exitinfo(vm, vcpuid); + vmexit->rip = rip; + vmexit->inst_length = 0; + vmexit->exitcode = VM_EXITCODE_REQIDLE; + vmm_stat_incr(vm, vcpuid, VMEXIT_REQIDLE, 1); +} + +void vm_exit_astpending(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_BOGUS; vmm_stat_incr(vm, vcpuid, VMEXIT_ASTPENDING, 1); } int vm_run(struct vm *vm, struct vm_run *vmrun) { + struct vm_eventinfo evinfo; int error, vcpuid; struct vcpu *vcpu; struct pcb *pcb; uint64_t tscval; struct vm_exit *vme; bool retu, intr_disabled; pmap_t pmap; - void *rptr, *sptr; vcpuid = vmrun->cpuid; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); if (!CPU_ISSET(vcpuid, &vm->active_cpus)) return (EINVAL); if (CPU_ISSET(vcpuid, &vm->suspended_cpus)) return (EINVAL); - rptr = &vm->rendezvous_func; - sptr = &vm->suspend; pmap = vmspace_pmap(vm->vmspace); vcpu = &vm->vcpu[vcpuid]; vme = &vcpu->exitinfo; + evinfo.rptr = &vm->rendezvous_func; + evinfo.sptr = &vm->suspend; + evinfo.iptr = &vcpu->reqidle; restart: critical_enter(); KASSERT(!CPU_ISSET(curcpu, &pmap->pm_active), ("vm_run: absurd pm_active")); tscval = rdtsc(); pcb = PCPU_GET(curpcb); set_pcb_flags(pcb, PCB_FULL_IRET); restore_guest_fpustate(vcpu); vcpu_require_state(vm, vcpuid, VCPU_RUNNING); - error = VMRUN(vm->cookie, vcpuid, vcpu->nextrip, pmap, rptr, sptr); + error = VMRUN(vm->cookie, vcpuid, vcpu->nextrip, pmap, &evinfo); vcpu_require_state(vm, vcpuid, VCPU_FROZEN); save_guest_fpustate(vcpu); vmm_stat_incr(vm, vcpuid, VCPU_TOTAL_RUNTIME, rdtsc() - tscval); critical_exit(); if (error == 0) { retu = false; vcpu->nextrip = vme->rip + vme->inst_length; switch (vme->exitcode) { + case VM_EXITCODE_REQIDLE: + error = vm_handle_reqidle(vm, vcpuid, &retu); + break; case VM_EXITCODE_SUSPENDED: error = vm_handle_suspend(vm, vcpuid, &retu); break; case VM_EXITCODE_IOAPIC_EOI: vioapic_process_eoi(vm, vcpuid, vme->u.ioapic_eoi.vector); break; case VM_EXITCODE_RENDEZVOUS: vm_handle_rendezvous(vm, vcpuid); error = 0; break; case VM_EXITCODE_HLT: intr_disabled = ((vme->u.hlt.rflags & PSL_I) == 0); error = vm_handle_hlt(vm, vcpuid, intr_disabled, &retu); break; case VM_EXITCODE_PAGING: error = vm_handle_paging(vm, vcpuid, &retu); break; case VM_EXITCODE_INST_EMUL: error = vm_handle_inst_emul(vm, vcpuid, &retu); break; case VM_EXITCODE_INOUT: case VM_EXITCODE_INOUT_STR: error = vm_handle_inout(vm, vcpuid, vme, &retu); break; case VM_EXITCODE_MONITOR: case VM_EXITCODE_MWAIT: vm_inject_ud(vm, vcpuid); break; default: retu = true; /* handled in userland */ break; } } if (error == 0 && retu == false) goto restart; + VCPU_CTR2(vm, vcpuid, "retu %d/%d", error, vme->exitcode); + /* copy the exit information */ bcopy(vme, &vmrun->vm_exit, sizeof(struct vm_exit)); return (error); } int vm_restart_instruction(void *arg, int vcpuid) { struct vm *vm; struct vcpu *vcpu; enum vcpu_state state; uint64_t rip; int error; vm = arg; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; state = vcpu_get_state(vm, vcpuid, NULL); if (state == VCPU_RUNNING) { /* * When a vcpu is "running" the next instruction is determined * by adding 'rip' and 'inst_length' in the vcpu's 'exitinfo'. * Thus setting 'inst_length' to zero will cause the current * instruction to be restarted. */ vcpu->exitinfo.inst_length = 0; VCPU_CTR1(vm, vcpuid, "restarting instruction at %#lx by " "setting inst_length to zero", vcpu->exitinfo.rip); } else if (state == VCPU_FROZEN) { /* * When a vcpu is "frozen" it is outside the critical section * around VMRUN() and 'nextrip' points to the next instruction. * Thus instruction restart is achieved by setting 'nextrip' * to the vcpu's %rip. */ error = vm_get_register(vm, vcpuid, VM_REG_GUEST_RIP, &rip); KASSERT(!error, ("%s: error %d getting rip", __func__, error)); VCPU_CTR2(vm, vcpuid, "restarting instruction by updating " "nextrip from %#lx to %#lx", vcpu->nextrip, rip); vcpu->nextrip = rip; } else { panic("%s: invalid state %d", __func__, state); } return (0); } int vm_exit_intinfo(struct vm *vm, int vcpuid, uint64_t info) { struct vcpu *vcpu; int type, vector; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; if (info & VM_INTINFO_VALID) { type = info & VM_INTINFO_TYPE; vector = info & 0xff; if (type == VM_INTINFO_NMI && vector != IDT_NMI) return (EINVAL); if (type == VM_INTINFO_HWEXCEPTION && vector >= 32) return (EINVAL); if (info & VM_INTINFO_RSVD) return (EINVAL); } else { info = 0; } VCPU_CTR2(vm, vcpuid, "%s: info1(%#lx)", __func__, info); vcpu->exitintinfo = info; return (0); } enum exc_class { EXC_BENIGN, EXC_CONTRIBUTORY, EXC_PAGEFAULT }; #define IDT_VE 20 /* Virtualization Exception (Intel specific) */ static enum exc_class exception_class(uint64_t info) { int type, vector; KASSERT(info & VM_INTINFO_VALID, ("intinfo must be valid: %#lx", info)); type = info & VM_INTINFO_TYPE; vector = info & 0xff; /* Table 6-4, "Interrupt and Exception Classes", Intel SDM, Vol 3 */ switch (type) { case VM_INTINFO_HWINTR: case VM_INTINFO_SWINTR: case VM_INTINFO_NMI: return (EXC_BENIGN); default: /* * Hardware exception. * * SVM and VT-x use identical type values to represent NMI, * hardware interrupt and software interrupt. * * SVM uses type '3' for all exceptions. VT-x uses type '3' * for exceptions except #BP and #OF. #BP and #OF use a type * value of '5' or '6'. Therefore we don't check for explicit * values of 'type' to classify 'intinfo' into a hardware * exception. */ break; } switch (vector) { case IDT_PF: case IDT_VE: return (EXC_PAGEFAULT); case IDT_DE: case IDT_TS: case IDT_NP: case IDT_SS: case IDT_GP: return (EXC_CONTRIBUTORY); default: return (EXC_BENIGN); } } static int nested_fault(struct vm *vm, int vcpuid, uint64_t info1, uint64_t info2, uint64_t *retinfo) { enum exc_class exc1, exc2; int type1, vector1; KASSERT(info1 & VM_INTINFO_VALID, ("info1 %#lx is not valid", info1)); KASSERT(info2 & VM_INTINFO_VALID, ("info2 %#lx is not valid", info2)); /* * If an exception occurs while attempting to call the double-fault * handler the processor enters shutdown mode (aka triple fault). */ type1 = info1 & VM_INTINFO_TYPE; vector1 = info1 & 0xff; if (type1 == VM_INTINFO_HWEXCEPTION && vector1 == IDT_DF) { VCPU_CTR2(vm, vcpuid, "triple fault: info1(%#lx), info2(%#lx)", info1, info2); vm_suspend(vm, VM_SUSPEND_TRIPLEFAULT); *retinfo = 0; return (0); } /* * Table 6-5 "Conditions for Generating a Double Fault", Intel SDM, Vol3 */ exc1 = exception_class(info1); exc2 = exception_class(info2); if ((exc1 == EXC_CONTRIBUTORY && exc2 == EXC_CONTRIBUTORY) || (exc1 == EXC_PAGEFAULT && exc2 != EXC_BENIGN)) { /* Convert nested fault into a double fault. */ *retinfo = IDT_DF; *retinfo |= VM_INTINFO_VALID | VM_INTINFO_HWEXCEPTION; *retinfo |= VM_INTINFO_DEL_ERRCODE; } else { /* Handle exceptions serially */ *retinfo = info2; } return (1); } static uint64_t vcpu_exception_intinfo(struct vcpu *vcpu) { uint64_t info = 0; if (vcpu->exception_pending) { info = vcpu->exc_vector & 0xff; info |= VM_INTINFO_VALID | VM_INTINFO_HWEXCEPTION; if (vcpu->exc_errcode_valid) { info |= VM_INTINFO_DEL_ERRCODE; info |= (uint64_t)vcpu->exc_errcode << 32; } } return (info); } int vm_entry_intinfo(struct vm *vm, int vcpuid, uint64_t *retinfo) { struct vcpu *vcpu; uint64_t info1, info2; int valid; KASSERT(vcpuid >= 0 && vcpuid < VM_MAXCPU, ("invalid vcpu %d", vcpuid)); vcpu = &vm->vcpu[vcpuid]; info1 = vcpu->exitintinfo; vcpu->exitintinfo = 0; info2 = 0; if (vcpu->exception_pending) { info2 = vcpu_exception_intinfo(vcpu); vcpu->exception_pending = 0; VCPU_CTR2(vm, vcpuid, "Exception %d delivered: %#lx", vcpu->exc_vector, info2); } if ((info1 & VM_INTINFO_VALID) && (info2 & VM_INTINFO_VALID)) { valid = nested_fault(vm, vcpuid, info1, info2, retinfo); } else if (info1 & VM_INTINFO_VALID) { *retinfo = info1; valid = 1; } else if (info2 & VM_INTINFO_VALID) { *retinfo = info2; valid = 1; } else { valid = 0; } if (valid) { VCPU_CTR4(vm, vcpuid, "%s: info1(%#lx), info2(%#lx), " "retinfo(%#lx)", __func__, info1, info2, *retinfo); } return (valid); } int vm_get_intinfo(struct vm *vm, int vcpuid, uint64_t *info1, uint64_t *info2) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; *info1 = vcpu->exitintinfo; *info2 = vcpu_exception_intinfo(vcpu); return (0); } int vm_inject_exception(struct vm *vm, int vcpuid, int vector, int errcode_valid, uint32_t errcode, int restart_instruction) { struct vcpu *vcpu; uint64_t regval; int error; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); if (vector < 0 || vector >= 32) return (EINVAL); /* * A double fault exception should never be injected directly into * the guest. It is a derived exception that results from specific * combinations of nested faults. */ if (vector == IDT_DF) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; if (vcpu->exception_pending) { VCPU_CTR2(vm, vcpuid, "Unable to inject exception %d due to " "pending exception %d", vector, vcpu->exc_vector); return (EBUSY); } if (errcode_valid) { /* * Exceptions don't deliver an error code in real mode. */ error = vm_get_register(vm, vcpuid, VM_REG_GUEST_CR0, ®val); KASSERT(!error, ("%s: error %d getting CR0", __func__, error)); if (!(regval & CR0_PE)) errcode_valid = 0; } /* * From section 26.6.1 "Interruptibility State" in Intel SDM: * * Event blocking by "STI" or "MOV SS" is cleared after guest executes * one instruction or incurs an exception. */ error = vm_set_register(vm, vcpuid, VM_REG_GUEST_INTR_SHADOW, 0); KASSERT(error == 0, ("%s: error %d clearing interrupt shadow", __func__, error)); if (restart_instruction) vm_restart_instruction(vm, vcpuid); vcpu->exception_pending = 1; vcpu->exc_vector = vector; vcpu->exc_errcode = errcode; vcpu->exc_errcode_valid = errcode_valid; VCPU_CTR1(vm, vcpuid, "Exception %d pending", vector); return (0); } void vm_inject_fault(void *vmarg, int vcpuid, int vector, int errcode_valid, int errcode) { struct vm *vm; int error, restart_instruction; vm = vmarg; restart_instruction = 1; error = vm_inject_exception(vm, vcpuid, vector, errcode_valid, errcode, restart_instruction); KASSERT(error == 0, ("vm_inject_exception error %d", error)); } void vm_inject_pf(void *vmarg, int vcpuid, int error_code, uint64_t cr2) { struct vm *vm; int error; vm = vmarg; VCPU_CTR2(vm, vcpuid, "Injecting page fault: error_code %#x, cr2 %#lx", error_code, cr2); error = vm_set_register(vm, vcpuid, VM_REG_GUEST_CR2, cr2); KASSERT(error == 0, ("vm_set_register(cr2) error %d", error)); vm_inject_fault(vm, vcpuid, IDT_PF, 1, error_code); } static VMM_STAT(VCPU_NMI_COUNT, "number of NMIs delivered to vcpu"); int vm_inject_nmi(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; vcpu->nmi_pending = 1; vcpu_notify_event(vm, vcpuid, false); return (0); } int vm_nmi_pending(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) panic("vm_nmi_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; return (vcpu->nmi_pending); } void vm_nmi_clear(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) panic("vm_nmi_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; if (vcpu->nmi_pending == 0) panic("vm_nmi_clear: inconsistent nmi_pending state"); vcpu->nmi_pending = 0; vmm_stat_incr(vm, vcpuid, VCPU_NMI_COUNT, 1); } static VMM_STAT(VCPU_EXTINT_COUNT, "number of ExtINTs delivered to vcpu"); int vm_inject_extint(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; vcpu->extint_pending = 1; vcpu_notify_event(vm, vcpuid, false); return (0); } int vm_extint_pending(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) panic("vm_extint_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; return (vcpu->extint_pending); } void vm_extint_clear(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) panic("vm_extint_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; if (vcpu->extint_pending == 0) panic("vm_extint_clear: inconsistent extint_pending state"); vcpu->extint_pending = 0; vmm_stat_incr(vm, vcpuid, VCPU_EXTINT_COUNT, 1); } int vm_get_capability(struct vm *vm, int vcpu, int type, int *retval) { if (vcpu < 0 || vcpu >= VM_MAXCPU) return (EINVAL); if (type < 0 || type >= VM_CAP_MAX) return (EINVAL); return (VMGETCAP(vm->cookie, vcpu, type, retval)); } int vm_set_capability(struct vm *vm, int vcpu, int type, int val) { if (vcpu < 0 || vcpu >= VM_MAXCPU) return (EINVAL); if (type < 0 || type >= VM_CAP_MAX) return (EINVAL); return (VMSETCAP(vm->cookie, vcpu, type, val)); } struct vlapic * vm_lapic(struct vm *vm, int cpu) { return (vm->vcpu[cpu].vlapic); } struct vioapic * vm_ioapic(struct vm *vm) { return (vm->vioapic); } struct vhpet * vm_hpet(struct vm *vm) { return (vm->vhpet); } boolean_t vmm_is_pptdev(int bus, int slot, int func) { int found, i, n; int b, s, f; char *val, *cp, *cp2; /* * XXX * The length of an environment variable is limited to 128 bytes which * puts an upper limit on the number of passthru devices that may be * specified using a single environment variable. * * Work around this by scanning multiple environment variable * names instead of a single one - yuck! */ const char *names[] = { "pptdevs", "pptdevs2", "pptdevs3", NULL }; /* set pptdevs="1/2/3 4/5/6 7/8/9 10/11/12" */ found = 0; for (i = 0; names[i] != NULL && !found; i++) { cp = val = kern_getenv(names[i]); while (cp != NULL && *cp != '\0') { if ((cp2 = strchr(cp, ' ')) != NULL) *cp2 = '\0'; n = sscanf(cp, "%d/%d/%d", &b, &s, &f); if (n == 3 && bus == b && slot == s && func == f) { found = 1; break; } if (cp2 != NULL) *cp2++ = ' '; cp = cp2; } freeenv(val); } return (found); } void * vm_iommu_domain(struct vm *vm) { return (vm->iommu); } int vcpu_set_state(struct vm *vm, int vcpuid, enum vcpu_state newstate, bool from_idle) { int error; struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) panic("vm_set_run_state: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; vcpu_lock(vcpu); - error = vcpu_set_state_locked(vcpu, newstate, from_idle); + error = vcpu_set_state_locked(vm, vcpuid, newstate, from_idle); vcpu_unlock(vcpu); return (error); } enum vcpu_state vcpu_get_state(struct vm *vm, int vcpuid, int *hostcpu) { struct vcpu *vcpu; enum vcpu_state state; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) panic("vm_get_run_state: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; vcpu_lock(vcpu); state = vcpu->state; if (hostcpu != NULL) *hostcpu = vcpu->hostcpu; vcpu_unlock(vcpu); return (state); } int vm_activate_cpu(struct vm *vm, int vcpuid) { if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); if (CPU_ISSET(vcpuid, &vm->active_cpus)) return (EBUSY); VCPU_CTR0(vm, vcpuid, "activated"); CPU_SET_ATOMIC(vcpuid, &vm->active_cpus); return (0); } cpuset_t vm_active_cpus(struct vm *vm) { return (vm->active_cpus); } cpuset_t vm_suspended_cpus(struct vm *vm) { return (vm->suspended_cpus); } void * vcpu_stats(struct vm *vm, int vcpuid) { return (vm->vcpu[vcpuid].stats); } int vm_get_x2apic_state(struct vm *vm, int vcpuid, enum x2apic_state *state) { if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); *state = vm->vcpu[vcpuid].x2apic_state; return (0); } int vm_set_x2apic_state(struct vm *vm, int vcpuid, enum x2apic_state state) { if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); if (state >= X2APIC_STATE_LAST) return (EINVAL); vm->vcpu[vcpuid].x2apic_state = state; vlapic_set_x2apic_state(vm, vcpuid, state); return (0); } /* * This function is called to ensure that a vcpu "sees" a pending event * as soon as possible: * - If the vcpu thread is sleeping then it is woken up. * - If the vcpu is running on a different host_cpu then an IPI will be directed * to the host_cpu to cause the vcpu to trap into the hypervisor. */ -void -vcpu_notify_event(struct vm *vm, int vcpuid, bool lapic_intr) +static void +vcpu_notify_event_locked(struct vcpu *vcpu, bool lapic_intr) { int hostcpu; - struct vcpu *vcpu; - vcpu = &vm->vcpu[vcpuid]; - - vcpu_lock(vcpu); hostcpu = vcpu->hostcpu; if (vcpu->state == VCPU_RUNNING) { KASSERT(hostcpu != NOCPU, ("vcpu running on invalid hostcpu")); if (hostcpu != curcpu) { if (lapic_intr) { vlapic_post_intr(vcpu->vlapic, hostcpu, vmm_ipinum); } else { ipi_cpu(hostcpu, vmm_ipinum); } } else { /* * If the 'vcpu' is running on 'curcpu' then it must * be sending a notification to itself (e.g. SELF_IPI). * The pending event will be picked up when the vcpu * transitions back to guest context. */ } } else { KASSERT(hostcpu == NOCPU, ("vcpu state %d not consistent " "with hostcpu %d", vcpu->state, hostcpu)); if (vcpu->state == VCPU_SLEEPING) wakeup_one(vcpu); } +} + +void +vcpu_notify_event(struct vm *vm, int vcpuid, bool lapic_intr) +{ + struct vcpu *vcpu = &vm->vcpu[vcpuid]; + + vcpu_lock(vcpu); + vcpu_notify_event_locked(vcpu, lapic_intr); vcpu_unlock(vcpu); } struct vmspace * vm_get_vmspace(struct vm *vm) { return (vm->vmspace); } int vm_apicid2vcpuid(struct vm *vm, int apicid) { /* * XXX apic id is assumed to be numerically identical to vcpu id */ return (apicid); } void vm_smp_rendezvous(struct vm *vm, int vcpuid, cpuset_t dest, vm_rendezvous_func_t func, void *arg) { int i; /* * Enforce that this function is called without any locks */ WITNESS_WARN(WARN_PANIC, NULL, "vm_smp_rendezvous"); KASSERT(vcpuid == -1 || (vcpuid >= 0 && vcpuid < VM_MAXCPU), ("vm_smp_rendezvous: invalid vcpuid %d", vcpuid)); restart: mtx_lock(&vm->rendezvous_mtx); if (vm->rendezvous_func != NULL) { /* * If a rendezvous is already in progress then we need to * call the rendezvous handler in case this 'vcpuid' is one * of the targets of the rendezvous. */ RENDEZVOUS_CTR0(vm, vcpuid, "Rendezvous already in progress"); mtx_unlock(&vm->rendezvous_mtx); vm_handle_rendezvous(vm, vcpuid); goto restart; } KASSERT(vm->rendezvous_func == NULL, ("vm_smp_rendezvous: previous " "rendezvous is still in progress")); RENDEZVOUS_CTR0(vm, vcpuid, "Initiating rendezvous"); vm->rendezvous_req_cpus = dest; CPU_ZERO(&vm->rendezvous_done_cpus); vm->rendezvous_arg = arg; vm_set_rendezvous_func(vm, func); mtx_unlock(&vm->rendezvous_mtx); /* * Wake up any sleeping vcpus and trigger a VM-exit in any running * vcpus so they handle the rendezvous as soon as possible. */ for (i = 0; i < VM_MAXCPU; i++) { if (CPU_ISSET(i, &dest)) vcpu_notify_event(vm, i, false); } vm_handle_rendezvous(vm, vcpuid); } struct vatpic * vm_atpic(struct vm *vm) { return (vm->vatpic); } struct vatpit * vm_atpit(struct vm *vm) { return (vm->vatpit); } struct vpmtmr * vm_pmtmr(struct vm *vm) { return (vm->vpmtmr); } struct vrtc * vm_rtc(struct vm *vm) { return (vm->vrtc); } enum vm_reg_name vm_segment_name(int seg) { static enum vm_reg_name seg_names[] = { VM_REG_GUEST_ES, VM_REG_GUEST_CS, VM_REG_GUEST_SS, VM_REG_GUEST_DS, VM_REG_GUEST_FS, VM_REG_GUEST_GS }; KASSERT(seg >= 0 && seg < nitems(seg_names), ("%s: invalid segment encoding %d", __func__, seg)); return (seg_names[seg]); } void vm_copy_teardown(struct vm *vm, int vcpuid, struct vm_copyinfo *copyinfo, int num_copyinfo) { int idx; for (idx = 0; idx < num_copyinfo; idx++) { if (copyinfo[idx].cookie != NULL) vm_gpa_release(copyinfo[idx].cookie); } bzero(copyinfo, num_copyinfo * sizeof(struct vm_copyinfo)); } int vm_copy_setup(struct vm *vm, int vcpuid, struct vm_guest_paging *paging, uint64_t gla, size_t len, int prot, struct vm_copyinfo *copyinfo, int num_copyinfo, int *fault) { int error, idx, nused; size_t n, off, remaining; void *hva, *cookie; uint64_t gpa; bzero(copyinfo, sizeof(struct vm_copyinfo) * num_copyinfo); nused = 0; remaining = len; while (remaining > 0) { KASSERT(nused < num_copyinfo, ("insufficient vm_copyinfo")); error = vm_gla2gpa(vm, vcpuid, paging, gla, prot, &gpa, fault); if (error || *fault) return (error); off = gpa & PAGE_MASK; n = min(remaining, PAGE_SIZE - off); copyinfo[nused].gpa = gpa; copyinfo[nused].len = n; remaining -= n; gla += n; nused++; } for (idx = 0; idx < nused; idx++) { hva = vm_gpa_hold(vm, copyinfo[idx].gpa, copyinfo[idx].len, prot, &cookie); if (hva == NULL) break; copyinfo[idx].hva = hva; copyinfo[idx].cookie = cookie; } if (idx != nused) { vm_copy_teardown(vm, vcpuid, copyinfo, num_copyinfo); return (EFAULT); } else { *fault = 0; return (0); } } void vm_copyin(struct vm *vm, int vcpuid, struct vm_copyinfo *copyinfo, void *kaddr, size_t len) { char *dst; int idx; dst = kaddr; idx = 0; while (len > 0) { bcopy(copyinfo[idx].hva, dst, copyinfo[idx].len); len -= copyinfo[idx].len; dst += copyinfo[idx].len; idx++; } } void vm_copyout(struct vm *vm, int vcpuid, const void *kaddr, struct vm_copyinfo *copyinfo, size_t len) { const char *src; int idx; src = kaddr; idx = 0; while (len > 0) { bcopy(src, copyinfo[idx].hva, copyinfo[idx].len); len -= copyinfo[idx].len; src += copyinfo[idx].len; idx++; } } /* * Return the amount of in-use and wired memory for the VM. Since * these are global stats, only return the values with for vCPU 0 */ VMM_STAT_DECLARE(VMM_MEM_RESIDENT); VMM_STAT_DECLARE(VMM_MEM_WIRED); static void vm_get_rescnt(struct vm *vm, int vcpu, struct vmm_stat_type *stat) { if (vcpu == 0) { vmm_stat_set(vm, vcpu, VMM_MEM_RESIDENT, PAGE_SIZE * vmspace_resident_count(vm->vmspace)); } } static void vm_get_wiredcnt(struct vm *vm, int vcpu, struct vmm_stat_type *stat) { if (vcpu == 0) { vmm_stat_set(vm, vcpu, VMM_MEM_WIRED, PAGE_SIZE * pmap_wired_count(vmspace_pmap(vm->vmspace))); } } VMM_STAT_FUNC(VMM_MEM_RESIDENT, "Resident memory", vm_get_rescnt); VMM_STAT_FUNC(VMM_MEM_WIRED, "Wired memory", vm_get_wiredcnt); Index: head/sys/amd64/vmm/vmm_stat.c =================================================================== --- head/sys/amd64/vmm/vmm_stat.c (revision 283656) +++ head/sys/amd64/vmm/vmm_stat.c (revision 283657) @@ -1,169 +1,170 @@ /*- * Copyright (c) 2011 NetApp, Inc. * 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, 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 NETAPP, INC ``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 NETAPP, INC OR CONTRIBUTORS 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. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include "vmm_util.h" #include "vmm_stat.h" /* * 'vst_num_elems' is the total number of addressable statistic elements * 'vst_num_types' is the number of unique statistic types * * It is always true that 'vst_num_elems' is greater than or equal to * 'vst_num_types'. This is because a stat type may represent more than * one element (for e.g. VMM_STAT_ARRAY). */ static int vst_num_elems, vst_num_types; static struct vmm_stat_type *vsttab[MAX_VMM_STAT_ELEMS]; static MALLOC_DEFINE(M_VMM_STAT, "vmm stat", "vmm stat"); #define vst_size ((size_t)vst_num_elems * sizeof(uint64_t)) void vmm_stat_register(void *arg) { struct vmm_stat_type *vst = arg; /* We require all stats to identify themselves with a description */ if (vst->desc == NULL) return; if (vst->scope == VMM_STAT_SCOPE_INTEL && !vmm_is_intel()) return; if (vst->scope == VMM_STAT_SCOPE_AMD && !vmm_is_amd()) return; if (vst_num_elems + vst->nelems >= MAX_VMM_STAT_ELEMS) { printf("Cannot accomodate vmm stat type \"%s\"!\n", vst->desc); return; } vst->index = vst_num_elems; vst_num_elems += vst->nelems; vsttab[vst_num_types++] = vst; } int vmm_stat_copy(struct vm *vm, int vcpu, int *num_stats, uint64_t *buf) { struct vmm_stat_type *vst; uint64_t *stats; int i; if (vcpu < 0 || vcpu >= VM_MAXCPU) return (EINVAL); /* Let stats functions update their counters */ for (i = 0; i < vst_num_types; i++) { vst = vsttab[i]; if (vst->func != NULL) (*vst->func)(vm, vcpu, vst); } /* Copy over the stats */ stats = vcpu_stats(vm, vcpu); for (i = 0; i < vst_num_elems; i++) buf[i] = stats[i]; *num_stats = vst_num_elems; return (0); } void * vmm_stat_alloc(void) { return (malloc(vst_size, M_VMM_STAT, M_WAITOK)); } void vmm_stat_init(void *vp) { bzero(vp, vst_size); } void vmm_stat_free(void *vp) { free(vp, M_VMM_STAT); } int vmm_stat_desc_copy(int index, char *buf, int bufsize) { int i; struct vmm_stat_type *vst; for (i = 0; i < vst_num_types; i++) { vst = vsttab[i]; if (index >= vst->index && index < vst->index + vst->nelems) { if (vst->nelems > 1) { snprintf(buf, bufsize, "%s[%d]", vst->desc, index - vst->index); } else { strlcpy(buf, vst->desc, bufsize); } return (0); /* found it */ } } return (EINVAL); } /* global statistics */ VMM_STAT(VCPU_MIGRATIONS, "vcpu migration across host cpus"); VMM_STAT(VMEXIT_COUNT, "total number of vm exits"); VMM_STAT(VMEXIT_EXTINT, "vm exits due to external interrupt"); VMM_STAT(VMEXIT_HLT, "number of times hlt was intercepted"); VMM_STAT(VMEXIT_CR_ACCESS, "number of times %cr access was intercepted"); VMM_STAT(VMEXIT_RDMSR, "number of times rdmsr was intercepted"); VMM_STAT(VMEXIT_WRMSR, "number of times wrmsr was intercepted"); VMM_STAT(VMEXIT_MTRAP, "number of monitor trap exits"); VMM_STAT(VMEXIT_PAUSE, "number of times pause was intercepted"); VMM_STAT(VMEXIT_INTR_WINDOW, "vm exits due to interrupt window opening"); VMM_STAT(VMEXIT_NMI_WINDOW, "vm exits due to nmi window opening"); VMM_STAT(VMEXIT_INOUT, "number of times in/out was intercepted"); VMM_STAT(VMEXIT_CPUID, "number of times cpuid was intercepted"); VMM_STAT(VMEXIT_NESTED_FAULT, "vm exits due to nested page fault"); VMM_STAT(VMEXIT_INST_EMUL, "vm exits for instruction emulation"); VMM_STAT(VMEXIT_UNKNOWN, "number of vm exits for unknown reason"); VMM_STAT(VMEXIT_ASTPENDING, "number of times astpending at exit"); +VMM_STAT(VMEXIT_REQIDLE, "number of times idle requested at exit"); VMM_STAT(VMEXIT_USERSPACE, "number of vm exits handled in userspace"); VMM_STAT(VMEXIT_RENDEZVOUS, "number of times rendezvous pending at exit"); VMM_STAT(VMEXIT_EXCEPTION, "number of vm exits due to exceptions"); Index: head/sys/amd64/vmm/vmm_stat.h =================================================================== --- head/sys/amd64/vmm/vmm_stat.h (revision 283656) +++ head/sys/amd64/vmm/vmm_stat.h (revision 283657) @@ -1,160 +1,161 @@ /*- * Copyright (c) 2011 NetApp, Inc. * 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, 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. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``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 REGENTS OR CONTRIBUTORS 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 _VMM_STAT_H_ #define _VMM_STAT_H_ struct vm; #define MAX_VMM_STAT_ELEMS 64 /* arbitrary */ enum vmm_stat_scope { VMM_STAT_SCOPE_ANY, VMM_STAT_SCOPE_INTEL, /* Intel VMX specific statistic */ VMM_STAT_SCOPE_AMD, /* AMD SVM specific statistic */ }; struct vmm_stat_type; typedef void (*vmm_stat_func_t)(struct vm *vm, int vcpu, struct vmm_stat_type *stat); struct vmm_stat_type { int index; /* position in the stats buffer */ int nelems; /* standalone or array */ const char *desc; /* description of statistic */ vmm_stat_func_t func; enum vmm_stat_scope scope; }; void vmm_stat_register(void *arg); #define VMM_STAT_FDEFINE(type, nelems, desc, func, scope) \ struct vmm_stat_type type[1] = { \ { -1, nelems, desc, func, scope } \ }; \ SYSINIT(type##_stat, SI_SUB_KLD, SI_ORDER_ANY, vmm_stat_register, type) #define VMM_STAT_DEFINE(type, nelems, desc, scope) \ VMM_STAT_FDEFINE(type, nelems, desc, NULL, scope) #define VMM_STAT_DECLARE(type) \ extern struct vmm_stat_type type[1] #define VMM_STAT(type, desc) \ VMM_STAT_DEFINE(type, 1, desc, VMM_STAT_SCOPE_ANY) #define VMM_STAT_INTEL(type, desc) \ VMM_STAT_DEFINE(type, 1, desc, VMM_STAT_SCOPE_INTEL) #define VMM_STAT_AMD(type, desc) \ VMM_STAT_DEFINE(type, 1, desc, VMM_STAT_SCOPE_AMD) #define VMM_STAT_FUNC(type, desc, func) \ VMM_STAT_FDEFINE(type, 1, desc, func, VMM_STAT_SCOPE_ANY) #define VMM_STAT_ARRAY(type, nelems, desc) \ VMM_STAT_DEFINE(type, nelems, desc, VMM_STAT_SCOPE_ANY) void *vmm_stat_alloc(void); void vmm_stat_init(void *vp); void vmm_stat_free(void *vp); /* * 'buf' should be at least fit 'MAX_VMM_STAT_TYPES' entries */ int vmm_stat_copy(struct vm *vm, int vcpu, int *num_stats, uint64_t *buf); int vmm_stat_desc_copy(int index, char *buf, int buflen); static void __inline vmm_stat_array_incr(struct vm *vm, int vcpu, struct vmm_stat_type *vst, int statidx, uint64_t x) { #ifdef VMM_KEEP_STATS uint64_t *stats; stats = vcpu_stats(vm, vcpu); if (vst->index >= 0 && statidx < vst->nelems) stats[vst->index + statidx] += x; #endif } static void __inline vmm_stat_array_set(struct vm *vm, int vcpu, struct vmm_stat_type *vst, int statidx, uint64_t val) { #ifdef VMM_KEEP_STATS uint64_t *stats; stats = vcpu_stats(vm, vcpu); if (vst->index >= 0 && statidx < vst->nelems) stats[vst->index + statidx] = val; #endif } static void __inline vmm_stat_incr(struct vm *vm, int vcpu, struct vmm_stat_type *vst, uint64_t x) { #ifdef VMM_KEEP_STATS vmm_stat_array_incr(vm, vcpu, vst, 0, x); #endif } static void __inline vmm_stat_set(struct vm *vm, int vcpu, struct vmm_stat_type *vst, uint64_t val) { #ifdef VMM_KEEP_STATS vmm_stat_array_set(vm, vcpu, vst, 0, val); #endif } VMM_STAT_DECLARE(VCPU_MIGRATIONS); VMM_STAT_DECLARE(VMEXIT_COUNT); VMM_STAT_DECLARE(VMEXIT_EXTINT); VMM_STAT_DECLARE(VMEXIT_HLT); VMM_STAT_DECLARE(VMEXIT_CR_ACCESS); VMM_STAT_DECLARE(VMEXIT_RDMSR); VMM_STAT_DECLARE(VMEXIT_WRMSR); VMM_STAT_DECLARE(VMEXIT_MTRAP); VMM_STAT_DECLARE(VMEXIT_PAUSE); VMM_STAT_DECLARE(VMEXIT_INTR_WINDOW); VMM_STAT_DECLARE(VMEXIT_NMI_WINDOW); VMM_STAT_DECLARE(VMEXIT_INOUT); VMM_STAT_DECLARE(VMEXIT_CPUID); VMM_STAT_DECLARE(VMEXIT_NESTED_FAULT); VMM_STAT_DECLARE(VMEXIT_INST_EMUL); VMM_STAT_DECLARE(VMEXIT_UNKNOWN); VMM_STAT_DECLARE(VMEXIT_ASTPENDING); VMM_STAT_DECLARE(VMEXIT_USERSPACE); VMM_STAT_DECLARE(VMEXIT_RENDEZVOUS); VMM_STAT_DECLARE(VMEXIT_EXCEPTION); +VMM_STAT_DECLARE(VMEXIT_REQIDLE); #endif Index: head/usr.sbin/bhyve/bhyverun.c =================================================================== --- head/usr.sbin/bhyve/bhyverun.c (revision 283656) +++ head/usr.sbin/bhyve/bhyverun.c (revision 283657) @@ -1,892 +1,904 @@ /*- * Copyright (c) 2011 NetApp, Inc. * 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, 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 NETAPP, INC ``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 NETAPP, INC OR CONTRIBUTORS 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. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "bhyverun.h" #include "acpi.h" #include "inout.h" #include "dbgport.h" #include "ioapic.h" #include "mem.h" #include "mevent.h" #include "mptbl.h" #include "pci_emul.h" #include "pci_irq.h" #include "pci_lpc.h" #include "smbiostbl.h" #include "xmsr.h" #include "spinup_ap.h" #include "rtc.h" #define GUEST_NIO_PORT 0x488 /* guest upcalls via i/o port */ #define MB (1024UL * 1024) #define GB (1024UL * MB) typedef int (*vmexit_handler_t)(struct vmctx *, struct vm_exit *, int *vcpu); extern int vmexit_task_switch(struct vmctx *, struct vm_exit *, int *vcpu); char *vmname; int guest_ncpus; char *guest_uuid_str; static int guest_vmexit_on_hlt, guest_vmexit_on_pause; static int virtio_msix = 1; static int x2apic_mode = 0; /* default is xAPIC */ static int strictio; static int strictmsr = 1; static int acpi; static char *progname; static const int BSP = 0; static cpuset_t cpumask; static void vm_loop(struct vmctx *ctx, int vcpu, uint64_t rip); static struct vm_exit vmexit[VM_MAXCPU]; struct bhyvestats { uint64_t vmexit_bogus; - uint64_t vmexit_bogus_switch; + uint64_t vmexit_reqidle; uint64_t vmexit_hlt; uint64_t vmexit_pause; uint64_t vmexit_mtrap; uint64_t vmexit_inst_emul; uint64_t cpu_switch_rotate; uint64_t cpu_switch_direct; } stats; struct mt_vmm_info { pthread_t mt_thr; struct vmctx *mt_ctx; int mt_vcpu; } mt_vmm_info[VM_MAXCPU]; static cpuset_t *vcpumap[VM_MAXCPU] = { NULL }; static void usage(int code) { fprintf(stderr, "Usage: %s [-abehuwxACHPWY] [-c vcpus] [-g ] [-l ]\n" " %*s [-m mem] [-p vcpu:hostcpu] [-s ] [-U uuid] \n" " -a: local apic is in xAPIC mode (deprecated)\n" " -A: create ACPI tables\n" " -c: # cpus (default 1)\n" " -C: include guest memory in core file\n" " -e: exit on unhandled I/O access\n" " -g: gdb port\n" " -h: help\n" " -H: vmexit from the guest on hlt\n" " -l: LPC device configuration\n" " -m: memory size in MB\n" " -p: pin 'vcpu' to 'hostcpu'\n" " -P: vmexit from the guest on pause\n" " -s: PCI slot config\n" " -u: RTC keeps UTC time\n" " -U: uuid\n" " -w: ignore unimplemented MSRs\n" " -W: force virtio to use single-vector MSI\n" " -x: local apic is in x2APIC mode\n" " -Y: disable MPtable generation\n", progname, (int)strlen(progname), ""); exit(code); } static int pincpu_parse(const char *opt) { int vcpu, pcpu; if (sscanf(opt, "%d:%d", &vcpu, &pcpu) != 2) { fprintf(stderr, "invalid format: %s\n", opt); return (-1); } if (vcpu < 0 || vcpu >= VM_MAXCPU) { fprintf(stderr, "vcpu '%d' outside valid range from 0 to %d\n", vcpu, VM_MAXCPU - 1); return (-1); } if (pcpu < 0 || pcpu >= CPU_SETSIZE) { fprintf(stderr, "hostcpu '%d' outside valid range from " "0 to %d\n", pcpu, CPU_SETSIZE - 1); return (-1); } if (vcpumap[vcpu] == NULL) { if ((vcpumap[vcpu] = malloc(sizeof(cpuset_t))) == NULL) { perror("malloc"); return (-1); } CPU_ZERO(vcpumap[vcpu]); } CPU_SET(pcpu, vcpumap[vcpu]); return (0); } void vm_inject_fault(void *arg, int vcpu, int vector, int errcode_valid, int errcode) { struct vmctx *ctx; int error, restart_instruction; ctx = arg; restart_instruction = 1; error = vm_inject_exception(ctx, vcpu, vector, errcode_valid, errcode, restart_instruction); assert(error == 0); } void * paddr_guest2host(struct vmctx *ctx, uintptr_t gaddr, size_t len) { return (vm_map_gpa(ctx, gaddr, len)); } int fbsdrun_vmexit_on_pause(void) { return (guest_vmexit_on_pause); } int fbsdrun_vmexit_on_hlt(void) { return (guest_vmexit_on_hlt); } int fbsdrun_virtio_msix(void) { return (virtio_msix); } static void * fbsdrun_start_thread(void *param) { char tname[MAXCOMLEN + 1]; struct mt_vmm_info *mtp; int vcpu; mtp = param; vcpu = mtp->mt_vcpu; snprintf(tname, sizeof(tname), "vcpu %d", vcpu); pthread_set_name_np(mtp->mt_thr, tname); vm_loop(mtp->mt_ctx, vcpu, vmexit[vcpu].rip); /* not reached */ exit(1); return (NULL); } void fbsdrun_addcpu(struct vmctx *ctx, int fromcpu, int newcpu, uint64_t rip) { int error; assert(fromcpu == BSP); /* * The 'newcpu' must be activated in the context of 'fromcpu'. If * vm_activate_cpu() is delayed until newcpu's pthread starts running * then vmm.ko is out-of-sync with bhyve and this can create a race * with vm_suspend(). */ error = vm_activate_cpu(ctx, newcpu); assert(error == 0); CPU_SET_ATOMIC(newcpu, &cpumask); /* * Set up the vmexit struct to allow execution to start * at the given RIP */ vmexit[newcpu].rip = rip; vmexit[newcpu].inst_length = 0; mt_vmm_info[newcpu].mt_ctx = ctx; mt_vmm_info[newcpu].mt_vcpu = newcpu; error = pthread_create(&mt_vmm_info[newcpu].mt_thr, NULL, fbsdrun_start_thread, &mt_vmm_info[newcpu]); assert(error == 0); } static int fbsdrun_deletecpu(struct vmctx *ctx, int vcpu) { if (!CPU_ISSET(vcpu, &cpumask)) { fprintf(stderr, "Attempting to delete unknown cpu %d\n", vcpu); exit(1); } CPU_CLR_ATOMIC(vcpu, &cpumask); return (CPU_EMPTY(&cpumask)); } static int vmexit_handle_notify(struct vmctx *ctx, struct vm_exit *vme, int *pvcpu, uint32_t eax) { #if BHYVE_DEBUG /* * put guest-driven debug here */ #endif return (VMEXIT_CONTINUE); } static int vmexit_inout(struct vmctx *ctx, struct vm_exit *vme, int *pvcpu) { int error; int bytes, port, in, out, string; int vcpu; vcpu = *pvcpu; port = vme->u.inout.port; bytes = vme->u.inout.bytes; string = vme->u.inout.string; in = vme->u.inout.in; out = !in; /* Extra-special case of host notifications */ if (out && port == GUEST_NIO_PORT) { error = vmexit_handle_notify(ctx, vme, pvcpu, vme->u.inout.eax); return (error); } error = emulate_inout(ctx, vcpu, vme, strictio); if (error) { fprintf(stderr, "Unhandled %s%c 0x%04x at 0x%lx\n", in ? "in" : "out", bytes == 1 ? 'b' : (bytes == 2 ? 'w' : 'l'), port, vmexit->rip); return (VMEXIT_ABORT); } else { return (VMEXIT_CONTINUE); } } static int vmexit_rdmsr(struct vmctx *ctx, struct vm_exit *vme, int *pvcpu) { uint64_t val; uint32_t eax, edx; int error; val = 0; error = emulate_rdmsr(ctx, *pvcpu, vme->u.msr.code, &val); if (error != 0) { fprintf(stderr, "rdmsr to register %#x on vcpu %d\n", vme->u.msr.code, *pvcpu); if (strictmsr) { vm_inject_gp(ctx, *pvcpu); return (VMEXIT_CONTINUE); } } eax = val; error = vm_set_register(ctx, *pvcpu, VM_REG_GUEST_RAX, eax); assert(error == 0); edx = val >> 32; error = vm_set_register(ctx, *pvcpu, VM_REG_GUEST_RDX, edx); assert(error == 0); return (VMEXIT_CONTINUE); } static int vmexit_wrmsr(struct vmctx *ctx, struct vm_exit *vme, int *pvcpu) { int error; error = emulate_wrmsr(ctx, *pvcpu, vme->u.msr.code, vme->u.msr.wval); if (error != 0) { fprintf(stderr, "wrmsr to register %#x(%#lx) on vcpu %d\n", vme->u.msr.code, vme->u.msr.wval, *pvcpu); if (strictmsr) { vm_inject_gp(ctx, *pvcpu); return (VMEXIT_CONTINUE); } } return (VMEXIT_CONTINUE); } static int vmexit_spinup_ap(struct vmctx *ctx, struct vm_exit *vme, int *pvcpu) { int newcpu; int retval = VMEXIT_CONTINUE; newcpu = spinup_ap(ctx, *pvcpu, vme->u.spinup_ap.vcpu, vme->u.spinup_ap.rip); return (retval); } #define DEBUG_EPT_MISCONFIG #ifdef DEBUG_EPT_MISCONFIG #define EXIT_REASON_EPT_MISCONFIG 49 #define VMCS_GUEST_PHYSICAL_ADDRESS 0x00002400 #define VMCS_IDENT(x) ((x) | 0x80000000) static uint64_t ept_misconfig_gpa, ept_misconfig_pte[4]; static int ept_misconfig_ptenum; #endif static int vmexit_vmx(struct vmctx *ctx, struct vm_exit *vmexit, int *pvcpu) { fprintf(stderr, "vm exit[%d]\n", *pvcpu); fprintf(stderr, "\treason\t\tVMX\n"); fprintf(stderr, "\trip\t\t0x%016lx\n", vmexit->rip); fprintf(stderr, "\tinst_length\t%d\n", vmexit->inst_length); fprintf(stderr, "\tstatus\t\t%d\n", vmexit->u.vmx.status); fprintf(stderr, "\texit_reason\t%u\n", vmexit->u.vmx.exit_reason); fprintf(stderr, "\tqualification\t0x%016lx\n", vmexit->u.vmx.exit_qualification); fprintf(stderr, "\tinst_type\t\t%d\n", vmexit->u.vmx.inst_type); fprintf(stderr, "\tinst_error\t\t%d\n", vmexit->u.vmx.inst_error); #ifdef DEBUG_EPT_MISCONFIG if (vmexit->u.vmx.exit_reason == EXIT_REASON_EPT_MISCONFIG) { vm_get_register(ctx, *pvcpu, VMCS_IDENT(VMCS_GUEST_PHYSICAL_ADDRESS), &ept_misconfig_gpa); vm_get_gpa_pmap(ctx, ept_misconfig_gpa, ept_misconfig_pte, &ept_misconfig_ptenum); fprintf(stderr, "\tEPT misconfiguration:\n"); fprintf(stderr, "\t\tGPA: %#lx\n", ept_misconfig_gpa); fprintf(stderr, "\t\tPTE(%d): %#lx %#lx %#lx %#lx\n", ept_misconfig_ptenum, ept_misconfig_pte[0], ept_misconfig_pte[1], ept_misconfig_pte[2], ept_misconfig_pte[3]); } #endif /* DEBUG_EPT_MISCONFIG */ return (VMEXIT_ABORT); } static int vmexit_svm(struct vmctx *ctx, struct vm_exit *vmexit, int *pvcpu) { fprintf(stderr, "vm exit[%d]\n", *pvcpu); fprintf(stderr, "\treason\t\tSVM\n"); fprintf(stderr, "\trip\t\t0x%016lx\n", vmexit->rip); fprintf(stderr, "\tinst_length\t%d\n", vmexit->inst_length); fprintf(stderr, "\texitcode\t%#lx\n", vmexit->u.svm.exitcode); fprintf(stderr, "\texitinfo1\t%#lx\n", vmexit->u.svm.exitinfo1); fprintf(stderr, "\texitinfo2\t%#lx\n", vmexit->u.svm.exitinfo2); return (VMEXIT_ABORT); } static int vmexit_bogus(struct vmctx *ctx, struct vm_exit *vmexit, int *pvcpu) { assert(vmexit->inst_length == 0); stats.vmexit_bogus++; return (VMEXIT_CONTINUE); } static int +vmexit_reqidle(struct vmctx *ctx, struct vm_exit *vmexit, int *pvcpu) +{ + + assert(vmexit->inst_length == 0); + + stats.vmexit_reqidle++; + + return (VMEXIT_CONTINUE); +} + +static int vmexit_hlt(struct vmctx *ctx, struct vm_exit *vmexit, int *pvcpu) { stats.vmexit_hlt++; /* * Just continue execution with the next instruction. We use * the HLT VM exit as a way to be friendly with the host * scheduler. */ return (VMEXIT_CONTINUE); } static int vmexit_pause(struct vmctx *ctx, struct vm_exit *vmexit, int *pvcpu) { stats.vmexit_pause++; return (VMEXIT_CONTINUE); } static int vmexit_mtrap(struct vmctx *ctx, struct vm_exit *vmexit, int *pvcpu) { assert(vmexit->inst_length == 0); stats.vmexit_mtrap++; return (VMEXIT_CONTINUE); } static int vmexit_inst_emul(struct vmctx *ctx, struct vm_exit *vmexit, int *pvcpu) { int err, i; struct vie *vie; stats.vmexit_inst_emul++; vie = &vmexit->u.inst_emul.vie; err = emulate_mem(ctx, *pvcpu, vmexit->u.inst_emul.gpa, vie, &vmexit->u.inst_emul.paging); if (err) { if (err == ESRCH) { fprintf(stderr, "Unhandled memory access to 0x%lx\n", vmexit->u.inst_emul.gpa); } fprintf(stderr, "Failed to emulate instruction ["); for (i = 0; i < vie->num_valid; i++) { fprintf(stderr, "0x%02x%s", vie->inst[i], i != (vie->num_valid - 1) ? " " : ""); } fprintf(stderr, "] at 0x%lx\n", vmexit->rip); return (VMEXIT_ABORT); } return (VMEXIT_CONTINUE); } static pthread_mutex_t resetcpu_mtx = PTHREAD_MUTEX_INITIALIZER; static pthread_cond_t resetcpu_cond = PTHREAD_COND_INITIALIZER; static int vmexit_suspend(struct vmctx *ctx, struct vm_exit *vmexit, int *pvcpu) { enum vm_suspend_how how; how = vmexit->u.suspended.how; fbsdrun_deletecpu(ctx, *pvcpu); if (*pvcpu != BSP) { pthread_mutex_lock(&resetcpu_mtx); pthread_cond_signal(&resetcpu_cond); pthread_mutex_unlock(&resetcpu_mtx); pthread_exit(NULL); } pthread_mutex_lock(&resetcpu_mtx); while (!CPU_EMPTY(&cpumask)) { pthread_cond_wait(&resetcpu_cond, &resetcpu_mtx); } pthread_mutex_unlock(&resetcpu_mtx); switch (how) { case VM_SUSPEND_RESET: exit(0); case VM_SUSPEND_POWEROFF: exit(1); case VM_SUSPEND_HALT: exit(2); case VM_SUSPEND_TRIPLEFAULT: exit(3); default: fprintf(stderr, "vmexit_suspend: invalid reason %d\n", how); exit(100); } return (0); /* NOTREACHED */ } static vmexit_handler_t handler[VM_EXITCODE_MAX] = { [VM_EXITCODE_INOUT] = vmexit_inout, [VM_EXITCODE_INOUT_STR] = vmexit_inout, [VM_EXITCODE_VMX] = vmexit_vmx, [VM_EXITCODE_SVM] = vmexit_svm, [VM_EXITCODE_BOGUS] = vmexit_bogus, + [VM_EXITCODE_REQIDLE] = vmexit_reqidle, [VM_EXITCODE_RDMSR] = vmexit_rdmsr, [VM_EXITCODE_WRMSR] = vmexit_wrmsr, [VM_EXITCODE_MTRAP] = vmexit_mtrap, [VM_EXITCODE_INST_EMUL] = vmexit_inst_emul, [VM_EXITCODE_SPINUP_AP] = vmexit_spinup_ap, [VM_EXITCODE_SUSPENDED] = vmexit_suspend, [VM_EXITCODE_TASK_SWITCH] = vmexit_task_switch, }; static void vm_loop(struct vmctx *ctx, int vcpu, uint64_t startrip) { int error, rc, prevcpu; enum vm_exitcode exitcode; cpuset_t active_cpus; if (vcpumap[vcpu] != NULL) { error = pthread_setaffinity_np(pthread_self(), sizeof(cpuset_t), vcpumap[vcpu]); assert(error == 0); } error = vm_active_cpus(ctx, &active_cpus); assert(CPU_ISSET(vcpu, &active_cpus)); error = vm_set_register(ctx, vcpu, VM_REG_GUEST_RIP, startrip); assert(error == 0); while (1) { error = vm_run(ctx, vcpu, &vmexit[vcpu]); if (error != 0) break; prevcpu = vcpu; exitcode = vmexit[vcpu].exitcode; if (exitcode >= VM_EXITCODE_MAX || handler[exitcode] == NULL) { fprintf(stderr, "vm_loop: unexpected exitcode 0x%x\n", exitcode); exit(1); } rc = (*handler[exitcode])(ctx, &vmexit[vcpu], &vcpu); switch (rc) { case VMEXIT_CONTINUE: break; case VMEXIT_ABORT: abort(); default: exit(1); } } fprintf(stderr, "vm_run error %d, errno %d\n", error, errno); } static int num_vcpus_allowed(struct vmctx *ctx) { int tmp, error; error = vm_get_capability(ctx, BSP, VM_CAP_UNRESTRICTED_GUEST, &tmp); /* * The guest is allowed to spinup more than one processor only if the * UNRESTRICTED_GUEST capability is available. */ if (error == 0) return (VM_MAXCPU); else return (1); } void fbsdrun_set_capabilities(struct vmctx *ctx, int cpu) { int err, tmp; if (fbsdrun_vmexit_on_hlt()) { err = vm_get_capability(ctx, cpu, VM_CAP_HALT_EXIT, &tmp); if (err < 0) { fprintf(stderr, "VM exit on HLT not supported\n"); exit(1); } vm_set_capability(ctx, cpu, VM_CAP_HALT_EXIT, 1); if (cpu == BSP) handler[VM_EXITCODE_HLT] = vmexit_hlt; } if (fbsdrun_vmexit_on_pause()) { /* * pause exit support required for this mode */ err = vm_get_capability(ctx, cpu, VM_CAP_PAUSE_EXIT, &tmp); if (err < 0) { fprintf(stderr, "SMP mux requested, no pause support\n"); exit(1); } vm_set_capability(ctx, cpu, VM_CAP_PAUSE_EXIT, 1); if (cpu == BSP) handler[VM_EXITCODE_PAUSE] = vmexit_pause; } if (x2apic_mode) err = vm_set_x2apic_state(ctx, cpu, X2APIC_ENABLED); else err = vm_set_x2apic_state(ctx, cpu, X2APIC_DISABLED); if (err) { fprintf(stderr, "Unable to set x2apic state (%d)\n", err); exit(1); } vm_set_capability(ctx, cpu, VM_CAP_ENABLE_INVPCID, 1); } int main(int argc, char *argv[]) { int c, error, gdb_port, err, bvmcons; int dump_guest_memory, max_vcpus, mptgen; int rtc_localtime; struct vmctx *ctx; uint64_t rip; size_t memsize; bvmcons = 0; dump_guest_memory = 0; progname = basename(argv[0]); gdb_port = 0; guest_ncpus = 1; memsize = 256 * MB; mptgen = 1; rtc_localtime = 1; while ((c = getopt(argc, argv, "abehuwxACHIPWYp:g:c:s:m:l:U:")) != -1) { switch (c) { case 'a': x2apic_mode = 0; break; case 'A': acpi = 1; break; case 'b': bvmcons = 1; break; case 'p': if (pincpu_parse(optarg) != 0) { errx(EX_USAGE, "invalid vcpu pinning " "configuration '%s'", optarg); } break; case 'c': guest_ncpus = atoi(optarg); break; case 'C': dump_guest_memory = 1; break; case 'g': gdb_port = atoi(optarg); break; case 'l': if (lpc_device_parse(optarg) != 0) { errx(EX_USAGE, "invalid lpc device " "configuration '%s'", optarg); } break; case 's': if (pci_parse_slot(optarg) != 0) exit(1); else break; case 'm': error = vm_parse_memsize(optarg, &memsize); if (error) errx(EX_USAGE, "invalid memsize '%s'", optarg); break; case 'H': guest_vmexit_on_hlt = 1; break; case 'I': /* * The "-I" option was used to add an ioapic to the * virtual machine. * * An ioapic is now provided unconditionally for each * virtual machine and this option is now deprecated. */ break; case 'P': guest_vmexit_on_pause = 1; break; case 'e': strictio = 1; break; case 'u': rtc_localtime = 0; break; case 'U': guest_uuid_str = optarg; break; case 'w': strictmsr = 0; break; case 'W': virtio_msix = 0; break; case 'x': x2apic_mode = 1; break; case 'Y': mptgen = 0; break; case 'h': usage(0); default: usage(1); } } argc -= optind; argv += optind; if (argc != 1) usage(1); vmname = argv[0]; ctx = vm_open(vmname); if (ctx == NULL) { perror("vm_open"); exit(1); } if (guest_ncpus < 1) { fprintf(stderr, "Invalid guest vCPUs (%d)\n", guest_ncpus); exit(1); } max_vcpus = num_vcpus_allowed(ctx); if (guest_ncpus > max_vcpus) { fprintf(stderr, "%d vCPUs requested but only %d available\n", guest_ncpus, max_vcpus); exit(1); } fbsdrun_set_capabilities(ctx, BSP); if (dump_guest_memory) vm_set_memflags(ctx, VM_MEM_F_INCORE); err = vm_setup_memory(ctx, memsize, VM_MMAP_ALL); if (err) { fprintf(stderr, "Unable to setup memory (%d)\n", err); exit(1); } error = init_msr(); if (error) { fprintf(stderr, "init_msr error %d", error); exit(1); } init_mem(); init_inout(); pci_irq_init(ctx); ioapic_init(ctx); rtc_init(ctx, rtc_localtime); sci_init(ctx); /* * Exit if a device emulation finds an error in it's initilization */ if (init_pci(ctx) != 0) exit(1); if (gdb_port != 0) init_dbgport(gdb_port); if (bvmcons) init_bvmcons(); error = vm_get_register(ctx, BSP, VM_REG_GUEST_RIP, &rip); assert(error == 0); /* * build the guest tables, MP etc. */ if (mptgen) { error = mptable_build(ctx, guest_ncpus); if (error) exit(1); } error = smbios_build(ctx); assert(error == 0); if (acpi) { error = acpi_build(ctx, guest_ncpus); assert(error == 0); } /* * Change the proc title to include the VM name. */ setproctitle("%s", vmname); /* * Add CPU 0 */ fbsdrun_addcpu(ctx, BSP, BSP, rip); /* * Head off to the main event dispatch loop */ mevent_dispatch(); exit(1); }