diff --git a/sys/amd64/include/vmm.h b/sys/amd64/include/vmm.h index 67730cc22980..bd6510b92527 100644 --- a/sys/amd64/include/vmm.h +++ b/sys/amd64/include/vmm.h @@ -1,821 +1,827 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * 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 #include #include struct vm_snapshot_meta; #ifdef _KERNEL SDT_PROVIDER_DECLARE(vmm); #endif 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_GUEST_DR0, VM_REG_GUEST_DR1, VM_REG_GUEST_DR2, VM_REG_GUEST_DR3, VM_REG_GUEST_DR6, VM_REG_GUEST_ENTRY_INST_LENGTH, 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) /* * The VM name has to fit into the pathname length constraints of devfs, * governed primarily by SPECNAMELEN. The length is the total number of * characters in the full path, relative to the mount point and not * including any leading '/' characters. * A prefix and a suffix are added to the name specified by the user. * The prefix is usually "vmm/" or "vmm.io/", but can be a few characters * longer for future use. * The suffix is a string that identifies a bootrom image or some similar * image that is attached to the VM. A separator character gets added to * the suffix automatically when generating the full path, so it must be * accounted for, reducing the effective length by 1. * The effective length of a VM name is 229 bytes for FreeBSD 13 and 37 * bytes for FreeBSD 12. A minimum length is set for safety and supports * a SPECNAMELEN as small as 32 on old systems. */ #define VM_MAX_PREFIXLEN 10 #define VM_MAX_SUFFIXLEN 15 #define VM_MIN_NAMELEN 6 #define VM_MAX_NAMELEN \ (SPECNAMELEN - VM_MAX_PREFIXLEN - VM_MAX_SUFFIXLEN - 1) #ifdef _KERNEL CTASSERT(VM_MAX_NAMELEN >= VM_MIN_NAMELEN); struct vcpu; struct vm; struct vm_exception; 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; enum snapshot_req; struct vm_eventinfo { - void *rptr; /* rendezvous cookie */ + cpuset_t *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 *vcpui, register_t rip, struct pmap *pmap, struct vm_eventinfo *info); typedef void (*vmi_cleanup_func_t)(void *vmi); typedef void * (*vmi_vcpu_init_func_t)(void *vmi, struct vcpu *vcpu, int vcpu_id); typedef void (*vmi_vcpu_cleanup_func_t)(void *vcpui); typedef int (*vmi_get_register_t)(void *vcpui, int num, uint64_t *retval); typedef int (*vmi_set_register_t)(void *vcpui, int num, uint64_t val); typedef int (*vmi_get_desc_t)(void *vcpui, int num, struct seg_desc *desc); typedef int (*vmi_set_desc_t)(void *vcpui, int num, struct seg_desc *desc); typedef int (*vmi_get_cap_t)(void *vcpui, int num, int *retval); typedef int (*vmi_set_cap_t)(void *vcpui, 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 *vcpui); typedef void (*vmi_vlapic_cleanup)(struct vlapic *vlapic); typedef int (*vmi_snapshot_t)(void *vmi, struct vm_snapshot_meta *meta); typedef int (*vmi_snapshot_vcpu_t)(void *vcpui, struct vm_snapshot_meta *meta); typedef int (*vmi_restore_tsc_t)(void *vcpui, uint64_t now); struct vmm_ops { vmm_init_func_t modinit; /* module wide initialization */ vmm_cleanup_func_t modcleanup; vmm_resume_func_t modresume; vmi_init_func_t init; /* vm-specific initialization */ vmi_run_func_t run; vmi_cleanup_func_t cleanup; vmi_vcpu_init_func_t vcpu_init; vmi_vcpu_cleanup_func_t vcpu_cleanup; vmi_get_register_t getreg; vmi_set_register_t setreg; vmi_get_desc_t getdesc; vmi_set_desc_t setdesc; vmi_get_cap_t getcap; vmi_set_cap_t setcap; vmi_vmspace_alloc vmspace_alloc; vmi_vmspace_free vmspace_free; vmi_vlapic_init vlapic_init; vmi_vlapic_cleanup vlapic_cleanup; /* checkpoint operations */ vmi_snapshot_t snapshot; vmi_snapshot_vcpu_t vcpu_snapshot; vmi_restore_tsc_t restore_tsc; }; extern const struct vmm_ops vmm_ops_intel; extern const struct vmm_ops vmm_ops_amd; extern u_int vm_maxcpu; /* maximum virtual cpus */ int vm_create(const char *name, struct vm **retvm); struct vcpu *vm_alloc_vcpu(struct vm *vm, int vcpuid); void vm_disable_vcpu_creation(struct vm *vm); void vm_slock_vcpus(struct vm *vm); void vm_unlock_vcpus(struct vm *vm); void vm_destroy(struct vm *vm); int vm_reinit(struct vm *vm); const char *vm_name(struct vm *vm); uint16_t vm_get_maxcpus(struct vm *vm); void vm_get_topology(struct vm *vm, uint16_t *sockets, uint16_t *cores, uint16_t *threads, uint16_t *maxcpus); int vm_set_topology(struct vm *vm, uint16_t sockets, uint16_t cores, uint16_t threads, uint16_t maxcpus); /* * APIs that modify the guest memory map require all vcpus to be frozen. */ void vm_slock_memsegs(struct vm *vm); void vm_xlock_memsegs(struct vm *vm); void vm_unlock_memsegs(struct vm *vm); int vm_mmap_memseg(struct vm *vm, vm_paddr_t gpa, int segid, vm_ooffset_t off, size_t len, int prot, int flags); int vm_munmap_memseg(struct vm *vm, vm_paddr_t gpa, size_t len); int vm_alloc_memseg(struct vm *vm, int ident, size_t len, bool sysmem); void vm_free_memseg(struct vm *vm, int ident); 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); 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); /* * APIs that inspect the guest memory map require only a *single* vcpu to * be frozen. This acts like a read lock on the guest memory map since any * modification requires *all* vcpus to be frozen. */ int vm_mmap_getnext(struct vm *vm, vm_paddr_t *gpa, int *segid, vm_ooffset_t *segoff, size_t *len, int *prot, int *flags); int vm_get_memseg(struct vm *vm, int ident, size_t *len, bool *sysmem, struct vm_object **objptr); vm_paddr_t vmm_sysmem_maxaddr(struct vm *vm); void *vm_gpa_hold(struct vcpu *vcpu, vm_paddr_t gpa, size_t len, int prot, void **cookie); void *vm_gpa_hold_global(struct vm *vm, vm_paddr_t gpa, size_t len, int prot, void **cookie); void *vm_gpa_hold_global(struct vm *vm, vm_paddr_t gpa, size_t len, int prot, void **cookie); void vm_gpa_release(void *cookie); bool vm_mem_allocated(struct vcpu *vcpu, vm_paddr_t gpa); int vm_get_register(struct vcpu *vcpu, int reg, uint64_t *retval); int vm_set_register(struct vcpu *vcpu, int reg, uint64_t val); int vm_get_seg_desc(struct vcpu *vcpu, int reg, struct seg_desc *ret_desc); int vm_set_seg_desc(struct vcpu *vcpu, int reg, struct seg_desc *desc); int vm_run(struct vcpu *vcpu, struct vm_exit *vme_user); int vm_suspend(struct vm *vm, enum vm_suspend_how how); int vm_inject_nmi(struct vcpu *vcpu); int vm_nmi_pending(struct vcpu *vcpu); void vm_nmi_clear(struct vcpu *vcpu); int vm_inject_extint(struct vcpu *vcpu); int vm_extint_pending(struct vcpu *vcpu); void vm_extint_clear(struct vcpu *vcpu); int vcpu_vcpuid(struct vcpu *vcpu); struct vm *vcpu_vm(struct vcpu *vcpu); struct vcpu *vm_vcpu(struct vm *vm, int cpu); struct vlapic *vm_lapic(struct vcpu *vcpu); struct vioapic *vm_ioapic(struct vm *vm); struct vhpet *vm_hpet(struct vm *vm); int vm_get_capability(struct vcpu *vcpu, int type, int *val); int vm_set_capability(struct vcpu *vcpu, int type, int val); int vm_get_x2apic_state(struct vcpu *vcpu, enum x2apic_state *state); int vm_set_x2apic_state(struct vcpu *vcpu, enum x2apic_state state); int vm_apicid2vcpuid(struct vm *vm, int apicid); int vm_activate_cpu(struct vcpu *vcpu); int vm_suspend_cpu(struct vm *vm, struct vcpu *vcpu); int vm_resume_cpu(struct vm *vm, struct vcpu *vcpu); int vm_restart_instruction(struct vcpu *vcpu); struct vm_exit *vm_exitinfo(struct vcpu *vcpu); void vm_exit_suspended(struct vcpu *vcpu, uint64_t rip); void vm_exit_debug(struct vcpu *vcpu, uint64_t rip); void vm_exit_rendezvous(struct vcpu *vcpu, uint64_t rip); void vm_exit_astpending(struct vcpu *vcpu, uint64_t rip); void vm_exit_reqidle(struct vcpu *vcpu, uint64_t rip); int vm_snapshot_req(struct vm *vm, struct vm_snapshot_meta *meta); int vm_restore_time(struct vm *vm); #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. * * 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 vcpu *vcpu, void *arg); int vm_smp_rendezvous(struct vcpu *vcpu, cpuset_t dest, vm_rendezvous_func_t func, void *arg); cpuset_t vm_active_cpus(struct vm *vm); cpuset_t vm_debug_cpus(struct vm *vm); cpuset_t vm_suspended_cpus(struct vm *vm); cpuset_t vm_start_cpus(struct vm *vm, const cpuset_t *tostart); void vm_await_start(struct vm *vm, const cpuset_t *waiting); #endif /* _SYS__CPUSET_H_ */ static __inline int -vcpu_rendezvous_pending(struct vm_eventinfo *info) +vcpu_rendezvous_pending(struct vcpu *vcpu, struct vm_eventinfo *info) { - - return (*((uintptr_t *)(info->rptr)) != 0); + /* + * This check isn't done with atomic operations or under a lock because + * there's no need to. If the vcpuid bit is set, the vcpu is part of a + * rendezvous and the bit won't be cleared until the vcpu enters the + * rendezvous. On rendezvous exit, the cpuset is cleared and the vcpu + * will see an empty cpuset. So, the races are harmless. + */ + return (CPU_ISSET(vcpu_vcpuid(vcpu), info->rptr)); } static __inline int vcpu_suspended(struct vm_eventinfo *info) { return (*info->sptr); } static __inline int vcpu_reqidle(struct vm_eventinfo *info) { return (*info->iptr); } int vcpu_debugged(struct vcpu *vcpu); /* * Return true if device indicated by bus/slot/func is supposed to be a * pci passthrough device. * * Return false otherwise. */ bool 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 vcpu *vcpu, enum vcpu_state state, bool from_idle); enum vcpu_state vcpu_get_state(struct vcpu *vcpu, int *hostcpu); static int __inline vcpu_is_running(struct vcpu *vcpu, int *hostcpu) { return (vcpu_get_state(vcpu, hostcpu) == VCPU_RUNNING); } #ifdef _SYS_PROC_H_ static int __inline vcpu_should_yield(struct vcpu *vcpu) { struct thread *td; td = curthread; return (td->td_ast != 0 || td->td_owepreempt != 0); } #endif void *vcpu_stats(struct vcpu *vcpu); void vcpu_notify_event(struct vcpu *vcpu, bool lapic_intr); struct vmspace *vm_get_vmspace(struct vm *vm); 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 vcpu *vcpu, 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 vcpu *vcpu, 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 vcpu *vcpu, uint64_t *info); int vm_get_intinfo(struct vcpu *vcpu, uint64_t *info1, uint64_t *info2); /* * Function used to keep track of the guest's TSC offset. The * offset is used by the virutalization extensions to provide a consistent * value for the Time Stamp Counter to the guest. */ void vm_set_tsc_offset(struct vcpu *vcpu, uint64_t offset); 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 Interpretation * 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 vcpu *vcpu, 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_copyinfo *copyinfo, int num_copyinfo); void vm_copyin(struct vm_copyinfo *copyinfo, void *kaddr, size_t len); void vm_copyout(const void *kaddr, struct vm_copyinfo *copyinfo, size_t len); int vcpu_trace_exceptions(struct vcpu *vcpu); int vcpu_trap_wbinvd(struct vcpu *vcpu); #endif /* KERNEL */ /* * 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_BPT_EXIT, VM_CAP_RDPID, VM_CAP_RDTSCP, VM_CAP_IPI_EXIT, 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, PAGING_MODE_64_LA57, }; 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; }; _Static_assert(sizeof(struct vie_op) == 4, "ABI"); _Static_assert(_Alignof(struct vie_op) == 2, "ABI"); #define VIE_INST_SIZE 15 struct vie { uint8_t inst[VIE_INST_SIZE]; /* instruction bytes */ uint8_t num_valid; /* size of the instruction */ /* The following fields are all zeroed upon restart. */ #define vie_startzero num_processed 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 */ vex_present:1, /* VEX prefixed */ vex_l:1, /* L bit */ index:4, /* SIB byte */ base:4; /* SIB byte */ uint8_t disp_bytes; uint8_t imm_bytes; uint8_t scale; uint8_t vex_reg:4, /* vvvv: first source register specifier */ vex_pp:2, /* pp */ _sparebits:2; uint8_t _sparebytes[2]; 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 */ uint8_t _sparebyte; struct vie_op op; /* opcode description */ }; _Static_assert(sizeof(struct vie) == 64, "ABI"); _Static_assert(__offsetof(struct vie, disp_bytes) == 22, "ABI"); _Static_assert(__offsetof(struct vie, scale) == 24, "ABI"); _Static_assert(__offsetof(struct vie, base_register) == 28, "ABI"); 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_DEBUG, VM_EXITCODE_VMINSN, VM_EXITCODE_BPT, VM_EXITCODE_IPI, 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 { int inst_length; } bpt; struct { uint32_t code; /* ecx value */ uint64_t wval; } msr; struct { int vcpu; uint64_t rip; } spinup_ap; struct { uint64_t rflags; uint64_t intr_status; } hlt; struct { int vector; } ioapic_eoi; struct { enum vm_suspend_how how; } suspended; struct { uint32_t mode; uint8_t vector; cpuset_t dmask; } ipi; struct vm_task_switch task_switch; } u; }; /* APIs to inject faults into the guest */ #ifdef _KERNEL void vm_inject_fault(struct vcpu *vcpu, int vector, int errcode_valid, int errcode); static __inline void vm_inject_ud(struct vcpu *vcpu) { vm_inject_fault(vcpu, IDT_UD, 0, 0); } static __inline void vm_inject_gp(struct vcpu *vcpu) { vm_inject_fault(vcpu, IDT_GP, 1, 0); } static __inline void vm_inject_ac(struct vcpu *vcpu, int errcode) { vm_inject_fault(vcpu, IDT_AC, 1, errcode); } static __inline void vm_inject_ss(struct vcpu *vcpu, int errcode) { vm_inject_fault(vcpu, IDT_SS, 1, errcode); } void vm_inject_pf(struct vcpu *vcpu, int error_code, uint64_t cr2); #else 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); #endif #endif /* _VMM_H_ */ diff --git a/sys/amd64/vmm/amd/svm.c b/sys/amd64/vmm/amd/svm.c index 2448501401e3..ee1154ef85b6 100644 --- a/sys/amd64/vmm/amd/svm.c +++ b/sys/amd64/vmm/amd/svm.c @@ -1,2663 +1,2663 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * 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 "opt_bhyve_snapshot.h" #include #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_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 | CTLFLAG_MPSAFE, 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_DR | \ 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"); static uint32_t svm_feature = ~0U; /* AMD SVM features. */ SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, features, CTLFLAG_RDTUN, &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_RDTUN, &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_getdesc(void *vcpui, int reg, struct seg_desc *desc); static int svm_setreg(void *vcpui, 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_modcleanup(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]; /* * The number of ASIDs can be configured to be less than what is * supported by the hardware but not more. */ if (nasid == 0 || nasid > regs[1]) 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_modinit(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_modresume(void) { svm_enable(NULL); } #ifdef BHYVE_SNAPSHOT void svm_set_tsc_offset(struct svm_vcpu *vcpu, uint64_t offset) { struct vmcb_ctrl *ctrl; ctrl = svm_get_vmcb_ctrl(vcpu); ctrl->tsc_offset = offset; svm_set_dirty(vcpu, VMCB_CACHE_I); SVM_CTR1(vcpu, "tsc offset changed to %#lx", offset); vm_set_tsc_offset(vcpu->vcpu, offset); } #endif /* 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 __diagused; 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_vcpu *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(vcpu); return (ctrl->intercept[idx] & bitmask ? 1 : 0); } static __inline void svm_set_intercept(struct svm_vcpu *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(vcpu); oldval = ctrl->intercept[idx]; if (enabled) ctrl->intercept[idx] |= bitmask; else ctrl->intercept[idx] &= ~bitmask; if (ctrl->intercept[idx] != oldval) { svm_set_dirty(vcpu, VMCB_CACHE_I); SVM_CTR3(vcpu, "intercept[%d] modified from %#x to %#x", idx, oldval, ctrl->intercept[idx]); } } static __inline void svm_disable_intercept(struct svm_vcpu *vcpu, int off, uint32_t bitmask) { svm_set_intercept(vcpu, off, bitmask, 0); } static __inline void svm_enable_intercept(struct svm_vcpu *vcpu, int off, uint32_t bitmask) { svm_set_intercept(vcpu, off, bitmask, 1); } static void vmcb_init(struct svm_softc *sc, struct svm_vcpu *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(vcpu); state = svm_get_vmcb_state(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(vcpu, VMCB_CR_INTCPT, mask); else svm_enable_intercept(vcpu, VMCB_CR_INTCPT, mask); } /* * Intercept everything when tracing guest exceptions otherwise * just intercept machine check exception. */ if (vcpu_trace_exceptions(vcpu->vcpu)) { for (n = 0; n < 32; n++) { /* * Skip unimplemented vectors in the exception bitmap. */ if (n == 2 || n == 9) { continue; } svm_enable_intercept(vcpu, VMCB_EXC_INTCPT, BIT(n)); } } else { svm_enable_intercept(vcpu, VMCB_EXC_INTCPT, BIT(IDT_MC)); } /* Intercept various events (for e.g. I/O, MSR and CPUID accesses) */ svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IO); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_MSR); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_CPUID); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INTR); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INIT); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_NMI); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SMI); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SHUTDOWN); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_FERR_FREEZE); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INVD); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INVLPGA); svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MONITOR); svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MWAIT); /* * Intercept SVM instructions since AMD enables them in guests otherwise. * Non-intercepted VMMCALL causes #UD, skip it. */ svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMLOAD); svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMSAVE); svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_STGI); svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_CLGI); svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_SKINIT); svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_ICEBP); if (vcpu_trap_wbinvd(vcpu->vcpu)) { svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_WBINVD); } /* * From section "Canonicalization and Consistency Checks" in APMv2 * the VMRUN intercept bit must be set to pass the consistency check. */ svm_enable_intercept(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); /* Set up DR6/7 to power-on state */ state->dr6 = DBREG_DR6_RESERVED1; state->dr7 = DBREG_DR7_RESERVED1; } /* * Initialize a virtual machine. */ static void * svm_init(struct vm *vm, pmap_t pmap) { struct svm_softc *svm_sc; svm_sc = malloc(sizeof (*svm_sc), M_SVM, M_WAITOK | M_ZERO); svm_sc->msr_bitmap = contigmalloc(SVM_MSR_BITMAP_SIZE, M_SVM, M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0); if (svm_sc->msr_bitmap == NULL) panic("contigmalloc of SVM MSR bitmap failed"); svm_sc->iopm_bitmap = contigmalloc(SVM_IO_BITMAP_SIZE, M_SVM, M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0); if (svm_sc->iopm_bitmap == NULL) panic("contigmalloc of SVM IO bitmap failed"); svm_sc->vm = vm; svm_sc->nptp = vtophys(pmap->pm_pmltop); /* * Intercept read and write accesses to all MSRs. */ memset(svm_sc->msr_bitmap, 0xFF, SVM_MSR_BITMAP_SIZE); /* * 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, SVM_IO_BITMAP_SIZE); return (svm_sc); } static void * svm_vcpu_init(void *vmi, struct vcpu *vcpu1, int vcpuid) { struct svm_softc *sc = vmi; struct svm_vcpu *vcpu; vcpu = malloc(sizeof(*vcpu), M_SVM, M_WAITOK | M_ZERO); vcpu->sc = sc; vcpu->vcpu = vcpu1; vcpu->vcpuid = vcpuid; vcpu->vmcb = malloc_aligned(sizeof(struct vmcb), PAGE_SIZE, M_SVM, M_WAITOK | M_ZERO); vcpu->nextrip = ~0; vcpu->lastcpu = NOCPU; vcpu->vmcb_pa = vtophys(vcpu->vmcb); vmcb_init(sc, vcpu, vtophys(sc->iopm_bitmap), vtophys(sc->msr_bitmap), sc->nptp); svm_msr_guest_init(sc, vcpu); return (vcpu); } /* * 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 __diagused; 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_vcpu *vcpu, int64_t info1, int in, struct vm_inout_str *vis) { int error __diagused, 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 = svm_getdesc(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_vcpu *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(vcpu); ctrl = svm_get_vmcb_ctrl(vcpu); regs = svm_get_guest_regctx(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(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(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 __diagused, 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_vcpu *vcpu, int intr_type, int vector, uint32_t error, bool ec_valid) { struct vmcb_ctrl *ctrl; ctrl = svm_get_vmcb_ctrl(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; SVM_CTR3(vcpu, "Injecting %s at vector %d errcode %#x", intrtype_to_str(intr_type), vector, error); } else { SVM_CTR2(vcpu, "Injecting %s at vector %d", intrtype_to_str(intr_type), vector); } } static void svm_update_virqinfo(struct svm_vcpu *vcpu) { struct vlapic *vlapic; struct vmcb_ctrl *ctrl; vlapic = vm_lapic(vcpu->vcpu); ctrl = svm_get_vmcb_ctrl(vcpu); /* Update %cr8 in the emulated vlapic */ vlapic_set_cr8(vlapic, ctrl->v_tpr); /* Virtual interrupt injection is not used. */ KASSERT(ctrl->v_intr_vector == 0, ("%s: invalid " "v_intr_vector %d", __func__, ctrl->v_intr_vector)); } static void svm_save_intinfo(struct svm_softc *svm_sc, struct svm_vcpu *vcpu) { struct vmcb_ctrl *ctrl; uint64_t intinfo; ctrl = svm_get_vmcb_ctrl(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. */ SVM_CTR2(vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n", intinfo, VMCB_EXITINTINFO_VECTOR(intinfo)); vmm_stat_incr(vcpu->vcpu, VCPU_EXITINTINFO, 1); vm_exit_intinfo(vcpu->vcpu, intinfo); } #ifdef INVARIANTS static __inline int vintr_intercept_enabled(struct svm_vcpu *vcpu) { return (svm_get_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR)); } #endif static __inline void enable_intr_window_exiting(struct svm_vcpu *vcpu) { struct vmcb_ctrl *ctrl; ctrl = svm_get_vmcb_ctrl(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(vcpu), ("%s: vintr intercept should be enabled", __func__)); return; } SVM_CTR0(vcpu, "Enable intr window exiting"); ctrl->v_irq = 1; ctrl->v_ign_tpr = 1; ctrl->v_intr_vector = 0; svm_set_dirty(vcpu, VMCB_CACHE_TPR); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR); } static __inline void disable_intr_window_exiting(struct svm_vcpu *vcpu) { struct vmcb_ctrl *ctrl; ctrl = svm_get_vmcb_ctrl(vcpu); if (!ctrl->v_irq && ctrl->v_intr_vector == 0) { KASSERT(!vintr_intercept_enabled(vcpu), ("%s: vintr intercept should be disabled", __func__)); return; } SVM_CTR0(vcpu, "Disable intr window exiting"); ctrl->v_irq = 0; ctrl->v_intr_vector = 0; svm_set_dirty(vcpu, VMCB_CACHE_TPR); svm_disable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR); } static int svm_modify_intr_shadow(struct svm_vcpu *vcpu, uint64_t val) { struct vmcb_ctrl *ctrl; int oldval, newval; ctrl = svm_get_vmcb_ctrl(vcpu); oldval = ctrl->intr_shadow; newval = val ? 1 : 0; if (newval != oldval) { ctrl->intr_shadow = newval; SVM_CTR1(vcpu, "Setting intr_shadow to %d", newval); } return (0); } static int svm_get_intr_shadow(struct svm_vcpu *vcpu, uint64_t *val) { struct vmcb_ctrl *ctrl; ctrl = svm_get_vmcb_ctrl(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_vcpu *vcpu) { int blocked; blocked = svm_get_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET); return (blocked); } static void enable_nmi_blocking(struct svm_vcpu *vcpu) { KASSERT(!nmi_blocked(vcpu), ("vNMI already blocked")); SVM_CTR0(vcpu, "vNMI blocking enabled"); svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET); } static void clear_nmi_blocking(struct svm_vcpu *vcpu) { int error __diagused; KASSERT(nmi_blocked(vcpu), ("vNMI already unblocked")); SVM_CTR0(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(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(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, struct svm_vcpu *vcpu, uint64_t newval, bool *retu) { struct vm_exit *vme; struct vmcb_state *state; uint64_t changed, lma, oldval; int error __diagused; state = svm_get_vmcb_state(vcpu); oldval = state->efer; SVM_CTR2(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(vcpu->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(vcpu->vcpu); vm_exit_svm(vme, VMCB_EXIT_MSR, 1, 0); *retu = true; return (0); } if (newval & EFER_FFXSR) { if (!vm_cpuid_capability(vcpu->vcpu, VCC_FFXSR)) goto gpf; } if (newval & EFER_TCE) { if (!vm_cpuid_capability(vcpu->vcpu, VCC_TCE)) goto gpf; } error = svm_setreg(vcpu, VM_REG_GUEST_EFER, newval); KASSERT(error == 0, ("%s: error %d updating efer", __func__, error)); return (0); gpf: vm_inject_gp(vcpu->vcpu); return (0); } static int emulate_wrmsr(struct svm_softc *sc, struct svm_vcpu *vcpu, u_int num, uint64_t val, bool *retu) { int error; if (lapic_msr(num)) error = lapic_wrmsr(vcpu->vcpu, num, val, retu); else if (num == MSR_EFER) error = svm_write_efer(sc, vcpu, val, retu); else error = svm_wrmsr(vcpu, num, val, retu); return (error); } static int emulate_rdmsr(struct svm_vcpu *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(vcpu->vcpu, num, &result, retu); else error = svm_rdmsr(vcpu, num, &result, retu); if (error == 0) { state = svm_get_vmcb_state(vcpu); ctx = svm_get_guest_regctx(vcpu); state->rax = result & 0xffffffff; ctx->sctx_rdx = result >> 32; } return (error); } #ifdef KTR static const char * exit_reason_to_str(uint64_t reason) { int i; static char reasonbuf[32]; static const struct { int reason; const char *str; } reasons[] = { { .reason = VMCB_EXIT_INVALID, .str = "invalvmcb" }, { .reason = VMCB_EXIT_SHUTDOWN, .str = "shutdown" }, { .reason = VMCB_EXIT_NPF, .str = "nptfault" }, { .reason = VMCB_EXIT_PAUSE, .str = "pause" }, { .reason = VMCB_EXIT_HLT, .str = "hlt" }, { .reason = VMCB_EXIT_CPUID, .str = "cpuid" }, { .reason = VMCB_EXIT_IO, .str = "inout" }, { .reason = VMCB_EXIT_MC, .str = "mchk" }, { .reason = VMCB_EXIT_INTR, .str = "extintr" }, { .reason = VMCB_EXIT_NMI, .str = "nmi" }, { .reason = VMCB_EXIT_VINTR, .str = "vintr" }, { .reason = VMCB_EXIT_MSR, .str = "msr" }, { .reason = VMCB_EXIT_IRET, .str = "iret" }, { .reason = VMCB_EXIT_MONITOR, .str = "monitor" }, { .reason = VMCB_EXIT_MWAIT, .str = "mwait" }, { .reason = VMCB_EXIT_VMRUN, .str = "vmrun" }, { .reason = VMCB_EXIT_VMMCALL, .str = "vmmcall" }, { .reason = VMCB_EXIT_VMLOAD, .str = "vmload" }, { .reason = VMCB_EXIT_VMSAVE, .str = "vmsave" }, { .reason = VMCB_EXIT_STGI, .str = "stgi" }, { .reason = VMCB_EXIT_CLGI, .str = "clgi" }, { .reason = VMCB_EXIT_SKINIT, .str = "skinit" }, { .reason = VMCB_EXIT_ICEBP, .str = "icebp" }, { .reason = VMCB_EXIT_INVD, .str = "invd" }, { .reason = VMCB_EXIT_INVLPGA, .str = "invlpga" }, }; for (i = 0; i < nitems(reasons); i++) { if (reasons[i].reason == reason) return (reasons[i].str); } 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, struct svm_vcpu *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 __diagused, errcode_valid, handled, idtvec, reflect; bool retu; ctx = svm_get_guest_regctx(vcpu); vmcb = svm_get_vmcb(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(vcpu->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(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(vcpu); handled = 1; break; case VMCB_EXIT_VINTR: /* interrupt window exiting */ vmm_stat_incr(vcpu->vcpu, VMEXIT_VINTR, 1); handled = 1; break; case VMCB_EXIT_INTR: /* external interrupt */ vmm_stat_incr(vcpu->vcpu, VMEXIT_EXTINT, 1); handled = 1; break; case VMCB_EXIT_NMI: /* external NMI */ handled = 1; break; case 0x40 ... 0x5F: vmm_stat_incr(vcpu->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; SVM_CTR0(vcpu, "Vectoring to MCE handler"); __asm __volatile("int $18"); break; case IDT_PF: error = svm_setreg(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. */ SVM_CTR2(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 */ SVM_CTR2(vcpu, "Reflecting exception " "%d/%#x into the guest", idtvec, (int)info1); error = vm_inject_exception(vcpu->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(vcpu->vcpu, VMEXIT_WRMSR, 1); val = (uint64_t)edx << 32 | eax; SVM_CTR2(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 { SVM_CTR1(vcpu, "rdmsr %#x", ecx); vmm_stat_incr(vcpu->vcpu, VMEXIT_RDMSR, 1); if (emulate_rdmsr(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(vcpu, vmexit); vmm_stat_incr(vcpu->vcpu, VMEXIT_INOUT, 1); break; case VMCB_EXIT_CPUID: vmm_stat_incr(vcpu->vcpu, VMEXIT_CPUID, 1); handled = x86_emulate_cpuid(vcpu->vcpu, &state->rax, &ctx->sctx_rbx, &ctx->sctx_rcx, &ctx->sctx_rdx); break; case VMCB_EXIT_HLT: vmm_stat_incr(vcpu->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(vcpu->vcpu, VMEXIT_PAUSE, 1); break; case VMCB_EXIT_NPF: /* EXITINFO2 contains the faulting guest physical address */ if (info1 & VMCB_NPF_INFO1_RSV) { SVM_CTR2(vcpu, "nested page fault with " "reserved bits set: info1(%#lx) info2(%#lx)", info1, info2); } else if (vm_mem_allocated(vcpu->vcpu, info2)) { vmexit->exitcode = VM_EXITCODE_PAGING; vmexit->u.paging.gpa = info2; vmexit->u.paging.fault_type = npf_fault_type(info1); vmm_stat_incr(vcpu->vcpu, VMEXIT_NESTED_FAULT, 1); SVM_CTR3(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(vcpu->vcpu, VMEXIT_INST_EMUL, 1); SVM_CTR3(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; case VMCB_EXIT_SHUTDOWN: case VMCB_EXIT_VMRUN: case VMCB_EXIT_VMMCALL: case VMCB_EXIT_VMLOAD: case VMCB_EXIT_VMSAVE: case VMCB_EXIT_STGI: case VMCB_EXIT_CLGI: case VMCB_EXIT_SKINIT: case VMCB_EXIT_ICEBP: case VMCB_EXIT_INVLPGA: vm_inject_ud(vcpu->vcpu); handled = 1; break; case VMCB_EXIT_INVD: case VMCB_EXIT_WBINVD: /* ignore exit */ handled = 1; break; default: vmm_stat_incr(vcpu->vcpu, VMEXIT_UNKNOWN, 1); break; } SVM_CTR4(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, struct svm_vcpu *vcpu) { uint64_t intinfo; if (!vm_entry_intinfo(vcpu->vcpu, &intinfo)) return; KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not " "valid: %#lx", __func__, intinfo)); svm_eventinject(vcpu, VMCB_EXITINTINFO_TYPE(intinfo), VMCB_EXITINTINFO_VECTOR(intinfo), VMCB_EXITINTINFO_EC(intinfo), VMCB_EXITINTINFO_EC_VALID(intinfo)); vmm_stat_incr(vcpu->vcpu, VCPU_INTINFO_INJECTED, 1); SVM_CTR1(vcpu, "Injected entry intinfo: %#lx", intinfo); } /* * Inject event to virtual cpu. */ static void svm_inj_interrupts(struct svm_softc *sc, struct svm_vcpu *vcpu, struct vlapic *vlapic) { struct vmcb_ctrl *ctrl; struct vmcb_state *state; uint8_t v_tpr; int vector, need_intr_window; int extint_pending; state = svm_get_vmcb_state(vcpu); ctrl = svm_get_vmcb_ctrl(vcpu); need_intr_window = 0; if (vcpu->nextrip != state->rip) { ctrl->intr_shadow = 0; SVM_CTR2(vcpu, "Guest interrupt blocking " "cleared due to rip change: %#lx/%#lx", vcpu->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(vcpu->vcpu)) { if (nmi_blocked(vcpu)) { /* * Can't inject another NMI if the guest has not * yet executed an "iret" after the last NMI. */ SVM_CTR0(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. */ SVM_CTR0(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. */ SVM_CTR1(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(vcpu->vcpu); /* Inject NMI, vector number is not used */ svm_eventinject(vcpu, VMCB_EVENTINJ_TYPE_NMI, IDT_NMI, 0, false); /* virtual NMI blocking is now in effect */ enable_nmi_blocking(vcpu); SVM_CTR0(vcpu, "Injecting vNMI"); } } extint_pending = vm_extint_pending(vcpu->vcpu); if (!extint_pending) { if (!vlapic_pending_intr(vlapic, &vector)) goto done; 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(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) { SVM_CTR2(vcpu, "Cannot inject vector %d due to " "rflags %#lx", vector, state->rflags); need_intr_window = 1; goto done; } if (ctrl->intr_shadow) { SVM_CTR1(vcpu, "Cannot inject vector %d due to " "interrupt shadow", vector); need_intr_window = 1; goto done; } if (ctrl->eventinj & VMCB_EVENTINJ_VALID) { SVM_CTR2(vcpu, "Cannot inject vector %d due to " "eventinj %#lx", vector, ctrl->eventinj); need_intr_window = 1; goto done; } svm_eventinject(vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false); if (!extint_pending) { vlapic_intr_accepted(vlapic, vector); } else { vm_extint_clear(vcpu->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) { SVM_CTR2(vcpu, "VMCB V_TPR changed from %#x to %#x", ctrl->v_tpr, v_tpr); ctrl->v_tpr = v_tpr; svm_set_dirty(vcpu, VMCB_CACHE_TPR); } 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(vcpu); } else { disable_intr_window_exiting(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 svm_pmap_activate(struct svm_vcpu *vcpu, pmap_t pmap) { struct vmcb_ctrl *ctrl; long eptgen; int cpu; bool alloc_asid; cpu = curcpu; CPU_SET_ATOMIC(cpu, &pmap->pm_active); smr_enter(pmap->pm_eptsmr); ctrl = svm_get_vmcb_ctrl(vcpu); /* * 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 = atomic_load_long(&pmap->pm_eptgen); ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING; if (vcpu->asid.gen != asid[cpu].gen) { alloc_asid = true; /* (c) and (d) */ } else if (vcpu->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[cpu].num >= nasid) { asid[cpu].num = 1; if (++asid[cpu].gen == 0) asid[cpu].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; } vcpu->asid.gen = asid[cpu].gen; vcpu->asid.num = asid[cpu].num; ctrl->asid = vcpu->asid.num; svm_set_dirty(vcpu, 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; } vcpu->eptgen = eptgen; KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero")); KASSERT(ctrl->asid == vcpu->asid.num, ("ASID mismatch: %u/%u", ctrl->asid, vcpu->asid.num)); } static void svm_pmap_deactivate(pmap_t pmap) { smr_exit(pmap->pm_eptsmr); CPU_CLR_ATOMIC(curcpu, &pmap->pm_active); } static __inline void disable_gintr(void) { __asm __volatile("clgi"); } static __inline void enable_gintr(void) { __asm __volatile("stgi"); } static __inline void svm_dr_enter_guest(struct svm_regctx *gctx) { /* Save host control debug registers. */ gctx->host_dr7 = rdr7(); gctx->host_debugctl = rdmsr(MSR_DEBUGCTLMSR); /* * Disable debugging in DR7 and DEBUGCTL to avoid triggering * exceptions in the host based on the guest DRx values. The * guest DR6, DR7, and DEBUGCTL are saved/restored in the * VMCB. */ load_dr7(0); wrmsr(MSR_DEBUGCTLMSR, 0); /* Save host debug registers. */ gctx->host_dr0 = rdr0(); gctx->host_dr1 = rdr1(); gctx->host_dr2 = rdr2(); gctx->host_dr3 = rdr3(); gctx->host_dr6 = rdr6(); /* Restore guest debug registers. */ load_dr0(gctx->sctx_dr0); load_dr1(gctx->sctx_dr1); load_dr2(gctx->sctx_dr2); load_dr3(gctx->sctx_dr3); } static __inline void svm_dr_leave_guest(struct svm_regctx *gctx) { /* Save guest debug registers. */ gctx->sctx_dr0 = rdr0(); gctx->sctx_dr1 = rdr1(); gctx->sctx_dr2 = rdr2(); gctx->sctx_dr3 = rdr3(); /* * Restore host debug registers. Restore DR7 and DEBUGCTL * last. */ load_dr0(gctx->host_dr0); load_dr1(gctx->host_dr1); load_dr2(gctx->host_dr2); load_dr3(gctx->host_dr3); load_dr6(gctx->host_dr6); wrmsr(MSR_DEBUGCTLMSR, gctx->host_debugctl); load_dr7(gctx->host_dr7); } /* * Start vcpu with specified RIP. */ static int svm_run(void *vcpui, register_t rip, pmap_t pmap, struct vm_eventinfo *evinfo) { struct svm_regctx *gctx; struct svm_softc *svm_sc; struct svm_vcpu *vcpu; struct vmcb_state *state; struct vmcb_ctrl *ctrl; struct vm_exit *vmexit; struct vlapic *vlapic; uint64_t vmcb_pa; int handled; uint16_t ldt_sel; vcpu = vcpui; svm_sc = vcpu->sc; state = svm_get_vmcb_state(vcpu); ctrl = svm_get_vmcb_ctrl(vcpu); vmexit = vm_exitinfo(vcpu->vcpu); vlapic = vm_lapic(vcpu->vcpu); gctx = svm_get_guest_regctx(vcpu); vmcb_pa = vcpu->vmcb_pa; if (vcpu->lastcpu != curcpu) { /* * Force new ASID allocation by invalidating the generation. */ vcpu->asid.gen = 0; /* * Invalidate the VMCB state cache by marking all fields dirty. */ svm_set_dirty(vcpu, 0xffffffff); /* * XXX * Setting 'vcpu->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. */ vcpu->lastcpu = curcpu; vmm_stat_incr(vcpu->vcpu, VCPU_MIGRATIONS, 1); } svm_msr_guest_enter(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(evinfo)) { enable_gintr(); vm_exit_suspended(vcpu->vcpu, state->rip); break; } - if (vcpu_rendezvous_pending(evinfo)) { + if (vcpu_rendezvous_pending(vcpu->vcpu, evinfo)) { enable_gintr(); vm_exit_rendezvous(vcpu->vcpu, state->rip); break; } if (vcpu_reqidle(evinfo)) { enable_gintr(); vm_exit_reqidle(vcpu->vcpu, state->rip); break; } /* We are asked to give the cpu by scheduler. */ if (vcpu_should_yield(vcpu->vcpu)) { enable_gintr(); vm_exit_astpending(vcpu->vcpu, state->rip); break; } if (vcpu_debugged(vcpu->vcpu)) { enable_gintr(); vm_exit_debug(vcpu->vcpu, state->rip); break; } /* * #VMEXIT resumes the host with the guest LDTR, so * save the current LDT selector so it can be restored * after an exit. The userspace hypervisor probably * doesn't use a LDT, but save and restore it to be * safe. */ ldt_sel = sldt(); svm_inj_interrupts(svm_sc, vcpu, vlapic); /* * Check the pmap generation and the ASID generation to * ensure that the vcpu does not use stale TLB mappings. */ svm_pmap_activate(vcpu, pmap); ctrl->vmcb_clean = vmcb_clean & ~vcpu->dirty; vcpu->dirty = 0; SVM_CTR1(vcpu, "vmcb clean %#x", ctrl->vmcb_clean); /* Launch Virtual Machine. */ SVM_CTR1(vcpu, "Resume execution at %#lx", state->rip); svm_dr_enter_guest(gctx); svm_launch(vmcb_pa, gctx, get_pcpu()); svm_dr_leave_guest(gctx); svm_pmap_deactivate(pmap); /* * 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(); /* Restore host LDTR. */ lldt(ldt_sel); /* #VMEXIT disables interrupts so re-enable them here. */ enable_gintr(); /* Update 'nextrip' */ vcpu->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(vcpu); return (0); } static void svm_vcpu_cleanup(void *vcpui) { struct svm_vcpu *vcpu = vcpui; free(vcpu->vmcb, M_SVM); free(vcpu, M_SVM); } static void svm_cleanup(void *vmi) { struct svm_softc *sc = vmi; contigfree(sc->iopm_bitmap, SVM_IO_BITMAP_SIZE, M_SVM); contigfree(sc->msr_bitmap, SVM_MSR_BITMAP_SIZE, M_SVM); 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); case VM_REG_GUEST_DR0: return (®ctx->sctx_dr0); case VM_REG_GUEST_DR1: return (®ctx->sctx_dr1); case VM_REG_GUEST_DR2: return (®ctx->sctx_dr2); case VM_REG_GUEST_DR3: return (®ctx->sctx_dr3); default: return (NULL); } } static int svm_getreg(void *vcpui, int ident, uint64_t *val) { struct svm_vcpu *vcpu; register_t *reg; vcpu = vcpui; if (ident == VM_REG_GUEST_INTR_SHADOW) { return (svm_get_intr_shadow(vcpu, val)); } if (vmcb_read(vcpu, ident, val) == 0) { return (0); } reg = swctx_regptr(svm_get_guest_regctx(vcpu), ident); if (reg != NULL) { *val = *reg; return (0); } SVM_CTR1(vcpu, "svm_getreg: unknown register %#x", ident); return (EINVAL); } static int svm_setreg(void *vcpui, int ident, uint64_t val) { struct svm_vcpu *vcpu; register_t *reg; vcpu = vcpui; if (ident == VM_REG_GUEST_INTR_SHADOW) { return (svm_modify_intr_shadow(vcpu, val)); } /* Do not permit user write access to VMCB fields by offset. */ if (!VMCB_ACCESS_OK(ident)) { if (vmcb_write(vcpu, ident, val) == 0) { return (0); } } reg = swctx_regptr(svm_get_guest_regctx(vcpu), ident); if (reg != NULL) { *reg = val; return (0); } if (ident == VM_REG_GUEST_ENTRY_INST_LENGTH) { /* Ignore. */ 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. */ SVM_CTR1(vcpu, "svm_setreg: unknown register %#x", ident); return (EINVAL); } static int svm_getdesc(void *vcpui, int reg, struct seg_desc *desc) { return (vmcb_getdesc(vcpui, reg, desc)); } static int svm_setdesc(void *vcpui, int reg, struct seg_desc *desc) { return (vmcb_setdesc(vcpui, reg, desc)); } #ifdef BHYVE_SNAPSHOT static int svm_snapshot_reg(void *vcpui, int ident, struct vm_snapshot_meta *meta) { int ret; uint64_t val; if (meta->op == VM_SNAPSHOT_SAVE) { ret = svm_getreg(vcpui, ident, &val); if (ret != 0) goto done; SNAPSHOT_VAR_OR_LEAVE(val, meta, ret, done); } else if (meta->op == VM_SNAPSHOT_RESTORE) { SNAPSHOT_VAR_OR_LEAVE(val, meta, ret, done); ret = svm_setreg(vcpui, ident, val); if (ret != 0) goto done; } else { ret = EINVAL; goto done; } done: return (ret); } #endif static int svm_setcap(void *vcpui, int type, int val) { struct svm_vcpu *vcpu; struct vlapic *vlapic; int error; vcpu = vcpui; error = 0; switch (type) { case VM_CAP_HALT_EXIT: svm_set_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_HLT, val); break; case VM_CAP_PAUSE_EXIT: svm_set_intercept(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; case VM_CAP_IPI_EXIT: vlapic = vm_lapic(vcpu->vcpu); vlapic->ipi_exit = val; break; default: error = ENOENT; break; } return (error); } static int svm_getcap(void *vcpui, int type, int *retval) { struct svm_vcpu *vcpu; struct vlapic *vlapic; int error; vcpu = vcpui; error = 0; switch (type) { case VM_CAP_HALT_EXIT: *retval = svm_get_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_HLT); break; case VM_CAP_PAUSE_EXIT: *retval = svm_get_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_PAUSE); break; case VM_CAP_UNRESTRICTED_GUEST: *retval = 1; /* unrestricted guest is always enabled */ break; case VM_CAP_IPI_EXIT: vlapic = vm_lapic(vcpu->vcpu); *retval = vlapic->ipi_exit; break; default: error = ENOENT; break; } return (error); } static struct vmspace * svm_vmspace_alloc(vm_offset_t min, vm_offset_t max) { return (svm_npt_alloc(min, max)); } static void svm_vmspace_free(struct vmspace *vmspace) { svm_npt_free(vmspace); } static struct vlapic * svm_vlapic_init(void *vcpui) { struct svm_vcpu *vcpu; struct vlapic *vlapic; vcpu = vcpui; vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO); vlapic->vm = vcpu->sc->vm; vlapic->vcpu = vcpu->vcpu; vlapic->vcpuid = vcpu->vcpuid; vlapic->apic_page = malloc_aligned(PAGE_SIZE, PAGE_SIZE, M_SVM_VLAPIC, M_WAITOK | M_ZERO); vlapic_init(vlapic); return (vlapic); } static void svm_vlapic_cleanup(struct vlapic *vlapic) { vlapic_cleanup(vlapic); free(vlapic->apic_page, M_SVM_VLAPIC); free(vlapic, M_SVM_VLAPIC); } #ifdef BHYVE_SNAPSHOT static int svm_snapshot(void *vmi, struct vm_snapshot_meta *meta) { if (meta->op == VM_SNAPSHOT_RESTORE) flush_by_asid(); return (0); } static int svm_vcpu_snapshot(void *vcpui, struct vm_snapshot_meta *meta) { struct svm_vcpu *vcpu; int err, running, hostcpu; vcpu = vcpui; err = 0; running = vcpu_is_running(vcpu->vcpu, &hostcpu); if (running && hostcpu != curcpu) { printf("%s: %s%d is running", __func__, vm_name(vcpu->sc->vm), vcpu->vcpuid); return (EINVAL); } err += svm_snapshot_reg(vcpu, VM_REG_GUEST_CR0, meta); err += svm_snapshot_reg(vcpu, VM_REG_GUEST_CR2, meta); err += svm_snapshot_reg(vcpu, VM_REG_GUEST_CR3, meta); err += svm_snapshot_reg(vcpu, VM_REG_GUEST_CR4, meta); err += svm_snapshot_reg(vcpu, VM_REG_GUEST_DR6, meta); err += svm_snapshot_reg(vcpu, VM_REG_GUEST_DR7, meta); err += svm_snapshot_reg(vcpu, VM_REG_GUEST_RAX, meta); err += svm_snapshot_reg(vcpu, VM_REG_GUEST_RSP, meta); err += svm_snapshot_reg(vcpu, VM_REG_GUEST_RIP, meta); err += svm_snapshot_reg(vcpu, VM_REG_GUEST_RFLAGS, meta); /* Guest segments */ /* ES */ err += svm_snapshot_reg(vcpu, VM_REG_GUEST_ES, meta); err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_ES, meta); /* CS */ err += svm_snapshot_reg(vcpu, VM_REG_GUEST_CS, meta); err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_CS, meta); /* SS */ err += svm_snapshot_reg(vcpu, VM_REG_GUEST_SS, meta); err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_SS, meta); /* DS */ err += svm_snapshot_reg(vcpu, VM_REG_GUEST_DS, meta); err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_DS, meta); /* FS */ err += svm_snapshot_reg(vcpu, VM_REG_GUEST_FS, meta); err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_FS, meta); /* GS */ err += svm_snapshot_reg(vcpu, VM_REG_GUEST_GS, meta); err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_GS, meta); /* TR */ err += svm_snapshot_reg(vcpu, VM_REG_GUEST_TR, meta); err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_TR, meta); /* LDTR */ err += svm_snapshot_reg(vcpu, VM_REG_GUEST_LDTR, meta); err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_LDTR, meta); /* EFER */ err += svm_snapshot_reg(vcpu, VM_REG_GUEST_EFER, meta); /* IDTR and GDTR */ err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_IDTR, meta); err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_GDTR, meta); /* Specific AMD registers */ err += svm_snapshot_reg(vcpu, VM_REG_GUEST_INTR_SHADOW, meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_CR_INTERCEPT, 4), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_DR_INTERCEPT, 4), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_EXC_INTERCEPT, 4), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_INST1_INTERCEPT, 4), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_INST2_INTERCEPT, 4), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_PAUSE_FILTHRESH, 2), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_PAUSE_FILCNT, 2), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_ASID, 4), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_TLB_CTRL, 4), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_VIRQ, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_EXIT_REASON, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_EXITINFO1, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_EXITINFO2, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_EXITINTINFO, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_NP_ENABLE, 1), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_AVIC_BAR, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_AVIC_PAGE, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_AVIC_LT, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_AVIC_PT, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_CPL, 1), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_STAR, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_LSTAR, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_CSTAR, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_SFMASK, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_KERNELGBASE, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_SYSENTER_CS, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_SYSENTER_ESP, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_SYSENTER_EIP, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_GUEST_PAT, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_DBGCTL, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_BR_FROM, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_BR_TO, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_INT_FROM, 8), meta); err += vmcb_snapshot_any(vcpu, VMCB_ACCESS(VMCB_OFF_INT_TO, 8), meta); if (err != 0) goto done; /* Snapshot swctx for virtual cpu */ SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rbp, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rbx, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rcx, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rdx, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rdi, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rsi, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r8, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r9, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r10, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r11, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r12, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r13, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r14, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r15, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr0, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr1, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr2, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr3, meta, err, done); /* Restore other svm_vcpu struct fields */ /* Restore NEXTRIP field */ SNAPSHOT_VAR_OR_LEAVE(vcpu->nextrip, meta, err, done); /* Restore lastcpu field */ SNAPSHOT_VAR_OR_LEAVE(vcpu->lastcpu, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->dirty, meta, err, done); /* Restore EPTGEN field - EPT is Extended Page Table */ SNAPSHOT_VAR_OR_LEAVE(vcpu->eptgen, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->asid.gen, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->asid.num, meta, err, done); /* Set all caches dirty */ if (meta->op == VM_SNAPSHOT_RESTORE) svm_set_dirty(vcpu, 0xffffffff); done: return (err); } static int svm_restore_tsc(void *vcpui, uint64_t offset) { struct svm_vcpu *vcpu = vcpui; svm_set_tsc_offset(vcpu, offset); return (0); } #endif const struct vmm_ops vmm_ops_amd = { .modinit = svm_modinit, .modcleanup = svm_modcleanup, .modresume = svm_modresume, .init = svm_init, .run = svm_run, .cleanup = svm_cleanup, .vcpu_init = svm_vcpu_init, .vcpu_cleanup = svm_vcpu_cleanup, .getreg = svm_getreg, .setreg = svm_setreg, .getdesc = svm_getdesc, .setdesc = svm_setdesc, .getcap = svm_getcap, .setcap = svm_setcap, .vmspace_alloc = svm_vmspace_alloc, .vmspace_free = svm_vmspace_free, .vlapic_init = svm_vlapic_init, .vlapic_cleanup = svm_vlapic_cleanup, #ifdef BHYVE_SNAPSHOT .snapshot = svm_snapshot, .vcpu_snapshot = svm_vcpu_snapshot, .restore_tsc = svm_restore_tsc, #endif }; diff --git a/sys/amd64/vmm/intel/vmx.c b/sys/amd64/vmm/intel/vmx.c index 4b65e254cc91..fa94c707001c 100644 --- a/sys/amd64/vmm/intel/vmx.c +++ b/sys/amd64/vmm/intel/vmx.c @@ -1,4261 +1,4261 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2011 NetApp, Inc. * All rights reserved. * Copyright (c) 2018 Joyent, Inc. * * 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 "opt_bhyve_snapshot.h" #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_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_SAVE_DEBUG_CONTROLS | \ VM_EXIT_HOST_LMA | \ VM_EXIT_SAVE_EFER | \ VM_EXIT_LOAD_EFER | \ VM_EXIT_ACKNOWLEDGE_INTERRUPT) #define VM_EXIT_CTLS_ZERO_SETTING 0 #define VM_ENTRY_CTLS_ONE_SETTING \ (VM_ENTRY_LOAD_DEBUG_CONTROLS | \ VM_ENTRY_LOAD_EFER) #define VM_ENTRY_CTLS_ZERO_SETTING \ (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"); bool vmx_have_msr_tsc_aux; SYSCTL_DECL(_hw_vmm); SYSCTL_NODE(_hw_vmm, OID_AUTO, vmx, CTLFLAG_RW | CTLFLAG_MPSAFE, 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 | CTLFLAG_MPSAFE, 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_wbinvd_exit; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, wbinvd_exit, CTLFLAG_RD, &cap_wbinvd_exit, 0, "WBINVD triggers a VM-exit"); static int cap_rdpid; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, rdpid, CTLFLAG_RD, &cap_rdpid, 0, "Guests are allowed to use RDPID"); static int cap_rdtscp; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, rdtscp, CTLFLAG_RD, &cap_rdtscp, 0, "Guests are allowed to use RDTSCP"); 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 tpr_shadowing; SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, tpr_shadowing, CTLFLAG_RD, &tpr_shadowing, 0, "TPR shadowing support"); 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); int guest_l1d_flush; SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, l1d_flush, CTLFLAG_RD, &guest_l1d_flush, 0, NULL); int guest_l1d_flush_sw; SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, l1d_flush_sw, CTLFLAG_RD, &guest_l1d_flush_sw, 0, NULL); static struct msr_entry msr_load_list[1] __aligned(16); /* * The definitions of SDT probes for VMX. */ SDT_PROBE_DEFINE3(vmm, vmx, exit, entry, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE4(vmm, vmx, exit, taskswitch, "struct vmx *", "int", "struct vm_exit *", "struct vm_task_switch *"); SDT_PROBE_DEFINE4(vmm, vmx, exit, craccess, "struct vmx *", "int", "struct vm_exit *", "uint64_t"); SDT_PROBE_DEFINE4(vmm, vmx, exit, rdmsr, "struct vmx *", "int", "struct vm_exit *", "uint32_t"); SDT_PROBE_DEFINE5(vmm, vmx, exit, wrmsr, "struct vmx *", "int", "struct vm_exit *", "uint32_t", "uint64_t"); SDT_PROBE_DEFINE3(vmm, vmx, exit, halt, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, mtrap, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, pause, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, intrwindow, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE4(vmm, vmx, exit, interrupt, "struct vmx *", "int", "struct vm_exit *", "uint32_t"); SDT_PROBE_DEFINE3(vmm, vmx, exit, nmiwindow, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, inout, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, cpuid, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE5(vmm, vmx, exit, exception, "struct vmx *", "int", "struct vm_exit *", "uint32_t", "int"); SDT_PROBE_DEFINE5(vmm, vmx, exit, nestedfault, "struct vmx *", "int", "struct vm_exit *", "uint64_t", "uint64_t"); SDT_PROBE_DEFINE4(vmm, vmx, exit, mmiofault, "struct vmx *", "int", "struct vm_exit *", "uint64_t"); SDT_PROBE_DEFINE3(vmm, vmx, exit, eoi, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, apicaccess, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE4(vmm, vmx, exit, apicwrite, "struct vmx *", "int", "struct vm_exit *", "struct vlapic *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, xsetbv, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, monitor, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, mwait, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE3(vmm, vmx, exit, vminsn, "struct vmx *", "int", "struct vm_exit *"); SDT_PROBE_DEFINE4(vmm, vmx, exit, unknown, "struct vmx *", "int", "struct vm_exit *", "uint32_t"); SDT_PROBE_DEFINE4(vmm, vmx, exit, return, "struct vmx *", "int", "struct vm_exit *", "int"); /* * 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 *vcpui, int reg, struct seg_desc *desc); static int vmx_getreg(void *vcpui, 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 BHYVE_SNAPSHOT static int vmx_restore_tsc(void *vcpui, uint64_t now); #endif static inline bool host_has_rdpid(void) { return ((cpu_stdext_feature2 & CPUID_STDEXT2_RDPID) != 0); } static inline bool host_has_rdtscp(void) { return ((amd_feature & AMDID_RDTSCP) != 0); } #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 uint16_t vpid_alloc(int vcpuid) { int x; /* * 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) return (0); /* * Try to allocate a unique VPID for each from the unit number * allocator. */ x = alloc_unr(vpid_unr); if (x == -1) { 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. */ return (vcpuid + 1); } return (x); } 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_modcleanup(void) { if (pirvec >= 0) lapic_ipi_free(pirvec); if (vpid_unr != NULL) { delete_unrhdr(vpid_unr); vpid_unr = NULL; } if (nmi_flush_l1d_sw == 1) nmi_flush_l1d_sw = 0; 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_modresume(void) { if (vmxon_enabled[curcpu]) vmxon(vmxon_region[curcpu]); } static int vmx_modinit(int ipinum) { int error; 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_modinit: 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_modinit: 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_modinit: 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_modinit: 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_modinit: 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_modinit: 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_modinit: 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_modinit: 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_wbinvd_exit = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED2_WBINVD_EXITING, 0, &tmp) == 0); /* * Check support for RDPID and/or RDTSCP. * * Support a pass-through-based implementation of these via the * "enable RDTSCP" VM-execution control and the "RDTSC exiting" * VM-execution control. * * The "enable RDTSCP" VM-execution control applies to both RDPID * and RDTSCP (see SDM volume 3, section 25.3, "Changes to * Instruction Behavior in VMX Non-root operation"); this is why * only this VM-execution control needs to be enabled in order to * enable passing through whichever of RDPID and/or RDTSCP are * supported by the host. * * The "RDTSC exiting" VM-execution control applies to both RDTSC * and RDTSCP (again, per SDM volume 3, section 25.3), and is * already set up for RDTSC and RDTSCP pass-through by the current * implementation of RDTSC. * * Although RDPID and RDTSCP are optional capabilities, since there * does not currently seem to be a use case for enabling/disabling * these via libvmmapi, choose not to support this and, instead, * just statically always enable or always disable this support * across all vCPUs on all VMs. (Note that there may be some * complications to providing this functionality, e.g., the MSR * bitmap is currently per-VM rather than per-vCPU while the * capability API wants to be able to control capabilities on a * per-vCPU basis). */ error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED2_ENABLE_RDTSCP, 0, &tmp); cap_rdpid = error == 0 && host_has_rdpid(); cap_rdtscp = error == 0 && host_has_rdtscp(); if (cap_rdpid || cap_rdtscp) { procbased_ctls2 |= PROCBASED2_ENABLE_RDTSCP; vmx_have_msr_tsc_aux = true; } 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 TPR shadow. */ error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_USE_TPR_SHADOW, 0, &tmp); if (error == 0) { tpr_shadowing = 1; TUNABLE_INT_FETCH("hw.vmm.vmx.use_tpr_shadowing", &tpr_shadowing); } if (tpr_shadowing) { procbased_ctls |= PROCBASED_USE_TPR_SHADOW; procbased_ctls &= ~PROCBASED_CR8_LOAD_EXITING; procbased_ctls &= ~PROCBASED_CR8_STORE_EXITING; } /* * 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); error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, procbased2_vid_bits, 0, &tmp); if (error == 0 && tpr_shadowing) { 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; /* * 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(pti ? &IDTVEC(justreturn1_pti) : &IDTVEC(justreturn)); if (pirvec < 0) { if (bootverbose) { printf("vmx_modinit: 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_modinit: ept initialization failed (%d)\n", error); return (error); } guest_l1d_flush = (cpu_ia32_arch_caps & IA32_ARCH_CAP_SKIP_L1DFL_VMENTRY) == 0; TUNABLE_INT_FETCH("hw.vmm.l1d_flush", &guest_l1d_flush); /* * L1D cache flush is enabled. Use IA32_FLUSH_CMD MSR when * available. Otherwise fall back to the software flush * method which loads enough data from the kernel text to * flush existing L1D content, both on VMX entry and on NMI * return. */ if (guest_l1d_flush) { if ((cpu_stdext_feature3 & CPUID_STDEXT3_L1D_FLUSH) == 0) { guest_l1d_flush_sw = 1; TUNABLE_INT_FETCH("hw.vmm.l1d_flush_sw", &guest_l1d_flush_sw); } if (guest_l1d_flush_sw) { if (nmi_flush_l1d_sw <= 1) nmi_flush_l1d_sw = 1; } else { msr_load_list[0].index = MSR_IA32_FLUSH_CMD; msr_load_list[0].val = IA32_FLUSH_CMD_L1D; } } /* * 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_init(struct vm *vm, pmap_t pmap) { int error __diagused; struct vmx *vmx; vmx = malloc(sizeof(struct vmx), M_VMX, M_WAITOK | M_ZERO); vmx->vm = vm; vmx->eptp = eptp(vtophys((vm_offset_t)pmap->pm_pmltop)); /* * 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); vmx->msr_bitmap = malloc_aligned(PAGE_SIZE, PAGE_SIZE, M_VMX, M_WAITOK | M_ZERO); 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. If the guest does a write * the "use TSC offsetting" execution control is enabled and the * difference between the host TSC and the guest TSC is written * into the TSC offset in the VMCS. * * Guest TSC_AUX support is enabled if any of guest RDPID and/or * guest RDTSCP support are enabled (since, as per Table 2-2 in SDM * volume 4, TSC_AUX is supported if any of RDPID and/or RDTSCP are * supported). If guest TSC_AUX support is enabled, TSC_AUX is * exposed read-only so that the VMM can do one fewer MSR read per * exit than if this register were exposed read-write; the guest * restore value can be updated during guest writes (expected to be * rare) instead of during all exits (common). */ 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) || ((cap_rdpid || cap_rdtscp) && guest_msr_ro(vmx, MSR_TSC_AUX))) panic("vmx_init: error setting guest msr access"); 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)); } vmx->pmap = pmap; return (vmx); } static void * vmx_vcpu_init(void *vmi, struct vcpu *vcpu1, int vcpuid) { struct vmx *vmx = vmi; struct vmcs *vmcs; struct vmx_vcpu *vcpu; uint32_t exc_bitmap; uint16_t vpid; int error; vpid = vpid_alloc(vcpuid); vcpu = malloc(sizeof(*vcpu), M_VMX, M_WAITOK | M_ZERO); vcpu->vmx = vmx; vcpu->vcpu = vcpu1; vcpu->vcpuid = vcpuid; vcpu->vmcs = malloc_aligned(sizeof(*vmcs), PAGE_SIZE, M_VMX, M_WAITOK | M_ZERO); vcpu->apic_page = malloc_aligned(PAGE_SIZE, PAGE_SIZE, M_VMX, M_WAITOK | M_ZERO); vcpu->pir_desc = malloc_aligned(sizeof(*vcpu->pir_desc), 64, M_VMX, M_WAITOK | M_ZERO); vmcs = vcpu->vmcs; vmcs->identifier = vmx_revision(); error = vmclear(vmcs); if (error != 0) { panic("vmx_init: vmclear error %d on vcpu %d\n", error, vcpuid); } vmx_msr_guest_init(vmx, vcpu); error = vmcs_init(vmcs); KASSERT(error == 0, ("vmcs_init error %d", error)); VMPTRLD(vmcs); error = 0; error += vmwrite(VMCS_HOST_RSP, (u_long)&vcpu->ctx); error += vmwrite(VMCS_EPTP, vmx->eptp); error += vmwrite(VMCS_PIN_BASED_CTLS, pinbased_ctls); error += vmwrite(VMCS_PRI_PROC_BASED_CTLS, procbased_ctls); if (vcpu_trap_wbinvd(vcpu->vcpu)) { KASSERT(cap_wbinvd_exit, ("WBINVD trap not available")); procbased_ctls2 |= PROCBASED2_WBINVD_EXITING; } 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); if (guest_l1d_flush && !guest_l1d_flush_sw) { vmcs_write(VMCS_ENTRY_MSR_LOAD, pmap_kextract( (vm_offset_t)&msr_load_list[0])); vmcs_write(VMCS_ENTRY_MSR_LOAD_COUNT, nitems(msr_load_list)); vmcs_write(VMCS_EXIT_MSR_STORE, 0); vmcs_write(VMCS_EXIT_MSR_STORE_COUNT, 0); } /* exception bitmap */ if (vcpu_trace_exceptions(vcpu->vcpu)) exc_bitmap = 0xffffffff; else exc_bitmap = 1 << IDT_MC; error += vmwrite(VMCS_EXCEPTION_BITMAP, exc_bitmap); vcpu->ctx.guest_dr6 = DBREG_DR6_RESERVED1; error += vmwrite(VMCS_GUEST_DR7, DBREG_DR7_RESERVED1); if (tpr_shadowing) { error += vmwrite(VMCS_VIRTUAL_APIC, vtophys(vcpu->apic_page)); } if (virtual_interrupt_delivery) { error += vmwrite(VMCS_APIC_ACCESS, APIC_ACCESS_ADDRESS); 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(vcpu->pir_desc)); } VMCLEAR(vmcs); KASSERT(error == 0, ("vmx_init: error customizing the vmcs")); vcpu->cap.set = 0; vcpu->cap.set |= cap_rdpid != 0 ? 1 << VM_CAP_RDPID : 0; vcpu->cap.set |= cap_rdtscp != 0 ? 1 << VM_CAP_RDTSCP : 0; vcpu->cap.proc_ctls = procbased_ctls; vcpu->cap.proc_ctls2 = procbased_ctls2; vcpu->cap.exc_bitmap = exc_bitmap; vcpu->state.nextrip = ~0; vcpu->state.lastcpu = NOCPU; vcpu->state.vpid = vpid; /* * 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); vcpu->ctx.pmap = vmx->pmap; return (vcpu); } static int vmx_handle_cpuid(struct vmx_vcpu *vcpu, struct vmxctx *vmxctx) { int handled; handled = x86_emulate_cpuid(vcpu->vcpu, (uint64_t *)&vmxctx->guest_rax, (uint64_t *)&vmxctx->guest_rbx, (uint64_t *)&vmxctx->guest_rcx, (uint64_t *)&vmxctx->guest_rdx); return (handled); } static __inline void vmx_run_trace(struct vmx_vcpu *vcpu) { VMX_CTR1(vcpu, "Resume execution at %#lx", vmcs_guest_rip()); } static __inline void vmx_exit_trace(struct vmx_vcpu *vcpu, uint64_t rip, uint32_t exit_reason, int handled) { VMX_CTR3(vcpu, "%s %s vmexit at 0x%0lx", handled ? "handled" : "unhandled", exit_reason_to_str(exit_reason), rip); } static __inline void vmx_astpending_trace(struct vmx_vcpu *vcpu, uint64_t rip) { VMX_CTR1(vcpu, "astpending vmexit at 0x%0lx", rip); } 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, struct vmx_vcpu *vcpu, pmap_t pmap, int running) { struct vmxstate *vmxstate; struct invvpid_desc invvpid_desc; vmxstate = &vcpu->state; 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->vcpuid)); /* * 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 (atomic_load_long(&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(vcpu->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(vcpu->vcpu, VCPU_INVVPID_SAVED, 1); } } static void vmx_set_pcpu_defaults(struct vmx *vmx, struct vmx_vcpu *vcpu, pmap_t pmap) { struct vmxstate *vmxstate; vmxstate = &vcpu->state; if (vmxstate->lastcpu == curcpu) return; vmxstate->lastcpu = curcpu; vmm_stat_incr(vcpu->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_vcpu *vcpu) { if ((vcpu->cap.proc_ctls & PROCBASED_INT_WINDOW_EXITING) == 0) { vcpu->cap.proc_ctls |= PROCBASED_INT_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vcpu->cap.proc_ctls); VMX_CTR0(vcpu, "Enabling interrupt window exiting"); } } static void __inline vmx_clear_int_window_exiting(struct vmx_vcpu *vcpu) { KASSERT((vcpu->cap.proc_ctls & PROCBASED_INT_WINDOW_EXITING) != 0, ("intr_window_exiting not set: %#x", vcpu->cap.proc_ctls)); vcpu->cap.proc_ctls &= ~PROCBASED_INT_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vcpu->cap.proc_ctls); VMX_CTR0(vcpu, "Disabling interrupt window exiting"); } static void __inline vmx_set_nmi_window_exiting(struct vmx_vcpu *vcpu) { if ((vcpu->cap.proc_ctls & PROCBASED_NMI_WINDOW_EXITING) == 0) { vcpu->cap.proc_ctls |= PROCBASED_NMI_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vcpu->cap.proc_ctls); VMX_CTR0(vcpu, "Enabling NMI window exiting"); } } static void __inline vmx_clear_nmi_window_exiting(struct vmx_vcpu *vcpu) { KASSERT((vcpu->cap.proc_ctls & PROCBASED_NMI_WINDOW_EXITING) != 0, ("nmi_window_exiting not set %#x", vcpu->cap.proc_ctls)); vcpu->cap.proc_ctls &= ~PROCBASED_NMI_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vcpu->cap.proc_ctls); VMX_CTR0(vcpu, "Disabling NMI window exiting"); } int vmx_set_tsc_offset(struct vmx_vcpu *vcpu, uint64_t offset) { int error; if ((vcpu->cap.proc_ctls & PROCBASED_TSC_OFFSET) == 0) { vcpu->cap.proc_ctls |= PROCBASED_TSC_OFFSET; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vcpu->cap.proc_ctls); VMX_CTR0(vcpu, "Enabling TSC offsetting"); } error = vmwrite(VMCS_TSC_OFFSET, offset); #ifdef BHYVE_SNAPSHOT if (error == 0) vm_set_tsc_offset(vcpu->vcpu, offset); #endif return (error); } #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_vcpu *vcpu) { uint32_t gi __diagused, 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); VMX_CTR0(vcpu, "Injecting vNMI"); /* Clear the request */ vm_nmi_clear(vcpu->vcpu); } static void vmx_inject_interrupts(struct vmx_vcpu *vcpu, struct vlapic *vlapic, uint64_t guestrip) { int vector, need_nmi_exiting, extint_pending; uint64_t rflags, entryinfo; uint32_t gi, info; if (vcpu->state.nextrip != guestrip) { gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); if (gi & HWINTR_BLOCKING) { VMX_CTR2(vcpu, "Guest interrupt blocking " "cleared due to rip change: %#lx/%#lx", vcpu->state.nextrip, guestrip); gi &= ~HWINTR_BLOCKING; vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi); } } if (vm_entry_intinfo(vcpu->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(vcpu->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(vcpu); need_nmi_exiting = 0; } else { VMX_CTR1(vcpu, "Cannot inject NMI " "due to VM-entry intr info %#x", info); } } else { VMX_CTR1(vcpu, "Cannot inject NMI due to " "Guest Interruptibility-state %#x", gi); } if (need_nmi_exiting) vmx_set_nmi_window_exiting(vcpu); } extint_pending = vm_extint_pending(vcpu->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 ((vcpu->cap.proc_ctls & PROCBASED_INT_WINDOW_EXITING) != 0) { VMX_CTR0(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(vcpu->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) { VMX_CTR2(vcpu, "Cannot inject vector %d due to " "rflags %#lx", vector, rflags); goto cantinject; } gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); if (gi & HWINTR_BLOCKING) { VMX_CTR2(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" */ VMX_CTR2(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(vcpu->vcpu); vatpic_intr_accepted(vcpu->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(vcpu); } VMX_CTR1(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(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_vcpu *vcpu) { uint32_t gi; VMX_CTR0(vcpu, "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_vcpu *vcpu) { uint32_t gi; VMX_CTR0(vcpu, "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_vcpu *vcpu) { uint32_t gi __diagused; 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, struct vmx_vcpu *vcpu, struct vm_exit *vmexit) { struct vmxctx *vmxctx; uint64_t xcrval; const struct xsave_limits *limits; vmxctx = &vcpu->ctx; 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(vcpu->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(vcpu->vcpu); return (HANDLED); } xcrval = vmxctx->guest_rdx << 32 | (vmxctx->guest_rax & 0xffffffff); if ((xcrval & ~limits->xcr0_allowed) != 0) { vm_inject_gp(vcpu->vcpu); return (HANDLED); } if (!(xcrval & XFEATURE_ENABLED_X87)) { vm_inject_gp(vcpu->vcpu); return (HANDLED); } /* AVX (YMM_Hi128) requires SSE. */ if (xcrval & XFEATURE_ENABLED_AVX && (xcrval & XFEATURE_AVX) != XFEATURE_AVX) { vm_inject_gp(vcpu->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(vcpu->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(vcpu->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_vcpu *vcpu, int ident) { const struct vmxctx *vmxctx; vmxctx = &vcpu->ctx; 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_vcpu *vcpu, int ident, uint64_t regval) { struct vmxctx *vmxctx; vmxctx = &vcpu->ctx; 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_vcpu *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(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_vcpu *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(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, struct vmx_vcpu *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(vcpu->vcpu); regnum = (exitqual >> 8) & 0xf; if (exitqual & 0x10) { cr8 = vlapic_get_cr8(vlapic); vmx_set_guest_reg(vcpu, regnum, cr8); } else { cr8 = vmx_get_guest_reg(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) { uint64_t cr4; if (!(vmcs_read(VMCS_GUEST_CR0) & CR0_PG)) return (PAGING_MODE_FLAT); cr4 = vmcs_read(VMCS_GUEST_CR4); if (!(cr4 & CR4_PAE)) return (PAGING_MODE_32); if (vmcs_read(VMCS_GUEST_IA32_EFER) & EFER_LME) { if (!(cr4 & CR4_LA57)) return (PAGING_MODE_64); return (PAGING_MODE_64_LA57); } else return (PAGING_MODE_PAE); } static uint64_t inout_str_index(struct vmx_vcpu *vcpu, int in) { uint64_t val; int error __diagused; enum vm_reg_name reg; reg = in ? VM_REG_GUEST_RDI : VM_REG_GUEST_RSI; error = vmx_getreg(vcpu, reg, &val); KASSERT(error == 0, ("%s: vmx_getreg error %d", __func__, error)); return (val); } static uint64_t inout_str_count(struct vmx_vcpu *vcpu, int rep) { uint64_t val; int error __diagused; if (rep) { error = vmx_getreg(vcpu, 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_vcpu *vcpu, uint32_t inst_info, int in, struct vm_inout_str *vis) { int error __diagused, 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(vcpu, 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 bool 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_vcpu *vcpu) { uint32_t proc_ctls2; proc_ctls2 = vcpu->cap.proc_ctls2; return ((proc_ctls2 & PROCBASED2_VIRTUALIZE_APIC_ACCESSES) ? 1 : 0); } static __inline int x2apic_virtualization(struct vmx_vcpu *vcpu) { uint32_t proc_ctls2; proc_ctls2 = vcpu->cap.proc_ctls2; return ((proc_ctls2 & PROCBASED2_VIRTUALIZE_X2APIC_MODE) ? 1 : 0); } static int vmx_handle_apic_write(struct vmx_vcpu *vcpu, 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(vcpu)) { /* * 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(vcpu) && 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_vcpu *vcpu, uint64_t gpa) { if (apic_access_virtualization(vcpu) && (gpa >= DEFAULT_APIC_BASE && gpa < DEFAULT_APIC_BASE + PAGE_SIZE)) return (true); else return (false); } static int vmx_handle_apic_access(struct vmx_vcpu *vcpu, struct vm_exit *vmexit) { uint64_t qual; int access_type, offset, allowed; if (!apic_access_virtualization(vcpu)) 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_vcpu *vcpu, u_int num, uint64_t val, bool *retu) { int error; if (lapic_msr(num)) error = lapic_wrmsr(vcpu->vcpu, num, val, retu); else error = vmx_wrmsr(vcpu, num, val, retu); return (error); } static int emulate_rdmsr(struct vmx_vcpu *vcpu, u_int num, bool *retu) { struct vmxctx *vmxctx; uint64_t result; uint32_t eax, edx; int error; if (lapic_msr(num)) error = lapic_rdmsr(vcpu->vcpu, num, &result, retu); else error = vmx_rdmsr(vcpu, num, &result, retu); if (error == 0) { eax = result; vmxctx = &vcpu->ctx; 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, struct vmx_vcpu *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; #ifdef KDTRACE_HOOKS int vcpuid; #endif bool retu; CTASSERT((PINBASED_CTLS_ONE_SETTING & PINBASED_VIRTUAL_NMI) != 0); CTASSERT((PINBASED_CTLS_ONE_SETTING & PINBASED_NMI_EXITING) != 0); handled = UNHANDLED; vmxctx = &vcpu->ctx; #ifdef KDTRACE_HOOKS vcpuid = vcpu->vcpuid; #endif qual = vmexit->u.vmx.exit_qualification; reason = vmexit->u.vmx.exit_reason; vmexit->exitcode = VM_EXITCODE_BOGUS; vmm_stat_incr(vcpu->vcpu, VMEXIT_COUNT, 1); SDT_PROBE3(vmm, vmx, exit, entry, vmx, vcpuid, vmexit); /* * 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)) { VMX_CTR0(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(vcpu->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(vcpu); else vmx_assert_nmi_blocking(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; SDT_PROBE4(vmm, vmx, exit, taskswitch, vmx, vcpuid, vmexit, ts); VMX_CTR4(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(vcpu->vcpu, VMEXIT_CR_ACCESS, 1); SDT_PROBE4(vmm, vmx, exit, craccess, vmx, vcpuid, vmexit, qual); switch (qual & 0xf) { case 0: handled = vmx_emulate_cr0_access(vcpu, qual); break; case 4: handled = vmx_emulate_cr4_access(vcpu, qual); break; case 8: handled = vmx_emulate_cr8_access(vmx, vcpu, qual); break; } break; case EXIT_REASON_RDMSR: vmm_stat_incr(vcpu->vcpu, VMEXIT_RDMSR, 1); retu = false; ecx = vmxctx->guest_rcx; VMX_CTR1(vcpu, "rdmsr 0x%08x", ecx); SDT_PROBE4(vmm, vmx, exit, rdmsr, vmx, vcpuid, vmexit, ecx); error = emulate_rdmsr(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(vcpu->vcpu, VMEXIT_WRMSR, 1); retu = false; eax = vmxctx->guest_rax; ecx = vmxctx->guest_rcx; edx = vmxctx->guest_rdx; VMX_CTR2(vcpu, "wrmsr 0x%08x value 0x%016lx", ecx, (uint64_t)edx << 32 | eax); SDT_PROBE5(vmm, vmx, exit, wrmsr, vmx, vmexit, vcpuid, ecx, (uint64_t)edx << 32 | eax); error = emulate_wrmsr(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(vcpu->vcpu, VMEXIT_HLT, 1); SDT_PROBE3(vmm, vmx, exit, halt, vmx, vcpuid, vmexit); vmexit->exitcode = VM_EXITCODE_HLT; vmexit->u.hlt.rflags = vmcs_read(VMCS_GUEST_RFLAGS); if (virtual_interrupt_delivery) vmexit->u.hlt.intr_status = vmcs_read(VMCS_GUEST_INTR_STATUS); else vmexit->u.hlt.intr_status = 0; break; case EXIT_REASON_MTF: vmm_stat_incr(vcpu->vcpu, VMEXIT_MTRAP, 1); SDT_PROBE3(vmm, vmx, exit, mtrap, vmx, vcpuid, vmexit); vmexit->exitcode = VM_EXITCODE_MTRAP; vmexit->inst_length = 0; break; case EXIT_REASON_PAUSE: vmm_stat_incr(vcpu->vcpu, VMEXIT_PAUSE, 1); SDT_PROBE3(vmm, vmx, exit, pause, vmx, vcpuid, vmexit); vmexit->exitcode = VM_EXITCODE_PAUSE; break; case EXIT_REASON_INTR_WINDOW: vmm_stat_incr(vcpu->vcpu, VMEXIT_INTR_WINDOW, 1); SDT_PROBE3(vmm, vmx, exit, intrwindow, vmx, vcpuid, vmexit); vmx_clear_int_window_exiting(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); SDT_PROBE4(vmm, vmx, exit, interrupt, vmx, vcpuid, vmexit, 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(vcpu->vcpu, VMEXIT_EXTINT, 1); return (1); case EXIT_REASON_NMI_WINDOW: SDT_PROBE3(vmm, vmx, exit, nmiwindow, vmx, vcpuid, vmexit); /* Exit to allow the pending virtual NMI to be injected */ if (vm_nmi_pending(vcpu->vcpu)) vmx_inject_nmi(vcpu); vmx_clear_nmi_window_exiting(vcpu); vmm_stat_incr(vcpu->vcpu, VMEXIT_NMI_WINDOW, 1); return (1); case EXIT_REASON_INOUT: vmm_stat_incr(vcpu->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(vcpu, in); vis->count = inout_str_count(vcpu, vis->inout.rep); vis->addrsize = inout_str_addrsize(inst_info); inout_str_seginfo(vcpu, inst_info, in, vis); } SDT_PROBE3(vmm, vmx, exit, inout, vmx, vcpuid, vmexit); break; case EXIT_REASON_CPUID: vmm_stat_incr(vcpu->vcpu, VMEXIT_CPUID, 1); SDT_PROBE3(vmm, vmx, exit, cpuid, vmx, vcpuid, vmexit); handled = vmx_handle_cpuid(vcpu, vmxctx); break; case EXIT_REASON_EXCEPTION: vmm_stat_incr(vcpu->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(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) { VMX_CTR0(vcpu, "Vectoring to MCE handler"); __asm __volatile("int $18"); return (1); } /* * If the hypervisor has requested user exits for * debug exceptions, bounce them out to userland. */ if (intr_type == VMCS_INTR_T_SWEXCEPTION && intr_vec == IDT_BP && (vcpu->cap.set & (1 << VM_CAP_BPT_EXIT))) { vmexit->exitcode = VM_EXITCODE_BPT; vmexit->u.bpt.inst_length = vmexit->inst_length; vmexit->inst_length = 0; break; } 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); } VMX_CTR2(vcpu, "Reflecting exception %d/%#x into " "the guest", intr_vec, errcode); SDT_PROBE5(vmm, vmx, exit, exception, vmx, vcpuid, vmexit, intr_vec, errcode); error = vm_inject_exception(vcpu->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(vcpu->vcpu, gpa) || apic_access_fault(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(vcpu->vcpu, VMEXIT_NESTED_FAULT, 1); SDT_PROBE5(vmm, vmx, exit, nestedfault, vmx, vcpuid, vmexit, gpa, qual); } else if (ept_emulation_fault(qual)) { vmexit_inst_emul(vmexit, gpa, vmcs_gla()); vmm_stat_incr(vcpu->vcpu, VMEXIT_INST_EMUL, 1); SDT_PROBE4(vmm, vmx, exit, mmiofault, vmx, vcpuid, vmexit, gpa); } /* * 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(vcpu); break; case EXIT_REASON_VIRTUALIZED_EOI: vmexit->exitcode = VM_EXITCODE_IOAPIC_EOI; vmexit->u.ioapic_eoi.vector = qual & 0xFF; SDT_PROBE3(vmm, vmx, exit, eoi, vmx, vcpuid, vmexit); vmexit->inst_length = 0; /* trap-like */ break; case EXIT_REASON_APIC_ACCESS: SDT_PROBE3(vmm, vmx, exit, apicaccess, vmx, vcpuid, vmexit); handled = vmx_handle_apic_access(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(vcpu->vcpu); SDT_PROBE4(vmm, vmx, exit, apicwrite, vmx, vcpuid, vmexit, vlapic); handled = vmx_handle_apic_write(vcpu, vlapic, qual); break; case EXIT_REASON_XSETBV: SDT_PROBE3(vmm, vmx, exit, xsetbv, vmx, vcpuid, vmexit); handled = vmx_emulate_xsetbv(vmx, vcpu, vmexit); break; case EXIT_REASON_MONITOR: SDT_PROBE3(vmm, vmx, exit, monitor, vmx, vcpuid, vmexit); vmexit->exitcode = VM_EXITCODE_MONITOR; break; case EXIT_REASON_MWAIT: SDT_PROBE3(vmm, vmx, exit, mwait, vmx, vcpuid, vmexit); vmexit->exitcode = VM_EXITCODE_MWAIT; break; case EXIT_REASON_TPR: vlapic = vm_lapic(vcpu->vcpu); vlapic_sync_tpr(vlapic); vmexit->inst_length = 0; handled = HANDLED; break; case EXIT_REASON_VMCALL: case EXIT_REASON_VMCLEAR: case EXIT_REASON_VMLAUNCH: case EXIT_REASON_VMPTRLD: case EXIT_REASON_VMPTRST: case EXIT_REASON_VMREAD: case EXIT_REASON_VMRESUME: case EXIT_REASON_VMWRITE: case EXIT_REASON_VMXOFF: case EXIT_REASON_VMXON: SDT_PROBE3(vmm, vmx, exit, vminsn, vmx, vcpuid, vmexit); vmexit->exitcode = VM_EXITCODE_VMINSN; break; case EXIT_REASON_INVD: case EXIT_REASON_WBINVD: /* ignore exit */ handled = HANDLED; break; default: SDT_PROBE4(vmm, vmx, exit, unknown, vmx, vcpuid, vmexit, reason); vmm_stat_incr(vcpu->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. */ } } SDT_PROBE4(vmm, vmx, exit, return, vmx, vcpuid, vmexit, handled); 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: 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_vcpu *vcpu, 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)); VMX_CTR0(vcpu, "Vectoring to NMI handler"); __asm __volatile("int $2"); } } static __inline void vmx_dr_enter_guest(struct vmxctx *vmxctx) { register_t rflags; /* Save host control debug registers. */ vmxctx->host_dr7 = rdr7(); vmxctx->host_debugctl = rdmsr(MSR_DEBUGCTLMSR); /* * Disable debugging in DR7 and DEBUGCTL to avoid triggering * exceptions in the host based on the guest DRx values. The * guest DR7 and DEBUGCTL are saved/restored in the VMCS. */ load_dr7(0); wrmsr(MSR_DEBUGCTLMSR, 0); /* * Disable single stepping the kernel to avoid corrupting the * guest DR6. A debugger might still be able to corrupt the * guest DR6 by setting a breakpoint after this point and then * single stepping. */ rflags = read_rflags(); vmxctx->host_tf = rflags & PSL_T; write_rflags(rflags & ~PSL_T); /* Save host debug registers. */ vmxctx->host_dr0 = rdr0(); vmxctx->host_dr1 = rdr1(); vmxctx->host_dr2 = rdr2(); vmxctx->host_dr3 = rdr3(); vmxctx->host_dr6 = rdr6(); /* Restore guest debug registers. */ load_dr0(vmxctx->guest_dr0); load_dr1(vmxctx->guest_dr1); load_dr2(vmxctx->guest_dr2); load_dr3(vmxctx->guest_dr3); load_dr6(vmxctx->guest_dr6); } static __inline void vmx_dr_leave_guest(struct vmxctx *vmxctx) { /* Save guest debug registers. */ vmxctx->guest_dr0 = rdr0(); vmxctx->guest_dr1 = rdr1(); vmxctx->guest_dr2 = rdr2(); vmxctx->guest_dr3 = rdr3(); vmxctx->guest_dr6 = rdr6(); /* * Restore host debug registers. Restore DR7, DEBUGCTL, and * PSL_T last. */ load_dr0(vmxctx->host_dr0); load_dr1(vmxctx->host_dr1); load_dr2(vmxctx->host_dr2); load_dr3(vmxctx->host_dr3); load_dr6(vmxctx->host_dr6); wrmsr(MSR_DEBUGCTLMSR, vmxctx->host_debugctl); load_dr7(vmxctx->host_dr7); write_rflags(read_rflags() | vmxctx->host_tf); } static __inline void vmx_pmap_activate(struct vmx *vmx, pmap_t pmap) { long eptgen; int cpu; cpu = curcpu; CPU_SET_ATOMIC(cpu, &pmap->pm_active); smr_enter(pmap->pm_eptsmr); eptgen = atomic_load_long(&pmap->pm_eptgen); if (eptgen != vmx->eptgen[cpu]) { vmx->eptgen[cpu] = eptgen; invept(INVEPT_TYPE_SINGLE_CONTEXT, (struct invept_desc){ .eptp = vmx->eptp, ._res = 0 }); } } static __inline void vmx_pmap_deactivate(struct vmx *vmx, pmap_t pmap) { smr_exit(pmap->pm_eptsmr); CPU_CLR_ATOMIC(curcpu, &pmap->pm_active); } static int vmx_run(void *vcpui, register_t rip, pmap_t pmap, struct vm_eventinfo *evinfo) { int rc, handled, launched; struct vmx *vmx; struct vmx_vcpu *vcpu; struct vmxctx *vmxctx; struct vmcs *vmcs; struct vm_exit *vmexit; struct vlapic *vlapic; uint32_t exit_reason; struct region_descriptor gdtr, idtr; uint16_t ldt_sel; vcpu = vcpui; vmx = vcpu->vmx; vmcs = vcpu->vmcs; vmxctx = &vcpu->ctx; vlapic = vm_lapic(vcpu->vcpu); vmexit = vm_exitinfo(vcpu->vcpu); launched = 0; KASSERT(vmxctx->pmap == pmap, ("pmap %p different than ctx pmap %p", pmap, vmxctx->pmap)); vmx_msr_guest_enter(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_init(). */ 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(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(evinfo)) { enable_intr(); vm_exit_suspended(vcpu->vcpu, rip); break; } - if (vcpu_rendezvous_pending(evinfo)) { + if (vcpu_rendezvous_pending(vcpu->vcpu, evinfo)) { enable_intr(); vm_exit_rendezvous(vcpu->vcpu, rip); break; } if (vcpu_reqidle(evinfo)) { enable_intr(); vm_exit_reqidle(vcpu->vcpu, rip); break; } if (vcpu_should_yield(vcpu->vcpu)) { enable_intr(); vm_exit_astpending(vcpu->vcpu, rip); vmx_astpending_trace(vcpu, rip); handled = HANDLED; break; } if (vcpu_debugged(vcpu->vcpu)) { enable_intr(); vm_exit_debug(vcpu->vcpu, rip); break; } /* * If TPR Shadowing is enabled, the TPR Threshold * must be updated right before entering the guest. */ if (tpr_shadowing && !virtual_interrupt_delivery) { if ((vcpu->cap.proc_ctls & PROCBASED_USE_TPR_SHADOW) != 0) { vmcs_write(VMCS_TPR_THRESHOLD, vlapic_get_cr8(vlapic)); } } /* * VM exits restore the base address but not the * limits of GDTR and IDTR. The VMCS only stores the * base address, so VM exits set the limits to 0xffff. * Save and restore the full GDTR and IDTR to restore * the limits. * * The VMCS does not save the LDTR at all, and VM * exits clear LDTR as if a NULL selector were loaded. * The userspace hypervisor probably doesn't use a * LDT, but save and restore it to be safe. */ sgdt(&gdtr); sidt(&idtr); ldt_sel = sldt(); /* * The TSC_AUX MSR must be saved/restored while interrupts * are disabled so that it is not possible for the guest * TSC_AUX MSR value to be overwritten by the resume * portion of the IPI_SUSPEND codepath. This is why the * transition of this MSR is handled separately from those * handled by vmx_msr_guest_{enter,exit}(), which are ok to * be transitioned with preemption disabled but interrupts * enabled. * * These vmx_msr_guest_{enter,exit}_tsc_aux() calls can be * anywhere in this loop so long as they happen with * interrupts disabled. This location is chosen for * simplicity. */ vmx_msr_guest_enter_tsc_aux(vmx, vcpu); vmx_dr_enter_guest(vmxctx); /* * Mark the EPT as active on this host CPU and invalidate * EPTP-tagged TLB entries if required. */ vmx_pmap_activate(vmx, pmap); vmx_run_trace(vcpu); rc = vmx_enter_guest(vmxctx, vmx, launched); vmx_pmap_deactivate(vmx, pmap); vmx_dr_leave_guest(vmxctx); vmx_msr_guest_exit_tsc_aux(vmx, vcpu); bare_lgdt(&gdtr); lidt(&idtr); lldt(ldt_sel); /* 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' */ vcpu->state.nextrip = rip; if (rc == VMX_GUEST_VMEXIT) { vmx_exit_handle_nmi(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(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); } VMX_CTR1(vcpu, "returning from vmx_run: exitcode %d", vmexit->exitcode); VMCLEAR(vmcs); vmx_msr_guest_exit(vcpu); return (0); } static void vmx_vcpu_cleanup(void *vcpui) { struct vmx_vcpu *vcpu = vcpui; vpid_free(vcpu->state.vpid); free(vcpu->pir_desc, M_VMX); free(vcpu->apic_page, M_VMX); free(vcpu->vmcs, M_VMX); free(vcpu, M_VMX); } static void vmx_cleanup(void *vmi) { struct vmx *vmx = vmi; if (virtual_interrupt_delivery) vm_unmap_mmio(vmx->vm, DEFAULT_APIC_BASE, PAGE_SIZE); free(vmx->msr_bitmap, M_VMX); 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); case VM_REG_GUEST_DR0: return (&vmxctx->guest_dr0); case VM_REG_GUEST_DR1: return (&vmxctx->guest_dr1); case VM_REG_GUEST_DR2: return (&vmxctx->guest_dr2); case VM_REG_GUEST_DR3: return (&vmxctx->guest_dr3); case VM_REG_GUEST_DR6: return (&vmxctx->guest_dr6); 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_vcpu *vcpu, int running, uint64_t *retval) { uint64_t gi; int error; error = vmcs_getreg(vcpu->vmcs, running, VMCS_IDENT(VMCS_GUEST_INTERRUPTIBILITY), &gi); *retval = (gi & HWINTR_BLOCKING) ? 1 : 0; return (error); } static int vmx_modify_intr_shadow(struct vmx_vcpu *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 = vcpu->vmcs; 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: VMX_CTR2(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 *vcpui, int reg, uint64_t *retval) { int running, hostcpu; struct vmx_vcpu *vcpu = vcpui; struct vmx *vmx = vcpu->vmx; running = vcpu_is_running(vcpu->vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_getreg: %s%d is running", vm_name(vmx->vm), vcpu->vcpuid); if (reg == VM_REG_GUEST_INTR_SHADOW) return (vmx_get_intr_shadow(vcpu, running, retval)); if (vmxctx_getreg(&vcpu->ctx, reg, retval) == 0) return (0); return (vmcs_getreg(vcpu->vmcs, running, reg, retval)); } static int vmx_setreg(void *vcpui, int reg, uint64_t val) { int error, hostcpu, running, shadow; uint64_t ctls; pmap_t pmap; struct vmx_vcpu *vcpu = vcpui; struct vmx *vmx = vcpu->vmx; running = vcpu_is_running(vcpu->vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_setreg: %s%d is running", vm_name(vmx->vm), vcpu->vcpuid); if (reg == VM_REG_GUEST_INTR_SHADOW) return (vmx_modify_intr_shadow(vcpu, running, val)); if (vmxctx_setreg(&vcpu->ctx, reg, val) == 0) return (0); /* Do not permit user write access to VMCS fields by offset. */ if (reg < 0) return (EINVAL); error = vmcs_setreg(vcpu->vmcs, 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(vcpu->vmcs, running, VMCS_IDENT(VMCS_ENTRY_CTLS), &ctls); if (val & EFER_LMA) ctls |= VM_ENTRY_GUEST_LMA; else ctls &= ~VM_ENTRY_GUEST_LMA; vmcs_setreg(vcpu->vmcs, 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(vcpu->vmcs, 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 = vcpu->ctx.pmap; vmx_invvpid(vmx, vcpu, pmap, running); } } return (error); } static int vmx_getdesc(void *vcpui, int reg, struct seg_desc *desc) { int hostcpu, running; struct vmx_vcpu *vcpu = vcpui; struct vmx *vmx = vcpu->vmx; running = vcpu_is_running(vcpu->vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_getdesc: %s%d is running", vm_name(vmx->vm), vcpu->vcpuid); return (vmcs_getdesc(vcpu->vmcs, running, reg, desc)); } static int vmx_setdesc(void *vcpui, int reg, struct seg_desc *desc) { int hostcpu, running; struct vmx_vcpu *vcpu = vcpui; struct vmx *vmx = vcpu->vmx; running = vcpu_is_running(vcpu->vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_setdesc: %s%d is running", vm_name(vmx->vm), vcpu->vcpuid); return (vmcs_setdesc(vcpu->vmcs, running, reg, desc)); } static int vmx_getcap(void *vcpui, int type, int *retval) { struct vmx_vcpu *vcpu = vcpui; int vcap; int ret; ret = ENOENT; vcap = vcpu->cap.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_RDPID: if (cap_rdpid) ret = 0; break; case VM_CAP_RDTSCP: if (cap_rdtscp) 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; case VM_CAP_BPT_EXIT: case VM_CAP_IPI_EXIT: ret = 0; break; default: break; } if (ret == 0) *retval = (vcap & (1 << type)) ? 1 : 0; return (ret); } static int vmx_setcap(void *vcpui, int type, int val) { struct vmx_vcpu *vcpu = vcpui; struct vmcs *vmcs = vcpu->vmcs; struct vlapic *vlapic; 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 = &vcpu->cap.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 = &vcpu->cap.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 = &vcpu->cap.proc_ctls; baseval = *pptr; flag = PROCBASED_PAUSE_EXITING; reg = VMCS_PRI_PROC_BASED_CTLS; } break; case VM_CAP_RDPID: case VM_CAP_RDTSCP: if (cap_rdpid || cap_rdtscp) /* * Choose not to support enabling/disabling * RDPID/RDTSCP via libvmmapi since, as per the * discussion in vmx_modinit(), RDPID/RDTSCP are * either always enabled or always disabled. */ error = EOPNOTSUPP; break; case VM_CAP_UNRESTRICTED_GUEST: if (cap_unrestricted_guest) { retval = 0; pptr = &vcpu->cap.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 = &vcpu->cap.proc_ctls2; baseval = *pptr; flag = PROCBASED2_ENABLE_INVPCID; reg = VMCS_SEC_PROC_BASED_CTLS; } break; case VM_CAP_BPT_EXIT: retval = 0; /* Don't change the bitmap if we are tracing all exceptions. */ if (vcpu->cap.exc_bitmap != 0xffffffff) { pptr = &vcpu->cap.exc_bitmap; baseval = *pptr; flag = (1 << IDT_BP); reg = VMCS_EXCEPTION_BITMAP; } break; case VM_CAP_IPI_EXIT: retval = 0; vlapic = vm_lapic(vcpu->vcpu); vlapic->ipi_exit = val; break; default: break; } if (retval) return (retval); if (pptr != NULL) { if (val) { baseval |= flag; } else { baseval &= ~flag; } VMPTRLD(vmcs); error = vmwrite(reg, baseval); VMCLEAR(vmcs); if (error) return (error); /* * Update optional stored flags, and record * setting */ *pptr = baseval; } if (val) { vcpu->cap.set |= (1 << type); } else { vcpu->cap.set &= ~(1 << type); } return (0); } static struct vmspace * vmx_vmspace_alloc(vm_offset_t min, vm_offset_t max) { return (ept_vmspace_alloc(min, max)); } static void vmx_vmspace_free(struct vmspace *vmspace) { ept_vmspace_free(vmspace); } struct vlapic_vtx { struct vlapic vlapic; struct pir_desc *pir_desc; struct vmx_vcpu *vcpu; u_int pending_prio; }; #define VPR_PRIO_BIT(vpr) (1 << ((vpr) >> 4)) #define VMX_CTR_PIR(vlapic, pir_desc, notify, vector, level, msg) \ do { \ VLAPIC_CTR2(vlapic, msg " assert %s-triggered vector %d", \ level ? "level" : "edge", vector); \ VLAPIC_CTR1(vlapic, msg " pir0 0x%016lx", pir_desc->pir[0]); \ VLAPIC_CTR1(vlapic, msg " pir1 0x%016lx", pir_desc->pir[1]); \ VLAPIC_CTR1(vlapic, msg " pir2 0x%016lx", pir_desc->pir[2]); \ VLAPIC_CTR1(vlapic, msg " pir3 0x%016lx", pir_desc->pir[3]); \ VLAPIC_CTR1(vlapic, 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 = 0; 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); /* * A notification is required whenever the 'pending' bit makes a * transition from 0->1. * * Even if the 'pending' bit is already asserted, notification about * the incoming interrupt may still be necessary. For example, if a * vCPU is HLTed with a high PPR, a low priority interrupt would cause * the 0->1 'pending' transition with a notification, but the vCPU * would ignore the interrupt for the time being. The same vCPU would * need to then be notified if a high-priority interrupt arrived which * satisfied the PPR. * * The priorities of interrupts injected while 'pending' is asserted * are tracked in a custom bitfield 'pending_prio'. Should the * to-be-injected interrupt exceed the priorities already present, the * notification is sent. The priorities recorded in 'pending_prio' are * cleared whenever the 'pending' bit makes another 0->1 transition. */ if (atomic_cmpset_long(&pir_desc->pending, 0, 1) != 0) { notify = 1; vlapic_vtx->pending_prio = 0; } else { const u_int old_prio = vlapic_vtx->pending_prio; const u_int prio_bit = VPR_PRIO_BIT(vector & APIC_TPR_INT); if ((old_prio & prio_bit) == 0 && prio_bit > old_prio) { atomic_set_int(&vlapic_vtx->pending_prio, prio_bit); notify = 1; } } VMX_CTR_PIR(vlapic, 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) { /* * While a virtual interrupt may have already been * processed the actual delivery maybe pending the * interruptibility of the guest. Recognize a pending * interrupt by reevaluating virtual interrupts * following Section 29.2.1 in the Intel SDM Volume 3. */ struct vm_exit *vmexit; uint8_t rvi, ppr; vmexit = vm_exitinfo(vlapic->vcpu); KASSERT(vmexit->exitcode == VM_EXITCODE_HLT, ("vmx_pending_intr: exitcode not 'HLT'")); rvi = vmexit->u.hlt.intr_status & APIC_TPR_INT; lapic = vlapic->apic_page; ppr = lapic->ppr & APIC_TPR_INT; if (rvi > ppr) { return (1); } return (0); } /* * 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 & APIC_TPR_INT; if (ppr == 0) return (1); VLAPIC_CTR1(vlapic, "HLT with non-zero PPR %d", lapic->ppr); vpr = 0; for (i = 3; i >= 0; i--) { pirval = pir_desc->pir[i]; if (pirval != 0) { vpr = (i * 64 + flsl(pirval) - 1) & APIC_TPR_INT; break; } } /* * If the highest-priority pending interrupt falls short of the * processor priority of this vCPU, ensure that 'pending_prio' does not * have any stale bits which would preclude a higher-priority interrupt * from incurring a notification later. */ if (vpr <= ppr) { const u_int prio_bit = VPR_PRIO_BIT(vpr); const u_int old = vlapic_vtx->pending_prio; if (old > prio_bit && (old & prio_bit) == 0) { vlapic_vtx->pending_prio = prio_bit; } return (0); } return (1); } 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 vmcs *vmcs; uint64_t mask, val; KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d", vector)); KASSERT(!vcpu_is_running(vlapic->vcpu, NULL), ("vmx_set_tmr: vcpu cannot be running")); vlapic_vtx = (struct vlapic_vtx *)vlapic; vmcs = vlapic_vtx->vcpu->vmcs; 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_ts(struct vlapic *vlapic) { struct vlapic_vtx *vlapic_vtx; struct vmx_vcpu *vcpu; struct vmcs *vmcs; uint32_t proc_ctls; vlapic_vtx = (struct vlapic_vtx *)vlapic; vcpu = vlapic_vtx->vcpu; vmcs = vcpu->vmcs; proc_ctls = vcpu->cap.proc_ctls; proc_ctls &= ~PROCBASED_USE_TPR_SHADOW; proc_ctls |= PROCBASED_CR8_LOAD_EXITING; proc_ctls |= PROCBASED_CR8_STORE_EXITING; vcpu->cap.proc_ctls = proc_ctls; VMPTRLD(vmcs); vmcs_write(VMCS_PRI_PROC_BASED_CTLS, proc_ctls); VMCLEAR(vmcs); } static void vmx_enable_x2apic_mode_vid(struct vlapic *vlapic) { struct vlapic_vtx *vlapic_vtx; struct vmx *vmx; struct vmx_vcpu *vcpu; struct vmcs *vmcs; uint32_t proc_ctls2; int error __diagused; vlapic_vtx = (struct vlapic_vtx *)vlapic; vcpu = vlapic_vtx->vcpu; vmx = vcpu->vmx; vmcs = vcpu->vmcs; proc_ctls2 = vcpu->cap.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; vcpu->cap.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) { VLAPIC_CTR0(vlapic, "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); VLAPIC_CTR2(vlapic, "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 *vcpui) { struct vmx *vmx; struct vmx_vcpu *vcpu; struct vlapic *vlapic; struct vlapic_vtx *vlapic_vtx; vcpu = vcpui; vmx = vcpu->vmx; vlapic = malloc(sizeof(struct vlapic_vtx), M_VLAPIC, M_WAITOK | M_ZERO); vlapic->vm = vmx->vm; vlapic->vcpu = vcpu->vcpu; vlapic->vcpuid = vcpu->vcpuid; vlapic->apic_page = (struct LAPIC *)vcpu->apic_page; vlapic_vtx = (struct vlapic_vtx *)vlapic; vlapic_vtx->pir_desc = vcpu->pir_desc; vlapic_vtx->vcpu = vcpu; if (tpr_shadowing) { vlapic->ops.enable_x2apic_mode = vmx_enable_x2apic_mode_ts; } 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_vid; } if (posted_interrupts) vlapic->ops.post_intr = vmx_post_intr; vlapic_init(vlapic); return (vlapic); } static void vmx_vlapic_cleanup(struct vlapic *vlapic) { vlapic_cleanup(vlapic); free(vlapic, M_VLAPIC); } #ifdef BHYVE_SNAPSHOT static int vmx_snapshot(void *vmi, struct vm_snapshot_meta *meta) { return (0); } static int vmx_vcpu_snapshot(void *vcpui, struct vm_snapshot_meta *meta) { struct vmcs *vmcs; struct vmx *vmx; struct vmx_vcpu *vcpu; struct vmxctx *vmxctx; int err, run, hostcpu; err = 0; vcpu = vcpui; vmx = vcpu->vmx; vmcs = vcpu->vmcs; run = vcpu_is_running(vcpu->vcpu, &hostcpu); if (run && hostcpu != curcpu) { printf("%s: %s%d is running", __func__, vm_name(vmx->vm), vcpu->vcpuid); return (EINVAL); } err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_CR0, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_CR3, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_CR4, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_DR7, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_RSP, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_RIP, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_RFLAGS, meta); /* Guest segments */ err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_ES, meta); err += vmcs_snapshot_desc(vmcs, run, VM_REG_GUEST_ES, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_CS, meta); err += vmcs_snapshot_desc(vmcs, run, VM_REG_GUEST_CS, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_SS, meta); err += vmcs_snapshot_desc(vmcs, run, VM_REG_GUEST_SS, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_DS, meta); err += vmcs_snapshot_desc(vmcs, run, VM_REG_GUEST_DS, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_FS, meta); err += vmcs_snapshot_desc(vmcs, run, VM_REG_GUEST_FS, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_GS, meta); err += vmcs_snapshot_desc(vmcs, run, VM_REG_GUEST_GS, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_TR, meta); err += vmcs_snapshot_desc(vmcs, run, VM_REG_GUEST_TR, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_LDTR, meta); err += vmcs_snapshot_desc(vmcs, run, VM_REG_GUEST_LDTR, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_EFER, meta); err += vmcs_snapshot_desc(vmcs, run, VM_REG_GUEST_IDTR, meta); err += vmcs_snapshot_desc(vmcs, run, VM_REG_GUEST_GDTR, meta); /* Guest page tables */ err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_PDPTE0, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_PDPTE1, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_PDPTE2, meta); err += vmcs_snapshot_reg(vmcs, run, VM_REG_GUEST_PDPTE3, meta); /* Other guest state */ err += vmcs_snapshot_any(vmcs, run, VMCS_GUEST_IA32_SYSENTER_CS, meta); err += vmcs_snapshot_any(vmcs, run, VMCS_GUEST_IA32_SYSENTER_ESP, meta); err += vmcs_snapshot_any(vmcs, run, VMCS_GUEST_IA32_SYSENTER_EIP, meta); err += vmcs_snapshot_any(vmcs, run, VMCS_GUEST_INTERRUPTIBILITY, meta); err += vmcs_snapshot_any(vmcs, run, VMCS_GUEST_ACTIVITY, meta); err += vmcs_snapshot_any(vmcs, run, VMCS_ENTRY_CTLS, meta); err += vmcs_snapshot_any(vmcs, run, VMCS_EXIT_CTLS, meta); if (err != 0) goto done; SNAPSHOT_BUF_OR_LEAVE(vcpu->guest_msrs, sizeof(vcpu->guest_msrs), meta, err, done); vmxctx = &vcpu->ctx; SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rdi, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rsi, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rdx, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rcx, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r8, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r9, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rax, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rbx, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rbp, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r10, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r11, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r12, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r13, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r14, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r15, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_cr2, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_dr0, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_dr1, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_dr2, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_dr3, meta, err, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_dr6, meta, err, done); done: return (err); } static int vmx_restore_tsc(void *vcpui, uint64_t offset) { struct vmx_vcpu *vcpu = vcpui; struct vmcs *vmcs; struct vmx *vmx; int error, running, hostcpu; vmx = vcpu->vmx; vmcs = vcpu->vmcs; running = vcpu_is_running(vcpu->vcpu, &hostcpu); if (running && hostcpu != curcpu) { printf("%s: %s%d is running", __func__, vm_name(vmx->vm), vcpu->vcpuid); return (EINVAL); } if (!running) VMPTRLD(vmcs); error = vmx_set_tsc_offset(vcpu, offset); if (!running) VMCLEAR(vmcs); return (error); } #endif const struct vmm_ops vmm_ops_intel = { .modinit = vmx_modinit, .modcleanup = vmx_modcleanup, .modresume = vmx_modresume, .init = vmx_init, .run = vmx_run, .cleanup = vmx_cleanup, .vcpu_init = vmx_vcpu_init, .vcpu_cleanup = vmx_vcpu_cleanup, .getreg = vmx_getreg, .setreg = vmx_setreg, .getdesc = vmx_getdesc, .setdesc = vmx_setdesc, .getcap = vmx_getcap, .setcap = vmx_setcap, .vmspace_alloc = vmx_vmspace_alloc, .vmspace_free = vmx_vmspace_free, .vlapic_init = vmx_vlapic_init, .vlapic_cleanup = vmx_vlapic_cleanup, #ifdef BHYVE_SNAPSHOT .snapshot = vmx_snapshot, .vcpu_snapshot = vmx_vcpu_snapshot, .restore_tsc = vmx_restore_tsc, #endif }; diff --git a/sys/amd64/vmm/vmm.c b/sys/amd64/vmm/vmm.c index 24f97a9244f0..7327bf6f40be 100644 --- a/sys/amd64/vmm/vmm.c +++ b/sys/amd64/vmm/vmm.c @@ -1,2991 +1,2992 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * 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 "opt_bhyve_snapshot.h" #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 #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 vcpuid; /* (o) */ int hostcpu; /* (o) vcpu's host cpu */ int reqidle; /* (i) request vcpu to idle */ struct vm *vm; /* (o) */ void *cookie; /* (i) cpu-specific data */ 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 */ uint64_t tsc_offset; /* (o) TSC offsetting */ }; #define vcpu_lock_init(v) mtx_init(&((v)->mtx), "vcpu lock", 0, MTX_SPIN) #define vcpu_lock_destroy(v) mtx_destroy(&((v)->mtx)) #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 { size_t len; bool sysmem; struct vm_object *object; }; #define VM_MAX_MEMSEGS 4 struct mem_map { vm_paddr_t gpa; size_t len; vm_ooffset_t segoff; int segid; int prot; int flags; }; #define VM_MAX_MEMMAPS 8 /* * 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 * * Locking: * [m] mem_segs_lock * [r] rendezvous_mtx * [v] reads require one frozen vcpu, writes require freezing all vcpus */ 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 */ volatile cpuset_t debug_cpus; /* (i) vcpus stopped for debug */ cpuset_t startup_cpus; /* (i) [r] waiting for startup */ int suspend; /* (i) stop VM execution */ bool dying; /* (o) is dying */ 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) [r] rendezvous requested */ cpuset_t rendezvous_done_cpus; /* (x) [r] rendezvous finished */ void *rendezvous_arg; /* (x) [r] rendezvous func/arg */ vm_rendezvous_func_t rendezvous_func; struct mtx rendezvous_mtx; /* (o) rendezvous lock */ struct mem_map mem_maps[VM_MAX_MEMMAPS]; /* (i) [m+v] guest address space */ struct mem_seg mem_segs[VM_MAX_MEMSEGS]; /* (o) [m+v] guest memory regions */ struct vmspace *vmspace; /* (o) guest's address space */ char name[VM_MAX_NAMELEN+1]; /* (o) virtual machine name */ struct vcpu **vcpu; /* (o) guest vcpus */ /* The following describe the vm cpu topology */ uint16_t sockets; /* (o) num of sockets */ uint16_t cores; /* (o) num of cores/socket */ uint16_t threads; /* (o) num of threads/core */ uint16_t maxcpus; /* (o) max pluggable cpus */ struct sx mem_segs_lock; /* (o) */ struct sx vcpus_init_lock; /* (o) */ }; #define VMM_CTR0(vcpu, format) \ VCPU_CTR0((vcpu)->vm, (vcpu)->vcpuid, format) #define VMM_CTR1(vcpu, format, p1) \ VCPU_CTR1((vcpu)->vm, (vcpu)->vcpuid, format, p1) #define VMM_CTR2(vcpu, format, p1, p2) \ VCPU_CTR2((vcpu)->vm, (vcpu)->vcpuid, format, p1, p2) #define VMM_CTR3(vcpu, format, p1, p2, p3) \ VCPU_CTR3((vcpu)->vm, (vcpu)->vcpuid, format, p1, p2, p3) #define VMM_CTR4(vcpu, format, p1, p2, p3, p4) \ VCPU_CTR4((vcpu)->vm, (vcpu)->vcpuid, format, p1, p2, p3, p4) static int vmm_initialized; static void vmmops_panic(void); static void vmmops_panic(void) { panic("vmm_ops func called when !vmm_is_intel() && !vmm_is_svm()"); } #define DEFINE_VMMOPS_IFUNC(ret_type, opname, args) \ DEFINE_IFUNC(static, ret_type, vmmops_##opname, args) \ { \ if (vmm_is_intel()) \ return (vmm_ops_intel.opname); \ else if (vmm_is_svm()) \ return (vmm_ops_amd.opname); \ else \ return ((ret_type (*)args)vmmops_panic); \ } DEFINE_VMMOPS_IFUNC(int, modinit, (int ipinum)) DEFINE_VMMOPS_IFUNC(int, modcleanup, (void)) DEFINE_VMMOPS_IFUNC(void, modresume, (void)) DEFINE_VMMOPS_IFUNC(void *, init, (struct vm *vm, struct pmap *pmap)) DEFINE_VMMOPS_IFUNC(int, run, (void *vcpui, register_t rip, struct pmap *pmap, struct vm_eventinfo *info)) DEFINE_VMMOPS_IFUNC(void, cleanup, (void *vmi)) DEFINE_VMMOPS_IFUNC(void *, vcpu_init, (void *vmi, struct vcpu *vcpu, int vcpu_id)) DEFINE_VMMOPS_IFUNC(void, vcpu_cleanup, (void *vcpui)) DEFINE_VMMOPS_IFUNC(int, getreg, (void *vcpui, int num, uint64_t *retval)) DEFINE_VMMOPS_IFUNC(int, setreg, (void *vcpui, int num, uint64_t val)) DEFINE_VMMOPS_IFUNC(int, getdesc, (void *vcpui, int num, struct seg_desc *desc)) DEFINE_VMMOPS_IFUNC(int, setdesc, (void *vcpui, int num, struct seg_desc *desc)) DEFINE_VMMOPS_IFUNC(int, getcap, (void *vcpui, int num, int *retval)) DEFINE_VMMOPS_IFUNC(int, setcap, (void *vcpui, int num, int val)) DEFINE_VMMOPS_IFUNC(struct vmspace *, vmspace_alloc, (vm_offset_t min, vm_offset_t max)) DEFINE_VMMOPS_IFUNC(void, vmspace_free, (struct vmspace *vmspace)) DEFINE_VMMOPS_IFUNC(struct vlapic *, vlapic_init, (void *vcpui)) DEFINE_VMMOPS_IFUNC(void, vlapic_cleanup, (struct vlapic *vlapic)) #ifdef BHYVE_SNAPSHOT DEFINE_VMMOPS_IFUNC(int, snapshot, (void *vmi, struct vm_snapshot_meta *meta)) DEFINE_VMMOPS_IFUNC(int, vcpu_snapshot, (void *vcpui, struct vm_snapshot_meta *meta)) DEFINE_VMMOPS_IFUNC(int, restore_tsc, (void *vcpui, uint64_t now)) #endif #define fpu_start_emulating() load_cr0(rcr0() | CR0_TS) #define fpu_stop_emulating() clts() SDT_PROVIDER_DEFINE(vmm); 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 | CTLFLAG_MPSAFE, 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 trap_wbinvd; SYSCTL_INT(_hw_vmm, OID_AUTO, trap_wbinvd, CTLFLAG_RDTUN, &trap_wbinvd, 0, "WBINVD triggers a VM-exit"); u_int vm_maxcpu; SYSCTL_UINT(_hw_vmm, OID_AUTO, maxcpu, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &vm_maxcpu, 0, "Maximum number of vCPUs"); static void vm_free_memmap(struct vm *vm, int ident); static bool sysmem_mapping(struct vm *vm, struct mem_map *mm); static void vcpu_notify_event_locked(struct vcpu *vcpu, bool lapic_intr); /* * Upper limit on vm_maxcpu. Limited by use of uint16_t types for CPU * counts as well as range of vpid values for VT-x and by the capacity * of cpuset_t masks. The call to new_unrhdr() in vpid_init() in * vmx.c requires 'vm_maxcpu + 1 <= 0xffff', hence the '- 1' below. */ #define VM_MAXCPU MIN(0xffff - 1, CPU_SETSIZE) #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 vcpu *vcpu, bool destroy) { vmmops_vlapic_cleanup(vcpu->vlapic); vmmops_vcpu_cleanup(vcpu->cookie); vcpu->cookie = NULL; if (destroy) { vmm_stat_free(vcpu->stats); fpu_save_area_free(vcpu->guestfpu); vcpu_lock_destroy(vcpu); free(vcpu, M_VM); } } static struct vcpu * vcpu_alloc(struct vm *vm, int vcpu_id) { struct vcpu *vcpu; KASSERT(vcpu_id >= 0 && vcpu_id < vm->maxcpus, ("vcpu_init: invalid vcpu %d", vcpu_id)); vcpu = malloc(sizeof(*vcpu), M_VM, M_WAITOK | M_ZERO); vcpu_lock_init(vcpu); vcpu->state = VCPU_IDLE; vcpu->hostcpu = NOCPU; vcpu->vcpuid = vcpu_id; vcpu->vm = vm; vcpu->guestfpu = fpu_save_area_alloc(); vcpu->stats = vmm_stat_alloc(); vcpu->tsc_offset = 0; return (vcpu); } static void vcpu_init(struct vcpu *vcpu) { vcpu->cookie = vmmops_vcpu_init(vcpu->vm->cookie, vcpu, vcpu->vcpuid); vcpu->vlapic = vmmops_vlapic_init(vcpu->cookie); vm_set_x2apic_state(vcpu, 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 vcpu *vcpu) { return (trace_guest_exceptions); } int vcpu_trap_wbinvd(struct vcpu *vcpu) { return (trap_wbinvd); } struct vm_exit * vm_exitinfo(struct vcpu *vcpu) { return (&vcpu->exitinfo); } static int vmm_init(void) { int error; if (!vmm_is_hw_supported()) return (ENXIO); vm_maxcpu = mp_ncpus; TUNABLE_INT_FETCH("hw.vmm.maxcpu", &vm_maxcpu); if (vm_maxcpu > VM_MAXCPU) { printf("vmm: vm_maxcpu clamped to %u\n", VM_MAXCPU); vm_maxcpu = VM_MAXCPU; } if (vm_maxcpu == 0) vm_maxcpu = 1; vmm_host_state_init(); vmm_ipinum = lapic_ipi_alloc(pti ? &IDTVEC(justreturn1_pti) : &IDTVEC(justreturn)); if (vmm_ipinum < 0) vmm_ipinum = IPI_AST; error = vmm_mem_init(); if (error) return (error); vmm_resume_p = vmmops_modresume; return (vmmops_modinit(vmm_ipinum)); } static int vmm_handler(module_t mod, int what, void *arg) { int error; switch (what) { case MOD_LOAD: if (vmm_is_hw_supported()) { vmmdev_init(); error = vmm_init(); if (error == 0) vmm_initialized = 1; } else { error = ENXIO; } break; case MOD_UNLOAD: if (vmm_is_hw_supported()) { error = vmmdev_cleanup(); if (error == 0) { vmm_resume_p = NULL; iommu_cleanup(); if (vmm_ipinum != IPI_AST) lapic_ipi_free(vmm_ipinum); error = vmmops_modcleanup(); /* * Something bad happened - prevent new * VMs from being created */ if (error) vmm_initialized = 0; } } else { error = 0; } break; default: error = 0; break; } return (error); } static moduledata_t vmm_kmod = { "vmm", vmm_handler, NULL }; /* * vmm initialization has the following dependencies: * * - 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) { vm->cookie = vmmops_init(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); CPU_ZERO(&vm->debug_cpus); CPU_ZERO(&vm->startup_cpus); vm->suspend = 0; CPU_ZERO(&vm->suspended_cpus); if (!create) { for (int i = 0; i < vm->maxcpus; i++) { if (vm->vcpu[i] != NULL) vcpu_init(vm->vcpu[i]); } } } void vm_disable_vcpu_creation(struct vm *vm) { sx_xlock(&vm->vcpus_init_lock); vm->dying = true; sx_xunlock(&vm->vcpus_init_lock); } struct vcpu * vm_alloc_vcpu(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm_get_maxcpus(vm)) return (NULL); vcpu = atomic_load_ptr(&vm->vcpu[vcpuid]); if (__predict_true(vcpu != NULL)) return (vcpu); sx_xlock(&vm->vcpus_init_lock); vcpu = vm->vcpu[vcpuid]; if (vcpu == NULL && !vm->dying) { vcpu = vcpu_alloc(vm, vcpuid); vcpu_init(vcpu); /* * Ensure vCPU is fully created before updating pointer * to permit unlocked reads above. */ atomic_store_rel_ptr((uintptr_t *)&vm->vcpu[vcpuid], (uintptr_t)vcpu); } sx_xunlock(&vm->vcpus_init_lock); return (vcpu); } void vm_slock_vcpus(struct vm *vm) { sx_slock(&vm->vcpus_init_lock); } void vm_unlock_vcpus(struct vm *vm) { sx_unlock(&vm->vcpus_init_lock); } /* * The default CPU topology is a single thread per package. */ u_int cores_per_package = 1; u_int threads_per_core = 1; 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 || strnlen(name, VM_MAX_NAMELEN + 1) == VM_MAX_NAMELEN + 1) return (EINVAL); vmspace = vmmops_vmspace_alloc(0, VM_MAXUSER_ADDRESS_LA48); if (vmspace == NULL) return (ENOMEM); vm = malloc(sizeof(struct vm), M_VM, M_WAITOK | M_ZERO); strcpy(vm->name, name); vm->vmspace = vmspace; mtx_init(&vm->rendezvous_mtx, "vm rendezvous lock", 0, MTX_DEF); sx_init(&vm->mem_segs_lock, "vm mem_segs"); sx_init(&vm->vcpus_init_lock, "vm vcpus"); vm->vcpu = malloc(sizeof(*vm->vcpu) * vm_maxcpu, M_VM, M_WAITOK | M_ZERO); vm->sockets = 1; vm->cores = cores_per_package; /* XXX backwards compatibility */ vm->threads = threads_per_core; /* XXX backwards compatibility */ vm->maxcpus = vm_maxcpu; vm_init(vm, true); *retvm = vm; return (0); } void vm_get_topology(struct vm *vm, uint16_t *sockets, uint16_t *cores, uint16_t *threads, uint16_t *maxcpus) { *sockets = vm->sockets; *cores = vm->cores; *threads = vm->threads; *maxcpus = vm->maxcpus; } uint16_t vm_get_maxcpus(struct vm *vm) { return (vm->maxcpus); } int vm_set_topology(struct vm *vm, uint16_t sockets, uint16_t cores, uint16_t threads, uint16_t maxcpus __unused) { /* Ignore maxcpus. */ if ((sockets * cores * threads) > vm->maxcpus) return (EINVAL); vm->sockets = sockets; vm->cores = cores; vm->threads = threads; return(0); } static void vm_cleanup(struct vm *vm, bool destroy) { struct mem_map *mm; int i; if (destroy) vm_xlock_memsegs(vm); 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->maxcpus; i++) { if (vm->vcpu[i] != NULL) vcpu_cleanup(vm->vcpu[i], destroy); } vmmops_cleanup(vm->cookie); /* * System memory is removed from the guest address space only when * the VM is destroyed. This is because the mapping remains the same * across VM reset. * * Device memory can be relocated by the guest (e.g. using PCI BARs) * so those mappings are removed on a VM reset. */ for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (destroy || !sysmem_mapping(vm, mm)) vm_free_memmap(vm, i); } if (destroy) { for (i = 0; i < VM_MAX_MEMSEGS; i++) vm_free_memseg(vm, i); vm_unlock_memsegs(vm); vmmops_vmspace_free(vm->vmspace); vm->vmspace = NULL; free(vm->vcpu, M_VM); sx_destroy(&vm->vcpus_init_lock); sx_destroy(&vm->mem_segs_lock); mtx_destroy(&vm->rendezvous_mtx); } } 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); } void vm_slock_memsegs(struct vm *vm) { sx_slock(&vm->mem_segs_lock); } void vm_xlock_memsegs(struct vm *vm) { sx_xlock(&vm->mem_segs_lock); } void vm_unlock_memsegs(struct vm *vm) { sx_unlock(&vm->mem_segs_lock); } 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); } /* * Return 'true' if 'gpa' is allocated in the guest address space. * * This function is called in the context of a running vcpu which acts as * an implicit lock on 'vm->mem_maps[]'. */ bool vm_mem_allocated(struct vcpu *vcpu, vm_paddr_t gpa) { struct vm *vm = vcpu->vm; struct mem_map *mm; int i; #ifdef INVARIANTS int hostcpu, state; state = vcpu_get_state(vcpu, &hostcpu); KASSERT(state == VCPU_RUNNING && hostcpu == curcpu, ("%s: invalid vcpu state %d/%d", __func__, state, hostcpu)); #endif for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (mm->len != 0 && gpa >= mm->gpa && gpa < mm->gpa + mm->len) return (true); /* 'gpa' is sysmem or devmem */ } if (ppt_is_mmio(vm, gpa)) return (true); /* 'gpa' is pci passthru mmio */ return (false); } int vm_alloc_memseg(struct vm *vm, int ident, size_t len, bool sysmem) { struct mem_seg *seg; vm_object_t obj; sx_assert(&vm->mem_segs_lock, SX_XLOCKED); if (ident < 0 || ident >= VM_MAX_MEMSEGS) return (EINVAL); if (len == 0 || (len & PAGE_MASK)) return (EINVAL); seg = &vm->mem_segs[ident]; if (seg->object != NULL) { if (seg->len == len && seg->sysmem == sysmem) return (EEXIST); else return (EINVAL); } obj = vm_object_allocate(OBJT_SWAP, len >> PAGE_SHIFT); if (obj == NULL) return (ENOMEM); seg->len = len; seg->object = obj; seg->sysmem = sysmem; return (0); } int vm_get_memseg(struct vm *vm, int ident, size_t *len, bool *sysmem, vm_object_t *objptr) { struct mem_seg *seg; sx_assert(&vm->mem_segs_lock, SX_LOCKED); if (ident < 0 || ident >= VM_MAX_MEMSEGS) return (EINVAL); seg = &vm->mem_segs[ident]; if (len) *len = seg->len; if (sysmem) *sysmem = seg->sysmem; if (objptr) *objptr = seg->object; return (0); } void vm_free_memseg(struct vm *vm, int ident) { struct mem_seg *seg; KASSERT(ident >= 0 && ident < VM_MAX_MEMSEGS, ("%s: invalid memseg ident %d", __func__, ident)); seg = &vm->mem_segs[ident]; if (seg->object != NULL) { vm_object_deallocate(seg->object); bzero(seg, sizeof(struct mem_seg)); } } int vm_mmap_memseg(struct vm *vm, vm_paddr_t gpa, int segid, vm_ooffset_t first, size_t len, int prot, int flags) { struct mem_seg *seg; struct mem_map *m, *map; vm_ooffset_t last; int i, error; if (prot == 0 || (prot & ~(VM_PROT_ALL)) != 0) return (EINVAL); if (flags & ~VM_MEMMAP_F_WIRED) return (EINVAL); if (segid < 0 || segid >= VM_MAX_MEMSEGS) return (EINVAL); seg = &vm->mem_segs[segid]; if (seg->object == NULL) return (EINVAL); last = first + len; if (first < 0 || first >= last || last > seg->len) return (EINVAL); if ((gpa | first | last) & PAGE_MASK) return (EINVAL); map = NULL; for (i = 0; i < VM_MAX_MEMMAPS; i++) { m = &vm->mem_maps[i]; if (m->len == 0) { map = m; break; } } if (map == NULL) return (ENOSPC); error = vm_map_find(&vm->vmspace->vm_map, seg->object, first, &gpa, len, 0, VMFS_NO_SPACE, prot, prot, 0); if (error != KERN_SUCCESS) return (EFAULT); vm_object_reference(seg->object); if (flags & VM_MEMMAP_F_WIRED) { error = vm_map_wire(&vm->vmspace->vm_map, gpa, gpa + len, VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); if (error != KERN_SUCCESS) { vm_map_remove(&vm->vmspace->vm_map, gpa, gpa + len); return (error == KERN_RESOURCE_SHORTAGE ? ENOMEM : EFAULT); } } map->gpa = gpa; map->len = len; map->segoff = first; map->segid = segid; map->prot = prot; map->flags = flags; return (0); } int vm_munmap_memseg(struct vm *vm, vm_paddr_t gpa, size_t len) { struct mem_map *m; int i; for (i = 0; i < VM_MAX_MEMMAPS; i++) { m = &vm->mem_maps[i]; if (m->gpa == gpa && m->len == len && (m->flags & VM_MEMMAP_F_IOMMU) == 0) { vm_free_memmap(vm, i); return (0); } } return (EINVAL); } int vm_mmap_getnext(struct vm *vm, vm_paddr_t *gpa, int *segid, vm_ooffset_t *segoff, size_t *len, int *prot, int *flags) { struct mem_map *mm, *mmnext; int i; mmnext = NULL; for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (mm->len == 0 || mm->gpa < *gpa) continue; if (mmnext == NULL || mm->gpa < mmnext->gpa) mmnext = mm; } if (mmnext != NULL) { *gpa = mmnext->gpa; if (segid) *segid = mmnext->segid; if (segoff) *segoff = mmnext->segoff; if (len) *len = mmnext->len; if (prot) *prot = mmnext->prot; if (flags) *flags = mmnext->flags; return (0); } else { return (ENOENT); } } static void vm_free_memmap(struct vm *vm, int ident) { struct mem_map *mm; int error __diagused; mm = &vm->mem_maps[ident]; if (mm->len) { error = vm_map_remove(&vm->vmspace->vm_map, mm->gpa, mm->gpa + mm->len); KASSERT(error == KERN_SUCCESS, ("%s: vm_map_remove error %d", __func__, error)); bzero(mm, sizeof(struct mem_map)); } } static __inline bool sysmem_mapping(struct vm *vm, struct mem_map *mm) { if (mm->len != 0 && vm->mem_segs[mm->segid].sysmem) return (true); else return (false); } vm_paddr_t vmm_sysmem_maxaddr(struct vm *vm) { struct mem_map *mm; vm_paddr_t maxaddr; int i; maxaddr = 0; for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (sysmem_mapping(vm, mm)) { if (maxaddr < mm->gpa + mm->len) maxaddr = mm->gpa + mm->len; } } return (maxaddr); } static void vm_iommu_modify(struct vm *vm, bool map) { int i, sz; vm_paddr_t gpa, hpa; struct mem_map *mm; void *vp, *cookie, *host_domain; sz = PAGE_SIZE; host_domain = iommu_host_domain(); for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (!sysmem_mapping(vm, mm)) continue; if (map) { KASSERT((mm->flags & VM_MEMMAP_F_IOMMU) == 0, ("iommu map found invalid memmap %#lx/%#lx/%#x", mm->gpa, mm->len, mm->flags)); if ((mm->flags & VM_MEMMAP_F_WIRED) == 0) continue; mm->flags |= VM_MEMMAP_F_IOMMU; } else { if ((mm->flags & VM_MEMMAP_F_IOMMU) == 0) continue; mm->flags &= ~VM_MEMMAP_F_IOMMU; KASSERT((mm->flags & VM_MEMMAP_F_WIRED) != 0, ("iommu unmap found invalid memmap %#lx/%#lx/%#x", mm->gpa, mm->len, mm->flags)); } gpa = mm->gpa; while (gpa < mm->gpa + mm->len) { vp = vm_gpa_hold_global(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); } else { iommu_remove_mapping(vm->iommu, gpa, 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); return (0); } int vm_assign_pptdev(struct vm *vm, int bus, int slot, int func) { int error; vm_paddr_t maxaddr; /* Set up the IOMMU to do the 'gpa' to 'hpa' translation */ if (ppt_assigned_devices(vm) == 0) { KASSERT(vm->iommu == NULL, ("vm_assign_pptdev: iommu must be NULL")); maxaddr = vmm_sysmem_maxaddr(vm); vm->iommu = iommu_create_domain(maxaddr); if (vm->iommu == NULL) return (ENXIO); vm_iommu_map(vm); } error = ppt_assign_device(vm, bus, slot, func); return (error); } static void * _vm_gpa_hold(struct vm *vm, vm_paddr_t gpa, size_t len, int reqprot, void **cookie) { int i, count, pageoff; struct mem_map *mm; 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 = 0; for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (gpa >= mm->gpa && gpa < mm->gpa + mm->len) { count = vm_fault_quick_hold_pages(&vm->vmspace->vm_map, trunc_page(gpa), PAGE_SIZE, reqprot, &m, 1); break; } } if (count == 1) { *cookie = m; return ((void *)(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)) + pageoff)); } else { *cookie = NULL; return (NULL); } } void * vm_gpa_hold(struct vcpu *vcpu, vm_paddr_t gpa, size_t len, int reqprot, void **cookie) { #ifdef INVARIANTS /* * The current vcpu should be frozen to ensure 'vm_memmap[]' * stability. */ int state = vcpu_get_state(vcpu, NULL); KASSERT(state == VCPU_FROZEN, ("%s: invalid vcpu state %d", __func__, state)); #endif return (_vm_gpa_hold(vcpu->vm, gpa, len, reqprot, cookie)); } void * vm_gpa_hold_global(struct vm *vm, vm_paddr_t gpa, size_t len, int reqprot, void **cookie) { sx_assert(&vm->mem_segs_lock, SX_LOCKED); return (_vm_gpa_hold(vm, gpa, len, reqprot, cookie)); } void vm_gpa_release(void *cookie) { vm_page_t m = cookie; vm_page_unwire(m, PQ_ACTIVE); } int vm_get_register(struct vcpu *vcpu, int reg, uint64_t *retval) { if (reg >= VM_REG_LAST) return (EINVAL); return (vmmops_getreg(vcpu->cookie, reg, retval)); } int vm_set_register(struct vcpu *vcpu, int reg, uint64_t val) { int error; if (reg >= VM_REG_LAST) return (EINVAL); error = vmmops_setreg(vcpu->cookie, reg, val); if (error || reg != VM_REG_GUEST_RIP) return (error); /* Set 'nextrip' to match the value of %rip */ VMM_CTR1(vcpu, "Setting nextrip to %#lx", val); vcpu->nextrip = val; return (0); } static bool is_descriptor_table(int reg) { switch (reg) { case VM_REG_GUEST_IDTR: case VM_REG_GUEST_GDTR: return (true); default: return (false); } } static bool 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 vcpu *vcpu, int reg, struct seg_desc *desc) { if (!is_segment_register(reg) && !is_descriptor_table(reg)) return (EINVAL); return (vmmops_getdesc(vcpu->cookie, reg, desc)); } int vm_set_seg_desc(struct vcpu *vcpu, int reg, struct seg_desc *desc) { if (!is_segment_register(reg) && !is_descriptor_table(reg)) return (EINVAL); return (vmmops_setdesc(vcpu->cookie, 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, bool from_idle) { int error; 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) { vcpu->reqidle = 1; vcpu_notify_event_locked(vcpu, false); VMM_CTR1(vcpu, "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); VMM_CTR2(vcpu, "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 vcpu *vcpu, enum vcpu_state newstate) { int error; if ((error = vcpu_set_state(vcpu, 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) { int error; if ((error = vcpu_set_state_locked(vcpu, newstate, false)) != 0) panic("Error %d setting state to %d", error, newstate); } static int vm_handle_rendezvous(struct vcpu *vcpu) { struct vm *vm = vcpu->vm; struct thread *td; int error, vcpuid; error = 0; vcpuid = vcpu->vcpuid; td = curthread; 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->rendezvous_req_cpus, &vm->active_cpus); if (CPU_ISSET(vcpuid, &vm->rendezvous_req_cpus) && !CPU_ISSET(vcpuid, &vm->rendezvous_done_cpus)) { VMM_CTR0(vcpu, "Calling rendezvous func"); (*vm->rendezvous_func)(vcpu, vm->rendezvous_arg); CPU_SET(vcpuid, &vm->rendezvous_done_cpus); } if (CPU_CMP(&vm->rendezvous_req_cpus, &vm->rendezvous_done_cpus) == 0) { VMM_CTR0(vcpu, "Rendezvous completed"); + CPU_ZERO(&vm->rendezvous_req_cpus); vm->rendezvous_func = NULL; wakeup(&vm->rendezvous_func); break; } VMM_CTR0(vcpu, "Wait for rendezvous completion"); mtx_sleep(&vm->rendezvous_func, &vm->rendezvous_mtx, 0, "vmrndv", hz); if (td_ast_pending(td, TDA_SUSPEND)) { mtx_unlock(&vm->rendezvous_mtx); error = thread_check_susp(td, true); if (error != 0) return (error); mtx_lock(&vm->rendezvous_mtx); } } mtx_unlock(&vm->rendezvous_mtx); return (0); } /* * Emulate a guest 'hlt' by sleeping until the vcpu is ready to run. */ static int vm_handle_hlt(struct vcpu *vcpu, bool intr_disabled, bool *retu) { struct vm *vm = vcpu->vm; const char *wmesg; struct thread *td; int error, t, vcpuid, vcpu_halted, vm_halted; vcpuid = vcpu->vcpuid; vcpu_halted = 0; vm_halted = 0; error = 0; td = curthread; KASSERT(!CPU_ISSET(vcpuid, &vm->halted_cpus), ("vcpu already halted")); 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 vmmops_run() and before it acquired the * vcpu lock above. */ if (vm->rendezvous_func != NULL || vm->suspend || vcpu->reqidle) break; if (vm_nmi_pending(vcpu)) break; if (!intr_disabled) { if (vm_extint_pending(vcpu) || vlapic_pending_intr(vcpu->vlapic, NULL)) { break; } } /* Don't go to sleep if the vcpu thread needs to yield */ if (vcpu_should_yield(vcpu)) break; if (vcpu_debugged(vcpu)) 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"; VMM_CTR0(vcpu, "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); /* * 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); vmm_stat_incr(vcpu, VCPU_IDLE_TICKS, ticks - t); if (td_ast_pending(td, TDA_SUSPEND)) { vcpu_unlock(vcpu); error = thread_check_susp(td, false); if (error != 0) { if (vcpu_halted) { CPU_CLR_ATOMIC(vcpuid, &vm->halted_cpus); } return (error); } vcpu_lock(vcpu); } } 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 vcpu *vcpu, bool *retu) { struct vm *vm = vcpu->vm; int rv, ftype; struct vm_map *map; struct vm_exit *vme; 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) { VMM_CTR2(vcpu, "%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, NULL); VMM_CTR3(vcpu, "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 vcpu *vcpu, bool *retu) { struct vie *vie; 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; 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; VMM_CTR1(vcpu, "inst_emul fault accessing gpa %#lx", gpa); /* Fetch, decode and emulate the faulting instruction */ if (vie->num_valid == 0) { error = vmm_fetch_instruction(vcpu, 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(vcpu, gla, cpu_mode, cs_d, vie) != 0) { VMM_CTR1(vcpu, "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; VMM_CTR1(vcpu, "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(vcpu, gpa, vie, paging, mread, mwrite, retu); return (error); } static int vm_handle_suspend(struct vcpu *vcpu, bool *retu) { struct vm *vm = vcpu->vm; int error, i; struct thread *td; error = 0; td = curthread; CPU_SET_ATOMIC(vcpu->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 (error == 0) { if (CPU_CMP(&vm->suspended_cpus, &vm->active_cpus) == 0) { VMM_CTR0(vcpu, "All vcpus suspended"); break; } if (vm->rendezvous_func == NULL) { VMM_CTR0(vcpu, "Sleeping during suspend"); vcpu_require_state_locked(vcpu, VCPU_SLEEPING); msleep_spin(vcpu, &vcpu->mtx, "vmsusp", hz); vcpu_require_state_locked(vcpu, VCPU_FROZEN); if (td_ast_pending(td, TDA_SUSPEND)) { vcpu_unlock(vcpu); error = thread_check_susp(td, false); vcpu_lock(vcpu); } } else { VMM_CTR0(vcpu, "Rendezvous during suspend"); vcpu_unlock(vcpu); error = vm_handle_rendezvous(vcpu); vcpu_lock(vcpu); } } vcpu_unlock(vcpu); /* * Wakeup the other sleeping vcpus and return to userspace. */ for (i = 0; i < vm->maxcpus; i++) { if (CPU_ISSET(i, &vm->suspended_cpus)) { vcpu_notify_event(vm_vcpu(vm, i), false); } } *retu = true; return (error); } static int vm_handle_reqidle(struct vcpu *vcpu, bool *retu) { 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->maxcpus; i++) { if (CPU_ISSET(i, &vm->active_cpus)) vcpu_notify_event(vm_vcpu(vm, i), false); } return (0); } void vm_exit_suspended(struct vcpu *vcpu, uint64_t rip) { struct vm *vm = vcpu->vm; 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(vcpu); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_SUSPENDED; vmexit->u.suspended.how = vm->suspend; } void vm_exit_debug(struct vcpu *vcpu, uint64_t rip) { struct vm_exit *vmexit; vmexit = vm_exitinfo(vcpu); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_DEBUG; } void vm_exit_rendezvous(struct vcpu *vcpu, uint64_t rip) { struct vm_exit *vmexit; vmexit = vm_exitinfo(vcpu); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_RENDEZVOUS; vmm_stat_incr(vcpu, VMEXIT_RENDEZVOUS, 1); } void vm_exit_reqidle(struct vcpu *vcpu, uint64_t rip) { struct vm_exit *vmexit; vmexit = vm_exitinfo(vcpu); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_REQIDLE; vmm_stat_incr(vcpu, VMEXIT_REQIDLE, 1); } void vm_exit_astpending(struct vcpu *vcpu, uint64_t rip) { struct vm_exit *vmexit; vmexit = vm_exitinfo(vcpu); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_BOGUS; vmm_stat_incr(vcpu, VMEXIT_ASTPENDING, 1); } int vm_run(struct vcpu *vcpu, struct vm_exit *vme_user) { struct vm *vm = vcpu->vm; struct vm_eventinfo evinfo; int error, vcpuid; struct pcb *pcb; uint64_t tscval; struct vm_exit *vme; bool retu, intr_disabled; pmap_t pmap; vcpuid = vcpu->vcpuid; if (!CPU_ISSET(vcpuid, &vm->active_cpus)) return (EINVAL); if (CPU_ISSET(vcpuid, &vm->suspended_cpus)) return (EINVAL); pmap = vmspace_pmap(vm->vmspace); vme = &vcpu->exitinfo; - evinfo.rptr = &vm->rendezvous_func; + evinfo.rptr = &vm->rendezvous_req_cpus; 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(vcpu, VCPU_RUNNING); error = vmmops_run(vcpu->cookie, vcpu->nextrip, pmap, &evinfo); vcpu_require_state(vcpu, VCPU_FROZEN); save_guest_fpustate(vcpu); vmm_stat_incr(vcpu, 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(vcpu, &retu); break; case VM_EXITCODE_SUSPENDED: error = vm_handle_suspend(vcpu, &retu); break; case VM_EXITCODE_IOAPIC_EOI: vioapic_process_eoi(vm, vme->u.ioapic_eoi.vector); break; case VM_EXITCODE_RENDEZVOUS: error = vm_handle_rendezvous(vcpu); break; case VM_EXITCODE_HLT: intr_disabled = ((vme->u.hlt.rflags & PSL_I) == 0); error = vm_handle_hlt(vcpu, intr_disabled, &retu); break; case VM_EXITCODE_PAGING: error = vm_handle_paging(vcpu, &retu); break; case VM_EXITCODE_INST_EMUL: error = vm_handle_inst_emul(vcpu, &retu); break; case VM_EXITCODE_INOUT: case VM_EXITCODE_INOUT_STR: error = vm_handle_inout(vcpu, vme, &retu); break; case VM_EXITCODE_MONITOR: case VM_EXITCODE_MWAIT: case VM_EXITCODE_VMINSN: vm_inject_ud(vcpu); break; default: retu = true; /* handled in userland */ break; } } /* * VM_EXITCODE_INST_EMUL could access the apic which could transform the * exit code into VM_EXITCODE_IPI. */ if (error == 0 && vme->exitcode == VM_EXITCODE_IPI) { retu = false; error = vm_handle_ipi(vcpu, vme, &retu); } if (error == 0 && retu == false) goto restart; vmm_stat_incr(vcpu, VMEXIT_USERSPACE, 1); VMM_CTR2(vcpu, "retu %d/%d", error, vme->exitcode); /* copy the exit information */ *vme_user = *vme; return (error); } int vm_restart_instruction(struct vcpu *vcpu) { enum vcpu_state state; uint64_t rip; int error __diagused; state = vcpu_get_state(vcpu, 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; VMM_CTR1(vcpu, "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 vmmops_run() and 'nextrip' points to the next * instruction. Thus instruction restart is achieved by setting * 'nextrip' to the vcpu's %rip. */ error = vm_get_register(vcpu, VM_REG_GUEST_RIP, &rip); KASSERT(!error, ("%s: error %d getting rip", __func__, error)); VMM_CTR2(vcpu, "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 vcpu *vcpu, uint64_t info) { int type, vector; 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; } VMM_CTR2(vcpu, "%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 vcpu *vcpu, 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) { VMM_CTR2(vcpu, "triple fault: info1(%#lx), info2(%#lx)", info1, info2); vm_suspend(vcpu->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 vcpu *vcpu, uint64_t *retinfo) { uint64_t info1, info2; int valid; info1 = vcpu->exitintinfo; vcpu->exitintinfo = 0; info2 = 0; if (vcpu->exception_pending) { info2 = vcpu_exception_intinfo(vcpu); vcpu->exception_pending = 0; VMM_CTR2(vcpu, "Exception %d delivered: %#lx", vcpu->exc_vector, info2); } if ((info1 & VM_INTINFO_VALID) && (info2 & VM_INTINFO_VALID)) { valid = nested_fault(vcpu, 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) { VMM_CTR4(vcpu, "%s: info1(%#lx), info2(%#lx), " "retinfo(%#lx)", __func__, info1, info2, *retinfo); } return (valid); } int vm_get_intinfo(struct vcpu *vcpu, uint64_t *info1, uint64_t *info2) { *info1 = vcpu->exitintinfo; *info2 = vcpu_exception_intinfo(vcpu); return (0); } int vm_inject_exception(struct vcpu *vcpu, int vector, int errcode_valid, uint32_t errcode, int restart_instruction) { uint64_t regval; int error __diagused; 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); if (vcpu->exception_pending) { VMM_CTR2(vcpu, "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(vcpu, 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(vcpu, VM_REG_GUEST_INTR_SHADOW, 0); KASSERT(error == 0, ("%s: error %d clearing interrupt shadow", __func__, error)); if (restart_instruction) vm_restart_instruction(vcpu); vcpu->exception_pending = 1; vcpu->exc_vector = vector; vcpu->exc_errcode = errcode; vcpu->exc_errcode_valid = errcode_valid; VMM_CTR1(vcpu, "Exception %d pending", vector); return (0); } void vm_inject_fault(struct vcpu *vcpu, int vector, int errcode_valid, int errcode) { int error __diagused, restart_instruction; restart_instruction = 1; error = vm_inject_exception(vcpu, vector, errcode_valid, errcode, restart_instruction); KASSERT(error == 0, ("vm_inject_exception error %d", error)); } void vm_inject_pf(struct vcpu *vcpu, int error_code, uint64_t cr2) { int error __diagused; VMM_CTR2(vcpu, "Injecting page fault: error_code %#x, cr2 %#lx", error_code, cr2); error = vm_set_register(vcpu, VM_REG_GUEST_CR2, cr2); KASSERT(error == 0, ("vm_set_register(cr2) error %d", error)); vm_inject_fault(vcpu, IDT_PF, 1, error_code); } static VMM_STAT(VCPU_NMI_COUNT, "number of NMIs delivered to vcpu"); int vm_inject_nmi(struct vcpu *vcpu) { vcpu->nmi_pending = 1; vcpu_notify_event(vcpu, false); return (0); } int vm_nmi_pending(struct vcpu *vcpu) { return (vcpu->nmi_pending); } void vm_nmi_clear(struct vcpu *vcpu) { if (vcpu->nmi_pending == 0) panic("vm_nmi_clear: inconsistent nmi_pending state"); vcpu->nmi_pending = 0; vmm_stat_incr(vcpu, VCPU_NMI_COUNT, 1); } static VMM_STAT(VCPU_EXTINT_COUNT, "number of ExtINTs delivered to vcpu"); int vm_inject_extint(struct vcpu *vcpu) { vcpu->extint_pending = 1; vcpu_notify_event(vcpu, false); return (0); } int vm_extint_pending(struct vcpu *vcpu) { return (vcpu->extint_pending); } void vm_extint_clear(struct vcpu *vcpu) { if (vcpu->extint_pending == 0) panic("vm_extint_clear: inconsistent extint_pending state"); vcpu->extint_pending = 0; vmm_stat_incr(vcpu, VCPU_EXTINT_COUNT, 1); } int vm_get_capability(struct vcpu *vcpu, int type, int *retval) { if (type < 0 || type >= VM_CAP_MAX) return (EINVAL); return (vmmops_getcap(vcpu->cookie, type, retval)); } int vm_set_capability(struct vcpu *vcpu, int type, int val) { if (type < 0 || type >= VM_CAP_MAX) return (EINVAL); return (vmmops_setcap(vcpu->cookie, type, val)); } struct vm * vcpu_vm(struct vcpu *vcpu) { return (vcpu->vm); } int vcpu_vcpuid(struct vcpu *vcpu) { return (vcpu->vcpuid); } struct vcpu * vm_vcpu(struct vm *vm, int vcpuid) { return (vm->vcpu[vcpuid]); } struct vlapic * vm_lapic(struct vcpu *vcpu) { return (vcpu->vlapic); } struct vioapic * vm_ioapic(struct vm *vm) { return (vm->vioapic); } struct vhpet * vm_hpet(struct vm *vm) { return (vm->vhpet); } bool vmm_is_pptdev(int bus, int slot, int func) { int b, f, i, n, s; char *val, *cp, *cp2; bool found; /* * 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 = false; 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 = true; 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 vcpu *vcpu, enum vcpu_state newstate, bool from_idle) { int error; vcpu_lock(vcpu); error = vcpu_set_state_locked(vcpu, newstate, from_idle); vcpu_unlock(vcpu); return (error); } enum vcpu_state vcpu_get_state(struct vcpu *vcpu, int *hostcpu) { enum vcpu_state state; vcpu_lock(vcpu); state = vcpu->state; if (hostcpu != NULL) *hostcpu = vcpu->hostcpu; vcpu_unlock(vcpu); return (state); } int vm_activate_cpu(struct vcpu *vcpu) { struct vm *vm = vcpu->vm; if (CPU_ISSET(vcpu->vcpuid, &vm->active_cpus)) return (EBUSY); VMM_CTR0(vcpu, "activated"); CPU_SET_ATOMIC(vcpu->vcpuid, &vm->active_cpus); return (0); } int vm_suspend_cpu(struct vm *vm, struct vcpu *vcpu) { if (vcpu == NULL) { vm->debug_cpus = vm->active_cpus; for (int i = 0; i < vm->maxcpus; i++) { if (CPU_ISSET(i, &vm->active_cpus)) vcpu_notify_event(vm_vcpu(vm, i), false); } } else { if (!CPU_ISSET(vcpu->vcpuid, &vm->active_cpus)) return (EINVAL); CPU_SET_ATOMIC(vcpu->vcpuid, &vm->debug_cpus); vcpu_notify_event(vcpu, false); } return (0); } int vm_resume_cpu(struct vm *vm, struct vcpu *vcpu) { if (vcpu == NULL) { CPU_ZERO(&vm->debug_cpus); } else { if (!CPU_ISSET(vcpu->vcpuid, &vm->debug_cpus)) return (EINVAL); CPU_CLR_ATOMIC(vcpu->vcpuid, &vm->debug_cpus); } return (0); } int vcpu_debugged(struct vcpu *vcpu) { return (CPU_ISSET(vcpu->vcpuid, &vcpu->vm->debug_cpus)); } cpuset_t vm_active_cpus(struct vm *vm) { return (vm->active_cpus); } cpuset_t vm_debug_cpus(struct vm *vm) { return (vm->debug_cpus); } cpuset_t vm_suspended_cpus(struct vm *vm) { return (vm->suspended_cpus); } /* * Returns the subset of vCPUs in tostart that are awaiting startup. * These vCPUs are also marked as no longer awaiting startup. */ cpuset_t vm_start_cpus(struct vm *vm, const cpuset_t *tostart) { cpuset_t set; mtx_lock(&vm->rendezvous_mtx); CPU_AND(&set, &vm->startup_cpus, tostart); CPU_ANDNOT(&vm->startup_cpus, &vm->startup_cpus, &set); mtx_unlock(&vm->rendezvous_mtx); return (set); } void vm_await_start(struct vm *vm, const cpuset_t *waiting) { mtx_lock(&vm->rendezvous_mtx); CPU_OR(&vm->startup_cpus, &vm->startup_cpus, waiting); mtx_unlock(&vm->rendezvous_mtx); } void * vcpu_stats(struct vcpu *vcpu) { return (vcpu->stats); } int vm_get_x2apic_state(struct vcpu *vcpu, enum x2apic_state *state) { *state = vcpu->x2apic_state; return (0); } int vm_set_x2apic_state(struct vcpu *vcpu, enum x2apic_state state) { if (state >= X2APIC_STATE_LAST) return (EINVAL); vcpu->x2apic_state = state; vlapic_set_x2apic_state(vcpu, 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. */ static void vcpu_notify_event_locked(struct vcpu *vcpu, bool lapic_intr) { int hostcpu; 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 vcpu *vcpu, bool lapic_intr) { 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); } int vm_smp_rendezvous(struct vcpu *vcpu, cpuset_t dest, vm_rendezvous_func_t func, void *arg) { struct vm *vm = vcpu->vm; int error, i; /* * Enforce that this function is called without any locks */ WITNESS_WARN(WARN_PANIC, NULL, "vm_smp_rendezvous"); 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 'vcpu' is one * of the targets of the rendezvous. */ VMM_CTR0(vcpu, "Rendezvous already in progress"); mtx_unlock(&vm->rendezvous_mtx); error = vm_handle_rendezvous(vcpu); if (error != 0) return (error); goto restart; } KASSERT(vm->rendezvous_func == NULL, ("vm_smp_rendezvous: previous " "rendezvous is still in progress")); VMM_CTR0(vcpu, "Initiating rendezvous"); vm->rendezvous_req_cpus = dest; CPU_ZERO(&vm->rendezvous_done_cpus); vm->rendezvous_arg = arg; vm->rendezvous_func = 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->maxcpus; i++) { if (CPU_ISSET(i, &dest)) vcpu_notify_event(vm_vcpu(vm, i), false); } return (vm_handle_rendezvous(vcpu)); } 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_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 vcpu *vcpu, 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(vcpu, 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(vcpu, 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(copyinfo, num_copyinfo); return (EFAULT); } else { *fault = 0; return (0); } } void vm_copyin(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(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 vcpu *vcpu, struct vmm_stat_type *stat) { if (vcpu->vcpuid == 0) { vmm_stat_set(vcpu, VMM_MEM_RESIDENT, PAGE_SIZE * vmspace_resident_count(vcpu->vm->vmspace)); } } static void vm_get_wiredcnt(struct vcpu *vcpu, struct vmm_stat_type *stat) { if (vcpu->vcpuid == 0) { vmm_stat_set(vcpu, VMM_MEM_WIRED, PAGE_SIZE * pmap_wired_count(vmspace_pmap(vcpu->vm->vmspace))); } } VMM_STAT_FUNC(VMM_MEM_RESIDENT, "Resident memory", vm_get_rescnt); VMM_STAT_FUNC(VMM_MEM_WIRED, "Wired memory", vm_get_wiredcnt); #ifdef BHYVE_SNAPSHOT static int vm_snapshot_vcpus(struct vm *vm, struct vm_snapshot_meta *meta) { uint64_t tsc, now; int ret; struct vcpu *vcpu; uint16_t i, maxcpus; now = rdtsc(); maxcpus = vm_get_maxcpus(vm); for (i = 0; i < maxcpus; i++) { vcpu = vm->vcpu[i]; if (vcpu == NULL) continue; SNAPSHOT_VAR_OR_LEAVE(vcpu->x2apic_state, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->exitintinfo, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->exc_vector, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->exc_errcode_valid, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->exc_errcode, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->guest_xcr0, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->exitinfo, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->nextrip, meta, ret, done); /* * Save the absolute TSC value by adding now to tsc_offset. * * It will be turned turned back into an actual offset when the * TSC restore function is called */ tsc = now + vcpu->tsc_offset; SNAPSHOT_VAR_OR_LEAVE(tsc, meta, ret, done); } done: return (ret); } static int vm_snapshot_vm(struct vm *vm, struct vm_snapshot_meta *meta) { int ret; ret = vm_snapshot_vcpus(vm, meta); if (ret != 0) goto done; SNAPSHOT_VAR_OR_LEAVE(vm->startup_cpus, meta, ret, done); done: return (ret); } static int vm_snapshot_vcpu(struct vm *vm, struct vm_snapshot_meta *meta) { int error; struct vcpu *vcpu; uint16_t i, maxcpus; error = 0; maxcpus = vm_get_maxcpus(vm); for (i = 0; i < maxcpus; i++) { vcpu = vm->vcpu[i]; if (vcpu == NULL) continue; error = vmmops_vcpu_snapshot(vcpu->cookie, meta); if (error != 0) { printf("%s: failed to snapshot vmcs/vmcb data for " "vCPU: %d; error: %d\n", __func__, i, error); goto done; } } done: return (error); } /* * Save kernel-side structures to user-space for snapshotting. */ int vm_snapshot_req(struct vm *vm, struct vm_snapshot_meta *meta) { int ret = 0; switch (meta->dev_req) { case STRUCT_VMX: ret = vmmops_snapshot(vm->cookie, meta); break; case STRUCT_VMCX: ret = vm_snapshot_vcpu(vm, meta); break; case STRUCT_VM: ret = vm_snapshot_vm(vm, meta); break; case STRUCT_VIOAPIC: ret = vioapic_snapshot(vm_ioapic(vm), meta); break; case STRUCT_VLAPIC: ret = vlapic_snapshot(vm, meta); break; case STRUCT_VHPET: ret = vhpet_snapshot(vm_hpet(vm), meta); break; case STRUCT_VATPIC: ret = vatpic_snapshot(vm_atpic(vm), meta); break; case STRUCT_VATPIT: ret = vatpit_snapshot(vm_atpit(vm), meta); break; case STRUCT_VPMTMR: ret = vpmtmr_snapshot(vm_pmtmr(vm), meta); break; case STRUCT_VRTC: ret = vrtc_snapshot(vm_rtc(vm), meta); break; default: printf("%s: failed to find the requested type %#x\n", __func__, meta->dev_req); ret = (EINVAL); } return (ret); } void vm_set_tsc_offset(struct vcpu *vcpu, uint64_t offset) { vcpu->tsc_offset = offset; } int vm_restore_time(struct vm *vm) { int error; uint64_t now; struct vcpu *vcpu; uint16_t i, maxcpus; now = rdtsc(); error = vhpet_restore_time(vm_hpet(vm)); if (error) return (error); maxcpus = vm_get_maxcpus(vm); for (i = 0; i < maxcpus; i++) { vcpu = vm->vcpu[i]; if (vcpu == NULL) continue; error = vmmops_restore_tsc(vcpu->cookie, vcpu->tsc_offset - now); if (error) return (error); } return (0); } #endif