Index: head/sys/amd64/include/vmm.h =================================================================== --- head/sys/amd64/include/vmm.h (revision 364339) +++ head/sys/amd64/include/vmm.h (revision 364340) @@ -1,773 +1,775 @@ /*- * 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 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 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 */ int *sptr; /* suspend cookie */ int *iptr; /* reqidle cookie */ }; typedef int (*vmm_init_func_t)(int ipinum); typedef int (*vmm_cleanup_func_t)(void); typedef void (*vmm_resume_func_t)(void); typedef void * (*vmi_init_func_t)(struct vm *vm, struct pmap *pmap); typedef int (*vmi_run_func_t)(void *vmi, int vcpu, register_t rip, struct pmap *pmap, struct vm_eventinfo *info); typedef void (*vmi_cleanup_func_t)(void *vmi); typedef int (*vmi_get_register_t)(void *vmi, int vcpu, int num, uint64_t *retval); typedef int (*vmi_set_register_t)(void *vmi, int vcpu, int num, uint64_t val); typedef int (*vmi_get_desc_t)(void *vmi, int vcpu, int num, struct seg_desc *desc); typedef int (*vmi_set_desc_t)(void *vmi, int vcpu, int num, struct seg_desc *desc); typedef int (*vmi_get_cap_t)(void *vmi, int vcpu, int num, int *retval); typedef int (*vmi_set_cap_t)(void *vmi, int vcpu, int num, int val); typedef struct vmspace * (*vmi_vmspace_alloc)(vm_offset_t min, vm_offset_t max); typedef void (*vmi_vmspace_free)(struct vmspace *vmspace); typedef struct vlapic * (*vmi_vlapic_init)(void *vmi, int vcpu); typedef void (*vmi_vlapic_cleanup)(void *vmi, struct vlapic *vlapic); typedef int (*vmi_snapshot_t)(void *vmi, struct vm_snapshot_meta *meta); typedef int (*vmi_snapshot_vmcx_t)(void *vmi, struct vm_snapshot_meta *meta, int vcpu); typedef int (*vmi_restore_tsc_t)(void *vmi, int vcpuid, uint64_t now); struct vmm_ops { vmm_init_func_t init; /* module wide initialization */ vmm_cleanup_func_t cleanup; vmm_resume_func_t resume; vmi_init_func_t vminit; /* vm-specific initialization */ vmi_run_func_t vmrun; vmi_cleanup_func_t vmcleanup; vmi_get_register_t vmgetreg; vmi_set_register_t vmsetreg; vmi_get_desc_t vmgetdesc; vmi_set_desc_t vmsetdesc; vmi_get_cap_t vmgetcap; vmi_set_cap_t vmsetcap; vmi_vmspace_alloc vmspace_alloc; vmi_vmspace_free vmspace_free; vmi_vlapic_init vlapic_init; vmi_vlapic_cleanup vlapic_cleanup; /* checkpoint operations */ vmi_snapshot_t vmsnapshot; vmi_snapshot_vmcx_t vmcx_snapshot; vmi_restore_tsc_t vm_restore_tsc; }; extern struct vmm_ops vmm_ops_intel; extern struct vmm_ops vmm_ops_amd; int vm_create(const char *name, struct vm **retvm); void vm_destroy(struct vm *vm); int vm_reinit(struct vm *vm); const char *vm_name(struct vm *vm); 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. */ 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_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 vm *, int vcpuid, vm_paddr_t gpa, size_t len, int prot, void **cookie); void vm_gpa_release(void *cookie); bool vm_mem_allocated(struct vm *vm, int vcpuid, vm_paddr_t gpa); int vm_get_register(struct vm *vm, int vcpu, int reg, uint64_t *retval); int vm_set_register(struct vm *vm, int vcpu, int reg, uint64_t val); int vm_get_seg_desc(struct vm *vm, int vcpu, int reg, struct seg_desc *ret_desc); int vm_set_seg_desc(struct vm *vm, int vcpu, int reg, struct seg_desc *desc); int vm_run(struct vm *vm, struct vm_run *vmrun); int vm_suspend(struct vm *vm, enum vm_suspend_how how); int vm_inject_nmi(struct vm *vm, int vcpu); int vm_nmi_pending(struct vm *vm, int vcpuid); void vm_nmi_clear(struct vm *vm, int vcpuid); int vm_inject_extint(struct vm *vm, int vcpu); int vm_extint_pending(struct vm *vm, int vcpuid); void vm_extint_clear(struct vm *vm, int vcpuid); struct vlapic *vm_lapic(struct vm *vm, int cpu); struct vioapic *vm_ioapic(struct vm *vm); struct vhpet *vm_hpet(struct vm *vm); int vm_get_capability(struct vm *vm, int vcpu, int type, int *val); int vm_set_capability(struct vm *vm, int vcpu, int type, int val); int vm_get_x2apic_state(struct vm *vm, int vcpu, enum x2apic_state *state); int vm_set_x2apic_state(struct vm *vm, int vcpu, enum x2apic_state state); int vm_apicid2vcpuid(struct vm *vm, int apicid); int vm_activate_cpu(struct vm *vm, int vcpu); int vm_suspend_cpu(struct vm *vm, int vcpu); int vm_resume_cpu(struct vm *vm, int vcpu); struct vm_exit *vm_exitinfo(struct vm *vm, int vcpuid); void vm_exit_suspended(struct vm *vm, int vcpuid, uint64_t rip); void vm_exit_debug(struct vm *vm, int vcpuid, uint64_t rip); void vm_exit_rendezvous(struct vm *vm, int vcpuid, uint64_t rip); void vm_exit_astpending(struct vm *vm, int vcpuid, uint64_t rip); void vm_exit_reqidle(struct vm *vm, int vcpuid, uint64_t rip); 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. * * If the rendezvous is being initiated from a vcpu context then the * 'vcpuid' must refer to that vcpu, otherwise it should be set to -1. * * The caller cannot hold any locks when initiating the rendezvous. * * The implementation of this API may cause vcpus other than those specified * by 'dest' to be stalled. The caller should not rely on any vcpus making * forward progress when the rendezvous is in progress. */ typedef void (*vm_rendezvous_func_t)(struct vm *vm, int vcpuid, void *arg); int vm_smp_rendezvous(struct vm *vm, int vcpuid, cpuset_t dest, vm_rendezvous_func_t func, void *arg); cpuset_t vm_active_cpus(struct vm *vm); cpuset_t vm_debug_cpus(struct vm *vm); cpuset_t vm_suspended_cpus(struct vm *vm); #endif /* _SYS__CPUSET_H_ */ static __inline int vcpu_rendezvous_pending(struct vm_eventinfo *info) { return (*((uintptr_t *)(info->rptr)) != 0); } 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 vm *vm, int vcpuid); /* * 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 vm *vm, int vcpu, enum vcpu_state state, bool from_idle); enum vcpu_state vcpu_get_state(struct vm *vm, int vcpu, int *hostcpu); static int __inline vcpu_is_running(struct vm *vm, int vcpu, int *hostcpu) { return (vcpu_get_state(vm, vcpu, hostcpu) == VCPU_RUNNING); } #ifdef _SYS_PROC_H_ static int __inline vcpu_should_yield(struct vm *vm, int vcpu) { if (curthread->td_flags & (TDF_ASTPENDING | TDF_NEEDRESCHED)) return (1); else if (curthread->td_owepreempt) return (1); else return (0); } #endif void *vcpu_stats(struct vm *vm, int vcpu); void vcpu_notify_event(struct vm *vm, int vcpuid, bool lapic_intr); struct vmspace *vm_get_vmspace(struct vm *vm); struct vatpic *vm_atpic(struct vm *vm); struct vatpit *vm_atpit(struct vm *vm); struct vpmtmr *vm_pmtmr(struct vm *vm); struct vrtc *vm_rtc(struct vm *vm); /* * Inject exception 'vector' into the guest vcpu. This function returns 0 on * success and non-zero on failure. * * Wrapper functions like 'vm_inject_gp()' should be preferred to calling * this function directly because they enforce the trap-like or fault-like * behavior of an exception. * * This function should only be called in the context of the thread that is * executing this vcpu. */ int vm_inject_exception(struct vm *vm, int vcpuid, int vector, int err_valid, uint32_t errcode, int restart_instruction); /* * This function is called after a VM-exit that occurred during exception or * interrupt delivery through the IDT. The format of 'intinfo' is described * in Figure 15-1, "EXITINTINFO for All Intercepts", APM, Vol 2. * * If a VM-exit handler completes the event delivery successfully then it * should call vm_exit_intinfo() to extinguish the pending event. For e.g., * if the task switch emulation is triggered via a task gate then it should * call this function with 'intinfo=0' to indicate that the external event * is not pending anymore. * * Return value is 0 on success and non-zero on failure. */ int vm_exit_intinfo(struct vm *vm, int vcpuid, uint64_t intinfo); /* * This function is called before every VM-entry to retrieve a pending * event that should be injected into the guest. This function combines * nested events into a double or triple fault. * * Returns 0 if there are no events that need to be injected into the guest * and non-zero otherwise. */ int vm_entry_intinfo(struct vm *vm, int vcpuid, uint64_t *info); int vm_get_intinfo(struct vm *vm, int vcpuid, uint64_t *info1, uint64_t *info2); /* * 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. * * Return value is 0 on success and non-zero on failure. */ int vm_set_tsc_offset(struct vm *vm, int vcpu_id, 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 vm *vm, int vcpuid, struct vm_guest_paging *paging, uint64_t gla, size_t len, int prot, struct vm_copyinfo *copyinfo, int num_copyinfo, int *is_fault); void vm_copy_teardown(struct vm *vm, int vcpuid, struct vm_copyinfo *copyinfo, int num_copyinfo); void vm_copyin(struct vm *vm, int vcpuid, struct vm_copyinfo *copyinfo, void *kaddr, size_t len); void vm_copyout(struct vm *vm, int vcpuid, const void *kaddr, struct vm_copyinfo *copyinfo, size_t len); int vcpu_trace_exceptions(struct vm *vm, int vcpuid); #endif /* KERNEL */ #define VM_MAXCPU 16 /* maximum virtual cpus */ /* * Identifiers for optional vmm capabilities */ enum vm_cap_type { VM_CAP_HALT_EXIT, VM_CAP_MTRAP_EXIT, VM_CAP_PAUSE_EXIT, VM_CAP_UNRESTRICTED_GUEST, VM_CAP_ENABLE_INVPCID, VM_CAP_BPT_EXIT, + VM_CAP_RDPID, + VM_CAP_RDTSCP, VM_CAP_MAX }; enum vm_intr_trigger { EDGE_TRIGGER, LEVEL_TRIGGER }; /* * The 'access' field has the format specified in Table 21-2 of the Intel * Architecture Manual vol 3b. * * XXX The contents of the 'access' field are architecturally defined except * bit 16 - Segment Unusable. */ struct seg_desc { uint64_t base; uint32_t limit; uint32_t access; }; #define SEG_DESC_TYPE(access) ((access) & 0x001f) #define SEG_DESC_DPL(access) (((access) >> 5) & 0x3) #define SEG_DESC_PRESENT(access) (((access) & 0x0080) ? 1 : 0) #define SEG_DESC_DEF32(access) (((access) & 0x4000) ? 1 : 0) #define SEG_DESC_GRANULARITY(access) (((access) & 0x8000) ? 1 : 0) #define SEG_DESC_UNUSABLE(access) (((access) & 0x10000) ? 1 : 0) enum vm_cpu_mode { CPU_MODE_REAL, CPU_MODE_PROTECTED, CPU_MODE_COMPATIBILITY, /* IA-32E mode (CS.L = 0) */ CPU_MODE_64BIT, /* IA-32E mode (CS.L = 1) */ }; enum vm_paging_mode { PAGING_MODE_FLAT, PAGING_MODE_32, PAGING_MODE_PAE, PAGING_MODE_64, }; struct vm_guest_paging { uint64_t cr3; int cpl; enum vm_cpu_mode cpu_mode; enum vm_paging_mode paging_mode; }; /* * The data structures 'vie' and 'vie_op' are meant to be opaque to the * consumers of instruction decoding. The only reason why their contents * need to be exposed is because they are part of the 'vm_exit' structure. */ struct vie_op { uint8_t op_byte; /* actual opcode byte */ uint8_t op_type; /* type of operation (e.g. MOV) */ uint16_t op_flags; }; _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_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 vm_task_switch task_switch; } u; }; /* APIs to inject faults into the guest */ void vm_inject_fault(void *vm, int vcpuid, int vector, int errcode_valid, int errcode); static __inline void vm_inject_ud(void *vm, int vcpuid) { vm_inject_fault(vm, vcpuid, IDT_UD, 0, 0); } static __inline void vm_inject_gp(void *vm, int vcpuid) { vm_inject_fault(vm, vcpuid, IDT_GP, 1, 0); } static __inline void vm_inject_ac(void *vm, int vcpuid, int errcode) { vm_inject_fault(vm, vcpuid, IDT_AC, 1, errcode); } static __inline void vm_inject_ss(void *vm, int vcpuid, int errcode) { vm_inject_fault(vm, vcpuid, IDT_SS, 1, errcode); } void vm_inject_pf(void *vm, int vcpuid, int error_code, uint64_t cr2); int vm_restart_instruction(void *vm, int vcpuid); #endif /* _VMM_H_ */ Index: head/sys/amd64/vmm/intel/vmx.c =================================================================== --- head/sys/amd64/vmm/intel/vmx.c (revision 364339) +++ head/sys/amd64/vmm/intel/vmx.c (revision 364340) @@ -1,4058 +1,4165 @@ /*- * 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 "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"); 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_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 *arg, int vcpu, int reg, struct seg_desc *desc); static int vmx_getreg(void *arg, int vcpu, int reg, uint64_t *retval); static int vmxctx_setreg(struct vmxctx *vmxctx, int reg, uint64_t val); static void vmx_inject_pir(struct vlapic *vlapic); #ifdef BHYVE_SNAPSHOT static int vmx_restore_tsc(void *arg, int vcpu, 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 void vpid_alloc(uint16_t *vpid, int num) { int i, x; if (num <= 0 || num > VM_MAXCPU) panic("invalid number of vpids requested: %d", num); /* * If the "enable vpid" execution control is not enabled then the * VPID is required to be 0 for all vcpus. */ if ((procbased_ctls2 & PROCBASED2_ENABLE_VPID) == 0) { for (i = 0; i < num; i++) vpid[i] = 0; return; } /* * Allocate a unique VPID for each vcpu from the unit number allocator. */ for (i = 0; i < num; i++) { x = alloc_unr(vpid_unr); if (x == -1) break; else vpid[i] = x; } if (i < num) { atomic_add_int(&vpid_alloc_failed, 1); /* * If the unit number allocator does not have enough unique * VPIDs then we need to allocate from the [1,VM_MAXCPU] range. * * These VPIDs are not be unique across VMs but this does not * affect correctness because the combined mappings are also * tagged with the EP4TA which is unique for each VM. * * It is still sub-optimal because the invvpid will invalidate * combined mappings for a particular VPID across all EP4TAs. */ while (i-- > 0) vpid_free(vpid[i]); for (i = 0; i < num; i++) vpid[i] = i + 1; } } static void vpid_init(void) { /* * VPID 0 is required when the "enable VPID" execution control is * disabled. * * VPIDs [1,VM_MAXCPU] are used as the "overflow namespace" when the * unit number allocator does not have sufficient unique VPIDs to * satisfy the allocation. * * The remaining VPIDs are managed by the unit number allocator. */ vpid_unr = new_unrhdr(VM_MAXCPU + 1, 0xffff, NULL); } static void vmx_disable(void *arg __unused) { struct invvpid_desc invvpid_desc = { 0 }; struct invept_desc invept_desc = { 0 }; if (vmxon_enabled[curcpu]) { /* * See sections 25.3.3.3 and 25.3.3.4 in Intel Vol 3b. * * VMXON or VMXOFF are not required to invalidate any TLB * caching structures. This prevents potential retention of * cached information in the TLB between distinct VMX episodes. */ invvpid(INVVPID_TYPE_ALL_CONTEXTS, invvpid_desc); invept(INVEPT_TYPE_ALL_CONTEXTS, invept_desc); vmxoff(); } load_cr4(rcr4() & ~CR4_VMXE); } static int vmx_cleanup(void) { if (pirvec >= 0) lapic_ipi_free(pirvec); if (vpid_unr != NULL) { delete_unrhdr(vpid_unr); vpid_unr = NULL; } 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_restore(void) { if (vmxon_enabled[curcpu]) vmxon(vmxon_region[curcpu]); } static int vmx_init(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_init: processor does not support VMX operation\n"); return (ENXIO); } /* * Verify that MSR_IA32_FEATURE_CONTROL lock and VMXON enable bits * are set (bits 0 and 2 respectively). */ feature_control = rdmsr(MSR_IA32_FEATURE_CONTROL); if ((feature_control & IA32_FEATURE_CONTROL_LOCK) == 1 && (feature_control & IA32_FEATURE_CONTROL_VMX_EN) == 0) { printf("vmx_init: VMX operation disabled by BIOS\n"); return (ENXIO); } /* * Verify capabilities MSR_VMX_BASIC: * - bit 54 indicates support for INS/OUTS decoding */ basic = rdmsr(MSR_VMX_BASIC); if ((basic & (1UL << 54)) == 0) { printf("vmx_init: processor does not support desired basic " "capabilities\n"); return (EINVAL); } /* Check support for primary processor-based VM-execution controls */ error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_CTLS_ONE_SETTING, PROCBASED_CTLS_ZERO_SETTING, &procbased_ctls); if (error) { printf("vmx_init: processor does not support desired primary " "processor-based controls\n"); return (error); } /* Clear the processor-based ctl bits that are set on demand */ procbased_ctls &= ~PROCBASED_CTLS_WINDOW_SETTING; /* Check support for secondary processor-based VM-execution controls */ error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED_CTLS2_ONE_SETTING, PROCBASED_CTLS2_ZERO_SETTING, &procbased_ctls2); if (error) { printf("vmx_init: processor does not support desired secondary " "processor-based controls\n"); return (error); } /* Check support for VPID */ error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2, PROCBASED2_ENABLE_VPID, 0, &tmp); if (error == 0) procbased_ctls2 |= PROCBASED2_ENABLE_VPID; /* Check support for pin-based VM-execution controls */ error = vmx_set_ctlreg(MSR_VMX_PINBASED_CTLS, MSR_VMX_TRUE_PINBASED_CTLS, PINBASED_CTLS_ONE_SETTING, PINBASED_CTLS_ZERO_SETTING, &pinbased_ctls); if (error) { printf("vmx_init: processor does not support desired " "pin-based controls\n"); return (error); } /* Check support for VM-exit controls */ error = vmx_set_ctlreg(MSR_VMX_EXIT_CTLS, MSR_VMX_TRUE_EXIT_CTLS, VM_EXIT_CTLS_ONE_SETTING, VM_EXIT_CTLS_ZERO_SETTING, &exit_ctls); if (error) { printf("vmx_init: processor does not support desired " "exit controls\n"); return (error); } /* Check support for VM-entry controls */ error = vmx_set_ctlreg(MSR_VMX_ENTRY_CTLS, MSR_VMX_TRUE_ENTRY_CTLS, VM_ENTRY_CTLS_ONE_SETTING, VM_ENTRY_CTLS_ZERO_SETTING, &entry_ctls); if (error) { printf("vmx_init: processor does not support desired " "entry controls\n"); return (error); } /* * Check support for optional features by testing them * as individual bits */ cap_halt_exit = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_HLT_EXITING, 0, &tmp) == 0); cap_monitor_trap = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_PROCBASED_CTLS, PROCBASED_MTF, 0, &tmp) == 0); cap_pause_exit = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS, MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_PAUSE_EXITING, 0, &tmp) == 0); + /* + * 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; + 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_init: unable to allocate " "posted interrupt vector\n"); } } else { posted_interrupts = 1; TUNABLE_INT_FETCH("hw.vmm.vmx.use_apic_pir", &posted_interrupts); } } } if (posted_interrupts) pinbased_ctls |= PINBASED_POSTED_INTERRUPT; /* Initialize EPT */ error = ept_init(ipinum); if (error) { printf("vmx_init: ept initialization failed (%d)\n", error); return (error); } 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_vminit(struct vm *vm, pmap_t pmap) { uint16_t vpid[VM_MAXCPU]; int i, error; struct vmx *vmx; struct vmcs *vmcs; uint32_t exc_bitmap; uint16_t maxcpus; vmx = malloc(sizeof(struct vmx), M_VMX, M_WAITOK | M_ZERO); if ((uintptr_t)vmx & PAGE_MASK) { panic("malloc of struct vmx not aligned on %d byte boundary", PAGE_SIZE); } vmx->vm = vm; vmx->eptp = eptp(vtophys((vm_offset_t)pmap->pm_pml4)); /* * Clean up EPTP-tagged guest physical and combined mappings * * VMX transitions are not required to invalidate any guest physical * mappings. So, it may be possible for stale guest physical mappings * to be present in the processor TLBs. * * Combined mappings for this EP4TA are also invalidated for all VPIDs. */ ept_invalidate_mappings(vmx->eptp); msr_bitmap_initialize(vmx->msr_bitmap); /* * It is safe to allow direct access to MSR_GSBASE and MSR_FSBASE. * The guest FSBASE and GSBASE are saved and restored during * vm-exit and vm-entry respectively. The host FSBASE and GSBASE are * always restored from the vmcs host state area on vm-exit. * * The SYSENTER_CS/ESP/EIP MSRs are identical to FS/GSBASE in * how they are saved/restored so can be directly accessed by the * guest. * * MSR_EFER is saved and restored in the guest VMCS area on a * VM exit and entry respectively. It is also restored from the * host VMCS area on a VM exit. * * The TSC MSR is exposed read-only. Writes are disallowed as * that will impact the host TSC. 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)) + guest_msr_ro(vmx, MSR_TSC) || + ((cap_rdpid || cap_rdtscp) && guest_msr_ro(vmx, MSR_TSC_AUX))) panic("vmx_vminit: error setting guest msr access"); vpid_alloc(vpid, VM_MAXCPU); if (virtual_interrupt_delivery) { error = vm_map_mmio(vm, DEFAULT_APIC_BASE, PAGE_SIZE, APIC_ACCESS_ADDRESS); /* XXX this should really return an error to the caller */ KASSERT(error == 0, ("vm_map_mmio(apicbase) error %d", error)); } maxcpus = vm_get_maxcpus(vm); for (i = 0; i < maxcpus; i++) { vmcs = &vmx->vmcs[i]; vmcs->identifier = vmx_revision(); error = vmclear(vmcs); if (error != 0) { panic("vmx_vminit: vmclear error %d on vcpu %d\n", error, i); } vmx_msr_guest_init(vmx, i); error = vmcs_init(vmcs); KASSERT(error == 0, ("vmcs_init error %d", error)); VMPTRLD(vmcs); error = 0; error += vmwrite(VMCS_HOST_RSP, (u_long)&vmx->ctx[i]); error += vmwrite(VMCS_EPTP, vmx->eptp); error += vmwrite(VMCS_PIN_BASED_CTLS, pinbased_ctls); error += vmwrite(VMCS_PRI_PROC_BASED_CTLS, procbased_ctls); error += vmwrite(VMCS_SEC_PROC_BASED_CTLS, procbased_ctls2); error += vmwrite(VMCS_EXIT_CTLS, exit_ctls); error += vmwrite(VMCS_ENTRY_CTLS, entry_ctls); error += vmwrite(VMCS_MSR_BITMAP, vtophys(vmx->msr_bitmap)); error += vmwrite(VMCS_VPID, vpid[i]); 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(vm, i)) exc_bitmap = 0xffffffff; else exc_bitmap = 1 << IDT_MC; error += vmwrite(VMCS_EXCEPTION_BITMAP, exc_bitmap); vmx->ctx[i].guest_dr6 = DBREG_DR6_RESERVED1; error += vmwrite(VMCS_GUEST_DR7, DBREG_DR7_RESERVED1); if (tpr_shadowing) { error += vmwrite(VMCS_VIRTUAL_APIC, vtophys(&vmx->apic_page[i])); } 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(&vmx->pir_desc[i])); } VMCLEAR(vmcs); KASSERT(error == 0, ("vmx_vminit: error customizing the vmcs")); vmx->cap[i].set = 0; + vmx->cap[i].set |= cap_rdpid != 0 ? 1 << VM_CAP_RDPID : 0; + vmx->cap[i].set |= cap_rdtscp != 0 ? 1 << VM_CAP_RDTSCP : 0; vmx->cap[i].proc_ctls = procbased_ctls; vmx->cap[i].proc_ctls2 = procbased_ctls2; vmx->cap[i].exc_bitmap = exc_bitmap; vmx->state[i].nextrip = ~0; vmx->state[i].lastcpu = NOCPU; vmx->state[i].vpid = vpid[i]; /* * Set up the CR0/4 shadows, and init the read shadow * to the power-on register value from the Intel Sys Arch. * CR0 - 0x60000010 * CR4 - 0 */ error = vmx_setup_cr0_shadow(vmcs, 0x60000010); if (error != 0) panic("vmx_setup_cr0_shadow %d", error); error = vmx_setup_cr4_shadow(vmcs, 0); if (error != 0) panic("vmx_setup_cr4_shadow %d", error); vmx->ctx[i].pmap = pmap; } return (vmx); } static int vmx_handle_cpuid(struct vm *vm, int vcpu, struct vmxctx *vmxctx) { int handled, func; func = vmxctx->guest_rax; handled = x86_emulate_cpuid(vm, vcpu, (uint32_t*)(&vmxctx->guest_rax), (uint32_t*)(&vmxctx->guest_rbx), (uint32_t*)(&vmxctx->guest_rcx), (uint32_t*)(&vmxctx->guest_rdx)); return (handled); } static __inline void vmx_run_trace(struct vmx *vmx, int vcpu) { #ifdef KTR VCPU_CTR1(vmx->vm, vcpu, "Resume execution at %#lx", vmcs_guest_rip()); #endif } static __inline void vmx_exit_trace(struct vmx *vmx, int vcpu, uint64_t rip, uint32_t exit_reason, int handled) { #ifdef KTR VCPU_CTR3(vmx->vm, vcpu, "%s %s vmexit at 0x%0lx", handled ? "handled" : "unhandled", exit_reason_to_str(exit_reason), rip); #endif } static __inline void vmx_astpending_trace(struct vmx *vmx, int vcpu, uint64_t rip) { #ifdef KTR VCPU_CTR1(vmx->vm, vcpu, "astpending vmexit at 0x%0lx", rip); #endif } static VMM_STAT_INTEL(VCPU_INVVPID_SAVED, "Number of vpid invalidations saved"); static VMM_STAT_INTEL(VCPU_INVVPID_DONE, "Number of vpid invalidations done"); /* * Invalidate guest mappings identified by its vpid from the TLB. */ static __inline void vmx_invvpid(struct vmx *vmx, int vcpu, pmap_t pmap, int running) { struct vmxstate *vmxstate; struct invvpid_desc invvpid_desc; vmxstate = &vmx->state[vcpu]; if (vmxstate->vpid == 0) return; if (!running) { /* * Set the 'lastcpu' to an invalid host cpu. * * This will invalidate TLB entries tagged with the vcpu's * vpid the next time it runs via vmx_set_pcpu_defaults(). */ vmxstate->lastcpu = NOCPU; return; } KASSERT(curthread->td_critnest > 0, ("%s: vcpu %d running outside " "critical section", __func__, vcpu)); /* * Invalidate all mappings tagged with 'vpid' * * We do this because this vcpu was executing on a different host * cpu when it last ran. We do not track whether it invalidated * mappings associated with its 'vpid' during that run. So we must * assume that the mappings associated with 'vpid' on 'curcpu' are * stale and invalidate them. * * Note that we incur this penalty only when the scheduler chooses to * move the thread associated with this vcpu between host cpus. * * Note also that this will invalidate mappings tagged with 'vpid' * for "all" EP4TAs. */ if (pmap->pm_eptgen == vmx->eptgen[curcpu]) { invvpid_desc._res1 = 0; invvpid_desc._res2 = 0; invvpid_desc.vpid = vmxstate->vpid; invvpid_desc.linear_addr = 0; invvpid(INVVPID_TYPE_SINGLE_CONTEXT, invvpid_desc); vmm_stat_incr(vmx->vm, vcpu, VCPU_INVVPID_DONE, 1); } else { /* * The invvpid can be skipped if an invept is going to * be performed before entering the guest. The invept * will invalidate combined mappings tagged with * 'vmx->eptp' for all vpids. */ vmm_stat_incr(vmx->vm, vcpu, VCPU_INVVPID_SAVED, 1); } } static void vmx_set_pcpu_defaults(struct vmx *vmx, int vcpu, pmap_t pmap) { struct vmxstate *vmxstate; vmxstate = &vmx->state[vcpu]; if (vmxstate->lastcpu == curcpu) return; vmxstate->lastcpu = curcpu; vmm_stat_incr(vmx->vm, vcpu, VCPU_MIGRATIONS, 1); vmcs_write(VMCS_HOST_TR_BASE, vmm_get_host_trbase()); vmcs_write(VMCS_HOST_GDTR_BASE, vmm_get_host_gdtrbase()); vmcs_write(VMCS_HOST_GS_BASE, vmm_get_host_gsbase()); vmx_invvpid(vmx, vcpu, pmap, 1); } /* * We depend on 'procbased_ctls' to have the Interrupt Window Exiting bit set. */ CTASSERT((PROCBASED_CTLS_ONE_SETTING & PROCBASED_INT_WINDOW_EXITING) != 0); static void __inline vmx_set_int_window_exiting(struct vmx *vmx, int vcpu) { if ((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) == 0) { vmx->cap[vcpu].proc_ctls |= PROCBASED_INT_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Enabling interrupt window exiting"); } } static void __inline vmx_clear_int_window_exiting(struct vmx *vmx, int vcpu) { KASSERT((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) != 0, ("intr_window_exiting not set: %#x", vmx->cap[vcpu].proc_ctls)); vmx->cap[vcpu].proc_ctls &= ~PROCBASED_INT_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Disabling interrupt window exiting"); } static void __inline vmx_set_nmi_window_exiting(struct vmx *vmx, int vcpu) { if ((vmx->cap[vcpu].proc_ctls & PROCBASED_NMI_WINDOW_EXITING) == 0) { vmx->cap[vcpu].proc_ctls |= PROCBASED_NMI_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Enabling NMI window exiting"); } } static void __inline vmx_clear_nmi_window_exiting(struct vmx *vmx, int vcpu) { KASSERT((vmx->cap[vcpu].proc_ctls & PROCBASED_NMI_WINDOW_EXITING) != 0, ("nmi_window_exiting not set %#x", vmx->cap[vcpu].proc_ctls)); vmx->cap[vcpu].proc_ctls &= ~PROCBASED_NMI_WINDOW_EXITING; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Disabling NMI window exiting"); } int vmx_set_tsc_offset(struct vmx *vmx, int vcpu, uint64_t offset) { int error; if ((vmx->cap[vcpu].proc_ctls & PROCBASED_TSC_OFFSET) == 0) { vmx->cap[vcpu].proc_ctls |= PROCBASED_TSC_OFFSET; vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls); VCPU_CTR0(vmx->vm, vcpu, "Enabling TSC offsetting"); } error = vmwrite(VMCS_TSC_OFFSET, offset); #ifdef BHYVE_SNAPSHOT if (error == 0) error = vm_set_tsc_offset(vmx->vm, 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 *vmx, int vcpu) { uint32_t gi, info; gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); KASSERT((gi & NMI_BLOCKING) == 0, ("vmx_inject_nmi: invalid guest " "interruptibility-state %#x", gi)); info = vmcs_read(VMCS_ENTRY_INTR_INFO); KASSERT((info & VMCS_INTR_VALID) == 0, ("vmx_inject_nmi: invalid " "VM-entry interruption information %#x", info)); /* * Inject the virtual NMI. The vector must be the NMI IDT entry * or the VMCS entry check will fail. */ info = IDT_NMI | VMCS_INTR_T_NMI | VMCS_INTR_VALID; vmcs_write(VMCS_ENTRY_INTR_INFO, info); VCPU_CTR0(vmx->vm, vcpu, "Injecting vNMI"); /* Clear the request */ vm_nmi_clear(vmx->vm, vcpu); } static void vmx_inject_interrupts(struct vmx *vmx, int vcpu, struct vlapic *vlapic, uint64_t guestrip) { int vector, need_nmi_exiting, extint_pending; uint64_t rflags, entryinfo; uint32_t gi, info; if (vmx->state[vcpu].nextrip != guestrip) { gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); if (gi & HWINTR_BLOCKING) { VCPU_CTR2(vmx->vm, vcpu, "Guest interrupt blocking " "cleared due to rip change: %#lx/%#lx", vmx->state[vcpu].nextrip, guestrip); gi &= ~HWINTR_BLOCKING; vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi); } } if (vm_entry_intinfo(vmx->vm, vcpu, &entryinfo)) { KASSERT((entryinfo & VMCS_INTR_VALID) != 0, ("%s: entry " "intinfo is not valid: %#lx", __func__, entryinfo)); info = vmcs_read(VMCS_ENTRY_INTR_INFO); KASSERT((info & VMCS_INTR_VALID) == 0, ("%s: cannot inject " "pending exception: %#lx/%#x", __func__, entryinfo, info)); info = entryinfo; vector = info & 0xff; if (vector == IDT_BP || vector == IDT_OF) { /* * VT-x requires #BP and #OF to be injected as software * exceptions. */ info &= ~VMCS_INTR_T_MASK; info |= VMCS_INTR_T_SWEXCEPTION; } if (info & VMCS_INTR_DEL_ERRCODE) vmcs_write(VMCS_ENTRY_EXCEPTION_ERROR, entryinfo >> 32); vmcs_write(VMCS_ENTRY_INTR_INFO, info); } if (vm_nmi_pending(vmx->vm, vcpu)) { /* * If there are no conditions blocking NMI injection then * inject it directly here otherwise enable "NMI window * exiting" to inject it as soon as we can. * * We also check for STI_BLOCKING because some implementations * don't allow NMI injection in this case. If we are running * on a processor that doesn't have this restriction it will * immediately exit and the NMI will be injected in the * "NMI window exiting" handler. */ need_nmi_exiting = 1; gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); if ((gi & (HWINTR_BLOCKING | NMI_BLOCKING)) == 0) { info = vmcs_read(VMCS_ENTRY_INTR_INFO); if ((info & VMCS_INTR_VALID) == 0) { vmx_inject_nmi(vmx, vcpu); need_nmi_exiting = 0; } else { VCPU_CTR1(vmx->vm, vcpu, "Cannot inject NMI " "due to VM-entry intr info %#x", info); } } else { VCPU_CTR1(vmx->vm, vcpu, "Cannot inject NMI due to " "Guest Interruptibility-state %#x", gi); } if (need_nmi_exiting) vmx_set_nmi_window_exiting(vmx, vcpu); } extint_pending = vm_extint_pending(vmx->vm, vcpu); if (!extint_pending && virtual_interrupt_delivery) { vmx_inject_pir(vlapic); return; } /* * If interrupt-window exiting is already in effect then don't bother * checking for pending interrupts. This is just an optimization and * not needed for correctness. */ if ((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) != 0) { VCPU_CTR0(vmx->vm, vcpu, "Skip interrupt injection due to " "pending int_window_exiting"); return; } if (!extint_pending) { /* Ask the local apic for a vector to inject */ if (!vlapic_pending_intr(vlapic, &vector)) return; /* * From the Intel SDM, Volume 3, Section "Maskable * Hardware Interrupts": * - maskable interrupt vectors [16,255] can be delivered * through the local APIC. */ KASSERT(vector >= 16 && vector <= 255, ("invalid vector %d from local APIC", vector)); } else { /* Ask the legacy pic for a vector to inject */ vatpic_pending_intr(vmx->vm, &vector); /* * From the Intel SDM, Volume 3, Section "Maskable * Hardware Interrupts": * - maskable interrupt vectors [0,255] can be delivered * through the INTR pin. */ KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR", vector)); } /* Check RFLAGS.IF and the interruptibility state of the guest */ rflags = vmcs_read(VMCS_GUEST_RFLAGS); if ((rflags & PSL_I) == 0) { VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to " "rflags %#lx", vector, rflags); goto cantinject; } gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); if (gi & HWINTR_BLOCKING) { VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to " "Guest Interruptibility-state %#x", vector, gi); goto cantinject; } info = vmcs_read(VMCS_ENTRY_INTR_INFO); if (info & VMCS_INTR_VALID) { /* * This is expected and could happen for multiple reasons: * - A vectoring VM-entry was aborted due to astpending * - A VM-exit happened during event injection. * - An exception was injected above. * - An NMI was injected above or after "NMI window exiting" */ VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to " "VM-entry intr info %#x", vector, info); goto cantinject; } /* Inject the interrupt */ info = VMCS_INTR_T_HWINTR | VMCS_INTR_VALID; info |= vector; vmcs_write(VMCS_ENTRY_INTR_INFO, info); if (!extint_pending) { /* Update the Local APIC ISR */ vlapic_intr_accepted(vlapic, vector); } else { vm_extint_clear(vmx->vm, vcpu); vatpic_intr_accepted(vmx->vm, vector); /* * After we accepted the current ExtINT the PIC may * have posted another one. If that is the case, set * the Interrupt Window Exiting execution control so * we can inject that one too. * * Also, interrupt window exiting allows us to inject any * pending APIC vector that was preempted by the ExtINT * as soon as possible. This applies both for the software * emulated vlapic and the hardware assisted virtual APIC. */ vmx_set_int_window_exiting(vmx, vcpu); } VCPU_CTR1(vmx->vm, vcpu, "Injecting hwintr at vector %d", vector); return; cantinject: /* * Set the Interrupt Window Exiting execution control so we can inject * the interrupt as soon as blocking condition goes away. */ vmx_set_int_window_exiting(vmx, vcpu); } /* * If the Virtual NMIs execution control is '1' then the logical processor * tracks virtual-NMI blocking in the Guest Interruptibility-state field of * the VMCS. An IRET instruction in VMX non-root operation will remove any * virtual-NMI blocking. * * This unblocking occurs even if the IRET causes a fault. In this case the * hypervisor needs to restore virtual-NMI blocking before resuming the guest. */ static void vmx_restore_nmi_blocking(struct vmx *vmx, int vcpuid) { uint32_t gi; VCPU_CTR0(vmx->vm, vcpuid, "Restore Virtual-NMI blocking"); gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); gi |= VMCS_INTERRUPTIBILITY_NMI_BLOCKING; vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi); } static void vmx_clear_nmi_blocking(struct vmx *vmx, int vcpuid) { uint32_t gi; VCPU_CTR0(vmx->vm, vcpuid, "Clear Virtual-NMI blocking"); gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); gi &= ~VMCS_INTERRUPTIBILITY_NMI_BLOCKING; vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi); } static void vmx_assert_nmi_blocking(struct vmx *vmx, int vcpuid) { uint32_t gi; gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY); KASSERT(gi & VMCS_INTERRUPTIBILITY_NMI_BLOCKING, ("NMI blocking is not in effect %#x", gi)); } static int vmx_emulate_xsetbv(struct vmx *vmx, int vcpu, struct vm_exit *vmexit) { struct vmxctx *vmxctx; uint64_t xcrval; const struct xsave_limits *limits; vmxctx = &vmx->ctx[vcpu]; limits = vmm_get_xsave_limits(); /* * Note that the processor raises a GP# fault on its own if * xsetbv is executed for CPL != 0, so we do not have to * emulate that fault here. */ /* Only xcr0 is supported. */ if (vmxctx->guest_rcx != 0) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* We only handle xcr0 if both the host and guest have XSAVE enabled. */ if (!limits->xsave_enabled || !(vmcs_read(VMCS_GUEST_CR4) & CR4_XSAVE)) { vm_inject_ud(vmx->vm, vcpu); return (HANDLED); } xcrval = vmxctx->guest_rdx << 32 | (vmxctx->guest_rax & 0xffffffff); if ((xcrval & ~limits->xcr0_allowed) != 0) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } if (!(xcrval & XFEATURE_ENABLED_X87)) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* AVX (YMM_Hi128) requires SSE. */ if (xcrval & XFEATURE_ENABLED_AVX && (xcrval & XFEATURE_AVX) != XFEATURE_AVX) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* * AVX512 requires base AVX (YMM_Hi128) as well as OpMask, * ZMM_Hi256, and Hi16_ZMM. */ if (xcrval & XFEATURE_AVX512 && (xcrval & (XFEATURE_AVX512 | XFEATURE_AVX)) != (XFEATURE_AVX512 | XFEATURE_AVX)) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* * Intel MPX requires both bound register state flags to be * set. */ if (((xcrval & XFEATURE_ENABLED_BNDREGS) != 0) != ((xcrval & XFEATURE_ENABLED_BNDCSR) != 0)) { vm_inject_gp(vmx->vm, vcpu); return (HANDLED); } /* * This runs "inside" vmrun() with the guest's FPU state, so * modifying xcr0 directly modifies the guest's xcr0, not the * host's. */ load_xcr(0, xcrval); return (HANDLED); } static uint64_t vmx_get_guest_reg(struct vmx *vmx, int vcpu, int ident) { const struct vmxctx *vmxctx; vmxctx = &vmx->ctx[vcpu]; switch (ident) { case 0: return (vmxctx->guest_rax); case 1: return (vmxctx->guest_rcx); case 2: return (vmxctx->guest_rdx); case 3: return (vmxctx->guest_rbx); case 4: return (vmcs_read(VMCS_GUEST_RSP)); case 5: return (vmxctx->guest_rbp); case 6: return (vmxctx->guest_rsi); case 7: return (vmxctx->guest_rdi); case 8: return (vmxctx->guest_r8); case 9: return (vmxctx->guest_r9); case 10: return (vmxctx->guest_r10); case 11: return (vmxctx->guest_r11); case 12: return (vmxctx->guest_r12); case 13: return (vmxctx->guest_r13); case 14: return (vmxctx->guest_r14); case 15: return (vmxctx->guest_r15); default: panic("invalid vmx register %d", ident); } } static void vmx_set_guest_reg(struct vmx *vmx, int vcpu, int ident, uint64_t regval) { struct vmxctx *vmxctx; vmxctx = &vmx->ctx[vcpu]; switch (ident) { case 0: vmxctx->guest_rax = regval; break; case 1: vmxctx->guest_rcx = regval; break; case 2: vmxctx->guest_rdx = regval; break; case 3: vmxctx->guest_rbx = regval; break; case 4: vmcs_write(VMCS_GUEST_RSP, regval); break; case 5: vmxctx->guest_rbp = regval; break; case 6: vmxctx->guest_rsi = regval; break; case 7: vmxctx->guest_rdi = regval; break; case 8: vmxctx->guest_r8 = regval; break; case 9: vmxctx->guest_r9 = regval; break; case 10: vmxctx->guest_r10 = regval; break; case 11: vmxctx->guest_r11 = regval; break; case 12: vmxctx->guest_r12 = regval; break; case 13: vmxctx->guest_r13 = regval; break; case 14: vmxctx->guest_r14 = regval; break; case 15: vmxctx->guest_r15 = regval; break; default: panic("invalid vmx register %d", ident); } } static int vmx_emulate_cr0_access(struct vmx *vmx, int vcpu, uint64_t exitqual) { uint64_t crval, regval; /* We only handle mov to %cr0 at this time */ if ((exitqual & 0xf0) != 0x00) return (UNHANDLED); regval = vmx_get_guest_reg(vmx, vcpu, (exitqual >> 8) & 0xf); vmcs_write(VMCS_CR0_SHADOW, regval); crval = regval | cr0_ones_mask; crval &= ~cr0_zeros_mask; vmcs_write(VMCS_GUEST_CR0, crval); if (regval & CR0_PG) { uint64_t efer, entry_ctls; /* * If CR0.PG is 1 and EFER.LME is 1 then EFER.LMA and * the "IA-32e mode guest" bit in VM-entry control must be * equal. */ efer = vmcs_read(VMCS_GUEST_IA32_EFER); if (efer & EFER_LME) { efer |= EFER_LMA; vmcs_write(VMCS_GUEST_IA32_EFER, efer); entry_ctls = vmcs_read(VMCS_ENTRY_CTLS); entry_ctls |= VM_ENTRY_GUEST_LMA; vmcs_write(VMCS_ENTRY_CTLS, entry_ctls); } } return (HANDLED); } static int vmx_emulate_cr4_access(struct vmx *vmx, int vcpu, uint64_t exitqual) { uint64_t crval, regval; /* We only handle mov to %cr4 at this time */ if ((exitqual & 0xf0) != 0x00) return (UNHANDLED); regval = vmx_get_guest_reg(vmx, vcpu, (exitqual >> 8) & 0xf); vmcs_write(VMCS_CR4_SHADOW, regval); crval = regval | cr4_ones_mask; crval &= ~cr4_zeros_mask; vmcs_write(VMCS_GUEST_CR4, crval); return (HANDLED); } static int vmx_emulate_cr8_access(struct vmx *vmx, int vcpu, uint64_t exitqual) { struct vlapic *vlapic; uint64_t cr8; int regnum; /* We only handle mov %cr8 to/from a register at this time. */ if ((exitqual & 0xe0) != 0x00) { return (UNHANDLED); } vlapic = vm_lapic(vmx->vm, vcpu); regnum = (exitqual >> 8) & 0xf; if (exitqual & 0x10) { cr8 = vlapic_get_cr8(vlapic); vmx_set_guest_reg(vmx, vcpu, regnum, cr8); } else { cr8 = vmx_get_guest_reg(vmx, vcpu, regnum); vlapic_set_cr8(vlapic, cr8); } return (HANDLED); } /* * From section "Guest Register State" in the Intel SDM: CPL = SS.DPL */ static int vmx_cpl(void) { uint32_t ssar; ssar = vmcs_read(VMCS_GUEST_SS_ACCESS_RIGHTS); return ((ssar >> 5) & 0x3); } static enum vm_cpu_mode vmx_cpu_mode(void) { uint32_t csar; if (vmcs_read(VMCS_GUEST_IA32_EFER) & EFER_LMA) { csar = vmcs_read(VMCS_GUEST_CS_ACCESS_RIGHTS); if (csar & 0x2000) return (CPU_MODE_64BIT); /* CS.L = 1 */ else return (CPU_MODE_COMPATIBILITY); } else if (vmcs_read(VMCS_GUEST_CR0) & CR0_PE) { return (CPU_MODE_PROTECTED); } else { return (CPU_MODE_REAL); } } static enum vm_paging_mode vmx_paging_mode(void) { if (!(vmcs_read(VMCS_GUEST_CR0) & CR0_PG)) return (PAGING_MODE_FLAT); if (!(vmcs_read(VMCS_GUEST_CR4) & CR4_PAE)) return (PAGING_MODE_32); if (vmcs_read(VMCS_GUEST_IA32_EFER) & EFER_LME) return (PAGING_MODE_64); else return (PAGING_MODE_PAE); } static uint64_t inout_str_index(struct vmx *vmx, int vcpuid, int in) { uint64_t val; int error; enum vm_reg_name reg; reg = in ? VM_REG_GUEST_RDI : VM_REG_GUEST_RSI; error = vmx_getreg(vmx, vcpuid, reg, &val); KASSERT(error == 0, ("%s: vmx_getreg error %d", __func__, error)); return (val); } static uint64_t inout_str_count(struct vmx *vmx, int vcpuid, int rep) { uint64_t val; int error; if (rep) { error = vmx_getreg(vmx, vcpuid, VM_REG_GUEST_RCX, &val); KASSERT(!error, ("%s: vmx_getreg error %d", __func__, error)); } else { val = 1; } return (val); } static int inout_str_addrsize(uint32_t inst_info) { uint32_t size; size = (inst_info >> 7) & 0x7; switch (size) { case 0: return (2); /* 16 bit */ case 1: return (4); /* 32 bit */ case 2: return (8); /* 64 bit */ default: panic("%s: invalid size encoding %d", __func__, size); } } static void inout_str_seginfo(struct vmx *vmx, int vcpuid, uint32_t inst_info, int in, struct vm_inout_str *vis) { int error, s; if (in) { vis->seg_name = VM_REG_GUEST_ES; } else { s = (inst_info >> 15) & 0x7; vis->seg_name = vm_segment_name(s); } error = vmx_getdesc(vmx, vcpuid, vis->seg_name, &vis->seg_desc); KASSERT(error == 0, ("%s: vmx_getdesc error %d", __func__, error)); } static void vmx_paging_info(struct vm_guest_paging *paging) { paging->cr3 = vmcs_guest_cr3(); paging->cpl = vmx_cpl(); paging->cpu_mode = vmx_cpu_mode(); paging->paging_mode = vmx_paging_mode(); } static void vmexit_inst_emul(struct vm_exit *vmexit, uint64_t gpa, uint64_t gla) { struct vm_guest_paging *paging; uint32_t csar; paging = &vmexit->u.inst_emul.paging; vmexit->exitcode = VM_EXITCODE_INST_EMUL; vmexit->inst_length = 0; vmexit->u.inst_emul.gpa = gpa; vmexit->u.inst_emul.gla = gla; vmx_paging_info(paging); switch (paging->cpu_mode) { case CPU_MODE_REAL: vmexit->u.inst_emul.cs_base = vmcs_read(VMCS_GUEST_CS_BASE); vmexit->u.inst_emul.cs_d = 0; break; case CPU_MODE_PROTECTED: case CPU_MODE_COMPATIBILITY: vmexit->u.inst_emul.cs_base = vmcs_read(VMCS_GUEST_CS_BASE); csar = vmcs_read(VMCS_GUEST_CS_ACCESS_RIGHTS); vmexit->u.inst_emul.cs_d = SEG_DESC_DEF32(csar); break; default: vmexit->u.inst_emul.cs_base = 0; vmexit->u.inst_emul.cs_d = 0; break; } vie_init(&vmexit->u.inst_emul.vie, NULL, 0); } static int ept_fault_type(uint64_t ept_qual) { int fault_type; if (ept_qual & EPT_VIOLATION_DATA_WRITE) fault_type = VM_PROT_WRITE; else if (ept_qual & EPT_VIOLATION_INST_FETCH) fault_type = VM_PROT_EXECUTE; else fault_type= VM_PROT_READ; return (fault_type); } static 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 *vmx, int vcpuid) { uint32_t proc_ctls2; proc_ctls2 = vmx->cap[vcpuid].proc_ctls2; return ((proc_ctls2 & PROCBASED2_VIRTUALIZE_APIC_ACCESSES) ? 1 : 0); } static __inline int x2apic_virtualization(struct vmx *vmx, int vcpuid) { uint32_t proc_ctls2; proc_ctls2 = vmx->cap[vcpuid].proc_ctls2; return ((proc_ctls2 & PROCBASED2_VIRTUALIZE_X2APIC_MODE) ? 1 : 0); } static int vmx_handle_apic_write(struct vmx *vmx, int vcpuid, struct vlapic *vlapic, uint64_t qual) { int error, handled, offset; uint32_t *apic_regs, vector; bool retu; handled = HANDLED; offset = APIC_WRITE_OFFSET(qual); if (!apic_access_virtualization(vmx, vcpuid)) { /* * In general there should not be any APIC write VM-exits * unless APIC-access virtualization is enabled. * * However self-IPI virtualization can legitimately trigger * an APIC-write VM-exit so treat it specially. */ if (x2apic_virtualization(vmx, vcpuid) && offset == APIC_OFFSET_SELF_IPI) { apic_regs = (uint32_t *)(vlapic->apic_page); vector = apic_regs[APIC_OFFSET_SELF_IPI / 4]; vlapic_self_ipi_handler(vlapic, vector); return (HANDLED); } else return (UNHANDLED); } switch (offset) { case APIC_OFFSET_ID: vlapic_id_write_handler(vlapic); break; case APIC_OFFSET_LDR: vlapic_ldr_write_handler(vlapic); break; case APIC_OFFSET_DFR: vlapic_dfr_write_handler(vlapic); break; case APIC_OFFSET_SVR: vlapic_svr_write_handler(vlapic); break; case APIC_OFFSET_ESR: vlapic_esr_write_handler(vlapic); break; case APIC_OFFSET_ICR_LOW: retu = false; error = vlapic_icrlo_write_handler(vlapic, &retu); if (error != 0 || retu) handled = UNHANDLED; break; case APIC_OFFSET_CMCI_LVT: case APIC_OFFSET_TIMER_LVT ... APIC_OFFSET_ERROR_LVT: vlapic_lvt_write_handler(vlapic, offset); break; case APIC_OFFSET_TIMER_ICR: vlapic_icrtmr_write_handler(vlapic); break; case APIC_OFFSET_TIMER_DCR: vlapic_dcr_write_handler(vlapic); break; default: handled = UNHANDLED; break; } return (handled); } static bool apic_access_fault(struct vmx *vmx, int vcpuid, uint64_t gpa) { if (apic_access_virtualization(vmx, vcpuid) && (gpa >= DEFAULT_APIC_BASE && gpa < DEFAULT_APIC_BASE + PAGE_SIZE)) return (true); else return (false); } static int vmx_handle_apic_access(struct vmx *vmx, int vcpuid, struct vm_exit *vmexit) { uint64_t qual; int access_type, offset, allowed; if (!apic_access_virtualization(vmx, vcpuid)) return (UNHANDLED); qual = vmexit->u.vmx.exit_qualification; access_type = APIC_ACCESS_TYPE(qual); offset = APIC_ACCESS_OFFSET(qual); allowed = 0; if (access_type == 0) { /* * Read data access to the following registers is expected. */ switch (offset) { case APIC_OFFSET_APR: case APIC_OFFSET_PPR: case APIC_OFFSET_RRR: case APIC_OFFSET_CMCI_LVT: case APIC_OFFSET_TIMER_CCR: allowed = 1; break; default: break; } } else if (access_type == 1) { /* * Write data access to the following registers is expected. */ switch (offset) { case APIC_OFFSET_VER: case APIC_OFFSET_APR: case APIC_OFFSET_PPR: case APIC_OFFSET_RRR: case APIC_OFFSET_ISR0 ... APIC_OFFSET_ISR7: case APIC_OFFSET_TMR0 ... APIC_OFFSET_TMR7: case APIC_OFFSET_IRR0 ... APIC_OFFSET_IRR7: case APIC_OFFSET_CMCI_LVT: case APIC_OFFSET_TIMER_CCR: allowed = 1; break; default: break; } } if (allowed) { vmexit_inst_emul(vmexit, DEFAULT_APIC_BASE + offset, VIE_INVALID_GLA); } /* * Regardless of whether the APIC-access is allowed this handler * always returns UNHANDLED: * - if the access is allowed then it is handled by emulating the * instruction that caused the VM-exit (outside the critical section) * - if the access is not allowed then it will be converted to an * exitcode of VM_EXITCODE_VMX and will be dealt with in userland. */ return (UNHANDLED); } static enum task_switch_reason vmx_task_switch_reason(uint64_t qual) { int reason; reason = (qual >> 30) & 0x3; switch (reason) { case 0: return (TSR_CALL); case 1: return (TSR_IRET); case 2: return (TSR_JMP); case 3: return (TSR_IDT_GATE); default: panic("%s: invalid reason %d", __func__, reason); } } static int emulate_wrmsr(struct vmx *vmx, int vcpuid, u_int num, uint64_t val, bool *retu) { int error; if (lapic_msr(num)) error = lapic_wrmsr(vmx->vm, vcpuid, num, val, retu); else error = vmx_wrmsr(vmx, vcpuid, num, val, retu); return (error); } static int emulate_rdmsr(struct vmx *vmx, int vcpuid, u_int num, bool *retu) { struct vmxctx *vmxctx; uint64_t result; uint32_t eax, edx; int error; if (lapic_msr(num)) error = lapic_rdmsr(vmx->vm, vcpuid, num, &result, retu); else error = vmx_rdmsr(vmx, vcpuid, num, &result, retu); if (error == 0) { eax = result; vmxctx = &vmx->ctx[vcpuid]; error = vmxctx_setreg(vmxctx, VM_REG_GUEST_RAX, eax); KASSERT(error == 0, ("vmxctx_setreg(rax) error %d", error)); edx = result >> 32; error = vmxctx_setreg(vmxctx, VM_REG_GUEST_RDX, edx); KASSERT(error == 0, ("vmxctx_setreg(rdx) error %d", error)); } return (error); } static int vmx_exit_process(struct vmx *vmx, int vcpu, struct vm_exit *vmexit) { int error, errcode, errcode_valid, handled, in; struct vmxctx *vmxctx; struct vlapic *vlapic; struct vm_inout_str *vis; struct vm_task_switch *ts; uint32_t eax, ecx, edx, idtvec_info, idtvec_err, intr_info, inst_info; uint32_t intr_type, intr_vec, reason; uint64_t exitintinfo, qual, gpa; bool retu; CTASSERT((PINBASED_CTLS_ONE_SETTING & PINBASED_VIRTUAL_NMI) != 0); CTASSERT((PINBASED_CTLS_ONE_SETTING & PINBASED_NMI_EXITING) != 0); handled = UNHANDLED; vmxctx = &vmx->ctx[vcpu]; qual = vmexit->u.vmx.exit_qualification; reason = vmexit->u.vmx.exit_reason; vmexit->exitcode = VM_EXITCODE_BOGUS; vmm_stat_incr(vmx->vm, vcpu, VMEXIT_COUNT, 1); SDT_PROBE3(vmm, vmx, exit, entry, vmx, vcpu, 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)) { VCPU_CTR0(vmx->vm, vcpu, "Handling MCE during VM-entry"); __asm __volatile("int $18"); return (1); } /* * VM exits that can be triggered during event delivery need to * be handled specially by re-injecting the event if the IDT * vectoring information field's valid bit is set. * * See "Information for VM Exits During Event Delivery" in Intel SDM * for details. */ idtvec_info = vmcs_idt_vectoring_info(); if (idtvec_info & VMCS_IDT_VEC_VALID) { idtvec_info &= ~(1 << 12); /* clear undefined bit */ exitintinfo = idtvec_info; if (idtvec_info & VMCS_IDT_VEC_ERRCODE_VALID) { idtvec_err = vmcs_idt_vectoring_err(); exitintinfo |= (uint64_t)idtvec_err << 32; } error = vm_exit_intinfo(vmx->vm, vcpu, exitintinfo); KASSERT(error == 0, ("%s: vm_set_intinfo error %d", __func__, error)); /* * If 'virtual NMIs' are being used and the VM-exit * happened while injecting an NMI during the previous * VM-entry, then clear "blocking by NMI" in the * Guest Interruptibility-State so the NMI can be * reinjected on the subsequent VM-entry. * * However, if the NMI was being delivered through a task * gate, then the new task must start execution with NMIs * blocked so don't clear NMI blocking in this case. */ intr_type = idtvec_info & VMCS_INTR_T_MASK; if (intr_type == VMCS_INTR_T_NMI) { if (reason != EXIT_REASON_TASK_SWITCH) vmx_clear_nmi_blocking(vmx, vcpu); else vmx_assert_nmi_blocking(vmx, vcpu); } /* * Update VM-entry instruction length if the event being * delivered was a software interrupt or software exception. */ if (intr_type == VMCS_INTR_T_SWINTR || intr_type == VMCS_INTR_T_PRIV_SWEXCEPTION || intr_type == VMCS_INTR_T_SWEXCEPTION) { vmcs_write(VMCS_ENTRY_INST_LENGTH, vmexit->inst_length); } } switch (reason) { case EXIT_REASON_TASK_SWITCH: ts = &vmexit->u.task_switch; ts->tsssel = qual & 0xffff; ts->reason = vmx_task_switch_reason(qual); ts->ext = 0; ts->errcode_valid = 0; vmx_paging_info(&ts->paging); /* * If the task switch was due to a CALL, JMP, IRET, software * interrupt (INT n) or software exception (INT3, INTO), * then the saved %rip references the instruction that caused * the task switch. The instruction length field in the VMCS * is valid in this case. * * In all other cases (e.g., NMI, hardware exception) the * saved %rip is one that would have been saved in the old TSS * had the task switch completed normally so the instruction * length field is not needed in this case and is explicitly * set to 0. */ if (ts->reason == TSR_IDT_GATE) { KASSERT(idtvec_info & VMCS_IDT_VEC_VALID, ("invalid idtvec_info %#x for IDT task switch", idtvec_info)); intr_type = idtvec_info & VMCS_INTR_T_MASK; if (intr_type != VMCS_INTR_T_SWINTR && intr_type != VMCS_INTR_T_SWEXCEPTION && intr_type != VMCS_INTR_T_PRIV_SWEXCEPTION) { /* Task switch triggered by external event */ ts->ext = 1; vmexit->inst_length = 0; if (idtvec_info & VMCS_IDT_VEC_ERRCODE_VALID) { ts->errcode_valid = 1; ts->errcode = vmcs_idt_vectoring_err(); } } } vmexit->exitcode = VM_EXITCODE_TASK_SWITCH; SDT_PROBE4(vmm, vmx, exit, taskswitch, vmx, vcpu, vmexit, ts); VCPU_CTR4(vmx->vm, vcpu, "task switch reason %d, tss 0x%04x, " "%s errcode 0x%016lx", ts->reason, ts->tsssel, ts->ext ? "external" : "internal", ((uint64_t)ts->errcode << 32) | ts->errcode_valid); break; case EXIT_REASON_CR_ACCESS: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_CR_ACCESS, 1); SDT_PROBE4(vmm, vmx, exit, craccess, vmx, vcpu, vmexit, qual); switch (qual & 0xf) { case 0: handled = vmx_emulate_cr0_access(vmx, vcpu, qual); break; case 4: handled = vmx_emulate_cr4_access(vmx, vcpu, qual); break; case 8: handled = vmx_emulate_cr8_access(vmx, vcpu, qual); break; } break; case EXIT_REASON_RDMSR: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_RDMSR, 1); retu = false; ecx = vmxctx->guest_rcx; VCPU_CTR1(vmx->vm, vcpu, "rdmsr 0x%08x", ecx); SDT_PROBE4(vmm, vmx, exit, rdmsr, vmx, vcpu, vmexit, ecx); error = emulate_rdmsr(vmx, vcpu, ecx, &retu); if (error) { vmexit->exitcode = VM_EXITCODE_RDMSR; vmexit->u.msr.code = ecx; } else if (!retu) { handled = HANDLED; } else { /* Return to userspace with a valid exitcode */ KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, ("emulate_rdmsr retu with bogus exitcode")); } break; case EXIT_REASON_WRMSR: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_WRMSR, 1); retu = false; eax = vmxctx->guest_rax; ecx = vmxctx->guest_rcx; edx = vmxctx->guest_rdx; VCPU_CTR2(vmx->vm, vcpu, "wrmsr 0x%08x value 0x%016lx", ecx, (uint64_t)edx << 32 | eax); SDT_PROBE5(vmm, vmx, exit, wrmsr, vmx, vmexit, vcpu, ecx, (uint64_t)edx << 32 | eax); error = emulate_wrmsr(vmx, vcpu, ecx, (uint64_t)edx << 32 | eax, &retu); if (error) { vmexit->exitcode = VM_EXITCODE_WRMSR; vmexit->u.msr.code = ecx; vmexit->u.msr.wval = (uint64_t)edx << 32 | eax; } else if (!retu) { handled = HANDLED; } else { /* Return to userspace with a valid exitcode */ KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, ("emulate_wrmsr retu with bogus exitcode")); } break; case EXIT_REASON_HLT: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_HLT, 1); SDT_PROBE3(vmm, vmx, exit, halt, vmx, vcpu, 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(vmx->vm, vcpu, VMEXIT_MTRAP, 1); SDT_PROBE3(vmm, vmx, exit, mtrap, vmx, vcpu, vmexit); vmexit->exitcode = VM_EXITCODE_MTRAP; vmexit->inst_length = 0; break; case EXIT_REASON_PAUSE: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_PAUSE, 1); SDT_PROBE3(vmm, vmx, exit, pause, vmx, vcpu, vmexit); vmexit->exitcode = VM_EXITCODE_PAUSE; break; case EXIT_REASON_INTR_WINDOW: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INTR_WINDOW, 1); SDT_PROBE3(vmm, vmx, exit, intrwindow, vmx, vcpu, vmexit); vmx_clear_int_window_exiting(vmx, vcpu); return (1); case EXIT_REASON_EXT_INTR: /* * External interrupts serve only to cause VM exits and allow * the host interrupt handler to run. * * If this external interrupt triggers a virtual interrupt * to a VM, then that state will be recorded by the * host interrupt handler in the VM's softc. We will inject * this virtual interrupt during the subsequent VM enter. */ intr_info = vmcs_read(VMCS_EXIT_INTR_INFO); SDT_PROBE4(vmm, vmx, exit, interrupt, vmx, vcpu, 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(vmx->vm, vcpu, VMEXIT_EXTINT, 1); return (1); case EXIT_REASON_NMI_WINDOW: SDT_PROBE3(vmm, vmx, exit, nmiwindow, vmx, vcpu, vmexit); /* Exit to allow the pending virtual NMI to be injected */ if (vm_nmi_pending(vmx->vm, vcpu)) vmx_inject_nmi(vmx, vcpu); vmx_clear_nmi_window_exiting(vmx, vcpu); vmm_stat_incr(vmx->vm, vcpu, VMEXIT_NMI_WINDOW, 1); return (1); case EXIT_REASON_INOUT: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INOUT, 1); vmexit->exitcode = VM_EXITCODE_INOUT; vmexit->u.inout.bytes = (qual & 0x7) + 1; vmexit->u.inout.in = in = (qual & 0x8) ? 1 : 0; vmexit->u.inout.string = (qual & 0x10) ? 1 : 0; vmexit->u.inout.rep = (qual & 0x20) ? 1 : 0; vmexit->u.inout.port = (uint16_t)(qual >> 16); vmexit->u.inout.eax = (uint32_t)(vmxctx->guest_rax); if (vmexit->u.inout.string) { inst_info = vmcs_read(VMCS_EXIT_INSTRUCTION_INFO); vmexit->exitcode = VM_EXITCODE_INOUT_STR; vis = &vmexit->u.inout_str; vmx_paging_info(&vis->paging); vis->rflags = vmcs_read(VMCS_GUEST_RFLAGS); vis->cr0 = vmcs_read(VMCS_GUEST_CR0); vis->index = inout_str_index(vmx, vcpu, in); vis->count = inout_str_count(vmx, vcpu, vis->inout.rep); vis->addrsize = inout_str_addrsize(inst_info); inout_str_seginfo(vmx, vcpu, inst_info, in, vis); } SDT_PROBE3(vmm, vmx, exit, inout, vmx, vcpu, vmexit); break; case EXIT_REASON_CPUID: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_CPUID, 1); SDT_PROBE3(vmm, vmx, exit, cpuid, vmx, vcpu, vmexit); handled = vmx_handle_cpuid(vmx->vm, vcpu, vmxctx); break; case EXIT_REASON_EXCEPTION: vmm_stat_incr(vmx->vm, vcpu, VMEXIT_EXCEPTION, 1); intr_info = vmcs_read(VMCS_EXIT_INTR_INFO); KASSERT((intr_info & VMCS_INTR_VALID) != 0, ("VM exit interruption info invalid: %#x", intr_info)); intr_vec = intr_info & 0xff; intr_type = intr_info & VMCS_INTR_T_MASK; /* * If Virtual NMIs control is 1 and the VM-exit is due to a * fault encountered during the execution of IRET then we must * restore the state of "virtual-NMI blocking" before resuming * the guest. * * See "Resuming Guest Software after Handling an Exception". * See "Information for VM Exits Due to Vectored Events". */ if ((idtvec_info & VMCS_IDT_VEC_VALID) == 0 && (intr_vec != IDT_DF) && (intr_info & EXIT_QUAL_NMIUDTI) != 0) vmx_restore_nmi_blocking(vmx, vcpu); /* * The NMI has already been handled in vmx_exit_handle_nmi(). */ if (intr_type == VMCS_INTR_T_NMI) return (1); /* * Call the machine check handler by hand. Also don't reflect * the machine check back into the guest. */ if (intr_vec == IDT_MC) { VCPU_CTR0(vmx->vm, vcpu, "Vectoring to MCE handler"); __asm __volatile("int $18"); return (1); } /* * If 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 && (vmx->cap[vcpu].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); } VCPU_CTR2(vmx->vm, vcpu, "Reflecting exception %d/%#x into " "the guest", intr_vec, errcode); SDT_PROBE5(vmm, vmx, exit, exception, vmx, vcpu, vmexit, intr_vec, errcode); error = vm_inject_exception(vmx->vm, vcpu, intr_vec, errcode_valid, errcode, 0); KASSERT(error == 0, ("%s: vm_inject_exception error %d", __func__, error)); return (1); case EXIT_REASON_EPT_FAULT: /* * If 'gpa' lies within the address space allocated to * memory then this must be a nested page fault otherwise * this must be an instruction that accesses MMIO space. */ gpa = vmcs_gpa(); if (vm_mem_allocated(vmx->vm, vcpu, gpa) || apic_access_fault(vmx, vcpu, gpa)) { vmexit->exitcode = VM_EXITCODE_PAGING; vmexit->inst_length = 0; vmexit->u.paging.gpa = gpa; vmexit->u.paging.fault_type = ept_fault_type(qual); vmm_stat_incr(vmx->vm, vcpu, VMEXIT_NESTED_FAULT, 1); SDT_PROBE5(vmm, vmx, exit, nestedfault, vmx, vcpu, vmexit, gpa, qual); } else if (ept_emulation_fault(qual)) { vmexit_inst_emul(vmexit, gpa, vmcs_gla()); vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INST_EMUL, 1); SDT_PROBE4(vmm, vmx, exit, mmiofault, vmx, vcpu, 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(vmx, 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, vcpu, vmexit); vmexit->inst_length = 0; /* trap-like */ break; case EXIT_REASON_APIC_ACCESS: SDT_PROBE3(vmm, vmx, exit, apicaccess, vmx, vcpu, vmexit); handled = vmx_handle_apic_access(vmx, vcpu, vmexit); break; case EXIT_REASON_APIC_WRITE: /* * APIC-write VM exit is trap-like so the %rip is already * pointing to the next instruction. */ vmexit->inst_length = 0; vlapic = vm_lapic(vmx->vm, vcpu); SDT_PROBE4(vmm, vmx, exit, apicwrite, vmx, vcpu, vmexit, vlapic); handled = vmx_handle_apic_write(vmx, vcpu, vlapic, qual); break; case EXIT_REASON_XSETBV: SDT_PROBE3(vmm, vmx, exit, xsetbv, vmx, vcpu, vmexit); handled = vmx_emulate_xsetbv(vmx, vcpu, vmexit); break; case EXIT_REASON_MONITOR: SDT_PROBE3(vmm, vmx, exit, monitor, vmx, vcpu, vmexit); vmexit->exitcode = VM_EXITCODE_MONITOR; break; case EXIT_REASON_MWAIT: SDT_PROBE3(vmm, vmx, exit, mwait, vmx, vcpu, vmexit); vmexit->exitcode = VM_EXITCODE_MWAIT; break; case EXIT_REASON_TPR: vlapic = vm_lapic(vmx->vm, 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, vcpu, vmexit); vmexit->exitcode = VM_EXITCODE_VMINSN; break; default: SDT_PROBE4(vmm, vmx, exit, unknown, vmx, vcpu, vmexit, reason); vmm_stat_incr(vmx->vm, vcpu, VMEXIT_UNKNOWN, 1); break; } if (handled) { /* * It is possible that control is returned to userland * even though we were able to handle the VM exit in the * kernel. * * In such a case we want to make sure that the userland * restarts guest execution at the instruction *after* * the one we just processed. Therefore we update the * guest rip in the VMCS and in 'vmexit'. */ vmexit->rip += vmexit->inst_length; vmexit->inst_length = 0; vmcs_write(VMCS_GUEST_RIP, vmexit->rip); } else { if (vmexit->exitcode == VM_EXITCODE_BOGUS) { /* * If this VM exit was not claimed by anybody then * treat it as a generic VMX exit. */ vmexit->exitcode = VM_EXITCODE_VMX; vmexit->u.vmx.status = VM_SUCCESS; vmexit->u.vmx.inst_type = 0; vmexit->u.vmx.inst_error = 0; } else { /* * The exitcode and collateral have been populated. * The VM exit will be processed further in userland. */ } } SDT_PROBE4(vmm, vmx, exit, return, vmx, vcpu, 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: case VMX_INVEPT_ERROR: vmexit->u.vmx.inst_type = rc; break; default: panic("vm_exit_inst_error: vmx_enter_guest returned %d", rc); } } /* * If the NMI-exiting VM execution control is set to '1' then an NMI in * non-root operation causes a VM-exit. NMI blocking is in effect so it is * sufficient to simply vector to the NMI handler via a software interrupt. * However, this must be done before maskable interrupts are enabled * otherwise the "iret" issued by an interrupt handler will incorrectly * clear NMI blocking. */ static __inline void vmx_exit_handle_nmi(struct vmx *vmx, int vcpuid, struct vm_exit *vmexit) { uint32_t intr_info; KASSERT((read_rflags() & PSL_I) == 0, ("interrupts enabled")); if (vmexit->u.vmx.exit_reason != EXIT_REASON_EXCEPTION) return; intr_info = vmcs_read(VMCS_EXIT_INTR_INFO); KASSERT((intr_info & VMCS_INTR_VALID) != 0, ("VM exit interruption info invalid: %#x", intr_info)); if ((intr_info & VMCS_INTR_T_MASK) == VMCS_INTR_T_NMI) { KASSERT((intr_info & 0xff) == IDT_NMI, ("VM exit due " "to NMI has invalid vector: %#x", intr_info)); VCPU_CTR0(vmx->vm, vcpuid, "Vectoring to NMI handler"); __asm __volatile("int $2"); } } static __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 int vmx_run(void *arg, int vcpu, register_t rip, pmap_t pmap, struct vm_eventinfo *evinfo) { int rc, handled, launched; struct vmx *vmx; struct vm *vm; struct vmxctx *vmxctx; struct vmcs *vmcs; struct vm_exit *vmexit; struct vlapic *vlapic; uint32_t exit_reason; struct region_descriptor gdtr, idtr; uint16_t ldt_sel; vmx = arg; vm = vmx->vm; vmcs = &vmx->vmcs[vcpu]; vmxctx = &vmx->ctx[vcpu]; vlapic = vm_lapic(vm, vcpu); vmexit = vm_exitinfo(vm, vcpu); launched = 0; KASSERT(vmxctx->pmap == pmap, ("pmap %p different than ctx pmap %p", pmap, vmxctx->pmap)); vmx_msr_guest_enter(vmx, vcpu); VMPTRLD(vmcs); /* * XXX * We do this every time because we may setup the virtual machine * from a different process than the one that actually runs it. * * If the life of a virtual machine was spent entirely in the context * of a single process we could do this once in vmx_vminit(). */ vmcs_write(VMCS_HOST_CR3, rcr3()); vmcs_write(VMCS_GUEST_RIP, rip); vmx_set_pcpu_defaults(vmx, vcpu, pmap); do { KASSERT(vmcs_guest_rip() == rip, ("%s: vmcs guest rip mismatch " "%#lx/%#lx", __func__, vmcs_guest_rip(), rip)); handled = UNHANDLED; /* * Interrupts are disabled from this point on until the * guest starts executing. This is done for the following * reasons: * * If an AST is asserted on this thread after the check below, * then the IPI_AST notification will not be lost, because it * will cause a VM exit due to external interrupt as soon as * the guest state is loaded. * * A posted interrupt after 'vmx_inject_interrupts()' will * not be "lost" because it will be held pending in the host * APIC because interrupts are disabled. The pending interrupt * will be recognized as soon as the guest state is loaded. * * The same reasoning applies to the IPI generated by * pmap_invalidate_ept(). */ disable_intr(); vmx_inject_interrupts(vmx, vcpu, vlapic, rip); /* * Check for vcpu suspension after injecting events because * vmx_inject_interrupts() can suspend the vcpu due to a * triple fault. */ if (vcpu_suspended(evinfo)) { enable_intr(); vm_exit_suspended(vmx->vm, vcpu, rip); break; } if (vcpu_rendezvous_pending(evinfo)) { enable_intr(); vm_exit_rendezvous(vmx->vm, vcpu, rip); break; } if (vcpu_reqidle(evinfo)) { enable_intr(); vm_exit_reqidle(vmx->vm, vcpu, rip); break; } if (vcpu_should_yield(vm, vcpu)) { enable_intr(); vm_exit_astpending(vmx->vm, vcpu, rip); vmx_astpending_trace(vmx, vcpu, rip); handled = HANDLED; break; } if (vcpu_debugged(vm, vcpu)) { enable_intr(); vm_exit_debug(vmx->vm, 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 ((vmx->cap[vcpu].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_run_trace(vmx, vcpu); vmx_dr_enter_guest(vmxctx); rc = vmx_enter_guest(vmxctx, vmx, launched); 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' */ vmx->state[vcpu].nextrip = rip; if (rc == VMX_GUEST_VMEXIT) { vmx_exit_handle_nmi(vmx, vcpu, vmexit); enable_intr(); handled = vmx_exit_process(vmx, vcpu, vmexit); } else { enable_intr(); vmx_exit_inst_error(vmxctx, rc, vmexit); } launched = 1; vmx_exit_trace(vmx, vcpu, rip, exit_reason, handled); rip = vmexit->rip; } while (handled); /* * If a VM exit has been handled then the exitcode must be BOGUS * If a VM exit is not handled then the exitcode must not be BOGUS */ if ((handled && vmexit->exitcode != VM_EXITCODE_BOGUS) || (!handled && vmexit->exitcode == VM_EXITCODE_BOGUS)) { panic("Mismatch between handled (%d) and exitcode (%d)", handled, vmexit->exitcode); } if (!handled) vmm_stat_incr(vm, vcpu, VMEXIT_USERSPACE, 1); VCPU_CTR1(vm, vcpu, "returning from vmx_run: exitcode %d", vmexit->exitcode); VMCLEAR(vmcs); vmx_msr_guest_exit(vmx, vcpu); return (0); } static void vmx_vmcleanup(void *arg) { int i; struct vmx *vmx = arg; uint16_t maxcpus; if (apic_access_virtualization(vmx, 0)) vm_unmap_mmio(vmx->vm, DEFAULT_APIC_BASE, PAGE_SIZE); maxcpus = vm_get_maxcpus(vmx->vm); for (i = 0; i < maxcpus; i++) vpid_free(vmx->state[i].vpid); free(vmx, M_VMX); return; } static register_t * vmxctx_regptr(struct vmxctx *vmxctx, int reg) { switch (reg) { case VM_REG_GUEST_RAX: return (&vmxctx->guest_rax); case VM_REG_GUEST_RBX: return (&vmxctx->guest_rbx); case VM_REG_GUEST_RCX: return (&vmxctx->guest_rcx); case VM_REG_GUEST_RDX: return (&vmxctx->guest_rdx); case VM_REG_GUEST_RSI: return (&vmxctx->guest_rsi); case VM_REG_GUEST_RDI: return (&vmxctx->guest_rdi); case VM_REG_GUEST_RBP: return (&vmxctx->guest_rbp); case VM_REG_GUEST_R8: return (&vmxctx->guest_r8); case VM_REG_GUEST_R9: return (&vmxctx->guest_r9); case VM_REG_GUEST_R10: return (&vmxctx->guest_r10); case VM_REG_GUEST_R11: return (&vmxctx->guest_r11); case VM_REG_GUEST_R12: return (&vmxctx->guest_r12); case VM_REG_GUEST_R13: return (&vmxctx->guest_r13); case VM_REG_GUEST_R14: return (&vmxctx->guest_r14); case VM_REG_GUEST_R15: return (&vmxctx->guest_r15); case VM_REG_GUEST_CR2: return (&vmxctx->guest_cr2); 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 *vmx, int vcpu, int running, uint64_t *retval) { uint64_t gi; int error; error = vmcs_getreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(VMCS_GUEST_INTERRUPTIBILITY), &gi); *retval = (gi & HWINTR_BLOCKING) ? 1 : 0; return (error); } static int vmx_modify_intr_shadow(struct vmx *vmx, int vcpu, int running, uint64_t val) { struct vmcs *vmcs; uint64_t gi; int error, ident; /* * Forcing the vcpu into an interrupt shadow is not supported. */ if (val) { error = EINVAL; goto done; } vmcs = &vmx->vmcs[vcpu]; ident = VMCS_IDENT(VMCS_GUEST_INTERRUPTIBILITY); error = vmcs_getreg(vmcs, running, ident, &gi); if (error == 0) { gi &= ~HWINTR_BLOCKING; error = vmcs_setreg(vmcs, running, ident, gi); } done: VCPU_CTR2(vmx->vm, vcpu, "Setting intr_shadow to %#lx %s", val, error ? "failed" : "succeeded"); return (error); } static int vmx_shadow_reg(int reg) { int shreg; shreg = -1; switch (reg) { case VM_REG_GUEST_CR0: shreg = VMCS_CR0_SHADOW; break; case VM_REG_GUEST_CR4: shreg = VMCS_CR4_SHADOW; break; default: break; } return (shreg); } static int vmx_getreg(void *arg, int vcpu, int reg, uint64_t *retval) { int running, hostcpu; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_getreg: %s%d is running", vm_name(vmx->vm), vcpu); if (reg == VM_REG_GUEST_INTR_SHADOW) return (vmx_get_intr_shadow(vmx, vcpu, running, retval)); if (vmxctx_getreg(&vmx->ctx[vcpu], reg, retval) == 0) return (0); return (vmcs_getreg(&vmx->vmcs[vcpu], running, reg, retval)); } static int vmx_setreg(void *arg, int vcpu, int reg, uint64_t val) { int error, hostcpu, running, shadow; uint64_t ctls; pmap_t pmap; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_setreg: %s%d is running", vm_name(vmx->vm), vcpu); if (reg == VM_REG_GUEST_INTR_SHADOW) return (vmx_modify_intr_shadow(vmx, vcpu, running, val)); if (vmxctx_setreg(&vmx->ctx[vcpu], reg, val) == 0) return (0); error = vmcs_setreg(&vmx->vmcs[vcpu], running, reg, val); if (error == 0) { /* * If the "load EFER" VM-entry control is 1 then the * value of EFER.LMA must be identical to "IA-32e mode guest" * bit in the VM-entry control. */ if ((entry_ctls & VM_ENTRY_LOAD_EFER) != 0 && (reg == VM_REG_GUEST_EFER)) { vmcs_getreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(VMCS_ENTRY_CTLS), &ctls); if (val & EFER_LMA) ctls |= VM_ENTRY_GUEST_LMA; else ctls &= ~VM_ENTRY_GUEST_LMA; vmcs_setreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(VMCS_ENTRY_CTLS), ctls); } shadow = vmx_shadow_reg(reg); if (shadow > 0) { /* * Store the unmodified value in the shadow */ error = vmcs_setreg(&vmx->vmcs[vcpu], running, VMCS_IDENT(shadow), val); } if (reg == VM_REG_GUEST_CR3) { /* * Invalidate the guest vcpu's TLB mappings to emulate * the behavior of updating %cr3. * * XXX the processor retains global mappings when %cr3 * is updated but vmx_invvpid() does not. */ pmap = vmx->ctx[vcpu].pmap; vmx_invvpid(vmx, vcpu, pmap, running); } } return (error); } static int vmx_getdesc(void *arg, int vcpu, int reg, struct seg_desc *desc) { int hostcpu, running; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_getdesc: %s%d is running", vm_name(vmx->vm), vcpu); return (vmcs_getdesc(&vmx->vmcs[vcpu], running, reg, desc)); } static int vmx_setdesc(void *arg, int vcpu, int reg, struct seg_desc *desc) { int hostcpu, running; struct vmx *vmx = arg; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) panic("vmx_setdesc: %s%d is running", vm_name(vmx->vm), vcpu); return (vmcs_setdesc(&vmx->vmcs[vcpu], running, reg, desc)); } static int vmx_getcap(void *arg, int vcpu, int type, int *retval) { struct vmx *vmx = arg; int vcap; int ret; ret = ENOENT; vcap = vmx->cap[vcpu].set; switch (type) { case VM_CAP_HALT_EXIT: if (cap_halt_exit) ret = 0; break; case VM_CAP_PAUSE_EXIT: if (cap_pause_exit) ret = 0; break; case VM_CAP_MTRAP_EXIT: if (cap_monitor_trap) ret = 0; break; + case VM_CAP_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: ret = 0; break; default: break; } if (ret == 0) *retval = (vcap & (1 << type)) ? 1 : 0; return (ret); } static int vmx_setcap(void *arg, int vcpu, int type, int val) { struct vmx *vmx = arg; struct vmcs *vmcs = &vmx->vmcs[vcpu]; uint32_t baseval; uint32_t *pptr; int error; int flag; int reg; int retval; retval = ENOENT; pptr = NULL; switch (type) { case VM_CAP_HALT_EXIT: if (cap_halt_exit) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls; baseval = *pptr; flag = PROCBASED_HLT_EXITING; reg = VMCS_PRI_PROC_BASED_CTLS; } break; case VM_CAP_MTRAP_EXIT: if (cap_monitor_trap) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls; baseval = *pptr; flag = PROCBASED_MTF; reg = VMCS_PRI_PROC_BASED_CTLS; } break; case VM_CAP_PAUSE_EXIT: if (cap_pause_exit) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls; baseval = *pptr; flag = PROCBASED_PAUSE_EXITING; reg = VMCS_PRI_PROC_BASED_CTLS; } + break; + case VM_CAP_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_init(), RDPID/RDTSCP are + * either always enabled or always disabled. + */ + error = EOPNOTSUPP; break; case VM_CAP_UNRESTRICTED_GUEST: if (cap_unrestricted_guest) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls2; baseval = *pptr; flag = PROCBASED2_UNRESTRICTED_GUEST; reg = VMCS_SEC_PROC_BASED_CTLS; } break; case VM_CAP_ENABLE_INVPCID: if (cap_invpcid) { retval = 0; pptr = &vmx->cap[vcpu].proc_ctls2; baseval = *pptr; flag = PROCBASED2_ENABLE_INVPCID; reg = VMCS_SEC_PROC_BASED_CTLS; } break; case VM_CAP_BPT_EXIT: retval = 0; /* Don't change the bitmap if we are tracing all exceptions. */ if (vmx->cap[vcpu].exc_bitmap != 0xffffffff) { pptr = &vmx->cap[vcpu].exc_bitmap; baseval = *pptr; flag = (1 << IDT_BP); reg = VMCS_EXCEPTION_BITMAP; } 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) { vmx->cap[vcpu].set |= (1 << type); } else { vmx->cap[vcpu].set &= ~(1 << type); } return (0); } struct vlapic_vtx { struct vlapic vlapic; struct pir_desc *pir_desc; struct vmx *vmx; u_int pending_prio; }; #define VPR_PRIO_BIT(vpr) (1 << ((vpr) >> 4)) #define VMX_CTR_PIR(vm, vcpuid, pir_desc, notify, vector, level, msg) \ do { \ VCPU_CTR2(vm, vcpuid, msg " assert %s-triggered vector %d", \ level ? "level" : "edge", vector); \ VCPU_CTR1(vm, vcpuid, msg " pir0 0x%016lx", pir_desc->pir[0]); \ VCPU_CTR1(vm, vcpuid, msg " pir1 0x%016lx", pir_desc->pir[1]); \ VCPU_CTR1(vm, vcpuid, msg " pir2 0x%016lx", pir_desc->pir[2]); \ VCPU_CTR1(vm, vcpuid, msg " pir3 0x%016lx", pir_desc->pir[3]); \ VCPU_CTR1(vm, vcpuid, msg " notify: %s", notify ? "yes" : "no");\ } while (0) /* * vlapic->ops handlers that utilize the APICv hardware assist described in * Chapter 29 of the Intel SDM. */ static int vmx_set_intr_ready(struct vlapic *vlapic, int vector, bool level) { struct vlapic_vtx *vlapic_vtx; struct pir_desc *pir_desc; uint64_t mask; int idx, notify = 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->vm, vlapic->vcpuid, pir_desc, notify, vector, level, "vmx_set_intr_ready"); return (notify); } static int vmx_pending_intr(struct vlapic *vlapic, int *vecptr) { struct vlapic_vtx *vlapic_vtx; struct pir_desc *pir_desc; struct LAPIC *lapic; uint64_t pending, pirval; uint32_t ppr, vpr; int i; /* * This function is only expected to be called from the 'HLT' exit * handler which does not care about the vector that is pending. */ KASSERT(vecptr == NULL, ("vmx_pending_intr: vecptr must be NULL")); vlapic_vtx = (struct vlapic_vtx *)vlapic; pir_desc = vlapic_vtx->pir_desc; pending = atomic_load_acq_long(&pir_desc->pending); if (!pending) { /* * 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->vm, vlapic->vcpuid); 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); VCPU_CTR1(vlapic->vm, vlapic->vcpuid, "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 vmx *vmx; struct vmcs *vmcs; uint64_t mask, val; KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d", vector)); KASSERT(!vcpu_is_running(vlapic->vm, vlapic->vcpuid, NULL), ("vmx_set_tmr: vcpu cannot be running")); vlapic_vtx = (struct vlapic_vtx *)vlapic; vmx = vlapic_vtx->vmx; vmcs = &vmx->vmcs[vlapic->vcpuid]; mask = 1UL << (vector % 64); VMPTRLD(vmcs); val = vmcs_read(VMCS_EOI_EXIT(vector)); if (level) val |= mask; else val &= ~mask; vmcs_write(VMCS_EOI_EXIT(vector), val); VMCLEAR(vmcs); } static void vmx_enable_x2apic_mode_ts(struct vlapic *vlapic) { struct vmx *vmx; struct vmcs *vmcs; uint32_t proc_ctls; int vcpuid; vcpuid = vlapic->vcpuid; vmx = ((struct vlapic_vtx *)vlapic)->vmx; vmcs = &vmx->vmcs[vcpuid]; proc_ctls = vmx->cap[vcpuid].proc_ctls; proc_ctls &= ~PROCBASED_USE_TPR_SHADOW; proc_ctls |= PROCBASED_CR8_LOAD_EXITING; proc_ctls |= PROCBASED_CR8_STORE_EXITING; vmx->cap[vcpuid].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 vmx *vmx; struct vmcs *vmcs; uint32_t proc_ctls2; int vcpuid, error; vcpuid = vlapic->vcpuid; vmx = ((struct vlapic_vtx *)vlapic)->vmx; vmcs = &vmx->vmcs[vcpuid]; proc_ctls2 = vmx->cap[vcpuid].proc_ctls2; KASSERT((proc_ctls2 & PROCBASED2_VIRTUALIZE_APIC_ACCESSES) != 0, ("%s: invalid proc_ctls2 %#x", __func__, proc_ctls2)); proc_ctls2 &= ~PROCBASED2_VIRTUALIZE_APIC_ACCESSES; proc_ctls2 |= PROCBASED2_VIRTUALIZE_X2APIC_MODE; vmx->cap[vcpuid].proc_ctls2 = proc_ctls2; VMPTRLD(vmcs); vmcs_write(VMCS_SEC_PROC_BASED_CTLS, proc_ctls2); VMCLEAR(vmcs); if (vlapic->vcpuid == 0) { /* * The nested page table mappings are shared by all vcpus * so unmap the APIC access page just once. */ error = vm_unmap_mmio(vmx->vm, DEFAULT_APIC_BASE, PAGE_SIZE); KASSERT(error == 0, ("%s: vm_unmap_mmio error %d", __func__, error)); /* * The MSR bitmap is shared by all vcpus so modify it only * once in the context of vcpu 0. */ error = vmx_allow_x2apic_msrs(vmx); KASSERT(error == 0, ("%s: vmx_allow_x2apic_msrs error %d", __func__, error)); } } static void vmx_post_intr(struct vlapic *vlapic, int hostcpu) { ipi_cpu(hostcpu, pirvec); } /* * Transfer the pending interrupts in the PIR descriptor to the IRR * in the virtual APIC page. */ static void vmx_inject_pir(struct vlapic *vlapic) { struct vlapic_vtx *vlapic_vtx; struct pir_desc *pir_desc; struct LAPIC *lapic; uint64_t val, pirval; int rvi, pirbase = -1; uint16_t intr_status_old, intr_status_new; vlapic_vtx = (struct vlapic_vtx *)vlapic; pir_desc = vlapic_vtx->pir_desc; if (atomic_cmpset_long(&pir_desc->pending, 1, 0) == 0) { VCPU_CTR0(vlapic->vm, vlapic->vcpuid, "vmx_inject_pir: " "no posted interrupt pending"); return; } pirval = 0; pirbase = -1; lapic = vlapic->apic_page; val = atomic_readandclear_long(&pir_desc->pir[0]); if (val != 0) { lapic->irr0 |= val; lapic->irr1 |= val >> 32; pirbase = 0; pirval = val; } val = atomic_readandclear_long(&pir_desc->pir[1]); if (val != 0) { lapic->irr2 |= val; lapic->irr3 |= val >> 32; pirbase = 64; pirval = val; } val = atomic_readandclear_long(&pir_desc->pir[2]); if (val != 0) { lapic->irr4 |= val; lapic->irr5 |= val >> 32; pirbase = 128; pirval = val; } val = atomic_readandclear_long(&pir_desc->pir[3]); if (val != 0) { lapic->irr6 |= val; lapic->irr7 |= val >> 32; pirbase = 192; pirval = val; } VLAPIC_CTR_IRR(vlapic, "vmx_inject_pir"); /* * Update RVI so the processor can evaluate pending virtual * interrupts on VM-entry. * * It is possible for pirval to be 0 here, even though the * pending bit has been set. The scenario is: * CPU-Y is sending a posted interrupt to CPU-X, which * is running a guest and processing posted interrupts in h/w. * CPU-X will eventually exit and the state seen in s/w is * the pending bit set, but no PIR bits set. * * CPU-X CPU-Y * (vm running) (host running) * rx posted interrupt * CLEAR pending bit * SET PIR bit * READ/CLEAR PIR bits * SET pending bit * (vm exit) * pending bit set, PIR 0 */ if (pirval != 0) { rvi = pirbase + flsl(pirval) - 1; intr_status_old = vmcs_read(VMCS_GUEST_INTR_STATUS); intr_status_new = (intr_status_old & 0xFF00) | rvi; if (intr_status_new > intr_status_old) { vmcs_write(VMCS_GUEST_INTR_STATUS, intr_status_new); VCPU_CTR2(vlapic->vm, vlapic->vcpuid, "vmx_inject_pir: " "guest_intr_status changed from 0x%04x to 0x%04x", intr_status_old, intr_status_new); } } } static struct vlapic * vmx_vlapic_init(void *arg, int vcpuid) { struct vmx *vmx; struct vlapic *vlapic; struct vlapic_vtx *vlapic_vtx; vmx = arg; vlapic = malloc(sizeof(struct vlapic_vtx), M_VLAPIC, M_WAITOK | M_ZERO); vlapic->vm = vmx->vm; vlapic->vcpuid = vcpuid; vlapic->apic_page = (struct LAPIC *)&vmx->apic_page[vcpuid]; vlapic_vtx = (struct vlapic_vtx *)vlapic; vlapic_vtx->pir_desc = &vmx->pir_desc[vcpuid]; vlapic_vtx->vmx = vmx; if (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(void *arg, struct vlapic *vlapic) { vlapic_cleanup(vlapic); free(vlapic, M_VLAPIC); } #ifdef BHYVE_SNAPSHOT static int vmx_snapshot_vmi(void *arg, struct vm_snapshot_meta *meta) { struct vmx *vmx; struct vmxctx *vmxctx; int i; int ret; vmx = arg; KASSERT(vmx != NULL, ("%s: arg was NULL", __func__)); for (i = 0; i < VM_MAXCPU; i++) { SNAPSHOT_BUF_OR_LEAVE(vmx->guest_msrs[i], sizeof(vmx->guest_msrs[i]), meta, ret, done); vmxctx = &vmx->ctx[i]; SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rdi, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rsi, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rdx, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rcx, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r8, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r9, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rax, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rbx, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_rbp, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r10, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r11, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r12, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r13, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r14, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_r15, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_cr2, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_dr0, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_dr1, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_dr2, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_dr3, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vmxctx->guest_dr6, meta, ret, done); } done: return (ret); } static int vmx_snapshot_vmcx(void *arg, struct vm_snapshot_meta *meta, int vcpu) { struct vmcs *vmcs; struct vmx *vmx; int err, run, hostcpu; vmx = (struct vmx *)arg; err = 0; KASSERT(arg != NULL, ("%s: arg was NULL", __func__)); vmcs = &vmx->vmcs[vcpu]; run = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (run && hostcpu != curcpu) { printf("%s: %s%d is running", __func__, vm_name(vmx->vm), vcpu); 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); return (err); } static int vmx_restore_tsc(void *arg, int vcpu, uint64_t offset) { struct vmcs *vmcs; struct vmx *vmx = (struct vmx *)arg; int error, running, hostcpu; KASSERT(arg != NULL, ("%s: arg was NULL", __func__)); vmcs = &vmx->vmcs[vcpu]; running = vcpu_is_running(vmx->vm, vcpu, &hostcpu); if (running && hostcpu != curcpu) { printf("%s: %s%d is running", __func__, vm_name(vmx->vm), vcpu); return (EINVAL); } if (!running) VMPTRLD(vmcs); error = vmx_set_tsc_offset(vmx, vcpu, offset); if (!running) VMCLEAR(vmcs); return (error); } #endif struct vmm_ops vmm_ops_intel = { .init = vmx_init, .cleanup = vmx_cleanup, .resume = vmx_restore, .vminit = vmx_vminit, .vmrun = vmx_run, .vmcleanup = vmx_vmcleanup, .vmgetreg = vmx_getreg, .vmsetreg = vmx_setreg, .vmgetdesc = vmx_getdesc, .vmsetdesc = vmx_setdesc, .vmgetcap = vmx_getcap, .vmsetcap = vmx_setcap, .vmspace_alloc = ept_vmspace_alloc, .vmspace_free = ept_vmspace_free, .vlapic_init = vmx_vlapic_init, .vlapic_cleanup = vmx_vlapic_cleanup, #ifdef BHYVE_SNAPSHOT .vmsnapshot = vmx_snapshot_vmi, .vmcx_snapshot = vmx_snapshot_vmcx, .vm_restore_tsc = vmx_restore_tsc, #endif }; Index: head/sys/amd64/vmm/intel/vmx.h =================================================================== --- head/sys/amd64/vmm/intel/vmx.h (revision 364339) +++ head/sys/amd64/vmm/intel/vmx.h (revision 364340) @@ -1,156 +1,171 @@ /*- * 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 _VMX_H_ #define _VMX_H_ #include "vmcs.h" struct pmap; struct vmxctx { register_t guest_rdi; /* Guest state */ register_t guest_rsi; register_t guest_rdx; register_t guest_rcx; register_t guest_r8; register_t guest_r9; register_t guest_rax; register_t guest_rbx; register_t guest_rbp; register_t guest_r10; register_t guest_r11; register_t guest_r12; register_t guest_r13; register_t guest_r14; register_t guest_r15; register_t guest_cr2; register_t guest_dr0; register_t guest_dr1; register_t guest_dr2; register_t guest_dr3; register_t guest_dr6; register_t host_r15; /* Host state */ register_t host_r14; register_t host_r13; register_t host_r12; register_t host_rbp; register_t host_rsp; register_t host_rbx; register_t host_dr0; register_t host_dr1; register_t host_dr2; register_t host_dr3; register_t host_dr6; register_t host_dr7; uint64_t host_debugctl; int host_tf; int inst_fail_status; /* * The pmap needs to be deactivated in vmx_enter_guest() * so keep a copy of the 'pmap' in each vmxctx. */ struct pmap *pmap; }; struct vmxcap { int set; uint32_t proc_ctls; uint32_t proc_ctls2; uint32_t exc_bitmap; }; struct vmxstate { uint64_t nextrip; /* next instruction to be executed by guest */ int lastcpu; /* host cpu that this 'vcpu' last ran on */ uint16_t vpid; }; struct apic_page { uint32_t reg[PAGE_SIZE / 4]; }; CTASSERT(sizeof(struct apic_page) == PAGE_SIZE); /* Posted Interrupt Descriptor (described in section 29.6 of the Intel SDM) */ struct pir_desc { uint64_t pir[4]; uint64_t pending; uint64_t unused[3]; } __aligned(64); CTASSERT(sizeof(struct pir_desc) == 64); /* Index into the 'guest_msrs[]' array */ enum { IDX_MSR_LSTAR, IDX_MSR_CSTAR, IDX_MSR_STAR, IDX_MSR_SF_MASK, IDX_MSR_KGSBASE, IDX_MSR_PAT, + IDX_MSR_TSC_AUX, GUEST_MSR_NUM /* must be the last enumeration */ }; /* virtual machine softc */ struct vmx { struct vmcs vmcs[VM_MAXCPU]; /* one vmcs per virtual cpu */ struct apic_page apic_page[VM_MAXCPU]; /* one apic page per vcpu */ char msr_bitmap[PAGE_SIZE]; struct pir_desc pir_desc[VM_MAXCPU]; uint64_t guest_msrs[VM_MAXCPU][GUEST_MSR_NUM]; struct vmxctx ctx[VM_MAXCPU]; struct vmxcap cap[VM_MAXCPU]; struct vmxstate state[VM_MAXCPU]; uint64_t eptp; struct vm *vm; long eptgen[MAXCPU]; /* cached pmap->pm_eptgen */ }; CTASSERT((offsetof(struct vmx, vmcs) & PAGE_MASK) == 0); CTASSERT((offsetof(struct vmx, msr_bitmap) & PAGE_MASK) == 0); CTASSERT((offsetof(struct vmx, pir_desc[0]) & 63) == 0); #define VMX_GUEST_VMEXIT 0 #define VMX_VMRESUME_ERROR 1 #define VMX_VMLAUNCH_ERROR 2 #define VMX_INVEPT_ERROR 3 int vmx_enter_guest(struct vmxctx *ctx, struct vmx *vmx, int launched); void vmx_call_isr(uintptr_t entry); u_long vmx_fix_cr0(u_long cr0); u_long vmx_fix_cr4(u_long cr4); int vmx_set_tsc_offset(struct vmx *vmx, int vcpu, uint64_t offset); extern char vmx_exit_guest[]; extern char vmx_exit_guest_flush_rsb[]; + +static inline bool +vmx_have_msr_tsc_aux(struct vmx *vmx) +{ + int rdpid_rdtscp_bits = ((1 << VM_CAP_RDPID) | (1 << VM_CAP_RDTSCP)); + + /* + * Since the values of these bits are uniform across all vCPUs + * (see discussion in vmx_init() and initialization of these bits + * in vmx_vminit()), just always use vCPU-zero's capability set and + * remove the need to require a vcpuid argument. + */ + return ((vmx->cap[0].set & rdpid_rdtscp_bits) != 0); +} #endif Index: head/sys/amd64/vmm/intel/vmx_msr.c =================================================================== --- head/sys/amd64/vmm/intel/vmx_msr.c (revision 364339) +++ head/sys/amd64/vmm/intel/vmx_msr.c (revision 364340) @@ -1,482 +1,521 @@ /*- * 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 #include #include #include #include #include #include #include #include #include "vmx.h" #include "vmx_msr.h" +#include "x86.h" static bool vmx_ctl_allows_one_setting(uint64_t msr_val, int bitpos) { return ((msr_val & (1UL << (bitpos + 32))) != 0); } static bool vmx_ctl_allows_zero_setting(uint64_t msr_val, int bitpos) { return ((msr_val & (1UL << bitpos)) == 0); } uint32_t vmx_revision(void) { return (rdmsr(MSR_VMX_BASIC) & 0xffffffff); } /* * Generate a bitmask to be used for the VMCS execution control fields. * * The caller specifies what bits should be set to one in 'ones_mask' * and what bits should be set to zero in 'zeros_mask'. The don't-care * bits are set to the default value. The default values are obtained * based on "Algorithm 3" in Section 27.5.1 "Algorithms for Determining * VMX Capabilities". * * Returns zero on success and non-zero on error. */ int vmx_set_ctlreg(int ctl_reg, int true_ctl_reg, uint32_t ones_mask, uint32_t zeros_mask, uint32_t *retval) { int i; uint64_t val, trueval; bool true_ctls_avail, one_allowed, zero_allowed; /* We cannot ask the same bit to be set to both '1' and '0' */ if ((ones_mask ^ zeros_mask) != (ones_mask | zeros_mask)) return (EINVAL); true_ctls_avail = (rdmsr(MSR_VMX_BASIC) & (1UL << 55)) != 0; val = rdmsr(ctl_reg); if (true_ctls_avail) trueval = rdmsr(true_ctl_reg); /* step c */ else trueval = val; /* step a */ for (i = 0; i < 32; i++) { one_allowed = vmx_ctl_allows_one_setting(trueval, i); zero_allowed = vmx_ctl_allows_zero_setting(trueval, i); KASSERT(one_allowed || zero_allowed, ("invalid zero/one setting for bit %d of ctl 0x%0x, " "truectl 0x%0x\n", i, ctl_reg, true_ctl_reg)); if (zero_allowed && !one_allowed) { /* b(i),c(i) */ if (ones_mask & (1 << i)) return (EINVAL); *retval &= ~(1 << i); } else if (one_allowed && !zero_allowed) { /* b(i),c(i) */ if (zeros_mask & (1 << i)) return (EINVAL); *retval |= 1 << i; } else { if (zeros_mask & (1 << i)) /* b(ii),c(ii) */ *retval &= ~(1 << i); else if (ones_mask & (1 << i)) /* b(ii), c(ii) */ *retval |= 1 << i; else if (!true_ctls_avail) *retval &= ~(1 << i); /* b(iii) */ else if (vmx_ctl_allows_zero_setting(val, i))/* c(iii)*/ *retval &= ~(1 << i); else if (vmx_ctl_allows_one_setting(val, i)) /* c(iv) */ *retval |= 1 << i; else { panic("vmx_set_ctlreg: unable to determine " "correct value of ctl bit %d for msr " "0x%0x and true msr 0x%0x", i, ctl_reg, true_ctl_reg); } } } return (0); } void msr_bitmap_initialize(char *bitmap) { memset(bitmap, 0xff, PAGE_SIZE); } int msr_bitmap_change_access(char *bitmap, u_int msr, int access) { int byte, bit; if (msr <= 0x00001FFF) byte = msr / 8; else if (msr >= 0xC0000000 && msr <= 0xC0001FFF) byte = 1024 + (msr - 0xC0000000) / 8; else return (EINVAL); bit = msr & 0x7; if (access & MSR_BITMAP_ACCESS_READ) bitmap[byte] &= ~(1 << bit); else bitmap[byte] |= 1 << bit; byte += 2048; if (access & MSR_BITMAP_ACCESS_WRITE) bitmap[byte] &= ~(1 << bit); else bitmap[byte] |= 1 << bit; return (0); } static uint64_t misc_enable; static uint64_t platform_info; static uint64_t turbo_ratio_limit; static uint64_t host_msrs[GUEST_MSR_NUM]; static bool nehalem_cpu(void) { u_int family, model; /* * The family:model numbers belonging to the Nehalem microarchitecture * are documented in Section 35.5, Intel SDM dated Feb 2014. */ family = CPUID_TO_FAMILY(cpu_id); model = CPUID_TO_MODEL(cpu_id); if (family == 0x6) { switch (model) { case 0x1A: case 0x1E: case 0x1F: case 0x2E: return (true); default: break; } } return (false); } static bool westmere_cpu(void) { u_int family, model; /* * The family:model numbers belonging to the Westmere microarchitecture * are documented in Section 35.6, Intel SDM dated Feb 2014. */ family = CPUID_TO_FAMILY(cpu_id); model = CPUID_TO_MODEL(cpu_id); if (family == 0x6) { switch (model) { case 0x25: case 0x2C: return (true); default: break; } } return (false); } static bool pat_valid(uint64_t val) { int i, pa; /* * From Intel SDM: Table "Memory Types That Can Be Encoded With PAT" * * Extract PA0 through PA7 and validate that each one encodes a * valid memory type. */ for (i = 0; i < 8; i++) { pa = (val >> (i * 8)) & 0xff; if (pa == 2 || pa == 3 || pa >= 8) return (false); } return (true); } void vmx_msr_init(void) { uint64_t bus_freq, ratio; int i; /* * It is safe to cache the values of the following MSRs because * they don't change based on curcpu, curproc or curthread. */ host_msrs[IDX_MSR_LSTAR] = rdmsr(MSR_LSTAR); host_msrs[IDX_MSR_CSTAR] = rdmsr(MSR_CSTAR); host_msrs[IDX_MSR_STAR] = rdmsr(MSR_STAR); host_msrs[IDX_MSR_SF_MASK] = rdmsr(MSR_SF_MASK); /* * Initialize emulated MSRs */ misc_enable = rdmsr(MSR_IA32_MISC_ENABLE); /* * Set mandatory bits * 11: branch trace disabled * 12: PEBS unavailable * Clear unsupported features * 16: SpeedStep enable * 18: enable MONITOR FSM */ misc_enable |= (1 << 12) | (1 << 11); misc_enable &= ~((1 << 18) | (1 << 16)); if (nehalem_cpu() || westmere_cpu()) bus_freq = 133330000; /* 133Mhz */ else bus_freq = 100000000; /* 100Mhz */ /* * XXXtime * The ratio should really be based on the virtual TSC frequency as * opposed to the host TSC. */ ratio = (tsc_freq / bus_freq) & 0xff; /* * The register definition is based on the micro-architecture * but the following bits are always the same: * [15:8] Maximum Non-Turbo Ratio * [28] Programmable Ratio Limit for Turbo Mode * [29] Programmable TDC-TDP Limit for Turbo Mode * [47:40] Maximum Efficiency Ratio * * The other bits can be safely set to 0 on all * micro-architectures up to Haswell. */ platform_info = (ratio << 8) | (ratio << 40); /* * The number of valid bits in the MSR_TURBO_RATIO_LIMITx register is * dependent on the maximum cores per package supported by the micro- * architecture. For e.g., Westmere supports 6 cores per package and * uses the low 48 bits. Sandybridge support 8 cores per package and * uses up all 64 bits. * * However, the unused bits are reserved so we pretend that all bits * in this MSR are valid. */ for (i = 0; i < 8; i++) turbo_ratio_limit = (turbo_ratio_limit << 8) | ratio; } void vmx_msr_guest_init(struct vmx *vmx, int vcpuid) { uint64_t *guest_msrs; guest_msrs = vmx->guest_msrs[vcpuid]; /* * The permissions bitmap is shared between all vcpus so initialize it * once when initializing the vBSP. */ if (vcpuid == 0) { guest_msr_rw(vmx, MSR_LSTAR); guest_msr_rw(vmx, MSR_CSTAR); guest_msr_rw(vmx, MSR_STAR); guest_msr_rw(vmx, MSR_SF_MASK); guest_msr_rw(vmx, MSR_KGSBASE); } /* * Initialize guest IA32_PAT MSR with default value after reset. */ guest_msrs[IDX_MSR_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); return; } void vmx_msr_guest_enter(struct vmx *vmx, int vcpuid) { uint64_t *guest_msrs = vmx->guest_msrs[vcpuid]; /* Save host MSRs (in particular, KGSBASE) and restore guest MSRs */ update_pcb_bases(curpcb); wrmsr(MSR_LSTAR, guest_msrs[IDX_MSR_LSTAR]); wrmsr(MSR_CSTAR, guest_msrs[IDX_MSR_CSTAR]); wrmsr(MSR_STAR, guest_msrs[IDX_MSR_STAR]); wrmsr(MSR_SF_MASK, guest_msrs[IDX_MSR_SF_MASK]); wrmsr(MSR_KGSBASE, guest_msrs[IDX_MSR_KGSBASE]); } void +vmx_msr_guest_enter_tsc_aux(struct vmx *vmx, int vcpuid) +{ + uint64_t guest_tsc_aux = vmx->guest_msrs[vcpuid][IDX_MSR_TSC_AUX]; + uint32_t cpuid = PCPU_GET(cpuid); + + if (vmx_have_msr_tsc_aux(vmx) && (guest_tsc_aux != cpuid)) + wrmsr(MSR_TSC_AUX, guest_tsc_aux); +} + +void vmx_msr_guest_exit(struct vmx *vmx, int vcpuid) { uint64_t *guest_msrs = vmx->guest_msrs[vcpuid]; /* Save guest MSRs */ guest_msrs[IDX_MSR_LSTAR] = rdmsr(MSR_LSTAR); guest_msrs[IDX_MSR_CSTAR] = rdmsr(MSR_CSTAR); guest_msrs[IDX_MSR_STAR] = rdmsr(MSR_STAR); guest_msrs[IDX_MSR_SF_MASK] = rdmsr(MSR_SF_MASK); guest_msrs[IDX_MSR_KGSBASE] = rdmsr(MSR_KGSBASE); /* Restore host MSRs */ wrmsr(MSR_LSTAR, host_msrs[IDX_MSR_LSTAR]); wrmsr(MSR_CSTAR, host_msrs[IDX_MSR_CSTAR]); wrmsr(MSR_STAR, host_msrs[IDX_MSR_STAR]); wrmsr(MSR_SF_MASK, host_msrs[IDX_MSR_SF_MASK]); /* MSR_KGSBASE will be restored on the way back to userspace */ } +void +vmx_msr_guest_exit_tsc_aux(struct vmx *vmx, int vcpuid) +{ + uint64_t guest_tsc_aux = vmx->guest_msrs[vcpuid][IDX_MSR_TSC_AUX]; + uint32_t cpuid = PCPU_GET(cpuid); + + if (vmx_have_msr_tsc_aux(vmx) && (guest_tsc_aux != cpuid)) + /* + * Note that it is not necessary to save the guest value + * here; vmx->guest_msrs[vcpuid][IDX_MSR_TSC_AUX] always + * contains the current value since it is updated whenever + * the guest writes to it (which is expected to be very + * rare). + */ + wrmsr(MSR_TSC_AUX, cpuid); +} + int vmx_rdmsr(struct vmx *vmx, int vcpuid, u_int num, uint64_t *val, bool *retu) { const uint64_t *guest_msrs; int error; guest_msrs = vmx->guest_msrs[vcpuid]; error = 0; switch (num) { case MSR_MCG_CAP: case MSR_MCG_STATUS: *val = 0; break; case MSR_MTRRcap: case MSR_MTRRdefType: case MSR_MTRR4kBase ... MSR_MTRR4kBase + 8: case MSR_MTRR16kBase ... MSR_MTRR16kBase + 1: case MSR_MTRR64kBase: *val = 0; break; case MSR_IA32_MISC_ENABLE: *val = misc_enable; break; case MSR_PLATFORM_INFO: *val = platform_info; break; case MSR_TURBO_RATIO_LIMIT: case MSR_TURBO_RATIO_LIMIT1: *val = turbo_ratio_limit; break; case MSR_PAT: *val = guest_msrs[IDX_MSR_PAT]; break; default: error = EINVAL; break; } return (error); } int vmx_wrmsr(struct vmx *vmx, int vcpuid, u_int num, uint64_t val, bool *retu) { uint64_t *guest_msrs; uint64_t changed; int error; guest_msrs = vmx->guest_msrs[vcpuid]; error = 0; switch (num) { case MSR_MCG_CAP: case MSR_MCG_STATUS: break; /* ignore writes */ case MSR_MTRRcap: vm_inject_gp(vmx->vm, vcpuid); break; case MSR_MTRRdefType: case MSR_MTRR4kBase ... MSR_MTRR4kBase + 8: case MSR_MTRR16kBase ... MSR_MTRR16kBase + 1: case MSR_MTRR64kBase: break; /* Ignore writes */ case MSR_IA32_MISC_ENABLE: changed = val ^ misc_enable; /* * If the host has disabled the NX feature then the guest * also cannot use it. However, a Linux guest will try to * enable the NX feature by writing to the MISC_ENABLE MSR. * * This can be safely ignored because the memory management * code looks at CPUID.80000001H:EDX.NX to check if the * functionality is actually enabled. */ changed &= ~(1UL << 34); /* * Punt to userspace if any other bits are being modified. */ if (changed) error = EINVAL; break; case MSR_PAT: if (pat_valid(val)) guest_msrs[IDX_MSR_PAT] = val; else vm_inject_gp(vmx->vm, vcpuid); break; case MSR_TSC: error = vmx_set_tsc_offset(vmx, vcpuid, val - rdtsc()); + break; + case MSR_TSC_AUX: + if (vmx_have_msr_tsc_aux(vmx)) + /* + * vmx_msr_guest_enter_tsc_aux() will apply this + * value when it is called immediately before guest + * entry. + */ + guest_msrs[IDX_MSR_TSC_AUX] = val; + else + vm_inject_gp(vmx->vm, vcpuid); break; default: error = EINVAL; break; } return (error); } Index: head/sys/amd64/vmm/intel/vmx_msr.h =================================================================== --- head/sys/amd64/vmm/intel/vmx_msr.h (revision 364339) +++ head/sys/amd64/vmm/intel/vmx_msr.h (revision 364340) @@ -1,72 +1,74 @@ /*- * 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 _VMX_MSR_H_ #define _VMX_MSR_H_ struct vmx; void vmx_msr_init(void); void vmx_msr_guest_init(struct vmx *vmx, int vcpuid); +void vmx_msr_guest_enter_tsc_aux(struct vmx *vmx, int vcpuid); void vmx_msr_guest_enter(struct vmx *vmx, int vcpuid); void vmx_msr_guest_exit(struct vmx *vmx, int vcpuid); +void vmx_msr_guest_exit_tsc_aux(struct vmx *vmx, int vcpuid); int vmx_rdmsr(struct vmx *, int vcpuid, u_int num, uint64_t *val, bool *retu); int vmx_wrmsr(struct vmx *, int vcpuid, u_int num, uint64_t val, bool *retu); uint32_t vmx_revision(void); int vmx_set_ctlreg(int ctl_reg, int true_ctl_reg, uint32_t ones_mask, uint32_t zeros_mask, uint32_t *retval); /* * According to Section 21.10.4 "Software Access to Related Structures", * changes to data structures pointed to by the VMCS must be made only when * there is no logical processor with a current VMCS that points to the * data structure. * * This pretty much limits us to configuring the MSR bitmap before VMCS * initialization for SMP VMs. Unless of course we do it the hard way - which * would involve some form of synchronization between the vcpus to vmclear * all VMCSs' that point to the bitmap. */ #define MSR_BITMAP_ACCESS_NONE 0x0 #define MSR_BITMAP_ACCESS_READ 0x1 #define MSR_BITMAP_ACCESS_WRITE 0x2 #define MSR_BITMAP_ACCESS_RW (MSR_BITMAP_ACCESS_READ|MSR_BITMAP_ACCESS_WRITE) void msr_bitmap_initialize(char *bitmap); int msr_bitmap_change_access(char *bitmap, u_int msr, int access); #define guest_msr_rw(vmx, msr) \ msr_bitmap_change_access((vmx)->msr_bitmap, (msr), MSR_BITMAP_ACCESS_RW) #define guest_msr_ro(vmx, msr) \ msr_bitmap_change_access((vmx)->msr_bitmap, (msr), MSR_BITMAP_ACCESS_READ) #endif Index: head/sys/amd64/vmm/x86.c =================================================================== --- head/sys/amd64/vmm/x86.c (revision 364339) +++ head/sys/amd64/vmm/x86.c (revision 364340) @@ -1,631 +1,640 @@ /*- * 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 #include #include #include #include #include #include #include #include #include #include "vmm_host.h" #include "vmm_ktr.h" #include "vmm_util.h" #include "x86.h" SYSCTL_DECL(_hw_vmm); static SYSCTL_NODE(_hw_vmm, OID_AUTO, topology, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, NULL); #define CPUID_VM_HIGH 0x40000000 static const char bhyve_id[12] = "bhyve bhyve "; static uint64_t bhyve_xcpuids; SYSCTL_ULONG(_hw_vmm, OID_AUTO, bhyve_xcpuids, CTLFLAG_RW, &bhyve_xcpuids, 0, "Number of times an unknown cpuid leaf was accessed"); #if __FreeBSD_version < 1200060 /* Remove after 11 EOL helps MFCing */ extern u_int threads_per_core; SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, threads_per_core, CTLFLAG_RDTUN, &threads_per_core, 0, NULL); extern u_int cores_per_package; SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, cores_per_package, CTLFLAG_RDTUN, &cores_per_package, 0, NULL); #endif static int cpuid_leaf_b = 1; SYSCTL_INT(_hw_vmm_topology, OID_AUTO, cpuid_leaf_b, CTLFLAG_RDTUN, &cpuid_leaf_b, 0, NULL); /* * Round up to the next power of two, if necessary, and then take log2. * Returns -1 if argument is zero. */ static __inline int log2(u_int x) { return (fls(x << (1 - powerof2(x))) - 1); } int x86_emulate_cpuid(struct vm *vm, int vcpu_id, uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx) { const struct xsave_limits *limits; uint64_t cr4; - int error, enable_invpcid, level, width, x2apic_id; + int error, enable_invpcid, enable_rdpid, enable_rdtscp, level, + width, x2apic_id; unsigned int func, regs[4], logical_cpus; enum x2apic_state x2apic_state; uint16_t cores, maxcpus, sockets, threads; VCPU_CTR2(vm, vcpu_id, "cpuid %#x,%#x", *eax, *ecx); /* * Requests for invalid CPUID levels should map to the highest * available level instead. */ if (cpu_exthigh != 0 && *eax >= 0x80000000) { if (*eax > cpu_exthigh) *eax = cpu_exthigh; } else if (*eax >= 0x40000000) { if (*eax > CPUID_VM_HIGH) *eax = CPUID_VM_HIGH; } else if (*eax > cpu_high) { *eax = cpu_high; } func = *eax; /* * In general the approach used for CPU topology is to * advertise a flat topology where all CPUs are packages with * no multi-core or SMT. */ switch (func) { /* * Pass these through to the guest */ case CPUID_0000_0000: case CPUID_0000_0002: case CPUID_0000_0003: case CPUID_8000_0000: case CPUID_8000_0002: case CPUID_8000_0003: case CPUID_8000_0004: case CPUID_8000_0006: cpuid_count(*eax, *ecx, regs); break; case CPUID_8000_0008: cpuid_count(*eax, *ecx, regs); if (vmm_is_svm()) { /* * As on Intel (0000_0007:0, EDX), mask out * unsupported or unsafe AMD extended features * (8000_0008 EBX). */ regs[1] &= (AMDFEID_CLZERO | AMDFEID_IRPERF | AMDFEID_XSAVEERPTR); vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); /* * Here, width is ApicIdCoreIdSize, present on * at least Family 15h and newer. It * represents the "number of bits in the * initial apicid that indicate thread id * within a package." * * Our topo_probe_amd() uses it for * pkg_id_shift and other OSes may rely on it. */ width = MIN(0xF, log2(threads * cores)); if (width < 0x4) width = 0; logical_cpus = MIN(0xFF, threads * cores - 1); regs[2] = (width << AMDID_COREID_SIZE_SHIFT) | logical_cpus; } break; case CPUID_8000_0001: cpuid_count(*eax, *ecx, regs); /* * Hide SVM from guest. */ regs[2] &= ~AMDID2_SVM; /* * Don't advertise extended performance counter MSRs * to the guest. */ regs[2] &= ~AMDID2_PCXC; regs[2] &= ~AMDID2_PNXC; regs[2] &= ~AMDID2_PTSCEL2I; /* * Don't advertise Instruction Based Sampling feature. */ regs[2] &= ~AMDID2_IBS; /* NodeID MSR not available */ regs[2] &= ~AMDID2_NODE_ID; /* Don't advertise the OS visible workaround feature */ regs[2] &= ~AMDID2_OSVW; /* Hide mwaitx/monitorx capability from the guest */ regs[2] &= ~AMDID2_MWAITX; - /* - * Hide rdtscp/ia32_tsc_aux until we know how - * to deal with them. - */ - regs[3] &= ~AMDID_RDTSCP; + /* Advertise RDTSCP if it is enabled. */ + error = vm_get_capability(vm, vcpu_id, + VM_CAP_RDTSCP, &enable_rdtscp); + if (error == 0 && enable_rdtscp) + regs[3] |= AMDID_RDTSCP; + else + regs[3] &= ~AMDID_RDTSCP; break; case CPUID_8000_0007: /* * AMD uses this leaf to advertise the processor's * power monitoring and RAS capabilities. These * features are hardware-specific and exposing * them to a guest doesn't make a lot of sense. * * Intel uses this leaf only to advertise the * "Invariant TSC" feature with all other bits * being reserved (set to zero). */ regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; /* * "Invariant TSC" can be advertised to the guest if: * - host TSC frequency is invariant * - host TSCs are synchronized across physical cpus * * XXX This still falls short because the vcpu * can observe the TSC moving backwards as it * migrates across physical cpus. But at least * it should discourage the guest from using the * TSC to keep track of time. */ if (tsc_is_invariant && smp_tsc) regs[3] |= AMDPM_TSC_INVARIANT; break; case CPUID_8000_001D: /* AMD Cache topology, like 0000_0004 for Intel. */ if (!vmm_is_svm()) goto default_leaf; /* * Similar to Intel, generate a ficticious cache * topology for the guest with L3 shared by the * package, and L1 and L2 local to a core. */ vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); switch (*ecx) { case 0: logical_cpus = threads; level = 1; func = 1; /* data cache */ break; case 1: logical_cpus = threads; level = 2; func = 3; /* unified cache */ break; case 2: logical_cpus = threads * cores; level = 3; func = 3; /* unified cache */ break; default: logical_cpus = 0; level = 0; func = 0; break; } logical_cpus = MIN(0xfff, logical_cpus - 1); regs[0] = (logical_cpus << 14) | (1 << 8) | (level << 5) | func; regs[1] = (func > 0) ? (CACHE_LINE_SIZE - 1) : 0; regs[2] = 0; regs[3] = 0; break; case CPUID_8000_001E: /* * AMD Family 16h+ and Hygon Family 18h additional * identifiers. */ if (!vmm_is_svm() || CPUID_TO_FAMILY(cpu_id) < 0x16) goto default_leaf; vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); regs[0] = vcpu_id; threads = MIN(0xFF, threads - 1); regs[1] = (threads << 8) | (vcpu_id >> log2(threads + 1)); /* * XXX Bhyve topology cannot yet represent >1 node per * processor. */ regs[2] = 0; regs[3] = 0; break; case CPUID_0000_0001: do_cpuid(1, regs); error = vm_get_x2apic_state(vm, vcpu_id, &x2apic_state); if (error) { panic("x86_emulate_cpuid: error %d " "fetching x2apic state", error); } /* * Override the APIC ID only in ebx */ regs[1] &= ~(CPUID_LOCAL_APIC_ID); regs[1] |= (vcpu_id << CPUID_0000_0001_APICID_SHIFT); /* * Don't expose VMX, SpeedStep, TME or SMX capability. * Advertise x2APIC capability and Hypervisor guest. */ regs[2] &= ~(CPUID2_VMX | CPUID2_EST | CPUID2_TM2); regs[2] &= ~(CPUID2_SMX); regs[2] |= CPUID2_HV; if (x2apic_state != X2APIC_DISABLED) regs[2] |= CPUID2_X2APIC; else regs[2] &= ~CPUID2_X2APIC; /* * Only advertise CPUID2_XSAVE in the guest if * the host is using XSAVE. */ if (!(regs[2] & CPUID2_OSXSAVE)) regs[2] &= ~CPUID2_XSAVE; /* * If CPUID2_XSAVE is being advertised and the * guest has set CR4_XSAVE, set * CPUID2_OSXSAVE. */ regs[2] &= ~CPUID2_OSXSAVE; if (regs[2] & CPUID2_XSAVE) { error = vm_get_register(vm, vcpu_id, VM_REG_GUEST_CR4, &cr4); if (error) panic("x86_emulate_cpuid: error %d " "fetching %%cr4", error); if (cr4 & CR4_XSAVE) regs[2] |= CPUID2_OSXSAVE; } /* * Hide monitor/mwait until we know how to deal with * these instructions. */ regs[2] &= ~CPUID2_MON; /* * Hide the performance and debug features. */ regs[2] &= ~CPUID2_PDCM; /* * No TSC deadline support in the APIC yet */ regs[2] &= ~CPUID2_TSCDLT; /* * Hide thermal monitoring */ regs[3] &= ~(CPUID_ACPI | CPUID_TM); /* * Hide the debug store capability. */ regs[3] &= ~CPUID_DS; /* * Advertise the Machine Check and MTRR capability. * * Some guest OSes (e.g. Windows) will not boot if * these features are absent. */ regs[3] |= (CPUID_MCA | CPUID_MCE | CPUID_MTRR); vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); logical_cpus = threads * cores; regs[1] &= ~CPUID_HTT_CORES; regs[1] |= (logical_cpus & 0xff) << 16; regs[3] |= CPUID_HTT; break; case CPUID_0000_0004: cpuid_count(*eax, *ecx, regs); if (regs[0] || regs[1] || regs[2] || regs[3]) { vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); regs[0] &= 0x3ff; regs[0] |= (cores - 1) << 26; /* * Cache topology: * - L1 and L2 are shared only by the logical * processors in a single core. * - L3 and above are shared by all logical * processors in the package. */ logical_cpus = threads; level = (regs[0] >> 5) & 0x7; if (level >= 3) logical_cpus *= cores; regs[0] |= (logical_cpus - 1) << 14; } break; case CPUID_0000_0007: regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; /* leaf 0 */ if (*ecx == 0) { cpuid_count(*eax, *ecx, regs); /* Only leaf 0 is supported */ regs[0] = 0; /* * Expose known-safe features. */ regs[1] &= (CPUID_STDEXT_FSGSBASE | CPUID_STDEXT_BMI1 | CPUID_STDEXT_HLE | CPUID_STDEXT_AVX2 | CPUID_STDEXT_BMI2 | CPUID_STDEXT_ERMS | CPUID_STDEXT_RTM | CPUID_STDEXT_AVX512F | CPUID_STDEXT_RDSEED | CPUID_STDEXT_AVX512PF | CPUID_STDEXT_AVX512ER | CPUID_STDEXT_AVX512CD | CPUID_STDEXT_SHA); regs[2] = 0; regs[3] &= CPUID_STDEXT3_MD_CLEAR; + + /* Advertise RDPID if it is enabled. */ + error = vm_get_capability(vm, vcpu_id, + VM_CAP_RDPID, &enable_rdpid); + if (error == 0 && enable_rdpid) + regs[2] |= CPUID_STDEXT2_RDPID; /* Advertise INVPCID if it is enabled. */ error = vm_get_capability(vm, vcpu_id, VM_CAP_ENABLE_INVPCID, &enable_invpcid); if (error == 0 && enable_invpcid) regs[1] |= CPUID_STDEXT_INVPCID; } break; case CPUID_0000_0006: regs[0] = CPUTPM1_ARAT; regs[1] = 0; regs[2] = 0; regs[3] = 0; break; case CPUID_0000_000A: /* * Handle the access, but report 0 for * all options */ regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; break; case CPUID_0000_000B: /* * Intel processor topology enumeration */ if (vmm_is_intel()) { vm_get_topology(vm, &sockets, &cores, &threads, &maxcpus); if (*ecx == 0) { logical_cpus = threads; width = log2(logical_cpus); level = CPUID_TYPE_SMT; x2apic_id = vcpu_id; } if (*ecx == 1) { logical_cpus = threads * cores; width = log2(logical_cpus); level = CPUID_TYPE_CORE; x2apic_id = vcpu_id; } if (!cpuid_leaf_b || *ecx >= 2) { width = 0; logical_cpus = 0; level = 0; x2apic_id = 0; } regs[0] = width & 0x1f; regs[1] = logical_cpus & 0xffff; regs[2] = (level << 8) | (*ecx & 0xff); regs[3] = x2apic_id; } else { regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; } break; case CPUID_0000_000D: limits = vmm_get_xsave_limits(); if (!limits->xsave_enabled) { regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; break; } cpuid_count(*eax, *ecx, regs); switch (*ecx) { case 0: /* * Only permit the guest to use bits * that are active in the host in * %xcr0. Also, claim that the * maximum save area size is * equivalent to the host's current * save area size. Since this runs * "inside" of vmrun(), it runs with * the guest's xcr0, so the current * save area size is correct as-is. */ regs[0] &= limits->xcr0_allowed; regs[2] = limits->xsave_max_size; regs[3] &= (limits->xcr0_allowed >> 32); break; case 1: /* Only permit XSAVEOPT. */ regs[0] &= CPUID_EXTSTATE_XSAVEOPT; regs[1] = 0; regs[2] = 0; regs[3] = 0; break; default: /* * If the leaf is for a permitted feature, * pass through as-is, otherwise return * all zeroes. */ if (!(limits->xcr0_allowed & (1ul << *ecx))) { regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; } break; } break; case CPUID_0000_0015: /* * Don't report CPU TSC/Crystal ratio and clock * values since guests may use these to derive the * local APIC frequency.. */ regs[0] = 0; regs[1] = 0; regs[2] = 0; regs[3] = 0; break; case 0x40000000: regs[0] = CPUID_VM_HIGH; bcopy(bhyve_id, ®s[1], 4); bcopy(bhyve_id + 4, ®s[2], 4); bcopy(bhyve_id + 8, ®s[3], 4); break; default: default_leaf: /* * The leaf value has already been clamped so * simply pass this through, keeping count of * how many unhandled leaf values have been seen. */ atomic_add_long(&bhyve_xcpuids, 1); cpuid_count(*eax, *ecx, regs); break; } *eax = regs[0]; *ebx = regs[1]; *ecx = regs[2]; *edx = regs[3]; return (1); } bool vm_cpuid_capability(struct vm *vm, int vcpuid, enum vm_cpuid_capability cap) { bool rv; KASSERT(cap > 0 && cap < VCC_LAST, ("%s: invalid vm_cpu_capability %d", __func__, cap)); /* * Simply passthrough the capabilities of the host cpu for now. */ rv = false; switch (cap) { case VCC_NO_EXECUTE: if (amd_feature & AMDID_NX) rv = true; break; case VCC_FFXSR: if (amd_feature & AMDID_FFXSR) rv = true; break; case VCC_TCE: if (amd_feature2 & AMDID2_TCE) rv = true; break; default: panic("%s: unknown vm_cpu_capability %d", __func__, cap); } return (rv); }