/*
* This function is called to ensure that a vcpu "sees" a pending event
* as soon as possible:
* - If the vcpu thread is sleeping then it is woken up.
* - If the vcpu is running on a different host_cpu then an IPI will be directed
* to the host_cpu to cause the vcpu to trap into the hypervisor.
*/
void
vcpu_notify_event(struct vm *vm, int vcpuid, UNUSED bool lapic_intr)
{
struct vcpu *vcpu;
vcpu = &vm->vcpu[vcpuid];
vcpu_lock(vcpu);
if (vcpu->state == VCPU_RUNNING) {
VCPU_INTERRUPT(vcpuid);
/* FIXME */
// if (hostcpu != curcpu) {
// if (lapic_intr) {
// vlapic_post_intr(vcpu->vlapic, hostcpu,
// vmm_ipinum);
// } else {
// ipi_cpu(hostcpu, vmm_ipinum);
// }
// } else {
// /*
// * If the 'vcpu' is running on 'curcpu' then it must
// * be sending a notification to itself (e.g. SELF_IPI).
// * The pending event will be picked up when the vcpu
// * transitions back to guest context.
// */
// }
} else {
if (vcpu->state == VCPU_SLEEPING)
pthread_cond_signal(&vcpu->vcpu_sleep_cnd);
//wakeup_one(vcpu);
}
vcpu_unlock(vcpu);
}
/*
* This function is called to ensure that a vcpu "sees" a pending event
* as soon as possible:
* - If the vcpu thread is sleeping then it is woken up.
* - If the vcpu is running on a different host_cpu then an IPI will be directed
* to the host_cpu to cause the vcpu to trap into the hypervisor.
*/
void
vcpu_notify_event(struct vm *vm, int vcpuid, bool lapic_intr)
{
int hostcpu;
struct vcpu *vcpu;
vcpu = &vm->vcpu[vcpuid];
vcpu_lock(vcpu);
hostcpu = vcpu->hostcpu;
if (vcpu->state == VCPU_RUNNING) {
KASSERT(hostcpu != NOCPU, ("vcpu running on invalid hostcpu"));
if (hostcpu != curcpu) {
if (lapic_intr) {
vlapic_post_intr(vcpu->vlapic, hostcpu,
vmm_ipinum);
} else {
ipi_cpu(hostcpu, vmm_ipinum);
}
} else {
/*
* If the 'vcpu' is running on 'curcpu' then it must
* be sending a notification to itself (e.g. SELF_IPI).
* The pending event will be picked up when the vcpu
* transitions back to guest context.
*/
}
} else {
KASSERT(hostcpu == NOCPU, ("vcpu state %d not consistent "
"with hostcpu %d", vcpu->state, hostcpu));
if (vcpu->state == VCPU_SLEEPING)
wakeup_one(vcpu);
}
vcpu_unlock(vcpu);
}
void
vcpu_notify_event(struct vm *vm, int vcpuid, bool lapic_intr)
{
struct vcpu *vcpu = &vm->vcpu[vcpuid];
vcpu_lock(vcpu);
vcpu_notify_event_locked(vcpu, vcpuid, lapic_intr);
vcpu_unlock(vcpu);
}
/*
* Emulate a guest 'hlt' by sleeping until the vcpu is ready to run.
*/
static int
vm_handle_hlt(struct vm *vm, int vcpuid, bool intr_disabled, bool *retu)
{
struct vm_exit *vmexit;
struct vcpu *vcpu;
int t, timo, spindown;
vcpu = &vm->vcpu[vcpuid];
spindown = 0;
vcpu_lock(vcpu);
/*
* Do a final check for pending NMI or interrupts before
* really putting this thread to sleep.
*
* These interrupts could have happened any time after we
* returned from VMRUN() and before we grabbed the vcpu lock.
*/
if (!vm_nmi_pending(vm, vcpuid) &&
(intr_disabled || !vlapic_pending_intr(vcpu->vlapic, NULL))) {
t = ticks;
vcpu_require_state_locked(vcpu, VCPU_SLEEPING);
if (vlapic_enabled(vcpu->vlapic)) {
/*
* XXX msleep_spin() is not interruptible so use the
* 'timo' to put an upper bound on the sleep time.
*/
timo = hz;
msleep_spin(vcpu, &vcpu->mtx, "vmidle", timo);
} else {
/*
* Spindown the vcpu if the apic is disabled and it
* had entered the halted state.
*/
spindown = 1;
}
vcpu_require_state_locked(vcpu, VCPU_FROZEN);
vmm_stat_incr(vm, vcpuid, VCPU_IDLE_TICKS, ticks - t);
}
vcpu_unlock(vcpu);
/*
* Since 'vm_deactivate_cpu()' grabs a sleep mutex we must call it
* outside the confines of the vcpu spinlock.
*/
if (spindown) {
*retu = true;
vmexit = vm_exitinfo(vm, vcpuid);
vmexit->exitcode = VM_EXITCODE_SPINDOWN_CPU;
vm_deactivate_cpu(vm, vcpuid);
VCPU_CTR0(vm, vcpuid, "spinning down cpu");
}
return (0);
}
enum vcpu_state
vcpu_get_state(struct vm *vm, int vcpuid)
{
struct vcpu *vcpu;
enum vcpu_state state;
if (vcpuid < 0 || vcpuid >= VM_MAXCPU)
xhyve_abort("vm_get_run_state: invalid vcpuid %d\n", vcpuid);
vcpu = &vm->vcpu[vcpuid];
vcpu_lock(vcpu);
state = vcpu->state;
vcpu_unlock(vcpu);
return (state);
}
int
vcpu_set_state(struct vm *vm, int vcpuid, enum vcpu_state newstate,
bool from_idle)
{
int error;
struct vcpu *vcpu;
if (vcpuid < 0 || vcpuid >= VM_MAXCPU)
xhyve_abort("vm_set_run_state: invalid vcpuid %d\n", vcpuid);
vcpu = &vm->vcpu[vcpuid];
vcpu_lock(vcpu);
error = vcpu_set_state_locked(vcpu, newstate, from_idle);
vcpu_unlock(vcpu);
return (error);
}
enum vcpu_state
vcpu_get_state(struct vm *vm, int vcpuid, int *hostcpu)
{
struct vcpu *vcpu;
enum vcpu_state state;
if (vcpuid < 0 || vcpuid >= VM_MAXCPU)
panic("vm_get_run_state: invalid vcpuid %d", vcpuid);
vcpu = &vm->vcpu[vcpuid];
vcpu_lock(vcpu);
state = vcpu->state;
if (hostcpu != NULL)
*hostcpu = vcpu->hostcpu;
vcpu_unlock(vcpu);
return (state);
}
void
vm_inject_gp(struct vm *vm, int vcpuid)
{
struct vm_exception gpf = {
.vector = IDT_GP,
.error_code_valid = 1,
.error_code = 0
};
vm_inject_fault(vm, vcpuid, &gpf);
}
void
vm_inject_ud(struct vm *vm, int vcpuid)
{
struct vm_exception udf = {
.vector = IDT_UD,
.error_code_valid = 0
};
vm_inject_fault(vm, vcpuid, &udf);
}
static VMM_STAT(VCPU_NMI_COUNT, "number of NMIs delivered to vcpu");
int
vm_inject_nmi(struct vm *vm, int vcpuid)
{
struct vcpu *vcpu;
if (vcpuid < 0 || vcpuid >= VM_MAXCPU)
return (EINVAL);
vcpu = &vm->vcpu[vcpuid];
vcpu->nmi_pending = 1;
vcpu_notify_event(vm, vcpuid, false);
return (0);
}
int
vm_nmi_pending(struct vm *vm, int vcpuid)
{
struct vcpu *vcpu;
if (vcpuid < 0 || vcpuid >= VM_MAXCPU)
panic("vm_nmi_pending: invalid vcpuid %d", vcpuid);
vcpu = &vm->vcpu[vcpuid];
return (vcpu->nmi_pending);
}
void
vm_nmi_clear(struct vm *vm, int vcpuid)
{
struct vcpu *vcpu;
if (vcpuid < 0 || vcpuid >= VM_MAXCPU)
panic("vm_nmi_pending: invalid vcpuid %d", vcpuid);
vcpu = &vm->vcpu[vcpuid];
if (vcpu->nmi_pending == 0)
panic("vm_nmi_clear: inconsistent nmi_pending state");
vcpu->nmi_pending = 0;
vmm_stat_incr(vm, vcpuid, VCPU_NMI_COUNT, 1);
}
static VMM_STAT(VCPU_EXTINT_COUNT, "number of ExtINTs delivered to vcpu");
int
vm_inject_extint(struct vm *vm, int vcpuid)
{
struct vcpu *vcpu;
if (vcpuid < 0 || vcpuid >= VM_MAXCPU)
return (EINVAL);
vcpu = &vm->vcpu[vcpuid];
vcpu->extint_pending = 1;
vcpu_notify_event(vm, vcpuid, false);
return (0);
}
int
vm_extint_pending(struct vm *vm, int vcpuid)
{
struct vcpu *vcpu;
if (vcpuid < 0 || vcpuid >= VM_MAXCPU)
panic("vm_extint_pending: invalid vcpuid %d", vcpuid);
vcpu = &vm->vcpu[vcpuid];
return (vcpu->extint_pending);
}
void
vm_extint_clear(struct vm *vm, int vcpuid)
{
struct vcpu *vcpu;
if (vcpuid < 0 || vcpuid >= VM_MAXCPU)
panic("vm_extint_pending: invalid vcpuid %d", vcpuid);
vcpu = &vm->vcpu[vcpuid];
if (vcpu->extint_pending == 0)
panic("vm_extint_clear: inconsistent extint_pending state");
vcpu->extint_pending = 0;
vmm_stat_incr(vm, vcpuid, VCPU_EXTINT_COUNT, 1);
}
int
vm_get_capability(struct vm *vm, int vcpu, int type, int *retval)
{
if (vcpu < 0 || vcpu >= VM_MAXCPU)
return (EINVAL);
if (type < 0 || type >= VM_CAP_MAX)
return (EINVAL);
return (VMGETCAP(vm->cookie, vcpu, type, retval));
}
int
vm_set_capability(struct vm *vm, int vcpu, int type, int val)
{
if (vcpu < 0 || vcpu >= VM_MAXCPU)
return (EINVAL);
if (type < 0 || type >= VM_CAP_MAX)
return (EINVAL);
return (VMSETCAP(vm->cookie, vcpu, type, val));
}
uint64_t *
vm_guest_msrs(struct vm *vm, int cpu)
{
return (vm->vcpu[cpu].guest_msrs);
}
struct vlapic *
vm_lapic(struct vm *vm, int cpu)
{
return (vm->vcpu[cpu].vlapic);
}
struct vioapic *
vm_ioapic(struct vm *vm)
{
return (vm->vioapic);
}
struct vhpet *
vm_hpet(struct vm *vm)
{
return (vm->vhpet);
}
boolean_t
vmm_is_pptdev(int bus, int slot, int func)
{
int found, i, n;
int b, s, f;
char *val, *cp, *cp2;
/*
* XXX
* The length of an environment variable is limited to 128 bytes which
* puts an upper limit on the number of passthru devices that may be
* specified using a single environment variable.
*
* Work around this by scanning multiple environment variable
* names instead of a single one - yuck!
*/
const char *names[] = { "pptdevs", "pptdevs2", "pptdevs3", NULL };
/* set pptdevs="1/2/3 4/5/6 7/8/9 10/11/12" */
found = 0;
for (i = 0; names[i] != NULL && !found; i++) {
cp = val = getenv(names[i]);
while (cp != NULL && *cp != '\0') {
if ((cp2 = strchr(cp, ' ')) != NULL)
*cp2 = '\0';
n = sscanf(cp, "%d/%d/%d", &b, &s, &f);
if (n == 3 && bus == b && slot == s && func == f) {
found = 1;
break;
}
if (cp2 != NULL)
*cp2++ = ' ';
cp = cp2;
}
freeenv(val);
}
return (found);
}
void *
vm_iommu_domain(struct vm *vm)
{
return (vm->iommu);
}
int
vcpu_set_state(struct vm *vm, int vcpuid, enum vcpu_state newstate,
bool from_idle)
{
int error;
struct vcpu *vcpu;
if (vcpuid < 0 || vcpuid >= VM_MAXCPU)
panic("vm_set_run_state: invalid vcpuid %d", vcpuid);
vcpu = &vm->vcpu[vcpuid];
vcpu_lock(vcpu);
error = vcpu_set_state_locked(vcpu, newstate, from_idle);
vcpu_unlock(vcpu);
return (error);
}
/*
* Emulate a guest 'hlt' by sleeping until the vcpu is ready to run.
*/
static int
vm_handle_hlt(struct vm *vm, int vcpuid, bool intr_disabled)
{
struct vcpu *vcpu;
const char *wmesg;
int vcpu_halted, vm_halted;
const struct timespec ts = {.tv_sec = 1, .tv_nsec = 0}; /* 1 second */
KASSERT(!CPU_ISSET(((unsigned) vcpuid), &vm->halted_cpus),
("vcpu already halted"));
vcpu = &vm->vcpu[vcpuid];
vcpu_halted = 0;
vm_halted = 0;
vcpu_lock(vcpu);
while (1) {
/*
* Do a final check for pending NMI or interrupts before
* really putting this thread to sleep. Also check for
* software events that would cause this vcpu to wakeup.
*
* These interrupts/events could have happened after the
* vcpu returned from VMRUN() and before it acquired the
* vcpu lock above.
*/
if (vm->rendezvous_func != NULL || vm->suspend)
break;
if (vm_nmi_pending(vm, vcpuid))
break;
if (!intr_disabled) {
if (vm_extint_pending(vm, vcpuid) ||
vlapic_pending_intr(vcpu->vlapic, NULL)) {
break;
}
}
/*
* Some Linux guests implement "halt" by having all vcpus
* execute HLT with interrupts disabled. 'halted_cpus' keeps
* track of the vcpus that have entered this state. When all
* vcpus enter the halted state the virtual machine is halted.
*/
if (intr_disabled) {
wmesg = "vmhalt";
VCPU_CTR0(vm, vcpuid, "Halted");
if (!vcpu_halted && halt_detection_enabled) {
vcpu_halted = 1;
CPU_SET_ATOMIC(((unsigned) vcpuid), &vm->halted_cpus);
}
if (CPU_CMP(&vm->halted_cpus, &vm->active_cpus) == 0) {
vm_halted = 1;
break;
}
} else {
wmesg = "vmidle";
}
//t = ticks;
vcpu_require_state_locked(vcpu, VCPU_SLEEPING);
/*
* XXX msleep_spin() cannot be interrupted by signals so
* wake up periodically to check pending signals.
*/
pthread_mutex_lock(&vcpu->vcpu_sleep_mtx);
vcpu_unlock(vcpu);
pthread_cond_timedwait_relative_np(&vcpu->vcpu_sleep_cnd,
&vcpu->vcpu_sleep_mtx, &ts);
vcpu_lock(vcpu);
pthread_mutex_unlock(&vcpu->vcpu_sleep_mtx);
//msleep_spin(vcpu, &vcpu->mtx, wmesg, hz);
vcpu_require_state_locked(vcpu, VCPU_FROZEN);
//vmm_stat_incr(vm, vcpuid, VCPU_IDLE_TICKS, ticks - t);
}
if (vcpu_halted)
CPU_CLR_ATOMIC(((unsigned) vcpuid), &vm->halted_cpus);
vcpu_unlock(vcpu);
if (vm_halted)
vm_suspend(vm, VM_SUSPEND_HALT);
return (0);
}
static int
vm_handle_inst_emul(struct vm *vm, int vcpuid, bool *retu)
{
struct vie *vie;
struct vcpu *vcpu;
struct vm_exit *vme;
uint64_t gla, gpa, cs_base;
struct vm_guest_paging *paging;
mem_region_read_t mread;
mem_region_write_t mwrite;
enum vm_cpu_mode cpu_mode;
int cs_d, error, fault, length;
vcpu = &vm->vcpu[vcpuid];
vme = &vcpu->exitinfo;
gla = vme->u.inst_emul.gla;
gpa = vme->u.inst_emul.gpa;
cs_base = vme->u.inst_emul.cs_base;
cs_d = vme->u.inst_emul.cs_d;
vie = &vme->u.inst_emul.vie;
paging = &vme->u.inst_emul.paging;
cpu_mode = paging->cpu_mode;
VCPU_CTR1(vm, vcpuid, "inst_emul fault accessing gpa %#llx", gpa);
/* Fetch, decode and emulate the faulting instruction */
if (vie->num_valid == 0) {
/*
* If the instruction length is not known then assume a
* maximum size instruction.
*/
length = vme->inst_length ? vme->inst_length : VIE_INST_SIZE;
error = vmm_fetch_instruction(vm, vcpuid, paging, vme->rip +
cs_base, length, vie, &fault);
} else {
/*
* The instruction bytes have already been copied into 'vie'
*/
error = fault = 0;
}
if (error || fault)
return (error);
if (vmm_decode_instruction(vm, vcpuid, gla, cpu_mode, cs_d, vie) != 0) {
VCPU_CTR1(vm, vcpuid, "Error decoding instruction at %#llx",
vme->rip + cs_base);
*retu = true; /* dump instruction bytes in userspace */
return (0);
}
/*
* If the instruction length was not specified then update it now
* along with 'nextrip'.
*/
if (vme->inst_length == 0) {
vme->inst_length = vie->num_processed;
vcpu->nextrip += vie->num_processed;
}
/* return to userland unless this is an in-kernel emulated device */
if (gpa >= DEFAULT_APIC_BASE && gpa < DEFAULT_APIC_BASE + XHYVE_PAGE_SIZE) {
mread = lapic_mmio_read;
mwrite = lapic_mmio_write;
} else if (gpa >= VIOAPIC_BASE && gpa < VIOAPIC_BASE + VIOAPIC_SIZE) {
mread = vioapic_mmio_read;
mwrite = vioapic_mmio_write;
} else if (gpa >= VHPET_BASE && gpa < VHPET_BASE + VHPET_SIZE) {
mread = vhpet_mmio_read;
mwrite = vhpet_mmio_write;
} else {
*retu = true;
return (0);
}
error = vmm_emulate_instruction(vm, vcpuid, gpa, vie, paging,
mread, mwrite, retu);
return (error);
}
static int
vm_handle_suspend(struct vm *vm, int vcpuid, bool *retu)
{
int i, done;
struct vcpu *vcpu;
const struct timespec ts = {.tv_sec = 1, .tv_nsec = 0}; /* 1 second */
done = 0;
vcpu = &vm->vcpu[vcpuid];
CPU_SET_ATOMIC(((unsigned) vcpuid), &vm->suspended_cpus);
/*
* Wait until all 'active_cpus' have suspended themselves.
*
* Since a VM may be suspended at any time including when one or
* more vcpus are doing a rendezvous we need to call the rendezvous
* handler while we are waiting to prevent a deadlock.
*/
vcpu_lock(vcpu);
while (1) {
if (CPU_CMP(&vm->suspended_cpus, &vm->active_cpus) == 0) {
VCPU_CTR0(vm, vcpuid, "All vcpus suspended");
break;
}
if (vm->rendezvous_func == NULL) {
VCPU_CTR0(vm, vcpuid, "Sleeping during suspend");
vcpu_require_state_locked(vcpu, VCPU_SLEEPING);
pthread_mutex_lock(&vcpu->vcpu_sleep_mtx);
vcpu_unlock(vcpu);
pthread_cond_timedwait_relative_np(&vcpu->vcpu_sleep_cnd,
&vcpu->vcpu_sleep_mtx, &ts);
vcpu_lock(vcpu);
pthread_mutex_unlock(&vcpu->vcpu_sleep_mtx);
//msleep_spin(vcpu, &vcpu->mtx, "vmsusp", hz);
vcpu_require_state_locked(vcpu, VCPU_FROZEN);
} else {
VCPU_CTR0(vm, vcpuid, "Rendezvous during suspend");
vcpu_unlock(vcpu);
vm_handle_rendezvous(vm, vcpuid);
vcpu_lock(vcpu);
}
}
vcpu_unlock(vcpu);
/*
* Wakeup the other sleeping vcpus and return to userspace.
*/
for (i = 0; i < VM_MAXCPU; i++) {
if (CPU_ISSET(((unsigned) i), &vm->suspended_cpus)) {
vcpu_notify_event(vm, i, false);
}
}
*retu = true;
return (0);
}
int
vm_suspend(struct vm *vm, enum vm_suspend_how how)
{
int i;
if (how <= VM_SUSPEND_NONE || how >= VM_SUSPEND_LAST)
return (EINVAL);
if (atomic_cmpset_int(((volatile u_int *) &vm->suspend), 0, how) == 0) {
VM_CTR2(vm, "virtual machine already suspended %d/%d",
vm->suspend, how);
return (EALREADY);
}
VM_CTR1(vm, "virtual machine successfully suspended %d", how);
/*
* Notify all active vcpus that they are now suspended.
*/
for (i = 0; i < VM_MAXCPU; i++) {
if (CPU_ISSET(((unsigned) i), &vm->active_cpus))
vcpu_notify_event(vm, i, false);
}
return (0);
}
static int
vcpu_set_state_locked(struct vcpu *vcpu, enum vcpu_state newstate,
bool from_idle)
{
int error;
const struct timespec ts = {.tv_sec = 1, .tv_nsec = 0}; /* 1 second */
/*
* State transitions from the vmmdev_ioctl() must always begin from
* the VCPU_IDLE state. This guarantees that there is only a single
* ioctl() operating on a vcpu at any point.
*/
if (from_idle) {
while (vcpu->state != VCPU_IDLE) {
pthread_mutex_lock(&vcpu->state_sleep_mtx);
vcpu_unlock(vcpu);
pthread_cond_timedwait_relative_np(&vcpu->state_sleep_cnd,
&vcpu->state_sleep_mtx, &ts);
vcpu_lock(vcpu);
pthread_mutex_unlock(&vcpu->state_sleep_mtx);
//msleep_spin(&vcpu->state, &vcpu->mtx, "vmstat", hz);
}
} else {
KASSERT(vcpu->state != VCPU_IDLE, ("invalid transition from "
"vcpu idle state"));
}
/*
* The following state transitions are allowed:
* IDLE -> FROZEN -> IDLE
* FROZEN -> RUNNING -> FROZEN
* FROZEN -> SLEEPING -> FROZEN
*/
switch (vcpu->state) {
case VCPU_IDLE:
case VCPU_RUNNING:
case VCPU_SLEEPING:
error = (newstate != VCPU_FROZEN);
break;
case VCPU_FROZEN:
error = (newstate == VCPU_FROZEN);
break;
}
if (error)
return (EBUSY);
vcpu->state = newstate;
if (newstate == VCPU_IDLE)
pthread_cond_broadcast(&vcpu->state_sleep_cnd);
//wakeup(&vcpu->state);
return (0);
}
static void
vcpu_require_state(struct vm *vm, int vcpuid, enum vcpu_state newstate)
{
int error;
if ((error = vcpu_set_state(vm, vcpuid, newstate, false)) != 0)
xhyve_abort("Error %d setting state to %d\n", error, newstate);
}