/*
* mac_soft_ring_bind
*
* Bind a soft ring worker thread to supplied CPU.
*/
cpu_t *
mac_soft_ring_bind(mac_soft_ring_t *ringp, processorid_t cpuid)
{
cpu_t *cp;
boolean_t clear = B_FALSE;
ASSERT(MUTEX_HELD(&cpu_lock));
if (mac_soft_ring_thread_bind == 0) {
DTRACE_PROBE1(mac__soft__ring__no__cpu__bound,
mac_soft_ring_t *, ringp);
return (NULL);
}
cp = cpu_get(cpuid);
if (cp == NULL || !cpu_is_online(cp))
return (NULL);
mutex_enter(&ringp->s_ring_lock);
ringp->s_ring_state |= S_RING_BOUND;
if (ringp->s_ring_cpuid != -1)
clear = B_TRUE;
ringp->s_ring_cpuid = cpuid;
mutex_exit(&ringp->s_ring_lock);
if (clear)
thread_affinity_clear(ringp->s_ring_worker);
DTRACE_PROBE2(mac__soft__ring__cpu__bound, mac_soft_ring_t *,
ringp, processorid_t, cpuid);
thread_affinity_set(ringp->s_ring_worker, cpuid);
return (cp);
}
/*
* Top level routine to direct suspend/resume of a domain.
*/
void
xen_suspend_domain(void)
{
extern void rtcsync(void);
extern hrtime_t hres_last_tick;
mfn_t start_info_mfn;
ulong_t flags;
pfn_t pfn;
int i;
/*
* Check that we are happy to suspend on this hypervisor.
*/
if (xen_hypervisor_supports_solaris(XEN_SUSPEND_CHECK) == 0) {
cpr_err(CE_WARN, "Cannot suspend on this hypervisor "
"version: v%lu.%lu%s, need at least version v3.0.4 or "
"-xvm based hypervisor", XENVER_CURRENT(xv_major),
XENVER_CURRENT(xv_minor), XENVER_CURRENT(xv_ver));
return;
}
/*
* XXPV - Are we definitely OK to suspend by the time we've connected
* the handler?
*/
cpr_err(CE_NOTE, "Domain suspending for save/migrate");
SUSPEND_DEBUG("xen_suspend_domain\n");
/*
* suspend interrupts and devices
* XXPV - we use suspend/resume for both save/restore domains (like sun
* cpr) and for migration. Would be nice to know the difference if
* possible. For save/restore where down time may be a long time, we
* may want to do more of the things that cpr does. (i.e. notify user
* processes, shrink memory footprint for faster restore, etc.)
*/
xen_suspend_devices();
SUSPEND_DEBUG("xenbus_suspend\n");
xenbus_suspend();
pfn = hat_getpfnum(kas.a_hat, (caddr_t)xen_info);
start_info_mfn = pfn_to_mfn(pfn);
/*
* XXPV: cpu hotplug can hold this under a xenbus watch. Are we safe
* wrt xenbus being suspended here?
*/
mutex_enter(&cpu_lock);
/*
* Suspend must be done on vcpu 0, as no context for other CPUs is
* saved.
*
* XXPV - add to taskq API ?
*/
thread_affinity_set(curthread, 0);
kpreempt_disable();
SUSPEND_DEBUG("xen_start_migrate\n");
xen_start_migrate();
if (ncpus > 1)
suspend_cpus();
/*
* We can grab the ec_lock as it's a spinlock with a high SPL. Hence
* any holder would have dropped it to get through suspend_cpus().
*/
mutex_enter(&ec_lock);
/*
* From here on in, we can't take locks.
*/
SUSPEND_DEBUG("ec_suspend\n");
ec_suspend();
SUSPEND_DEBUG("gnttab_suspend\n");
gnttab_suspend();
flags = intr_clear();
xpv_time_suspend();
/*
* Currently, the hypervisor incorrectly fails to bring back
* powered-down VCPUs. Thus we need to record any powered-down VCPUs
* to prevent any attempts to operate on them. But we have to do this
* *after* the very first time we do ec_suspend().
*/
for (i = 1; i < ncpus; i++) {
if (cpu[i] == NULL)
continue;
if (cpu_get_state(cpu[i]) == P_POWEROFF)
CPUSET_ATOMIC_ADD(cpu_suspend_lost_set, i);
}
/*
* The dom0 save/migrate code doesn't automatically translate
* these into PFNs, but expects them to be, so we do it here.
* We don't use mfn_to_pfn() because so many OS services have
* been disabled at this point.
*/
xen_info->store_mfn = mfn_to_pfn_mapping[xen_info->store_mfn];
xen_info->console.domU.mfn =
mfn_to_pfn_mapping[xen_info->console.domU.mfn];
if (CPU->cpu_m.mcpu_vcpu_info->evtchn_upcall_mask == 0) {
prom_printf("xen_suspend_domain(): "
"CPU->cpu_m.mcpu_vcpu_info->evtchn_upcall_mask not set\n");
(void) HYPERVISOR_shutdown(SHUTDOWN_crash);
}
if (HYPERVISOR_update_va_mapping((uintptr_t)HYPERVISOR_shared_info,
0, UVMF_INVLPG)) {
prom_printf("xen_suspend_domain(): "
"HYPERVISOR_update_va_mapping() failed\n");
(void) HYPERVISOR_shutdown(SHUTDOWN_crash);
}
SUSPEND_DEBUG("HYPERVISOR_suspend\n");
/*
* At this point we suspend and sometime later resume.
*/
if (HYPERVISOR_suspend(start_info_mfn)) {
prom_printf("xen_suspend_domain(): "
"HYPERVISOR_suspend() failed\n");
(void) HYPERVISOR_shutdown(SHUTDOWN_crash);
}
/*
* Point HYPERVISOR_shared_info to its new value.
*/
if (HYPERVISOR_update_va_mapping((uintptr_t)HYPERVISOR_shared_info,
xen_info->shared_info | PT_NOCONSIST | PT_VALID | PT_WRITABLE,
UVMF_INVLPG))
(void) HYPERVISOR_shutdown(SHUTDOWN_crash);
if (xen_info->nr_pages != mfn_count) {
prom_printf("xen_suspend_domain(): number of pages"
" changed, was 0x%lx, now 0x%lx\n", mfn_count,
xen_info->nr_pages);
(void) HYPERVISOR_shutdown(SHUTDOWN_crash);
}
xpv_time_resume();
cached_max_mfn = 0;
SUSPEND_DEBUG("gnttab_resume\n");
gnttab_resume();
/* XXPV: add a note that this must be lockless. */
SUSPEND_DEBUG("ec_resume\n");
ec_resume();
intr_restore(flags);
if (ncpus > 1)
resume_cpus();
mutex_exit(&ec_lock);
xen_end_migrate();
mutex_exit(&cpu_lock);
/*
* Now we can take locks again.
*/
/*
* Force the tick value used for tv_nsec in hres_tick() to be up to
* date. rtcsync() will reset the hrestime value appropriately.
*/
hres_last_tick = xpv_gethrtime();
/*
* XXPV: we need to have resumed the CPUs since this takes locks, but
* can remote CPUs see bad state? Presumably yes. Should probably nest
* taking of todlock inside of cpu_lock, or vice versa, then provide an
* unlocked version. Probably need to call clkinitf to reset cpu freq
* and re-calibrate if we migrated to a different speed cpu. Also need
* to make a (re)init_cpu_info call to update processor info structs
* and device tree info. That remains to be written at the moment.
*/
rtcsync();
rebuild_mfn_list();
SUSPEND_DEBUG("xenbus_resume\n");
xenbus_resume();
SUSPEND_DEBUG("xenbus_resume_devices\n");
xen_resume_devices();
thread_affinity_clear(curthread);
kpreempt_enable();
SUSPEND_DEBUG("finished xen_suspend_domain\n");
/*
* We have restarted our suspended domain, update the hypervisor
* details. NB: This must be done at the end of this function,
* since we need the domain to be completely resumed before
* these functions will work correctly.
*/
xen_set_version(XENVER_CURRENT_IDX);
/*
* We can check and report a warning, but we don't stop the
* process.
*/
if (xen_hypervisor_supports_solaris(XEN_SUSPEND_CHECK) == 0)
cmn_err(CE_WARN, "Found hypervisor version: v%lu.%lu%s "
"but need at least version v3.0.4",
XENVER_CURRENT(xv_major), XENVER_CURRENT(xv_minor),
XENVER_CURRENT(xv_ver));
cmn_err(CE_NOTE, "domain restore/migrate completed");
}
kern_return_t
thread_policy_set_internal(
thread_t thread,
thread_policy_flavor_t flavor,
thread_policy_t policy_info,
mach_msg_type_number_t count)
{
kern_return_t result = KERN_SUCCESS;
spl_t s;
thread_mtx_lock(thread);
if (!thread->active) {
thread_mtx_unlock(thread);
return (KERN_TERMINATED);
}
switch (flavor) {
case THREAD_EXTENDED_POLICY:
{
boolean_t timeshare = TRUE;
if (count >= THREAD_EXTENDED_POLICY_COUNT) {
thread_extended_policy_t info;
info = (thread_extended_policy_t)policy_info;
timeshare = info->timeshare;
}
sched_mode_t mode = (timeshare == TRUE) ? TH_MODE_TIMESHARE : TH_MODE_FIXED;
s = splsched();
thread_lock(thread);
thread_set_user_sched_mode_and_recompute_pri(thread, mode);
thread_unlock(thread);
splx(s);
sfi_reevaluate(thread);
break;
}
case THREAD_TIME_CONSTRAINT_POLICY:
{
thread_time_constraint_policy_t info;
if (count < THREAD_TIME_CONSTRAINT_POLICY_COUNT) {
result = KERN_INVALID_ARGUMENT;
break;
}
info = (thread_time_constraint_policy_t)policy_info;
if ( info->constraint < info->computation ||
info->computation > max_rt_quantum ||
info->computation < min_rt_quantum ) {
result = KERN_INVALID_ARGUMENT;
break;
}
s = splsched();
thread_lock(thread);
thread->realtime.period = info->period;
thread->realtime.computation = info->computation;
thread->realtime.constraint = info->constraint;
thread->realtime.preemptible = info->preemptible;
thread_set_user_sched_mode_and_recompute_pri(thread, TH_MODE_REALTIME);
thread_unlock(thread);
splx(s);
sfi_reevaluate(thread);
break;
}
case THREAD_PRECEDENCE_POLICY:
{
thread_precedence_policy_t info;
if (count < THREAD_PRECEDENCE_POLICY_COUNT) {
result = KERN_INVALID_ARGUMENT;
break;
}
info = (thread_precedence_policy_t)policy_info;
s = splsched();
thread_lock(thread);
thread->importance = info->importance;
thread_recompute_priority(thread);
thread_unlock(thread);
splx(s);
break;
}
case THREAD_AFFINITY_POLICY:
{
thread_affinity_policy_t info;
if (!thread_affinity_is_supported()) {
result = KERN_NOT_SUPPORTED;
break;
}
if (count < THREAD_AFFINITY_POLICY_COUNT) {
result = KERN_INVALID_ARGUMENT;
break;
}
info = (thread_affinity_policy_t) policy_info;
/*
* Unlock the thread mutex here and
* return directly after calling thread_affinity_set().
* This is necessary for correct lock ordering because
* thread_affinity_set() takes the task lock.
*/
thread_mtx_unlock(thread);
return thread_affinity_set(thread, info->affinity_tag);
}
case THREAD_THROUGHPUT_QOS_POLICY:
{
thread_throughput_qos_policy_t info = (thread_throughput_qos_policy_t) policy_info;
int tqos;
if (count < THREAD_LATENCY_QOS_POLICY_COUNT) {
result = KERN_INVALID_ARGUMENT;
break;
}
if ((result = qos_throughput_policy_validate(info->thread_throughput_qos_tier)) !=
KERN_SUCCESS) {
break;
}
tqos = qos_extract(info->thread_throughput_qos_tier);
thread->effective_policy.t_through_qos = tqos;
}
break;
case THREAD_LATENCY_QOS_POLICY:
{
thread_latency_qos_policy_t info = (thread_latency_qos_policy_t) policy_info;
int lqos;
if (count < THREAD_THROUGHPUT_QOS_POLICY_COUNT) {
result = KERN_INVALID_ARGUMENT;
break;
}
if ((result = qos_latency_policy_validate(info->thread_latency_qos_tier)) !=
KERN_SUCCESS) {
break;
}
lqos = qos_extract(info->thread_latency_qos_tier);
/* The expected use cases (opt-in) of per-thread latency QoS would seem to
* preclude any requirement at present to re-evaluate timers on a thread level
* latency QoS change.
*/
thread->effective_policy.t_latency_qos = lqos;
}
break;
case THREAD_QOS_POLICY:
case THREAD_QOS_POLICY_OVERRIDE:
{
thread_qos_policy_t info = (thread_qos_policy_t)policy_info;
if (count < THREAD_QOS_POLICY_COUNT) {
result = KERN_INVALID_ARGUMENT;
break;
}
if (info->qos_tier < 0 || info->qos_tier >= THREAD_QOS_LAST) {
result = KERN_INVALID_ARGUMENT;
break;
}
if (info->tier_importance > 0 || info->tier_importance < THREAD_QOS_MIN_TIER_IMPORTANCE) {
result = KERN_INVALID_ARGUMENT;
break;
}
if (info->qos_tier == THREAD_QOS_UNSPECIFIED && info->tier_importance != 0) {
result = KERN_INVALID_ARGUMENT;
break;
}
/*
* Going into task policy requires the task mutex,
* because of the way synchronization against the IO policy
* subsystem works.
*
* We need to move thread policy to the thread mutex instead.
* <rdar://problem/15831652> separate thread policy from task policy
*/
if (flavor == THREAD_QOS_POLICY_OVERRIDE) {
int strongest_override = info->qos_tier;
if (info->qos_tier != THREAD_QOS_UNSPECIFIED &&
thread->requested_policy.thrp_qos_override != THREAD_QOS_UNSPECIFIED)
strongest_override = MAX(thread->requested_policy.thrp_qos_override, info->qos_tier);
thread_mtx_unlock(thread);
/* There is a race here. To be closed in <rdar://problem/15831652> separate thread policy from task policy */
proc_set_task_policy(thread->task, thread, TASK_POLICY_ATTRIBUTE, TASK_POLICY_QOS_OVERRIDE, strongest_override);
return (result);
}
thread_mtx_unlock(thread);
proc_set_task_policy2(thread->task, thread, TASK_POLICY_ATTRIBUTE, TASK_POLICY_QOS_AND_RELPRIO, info->qos_tier, -info->tier_importance);
thread_mtx_lock(thread);
if (!thread->active) {
thread_mtx_unlock(thread);
return (KERN_TERMINATED);
}
break;
}
default:
result = KERN_INVALID_ARGUMENT;
break;
}
thread_mtx_unlock(thread);
return (result);
}
kern_return_t
thread_policy_set(
thread_t thread,
thread_policy_flavor_t flavor,
thread_policy_t policy_info,
mach_msg_type_number_t count)
{
kern_return_t result = KERN_SUCCESS;
spl_t s;
if (thread == THREAD_NULL)
return (KERN_INVALID_ARGUMENT);
thread_mtx_lock(thread);
if (!thread->active) {
thread_mtx_unlock(thread);
return (KERN_TERMINATED);
}
if (thread->static_param) {
thread_mtx_unlock(thread);
return (KERN_SUCCESS);
}
switch (flavor) {
case THREAD_EXTENDED_POLICY:
{
boolean_t timeshare = TRUE;
if (count >= THREAD_EXTENDED_POLICY_COUNT) {
thread_extended_policy_t info;
info = (thread_extended_policy_t)policy_info;
timeshare = info->timeshare;
}
s = splsched();
thread_lock(thread);
if (!(thread->sched_mode & TH_MODE_FAILSAFE)) {
integer_t oldmode = (thread->sched_mode & TH_MODE_TIMESHARE);
thread->sched_mode &= ~TH_MODE_REALTIME;
if (timeshare && !oldmode) {
thread->sched_mode |= TH_MODE_TIMESHARE;
if ((thread->state & (TH_RUN|TH_IDLE)) == TH_RUN)
sched_share_incr();
}
else
if (!timeshare && oldmode) {
thread->sched_mode &= ~TH_MODE_TIMESHARE;
if ((thread->state & (TH_RUN|TH_IDLE)) == TH_RUN)
sched_share_decr();
}
thread_recompute_priority(thread);
}
else {
thread->safe_mode &= ~TH_MODE_REALTIME;
if (timeshare)
thread->safe_mode |= TH_MODE_TIMESHARE;
else
thread->safe_mode &= ~TH_MODE_TIMESHARE;
}
thread_unlock(thread);
splx(s);
break;
}
case THREAD_TIME_CONSTRAINT_POLICY:
{
thread_time_constraint_policy_t info;
if (count < THREAD_TIME_CONSTRAINT_POLICY_COUNT) {
result = KERN_INVALID_ARGUMENT;
break;
}
info = (thread_time_constraint_policy_t)policy_info;
if ( info->constraint < info->computation ||
info->computation > max_rt_quantum ||
info->computation < min_rt_quantum ) {
result = KERN_INVALID_ARGUMENT;
break;
}
s = splsched();
thread_lock(thread);
thread->realtime.period = info->period;
thread->realtime.computation = info->computation;
thread->realtime.constraint = info->constraint;
thread->realtime.preemptible = info->preemptible;
if (!(thread->sched_mode & TH_MODE_FAILSAFE)) {
if (thread->sched_mode & TH_MODE_TIMESHARE) {
thread->sched_mode &= ~TH_MODE_TIMESHARE;
if ((thread->state & (TH_RUN|TH_IDLE)) == TH_RUN)
sched_share_decr();
}
thread->sched_mode |= TH_MODE_REALTIME;
thread_recompute_priority(thread);
}
else {
thread->safe_mode &= ~TH_MODE_TIMESHARE;
thread->safe_mode |= TH_MODE_REALTIME;
}
thread_unlock(thread);
splx(s);
break;
}
case THREAD_PRECEDENCE_POLICY:
{
thread_precedence_policy_t info;
if (count < THREAD_PRECEDENCE_POLICY_COUNT) {
result = KERN_INVALID_ARGUMENT;
break;
}
info = (thread_precedence_policy_t)policy_info;
s = splsched();
thread_lock(thread);
thread->importance = info->importance;
thread_recompute_priority(thread);
thread_unlock(thread);
splx(s);
break;
}
case THREAD_AFFINITY_POLICY:
{
thread_affinity_policy_t info;
if (!thread_affinity_is_supported()) {
result = KERN_NOT_SUPPORTED;
break;
}
if (count < THREAD_AFFINITY_POLICY_COUNT) {
result = KERN_INVALID_ARGUMENT;
break;
}
info = (thread_affinity_policy_t) policy_info;
/*
* Unlock the thread mutex here and
* return directly after calling thread_affinity_set().
* This is necessary for correct lock ordering because
* thread_affinity_set() takes the task lock.
*/
thread_mtx_unlock(thread);
return thread_affinity_set(thread, info->affinity_tag);
}
default:
result = KERN_INVALID_ARGUMENT;
break;
}
thread_mtx_unlock(thread);
return (result);
}
/*
* Top level routine to direct suspend/resume of a domain.
*/
void
xen_suspend_domain(void)
{
extern void rtcsync(void);
extern void ec_resume(void);
extern kmutex_t ec_lock;
struct xen_add_to_physmap xatp;
ulong_t flags;
int err;
cmn_err(CE_NOTE, "Domain suspending for save/migrate");
SUSPEND_DEBUG("xen_suspend_domain\n");
/*
* We only want to suspend the PV devices, since the emulated devices
* are suspended by saving the emulated device state. The PV devices
* are all children of the xpvd nexus device. So we search the
* device tree for the xpvd node to use as the root of the tree to
* be suspended.
*/
if (xpvd_dip == NULL)
ddi_walk_devs(ddi_root_node(), check_xpvd, NULL);
/*
* suspend interrupts and devices
*/
if (xpvd_dip != NULL)
(void) xen_suspend_devices(ddi_get_child(xpvd_dip));
else
cmn_err(CE_WARN, "No PV devices found to suspend");
SUSPEND_DEBUG("xenbus_suspend\n");
xenbus_suspend();
mutex_enter(&cpu_lock);
/*
* Suspend on vcpu 0
*/
thread_affinity_set(curthread, 0);
kpreempt_disable();
if (ncpus > 1)
pause_cpus(NULL, NULL);
/*
* We can grab the ec_lock as it's a spinlock with a high SPL. Hence
* any holder would have dropped it to get through pause_cpus().
*/
mutex_enter(&ec_lock);
/*
* From here on in, we can't take locks.
*/
flags = intr_clear();
SUSPEND_DEBUG("HYPERVISOR_suspend\n");
/*
* At this point we suspend and sometime later resume.
* Note that this call may return with an indication of a cancelled
* for now no matter ehat the return we do a full resume of all
* suspended drivers, etc.
*/
(void) HYPERVISOR_shutdown(SHUTDOWN_suspend);
/*
* Point HYPERVISOR_shared_info to the proper place.
*/
xatp.domid = DOMID_SELF;
xatp.idx = 0;
xatp.space = XENMAPSPACE_shared_info;
xatp.gpfn = xen_shared_info_frame;
if ((err = HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp)) != 0)
panic("Could not set shared_info page. error: %d", err);
SUSPEND_DEBUG("gnttab_resume\n");
gnttab_resume();
SUSPEND_DEBUG("ec_resume\n");
ec_resume();
intr_restore(flags);
if (ncpus > 1)
start_cpus();
mutex_exit(&ec_lock);
mutex_exit(&cpu_lock);
/*
* Now we can take locks again.
*/
rtcsync();
SUSPEND_DEBUG("xenbus_resume\n");
xenbus_resume();
SUSPEND_DEBUG("xen_resume_devices\n");
if (xpvd_dip != NULL)
(void) xen_resume_devices(ddi_get_child(xpvd_dip), 0);
thread_affinity_clear(curthread);
kpreempt_enable();
SUSPEND_DEBUG("finished xen_suspend_domain\n");
cmn_err(CE_NOTE, "domain restore/migrate completed");
}
