static int xen_suspend(void *data)
{
int *cancelled = data;
int err;
BUG_ON(!irqs_disabled());
load_cr3(swapper_pg_dir);
err = device_power_down(PMSG_SUSPEND);
if (err) {
printk(KERN_ERR "xen_suspend: device_power_down failed: %d\n",
err);
return err;
}
xen_mm_pin_all();
gnttab_suspend();
xen_pre_suspend();
/*
* This hypercall returns 1 if suspend was cancelled
* or the domain was merely checkpointed, and 0 if it
* is resuming in a new domain.
*/
*cancelled = HYPERVISOR_suspend(virt_to_mfn(xen_start_info));
xen_post_suspend(*cancelled);
gnttab_resume();
xen_mm_unpin_all();
device_power_up(PMSG_RESUME);
if (!*cancelled) {
xen_irq_resume();
xen_console_resume();
xen_timer_resume();
}
return 0;
}
static int xen_suspend(void *data)
{
struct suspend_info *si = data;
int err;
BUG_ON(!irqs_disabled());
err = syscore_suspend();
if (err) {
pr_err("%s: system core suspend failed: %d\n", __func__, err);
return err;
}
gnttab_suspend();
xen_arch_pre_suspend();
/*
* This hypercall returns 1 if suspend was cancelled
* or the domain was merely checkpointed, and 0 if it
* is resuming in a new domain.
*/
si->cancelled = HYPERVISOR_suspend(xen_pv_domain()
? virt_to_mfn(xen_start_info)
: 0);
xen_arch_post_suspend(si->cancelled);
gnttab_resume();
if (!si->cancelled) {
xen_irq_resume();
xen_console_resume();
xen_timer_resume();
}
syscore_resume();
return 0;
}
static int __do_suspend(void *ignore)
{
int err;
extern void time_resume(void);
BUG_ON(smp_processor_id() != 0);
BUG_ON(in_interrupt());
#if defined(__i386__) || defined(__x86_64__)
if (xen_feature(XENFEAT_auto_translated_physmap)) {
printk(KERN_WARNING "Cannot suspend in "
"auto_translated_physmap mode.\n");
return -EOPNOTSUPP;
}
#endif
err = smp_suspend();
if (err)
return err;
xenbus_suspend();
preempt_disable();
#ifdef __i386__
kmem_cache_shrink(pgd_cache);
#endif
mm_pin_all();
local_irq_disable();
preempt_enable();
gnttab_suspend();
pre_suspend();
/*
* We'll stop somewhere inside this hypercall. When it returns,
* we'll start resuming after the restore.
*/
HYPERVISOR_suspend(virt_to_mfn(xen_start_info));
shutting_down = SHUTDOWN_INVALID;
post_suspend();
gnttab_resume();
irq_resume();
time_resume();
switch_idle_mm();
local_irq_enable();
xencons_resume();
xenbus_resume();
smp_resume();
return err;
}
/*
* 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");
}
static void xen_pre_suspend(void)
{
xen_mm_pin_all();
gnttab_suspend();
xen_arch_pre_suspend();
}
static int __do_suspend(void *ignore)
{
int i, j, k, fpp, err;
extern unsigned long max_pfn;
extern unsigned long *pfn_to_mfn_frame_list_list;
extern unsigned long *pfn_to_mfn_frame_list[];
extern void time_resume(void);
BUG_ON(smp_processor_id() != 0);
BUG_ON(in_interrupt());
if (xen_feature(XENFEAT_auto_translated_physmap)) {
printk(KERN_WARNING "Cannot suspend in "
"auto_translated_physmap mode.\n");
return -EOPNOTSUPP;
}
err = smp_suspend();
if (err)
return err;
xenbus_suspend();
preempt_disable();
#ifdef __i386__
kmem_cache_shrink(pgd_cache);
#endif
mm_pin_all();
__cli();
preempt_enable();
gnttab_suspend();
HYPERVISOR_shared_info = (shared_info_t *)empty_zero_page;
clear_fixmap(FIX_SHARED_INFO);
xen_start_info->store_mfn = mfn_to_pfn(xen_start_info->store_mfn);
xen_start_info->console_mfn = mfn_to_pfn(xen_start_info->console_mfn);
/*
* We'll stop somewhere inside this hypercall. When it returns,
* we'll start resuming after the restore.
*/
HYPERVISOR_suspend(virt_to_mfn(xen_start_info));
shutting_down = SHUTDOWN_INVALID;
set_fixmap(FIX_SHARED_INFO, xen_start_info->shared_info);
HYPERVISOR_shared_info = (shared_info_t *)fix_to_virt(FIX_SHARED_INFO);
memset(empty_zero_page, 0, PAGE_SIZE);
HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
virt_to_mfn(pfn_to_mfn_frame_list_list);
fpp = PAGE_SIZE/sizeof(unsigned long);
for (i = 0, j = 0, k = -1; i < max_pfn; i += fpp, j++) {
if ((j % fpp) == 0) {
k++;
pfn_to_mfn_frame_list_list[k] =
virt_to_mfn(pfn_to_mfn_frame_list[k]);
j = 0;
}
pfn_to_mfn_frame_list[k][j] =
virt_to_mfn(&phys_to_machine_mapping[i]);
}
HYPERVISOR_shared_info->arch.max_pfn = max_pfn;
gnttab_resume();
irq_resume();
time_resume();
switch_idle_mm();
__sti();
xencons_resume();
xenbus_resume();
smp_resume();
return err;
}
/* Full PV mode suspension. */
static void
xctrl_suspend()
{
int i, j, k, fpp, suspend_cancelled;
unsigned long max_pfn, start_info_mfn;
EVENTHANDLER_INVOKE(power_suspend);
#ifdef SMP
struct thread *td;
cpuset_t map;
u_int cpuid;
/*
* Bind us to CPU 0 and stop any other VCPUs.
*/
td = curthread;
thread_lock(td);
sched_bind(td, 0);
thread_unlock(td);
cpuid = PCPU_GET(cpuid);
KASSERT(cpuid == 0, ("xen_suspend: not running on cpu 0"));
map = all_cpus;
CPU_CLR(cpuid, &map);
CPU_NAND(&map, &stopped_cpus);
if (!CPU_EMPTY(&map))
stop_cpus(map);
#endif
/*
* Be sure to hold Giant across DEVICE_SUSPEND/RESUME since non-MPSAFE
* drivers need this.
*/
mtx_lock(&Giant);
if (DEVICE_SUSPEND(root_bus) != 0) {
mtx_unlock(&Giant);
printf("%s: device_suspend failed\n", __func__);
#ifdef SMP
if (!CPU_EMPTY(&map))
restart_cpus(map);
#endif
return;
}
mtx_unlock(&Giant);
local_irq_disable();
xencons_suspend();
gnttab_suspend();
intr_suspend();
max_pfn = HYPERVISOR_shared_info->arch.max_pfn;
void *shared_info = HYPERVISOR_shared_info;
HYPERVISOR_shared_info = NULL;
pmap_kremove((vm_offset_t) shared_info);
PT_UPDATES_FLUSH();
xen_start_info->store_mfn = MFNTOPFN(xen_start_info->store_mfn);
xen_start_info->console.domU.mfn = MFNTOPFN(xen_start_info->console.domU.mfn);
/*
* We'll stop somewhere inside this hypercall. When it returns,
* we'll start resuming after the restore.
*/
start_info_mfn = VTOMFN(xen_start_info);
pmap_suspend();
suspend_cancelled = HYPERVISOR_suspend(start_info_mfn);
pmap_resume();
pmap_kenter_ma((vm_offset_t) shared_info, xen_start_info->shared_info);
HYPERVISOR_shared_info = shared_info;
HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
VTOMFN(xen_pfn_to_mfn_frame_list_list);
fpp = PAGE_SIZE/sizeof(unsigned long);
for (i = 0, j = 0, k = -1; i < max_pfn; i += fpp, j++) {
if ((j % fpp) == 0) {
k++;
xen_pfn_to_mfn_frame_list_list[k] =
VTOMFN(xen_pfn_to_mfn_frame_list[k]);
j = 0;
}
xen_pfn_to_mfn_frame_list[k][j] =
VTOMFN(&xen_phys_machine[i]);
}
HYPERVISOR_shared_info->arch.max_pfn = max_pfn;
gnttab_resume();
intr_resume(suspend_cancelled != 0);
local_irq_enable();
xencons_resume();
#ifdef CONFIG_SMP
for_each_cpu(i)
vcpu_prepare(i);
#endif
/*
* Only resume xenbus /after/ we've prepared our VCPUs; otherwise
* the VCPU hotplug callback can race with our vcpu_prepare
*/
mtx_lock(&Giant);
DEVICE_RESUME(root_bus);
mtx_unlock(&Giant);
#ifdef SMP
thread_lock(curthread);
sched_unbind(curthread);
thread_unlock(curthread);
if (!CPU_EMPTY(&map))
restart_cpus(map);
#endif
EVENTHANDLER_INVOKE(power_resume);
}