1 // SPDX-License-Identifier: GPL-2.0
3 * Xen time implementation.
5 * This is implemented in terms of a clocksource driver which uses
6 * the hypervisor clock as a nanosecond timebase, and a clockevent
7 * driver which uses the hypervisor's timer mechanism.
11 #include <linux/kernel.h>
12 #include <linux/interrupt.h>
13 #include <linux/clocksource.h>
14 #include <linux/clockchips.h>
15 #include <linux/gfp.h>
16 #include <linux/slab.h>
17 #include <linux/pvclock_gtod.h>
18 #include <linux/timekeeper_internal.h>
20 #include <asm/pvclock.h>
21 #include <asm/xen/hypervisor.h>
22 #include <asm/xen/hypercall.h>
24 #include <xen/events.h>
25 #include <xen/features.h>
26 #include <xen/interface/xen.h>
27 #include <xen/interface/vcpu.h>
31 /* Xen may fire a timer up to this many ns early */
32 #define TIMER_SLOP 100000
34 static u64 xen_sched_clock_offset __read_mostly;
36 /* Get the TSC speed from Xen */
37 static unsigned long xen_tsc_khz(void)
39 struct pvclock_vcpu_time_info *info =
40 &HYPERVISOR_shared_info->vcpu_info[0].time;
42 return pvclock_tsc_khz(info);
45 static u64 xen_clocksource_read(void)
47 struct pvclock_vcpu_time_info *src;
50 preempt_disable_notrace();
51 src = &__this_cpu_read(xen_vcpu)->time;
52 ret = pvclock_clocksource_read(src);
53 preempt_enable_notrace();
57 static u64 xen_clocksource_get_cycles(struct clocksource *cs)
59 return xen_clocksource_read();
62 static u64 xen_sched_clock(void)
64 return xen_clocksource_read() - xen_sched_clock_offset;
67 static void xen_read_wallclock(struct timespec64 *ts)
69 struct shared_info *s = HYPERVISOR_shared_info;
70 struct pvclock_wall_clock *wall_clock = &(s->wc);
71 struct pvclock_vcpu_time_info *vcpu_time;
73 vcpu_time = &get_cpu_var(xen_vcpu)->time;
74 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
75 put_cpu_var(xen_vcpu);
78 static void xen_get_wallclock(struct timespec64 *now)
80 xen_read_wallclock(now);
83 static int xen_set_wallclock(const struct timespec64 *now)
88 static int xen_pvclock_gtod_notify(struct notifier_block *nb,
89 unsigned long was_set, void *priv)
91 /* Protected by the calling core code serialization */
92 static struct timespec64 next_sync;
94 struct xen_platform_op op;
95 struct timespec64 now;
96 struct timekeeper *tk = priv;
97 static bool settime64_supported = true;
100 now.tv_sec = tk->xtime_sec;
101 now.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
104 * We only take the expensive HV call when the clock was set
105 * or when the 11 minutes RTC synchronization time elapsed.
107 if (!was_set && timespec64_compare(&now, &next_sync) < 0)
111 if (settime64_supported) {
112 op.cmd = XENPF_settime64;
113 op.u.settime64.mbz = 0;
114 op.u.settime64.secs = now.tv_sec;
115 op.u.settime64.nsecs = now.tv_nsec;
116 op.u.settime64.system_time = xen_clocksource_read();
118 op.cmd = XENPF_settime32;
119 op.u.settime32.secs = now.tv_sec;
120 op.u.settime32.nsecs = now.tv_nsec;
121 op.u.settime32.system_time = xen_clocksource_read();
124 ret = HYPERVISOR_platform_op(&op);
126 if (ret == -ENOSYS && settime64_supported) {
127 settime64_supported = false;
134 * Move the next drift compensation time 11 minutes
135 * ahead. That's emulating the sync_cmos_clock() update for
139 next_sync.tv_sec += 11 * 60;
144 static struct notifier_block xen_pvclock_gtod_notifier = {
145 .notifier_call = xen_pvclock_gtod_notify,
148 static struct clocksource xen_clocksource __read_mostly = {
151 .read = xen_clocksource_get_cycles,
153 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
157 Xen clockevent implementation
159 Xen has two clockevent implementations:
161 The old timer_op one works with all released versions of Xen prior
162 to version 3.0.4. This version of the hypervisor provides a
163 single-shot timer with nanosecond resolution. However, sharing the
164 same event channel is a 100Hz tick which is delivered while the
165 vcpu is running. We don't care about or use this tick, but it will
166 cause the core time code to think the timer fired too soon, and
167 will end up resetting it each time. It could be filtered, but
168 doing so has complications when the ktime clocksource is not yet
169 the xen clocksource (ie, at boot time).
171 The new vcpu_op-based timer interface allows the tick timer period
172 to be changed or turned off. The tick timer is not useful as a
173 periodic timer because events are only delivered to running vcpus.
174 The one-shot timer can report when a timeout is in the past, so
175 set_next_event is capable of returning -ETIME when appropriate.
176 This interface is used when available.
181 Get a hypervisor absolute time. In theory we could maintain an
182 offset between the kernel's time and the hypervisor's time, and
183 apply that to a kernel's absolute timeout. Unfortunately the
184 hypervisor and kernel times can drift even if the kernel is using
185 the Xen clocksource, because ntp can warp the kernel's clocksource.
187 static s64 get_abs_timeout(unsigned long delta)
189 return xen_clocksource_read() + delta;
192 static int xen_timerop_shutdown(struct clock_event_device *evt)
195 HYPERVISOR_set_timer_op(0);
200 static int xen_timerop_set_next_event(unsigned long delta,
201 struct clock_event_device *evt)
203 WARN_ON(!clockevent_state_oneshot(evt));
205 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
208 /* We may have missed the deadline, but there's no real way of
209 knowing for sure. If the event was in the past, then we'll
210 get an immediate interrupt. */
215 static const struct clock_event_device xen_timerop_clockevent = {
217 .features = CLOCK_EVT_FEAT_ONESHOT,
219 .max_delta_ns = 0xffffffff,
220 .max_delta_ticks = 0xffffffff,
221 .min_delta_ns = TIMER_SLOP,
222 .min_delta_ticks = TIMER_SLOP,
228 .set_state_shutdown = xen_timerop_shutdown,
229 .set_next_event = xen_timerop_set_next_event,
232 static int xen_vcpuop_shutdown(struct clock_event_device *evt)
234 int cpu = smp_processor_id();
236 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, xen_vcpu_nr(cpu),
238 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
245 static int xen_vcpuop_set_oneshot(struct clock_event_device *evt)
247 int cpu = smp_processor_id();
249 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
256 static int xen_vcpuop_set_next_event(unsigned long delta,
257 struct clock_event_device *evt)
259 int cpu = smp_processor_id();
260 struct vcpu_set_singleshot_timer single;
263 WARN_ON(!clockevent_state_oneshot(evt));
265 single.timeout_abs_ns = get_abs_timeout(delta);
266 /* Get an event anyway, even if the timeout is already expired */
269 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, xen_vcpu_nr(cpu),
276 static const struct clock_event_device xen_vcpuop_clockevent = {
278 .features = CLOCK_EVT_FEAT_ONESHOT,
280 .max_delta_ns = 0xffffffff,
281 .max_delta_ticks = 0xffffffff,
282 .min_delta_ns = TIMER_SLOP,
283 .min_delta_ticks = TIMER_SLOP,
289 .set_state_shutdown = xen_vcpuop_shutdown,
290 .set_state_oneshot = xen_vcpuop_set_oneshot,
291 .set_next_event = xen_vcpuop_set_next_event,
294 static const struct clock_event_device *xen_clockevent =
295 &xen_timerop_clockevent;
297 struct xen_clock_event_device {
298 struct clock_event_device evt;
301 static DEFINE_PER_CPU(struct xen_clock_event_device, xen_clock_events) = { .evt.irq = -1 };
303 static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
305 struct clock_event_device *evt = this_cpu_ptr(&xen_clock_events.evt);
309 if (evt->event_handler) {
310 evt->event_handler(evt);
317 void xen_teardown_timer(int cpu)
319 struct clock_event_device *evt;
320 evt = &per_cpu(xen_clock_events, cpu).evt;
323 unbind_from_irqhandler(evt->irq, NULL);
328 void xen_setup_timer(int cpu)
330 struct xen_clock_event_device *xevt = &per_cpu(xen_clock_events, cpu);
331 struct clock_event_device *evt = &xevt->evt;
334 WARN(evt->irq >= 0, "IRQ%d for CPU%d is already allocated\n", evt->irq, cpu);
336 xen_teardown_timer(cpu);
338 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
340 snprintf(xevt->name, sizeof(xevt->name), "timer%d", cpu);
342 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
343 IRQF_PERCPU|IRQF_NOBALANCING|IRQF_TIMER|
344 IRQF_FORCE_RESUME|IRQF_EARLY_RESUME,
346 (void)xen_set_irq_priority(irq, XEN_IRQ_PRIORITY_MAX);
348 memcpy(evt, xen_clockevent, sizeof(*evt));
350 evt->cpumask = cpumask_of(cpu);
355 void xen_setup_cpu_clockevents(void)
357 clockevents_register_device(this_cpu_ptr(&xen_clock_events.evt));
360 void xen_timer_resume(void)
366 if (xen_clockevent != &xen_vcpuop_clockevent)
369 for_each_online_cpu(cpu) {
370 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer,
371 xen_vcpu_nr(cpu), NULL))
376 static const struct pv_time_ops xen_time_ops __initconst = {
377 .sched_clock = xen_sched_clock,
378 .steal_clock = xen_steal_clock,
381 static struct pvclock_vsyscall_time_info *xen_clock __read_mostly;
383 void xen_save_time_memory_area(void)
385 struct vcpu_register_time_memory_area t;
393 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
395 pr_notice("Cannot save secondary vcpu_time_info (err %d)",
398 clear_page(xen_clock);
401 void xen_restore_time_memory_area(void)
403 struct vcpu_register_time_memory_area t;
409 t.addr.v = &xen_clock->pvti;
411 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
414 * We don't disable VCLOCK_PVCLOCK entirely if it fails to register the
415 * secondary time info with Xen or if we migrated to a host without the
416 * necessary flags. On both of these cases what happens is either
417 * process seeing a zeroed out pvti or seeing no PVCLOCK_TSC_STABLE_BIT
418 * bit set. Userspace checks the latter and if 0, it discards the data
419 * in pvti and fallbacks to a system call for a reliable timestamp.
422 pr_notice("Cannot restore secondary vcpu_time_info (err %d)",
426 static void xen_setup_vsyscall_time_info(void)
428 struct vcpu_register_time_memory_area t;
429 struct pvclock_vsyscall_time_info *ti;
432 ti = (struct pvclock_vsyscall_time_info *)get_zeroed_page(GFP_KERNEL);
436 t.addr.v = &ti->pvti;
438 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
440 pr_notice("xen: VCLOCK_PVCLOCK not supported (err %d)\n", ret);
441 free_page((unsigned long)ti);
446 * If primary time info had this bit set, secondary should too since
447 * it's the same data on both just different memory regions. But we
448 * still check it in case hypervisor is buggy.
450 if (!(ti->pvti.flags & PVCLOCK_TSC_STABLE_BIT)) {
452 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area,
455 free_page((unsigned long)ti);
457 pr_notice("xen: VCLOCK_PVCLOCK not supported (tsc unstable)\n");
462 pvclock_set_pvti_cpu0_va(xen_clock);
464 xen_clocksource.archdata.vclock_mode = VCLOCK_PVCLOCK;
467 static void __init xen_time_init(void)
469 struct pvclock_vcpu_time_info *pvti;
470 int cpu = smp_processor_id();
471 struct timespec64 tp;
473 /* As Dom0 is never moved, no penalty on using TSC there */
474 if (xen_initial_domain())
475 xen_clocksource.rating = 275;
477 clocksource_register_hz(&xen_clocksource, NSEC_PER_SEC);
479 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
481 /* Successfully turned off 100Hz tick, so we have the
482 vcpuop-based timer interface */
483 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
484 xen_clockevent = &xen_vcpuop_clockevent;
487 /* Set initial system time with full resolution */
488 xen_read_wallclock(&tp);
489 do_settimeofday64(&tp);
491 setup_force_cpu_cap(X86_FEATURE_TSC);
494 * We check ahead on the primary time info if this
495 * bit is supported hence speeding up Xen clocksource.
497 pvti = &__this_cpu_read(xen_vcpu)->time;
498 if (pvti->flags & PVCLOCK_TSC_STABLE_BIT) {
499 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
500 xen_setup_vsyscall_time_info();
503 xen_setup_runstate_info(cpu);
504 xen_setup_timer(cpu);
505 xen_setup_cpu_clockevents();
507 xen_time_setup_guest();
509 if (xen_initial_domain())
510 pvclock_gtod_register_notifier(&xen_pvclock_gtod_notifier);
513 void __init xen_init_time_ops(void)
515 xen_sched_clock_offset = xen_clocksource_read();
516 pv_ops.time = xen_time_ops;
518 x86_init.timers.timer_init = xen_time_init;
519 x86_init.timers.setup_percpu_clockev = x86_init_noop;
520 x86_cpuinit.setup_percpu_clockev = x86_init_noop;
522 x86_platform.calibrate_tsc = xen_tsc_khz;
523 x86_platform.get_wallclock = xen_get_wallclock;
524 /* Dom0 uses the native method to set the hardware RTC. */
525 if (!xen_initial_domain())
526 x86_platform.set_wallclock = xen_set_wallclock;
529 #ifdef CONFIG_XEN_PVHVM
530 static void xen_hvm_setup_cpu_clockevents(void)
532 int cpu = smp_processor_id();
533 xen_setup_runstate_info(cpu);
535 * xen_setup_timer(cpu) - snprintf is bad in atomic context. Hence
536 * doing it xen_hvm_cpu_notify (which gets called by smp_init during
537 * early bootup and also during CPU hotplug events).
539 xen_setup_cpu_clockevents();
542 void __init xen_hvm_init_time_ops(void)
545 * vector callback is needed otherwise we cannot receive interrupts
546 * on cpu > 0 and at this point we don't know how many cpus are
549 if (!xen_have_vector_callback)
552 if (!xen_feature(XENFEAT_hvm_safe_pvclock)) {
553 pr_info("Xen doesn't support pvclock on HVM, disable pv timer");
557 xen_sched_clock_offset = xen_clocksource_read();
558 pv_ops.time = xen_time_ops;
559 x86_init.timers.setup_percpu_clockev = xen_time_init;
560 x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents;
562 x86_platform.calibrate_tsc = xen_tsc_khz;
563 x86_platform.get_wallclock = xen_get_wallclock;
564 x86_platform.set_wallclock = xen_set_wallclock;