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 /* Minimum amount of time until next clock event fires */
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 int xen_cs_enable(struct clocksource *cs)
150 vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK);
154 static struct clocksource xen_clocksource __read_mostly = {
157 .read = xen_clocksource_get_cycles,
158 .mask = CLOCKSOURCE_MASK(64),
159 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
160 .enable = xen_cs_enable,
164 Xen clockevent implementation
166 Xen has two clockevent implementations:
168 The old timer_op one works with all released versions of Xen prior
169 to version 3.0.4. This version of the hypervisor provides a
170 single-shot timer with nanosecond resolution. However, sharing the
171 same event channel is a 100Hz tick which is delivered while the
172 vcpu is running. We don't care about or use this tick, but it will
173 cause the core time code to think the timer fired too soon, and
174 will end up resetting it each time. It could be filtered, but
175 doing so has complications when the ktime clocksource is not yet
176 the xen clocksource (ie, at boot time).
178 The new vcpu_op-based timer interface allows the tick timer period
179 to be changed or turned off. The tick timer is not useful as a
180 periodic timer because events are only delivered to running vcpus.
181 The one-shot timer can report when a timeout is in the past, so
182 set_next_event is capable of returning -ETIME when appropriate.
183 This interface is used when available.
188 Get a hypervisor absolute time. In theory we could maintain an
189 offset between the kernel's time and the hypervisor's time, and
190 apply that to a kernel's absolute timeout. Unfortunately the
191 hypervisor and kernel times can drift even if the kernel is using
192 the Xen clocksource, because ntp can warp the kernel's clocksource.
194 static s64 get_abs_timeout(unsigned long delta)
196 return xen_clocksource_read() + delta;
199 static int xen_timerop_shutdown(struct clock_event_device *evt)
202 HYPERVISOR_set_timer_op(0);
207 static int xen_timerop_set_next_event(unsigned long delta,
208 struct clock_event_device *evt)
210 WARN_ON(!clockevent_state_oneshot(evt));
212 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
215 /* We may have missed the deadline, but there's no real way of
216 knowing for sure. If the event was in the past, then we'll
217 get an immediate interrupt. */
222 static struct clock_event_device xen_timerop_clockevent __ro_after_init = {
224 .features = CLOCK_EVT_FEAT_ONESHOT,
226 .max_delta_ns = 0xffffffff,
227 .max_delta_ticks = 0xffffffff,
228 .min_delta_ns = TIMER_SLOP,
229 .min_delta_ticks = TIMER_SLOP,
235 .set_state_shutdown = xen_timerop_shutdown,
236 .set_next_event = xen_timerop_set_next_event,
239 static int xen_vcpuop_shutdown(struct clock_event_device *evt)
241 int cpu = smp_processor_id();
243 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, xen_vcpu_nr(cpu),
245 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
252 static int xen_vcpuop_set_oneshot(struct clock_event_device *evt)
254 int cpu = smp_processor_id();
256 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
263 static int xen_vcpuop_set_next_event(unsigned long delta,
264 struct clock_event_device *evt)
266 int cpu = smp_processor_id();
267 struct vcpu_set_singleshot_timer single;
270 WARN_ON(!clockevent_state_oneshot(evt));
272 single.timeout_abs_ns = get_abs_timeout(delta);
273 /* Get an event anyway, even if the timeout is already expired */
276 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, xen_vcpu_nr(cpu),
283 static struct clock_event_device xen_vcpuop_clockevent __ro_after_init = {
285 .features = CLOCK_EVT_FEAT_ONESHOT,
287 .max_delta_ns = 0xffffffff,
288 .max_delta_ticks = 0xffffffff,
289 .min_delta_ns = TIMER_SLOP,
290 .min_delta_ticks = TIMER_SLOP,
296 .set_state_shutdown = xen_vcpuop_shutdown,
297 .set_state_oneshot = xen_vcpuop_set_oneshot,
298 .set_next_event = xen_vcpuop_set_next_event,
301 static const struct clock_event_device *xen_clockevent =
302 &xen_timerop_clockevent;
304 struct xen_clock_event_device {
305 struct clock_event_device evt;
308 static DEFINE_PER_CPU(struct xen_clock_event_device, xen_clock_events) = { .evt.irq = -1 };
310 static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
312 struct clock_event_device *evt = this_cpu_ptr(&xen_clock_events.evt);
316 if (evt->event_handler) {
317 evt->event_handler(evt);
324 void xen_teardown_timer(int cpu)
326 struct clock_event_device *evt;
327 evt = &per_cpu(xen_clock_events, cpu).evt;
330 unbind_from_irqhandler(evt->irq, NULL);
335 void xen_setup_timer(int cpu)
337 struct xen_clock_event_device *xevt = &per_cpu(xen_clock_events, cpu);
338 struct clock_event_device *evt = &xevt->evt;
341 WARN(evt->irq >= 0, "IRQ%d for CPU%d is already allocated\n", evt->irq, cpu);
343 xen_teardown_timer(cpu);
345 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
347 snprintf(xevt->name, sizeof(xevt->name), "timer%d", cpu);
349 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
350 IRQF_PERCPU|IRQF_NOBALANCING|IRQF_TIMER|
351 IRQF_FORCE_RESUME|IRQF_EARLY_RESUME,
353 (void)xen_set_irq_priority(irq, XEN_IRQ_PRIORITY_MAX);
355 memcpy(evt, xen_clockevent, sizeof(*evt));
357 evt->cpumask = cpumask_of(cpu);
362 void xen_setup_cpu_clockevents(void)
364 clockevents_register_device(this_cpu_ptr(&xen_clock_events.evt));
367 void xen_timer_resume(void)
371 if (xen_clockevent != &xen_vcpuop_clockevent)
374 for_each_online_cpu(cpu) {
375 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer,
376 xen_vcpu_nr(cpu), NULL))
381 static const struct pv_time_ops xen_time_ops __initconst = {
382 .sched_clock = xen_sched_clock,
383 .steal_clock = xen_steal_clock,
386 static struct pvclock_vsyscall_time_info *xen_clock __read_mostly;
387 static u64 xen_clock_value_saved;
389 void xen_save_time_memory_area(void)
391 struct vcpu_register_time_memory_area t;
394 xen_clock_value_saved = xen_clocksource_read() - xen_sched_clock_offset;
401 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
403 pr_notice("Cannot save secondary vcpu_time_info (err %d)",
406 clear_page(xen_clock);
409 void xen_restore_time_memory_area(void)
411 struct vcpu_register_time_memory_area t;
417 t.addr.v = &xen_clock->pvti;
419 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
422 * We don't disable VDSO_CLOCKMODE_PVCLOCK entirely if it fails to
423 * register the secondary time info with Xen or if we migrated to a
424 * host without the necessary flags. On both of these cases what
425 * happens is either process seeing a zeroed out pvti or seeing no
426 * PVCLOCK_TSC_STABLE_BIT bit set. Userspace checks the latter and
427 * if 0, it discards the data in pvti and fallbacks to a system
428 * call for a reliable timestamp.
431 pr_notice("Cannot restore secondary vcpu_time_info (err %d)",
435 /* Need pvclock_resume() before using xen_clocksource_read(). */
437 xen_sched_clock_offset = xen_clocksource_read() - xen_clock_value_saved;
440 static void xen_setup_vsyscall_time_info(void)
442 struct vcpu_register_time_memory_area t;
443 struct pvclock_vsyscall_time_info *ti;
446 ti = (struct pvclock_vsyscall_time_info *)get_zeroed_page(GFP_KERNEL);
450 t.addr.v = &ti->pvti;
452 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
454 pr_notice("xen: VDSO_CLOCKMODE_PVCLOCK not supported (err %d)\n", ret);
455 free_page((unsigned long)ti);
460 * If primary time info had this bit set, secondary should too since
461 * it's the same data on both just different memory regions. But we
462 * still check it in case hypervisor is buggy.
464 if (!(ti->pvti.flags & PVCLOCK_TSC_STABLE_BIT)) {
466 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area,
469 free_page((unsigned long)ti);
471 pr_notice("xen: VDSO_CLOCKMODE_PVCLOCK not supported (tsc unstable)\n");
476 pvclock_set_pvti_cpu0_va(xen_clock);
478 xen_clocksource.vdso_clock_mode = VDSO_CLOCKMODE_PVCLOCK;
481 static void __init xen_time_init(void)
483 struct pvclock_vcpu_time_info *pvti;
484 int cpu = smp_processor_id();
485 struct timespec64 tp;
487 /* As Dom0 is never moved, no penalty on using TSC there */
488 if (xen_initial_domain())
489 xen_clocksource.rating = 275;
491 clocksource_register_hz(&xen_clocksource, NSEC_PER_SEC);
493 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
495 /* Successfully turned off 100Hz tick, so we have the
496 vcpuop-based timer interface */
497 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
498 xen_clockevent = &xen_vcpuop_clockevent;
501 /* Set initial system time with full resolution */
502 xen_read_wallclock(&tp);
503 do_settimeofday64(&tp);
505 setup_force_cpu_cap(X86_FEATURE_TSC);
508 * We check ahead on the primary time info if this
509 * bit is supported hence speeding up Xen clocksource.
511 pvti = &__this_cpu_read(xen_vcpu)->time;
512 if (pvti->flags & PVCLOCK_TSC_STABLE_BIT) {
513 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
514 xen_setup_vsyscall_time_info();
517 xen_setup_runstate_info(cpu);
518 xen_setup_timer(cpu);
519 xen_setup_cpu_clockevents();
521 xen_time_setup_guest();
523 if (xen_initial_domain())
524 pvclock_gtod_register_notifier(&xen_pvclock_gtod_notifier);
527 void __init xen_init_time_ops(void)
529 xen_sched_clock_offset = xen_clocksource_read();
530 pv_ops.time = xen_time_ops;
532 x86_init.timers.timer_init = xen_time_init;
533 x86_init.timers.setup_percpu_clockev = x86_init_noop;
534 x86_cpuinit.setup_percpu_clockev = x86_init_noop;
536 x86_platform.calibrate_tsc = xen_tsc_khz;
537 x86_platform.get_wallclock = xen_get_wallclock;
538 /* Dom0 uses the native method to set the hardware RTC. */
539 if (!xen_initial_domain())
540 x86_platform.set_wallclock = xen_set_wallclock;
543 #ifdef CONFIG_XEN_PVHVM
544 static void xen_hvm_setup_cpu_clockevents(void)
546 int cpu = smp_processor_id();
547 xen_setup_runstate_info(cpu);
549 * xen_setup_timer(cpu) - snprintf is bad in atomic context. Hence
550 * doing it xen_hvm_cpu_notify (which gets called by smp_init during
551 * early bootup and also during CPU hotplug events).
553 xen_setup_cpu_clockevents();
556 void __init xen_hvm_init_time_ops(void)
559 * vector callback is needed otherwise we cannot receive interrupts
560 * on cpu > 0 and at this point we don't know how many cpus are
563 if (!xen_have_vector_callback)
566 if (!xen_feature(XENFEAT_hvm_safe_pvclock)) {
567 pr_info("Xen doesn't support pvclock on HVM, disable pv timer");
571 xen_sched_clock_offset = xen_clocksource_read();
572 pv_ops.time = xen_time_ops;
573 x86_init.timers.setup_percpu_clockev = xen_time_init;
574 x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents;
576 x86_platform.calibrate_tsc = xen_tsc_khz;
577 x86_platform.get_wallclock = xen_get_wallclock;
578 x86_platform.set_wallclock = xen_set_wallclock;
582 /* Kernel parameter to specify Xen timer slop */
583 static int __init parse_xen_timer_slop(char *ptr)
585 unsigned long slop = memparse(ptr, NULL);
587 xen_timerop_clockevent.min_delta_ns = slop;
588 xen_timerop_clockevent.min_delta_ticks = slop;
589 xen_vcpuop_clockevent.min_delta_ns = slop;
590 xen_vcpuop_clockevent.min_delta_ticks = slop;
594 early_param("xen_timer_slop", parse_xen_timer_slop);