Loading...
1/*
2 * Xen time implementation.
3 *
4 * This is implemented in terms of a clocksource driver which uses
5 * the hypervisor clock as a nanosecond timebase, and a clockevent
6 * driver which uses the hypervisor's timer mechanism.
7 *
8 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
9 */
10#include <linux/kernel.h>
11#include <linux/interrupt.h>
12#include <linux/clocksource.h>
13#include <linux/clockchips.h>
14#include <linux/kernel_stat.h>
15#include <linux/math64.h>
16#include <linux/gfp.h>
17#include <linux/slab.h>
18#include <linux/pvclock_gtod.h>
19
20#include <asm/pvclock.h>
21#include <asm/xen/hypervisor.h>
22#include <asm/xen/hypercall.h>
23
24#include <xen/events.h>
25#include <xen/features.h>
26#include <xen/interface/xen.h>
27#include <xen/interface/vcpu.h>
28
29#include "xen-ops.h"
30
31/* Xen may fire a timer up to this many ns early */
32#define TIMER_SLOP 100000
33#define NS_PER_TICK (1000000000LL / HZ)
34
35/* runstate info updated by Xen */
36static DEFINE_PER_CPU(struct vcpu_runstate_info, xen_runstate);
37
38/* snapshots of runstate info */
39static DEFINE_PER_CPU(struct vcpu_runstate_info, xen_runstate_snapshot);
40
41/* unused ns of stolen time */
42static DEFINE_PER_CPU(u64, xen_residual_stolen);
43
44/* return an consistent snapshot of 64-bit time/counter value */
45static u64 get64(const u64 *p)
46{
47 u64 ret;
48
49 if (BITS_PER_LONG < 64) {
50 u32 *p32 = (u32 *)p;
51 u32 h, l;
52
53 /*
54 * Read high then low, and then make sure high is
55 * still the same; this will only loop if low wraps
56 * and carries into high.
57 * XXX some clean way to make this endian-proof?
58 */
59 do {
60 h = p32[1];
61 barrier();
62 l = p32[0];
63 barrier();
64 } while (p32[1] != h);
65
66 ret = (((u64)h) << 32) | l;
67 } else
68 ret = *p;
69
70 return ret;
71}
72
73/*
74 * Runstate accounting
75 */
76static void get_runstate_snapshot(struct vcpu_runstate_info *res)
77{
78 u64 state_time;
79 struct vcpu_runstate_info *state;
80
81 BUG_ON(preemptible());
82
83 state = &__get_cpu_var(xen_runstate);
84
85 /*
86 * The runstate info is always updated by the hypervisor on
87 * the current CPU, so there's no need to use anything
88 * stronger than a compiler barrier when fetching it.
89 */
90 do {
91 state_time = get64(&state->state_entry_time);
92 barrier();
93 *res = *state;
94 barrier();
95 } while (get64(&state->state_entry_time) != state_time);
96}
97
98/* return true when a vcpu could run but has no real cpu to run on */
99bool xen_vcpu_stolen(int vcpu)
100{
101 return per_cpu(xen_runstate, vcpu).state == RUNSTATE_runnable;
102}
103
104void xen_setup_runstate_info(int cpu)
105{
106 struct vcpu_register_runstate_memory_area area;
107
108 area.addr.v = &per_cpu(xen_runstate, cpu);
109
110 if (HYPERVISOR_vcpu_op(VCPUOP_register_runstate_memory_area,
111 cpu, &area))
112 BUG();
113}
114
115static void do_stolen_accounting(void)
116{
117 struct vcpu_runstate_info state;
118 struct vcpu_runstate_info *snap;
119 s64 runnable, offline, stolen;
120 cputime_t ticks;
121
122 get_runstate_snapshot(&state);
123
124 WARN_ON(state.state != RUNSTATE_running);
125
126 snap = &__get_cpu_var(xen_runstate_snapshot);
127
128 /* work out how much time the VCPU has not been runn*ing* */
129 runnable = state.time[RUNSTATE_runnable] - snap->time[RUNSTATE_runnable];
130 offline = state.time[RUNSTATE_offline] - snap->time[RUNSTATE_offline];
131
132 *snap = state;
133
134 /* Add the appropriate number of ticks of stolen time,
135 including any left-overs from last time. */
136 stolen = runnable + offline + __this_cpu_read(xen_residual_stolen);
137
138 if (stolen < 0)
139 stolen = 0;
140
141 ticks = iter_div_u64_rem(stolen, NS_PER_TICK, &stolen);
142 __this_cpu_write(xen_residual_stolen, stolen);
143 account_steal_ticks(ticks);
144}
145
146/* Get the TSC speed from Xen */
147static unsigned long xen_tsc_khz(void)
148{
149 struct pvclock_vcpu_time_info *info =
150 &HYPERVISOR_shared_info->vcpu_info[0].time;
151
152 return pvclock_tsc_khz(info);
153}
154
155cycle_t xen_clocksource_read(void)
156{
157 struct pvclock_vcpu_time_info *src;
158 cycle_t ret;
159
160 preempt_disable_notrace();
161 src = &__get_cpu_var(xen_vcpu)->time;
162 ret = pvclock_clocksource_read(src);
163 preempt_enable_notrace();
164 return ret;
165}
166
167static cycle_t xen_clocksource_get_cycles(struct clocksource *cs)
168{
169 return xen_clocksource_read();
170}
171
172static void xen_read_wallclock(struct timespec *ts)
173{
174 struct shared_info *s = HYPERVISOR_shared_info;
175 struct pvclock_wall_clock *wall_clock = &(s->wc);
176 struct pvclock_vcpu_time_info *vcpu_time;
177
178 vcpu_time = &get_cpu_var(xen_vcpu)->time;
179 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
180 put_cpu_var(xen_vcpu);
181}
182
183static void xen_get_wallclock(struct timespec *now)
184{
185 xen_read_wallclock(now);
186}
187
188static int xen_set_wallclock(const struct timespec *now)
189{
190 return -1;
191}
192
193static int xen_pvclock_gtod_notify(struct notifier_block *nb,
194 unsigned long was_set, void *priv)
195{
196 /* Protected by the calling core code serialization */
197 static struct timespec next_sync;
198
199 struct xen_platform_op op;
200 struct timespec now;
201
202 now = __current_kernel_time();
203
204 /*
205 * We only take the expensive HV call when the clock was set
206 * or when the 11 minutes RTC synchronization time elapsed.
207 */
208 if (!was_set && timespec_compare(&now, &next_sync) < 0)
209 return NOTIFY_OK;
210
211 op.cmd = XENPF_settime;
212 op.u.settime.secs = now.tv_sec;
213 op.u.settime.nsecs = now.tv_nsec;
214 op.u.settime.system_time = xen_clocksource_read();
215
216 (void)HYPERVISOR_dom0_op(&op);
217
218 /*
219 * Move the next drift compensation time 11 minutes
220 * ahead. That's emulating the sync_cmos_clock() update for
221 * the hardware RTC.
222 */
223 next_sync = now;
224 next_sync.tv_sec += 11 * 60;
225
226 return NOTIFY_OK;
227}
228
229static struct notifier_block xen_pvclock_gtod_notifier = {
230 .notifier_call = xen_pvclock_gtod_notify,
231};
232
233static struct clocksource xen_clocksource __read_mostly = {
234 .name = "xen",
235 .rating = 400,
236 .read = xen_clocksource_get_cycles,
237 .mask = ~0,
238 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
239};
240
241/*
242 Xen clockevent implementation
243
244 Xen has two clockevent implementations:
245
246 The old timer_op one works with all released versions of Xen prior
247 to version 3.0.4. This version of the hypervisor provides a
248 single-shot timer with nanosecond resolution. However, sharing the
249 same event channel is a 100Hz tick which is delivered while the
250 vcpu is running. We don't care about or use this tick, but it will
251 cause the core time code to think the timer fired too soon, and
252 will end up resetting it each time. It could be filtered, but
253 doing so has complications when the ktime clocksource is not yet
254 the xen clocksource (ie, at boot time).
255
256 The new vcpu_op-based timer interface allows the tick timer period
257 to be changed or turned off. The tick timer is not useful as a
258 periodic timer because events are only delivered to running vcpus.
259 The one-shot timer can report when a timeout is in the past, so
260 set_next_event is capable of returning -ETIME when appropriate.
261 This interface is used when available.
262*/
263
264
265/*
266 Get a hypervisor absolute time. In theory we could maintain an
267 offset between the kernel's time and the hypervisor's time, and
268 apply that to a kernel's absolute timeout. Unfortunately the
269 hypervisor and kernel times can drift even if the kernel is using
270 the Xen clocksource, because ntp can warp the kernel's clocksource.
271*/
272static s64 get_abs_timeout(unsigned long delta)
273{
274 return xen_clocksource_read() + delta;
275}
276
277static void xen_timerop_set_mode(enum clock_event_mode mode,
278 struct clock_event_device *evt)
279{
280 switch (mode) {
281 case CLOCK_EVT_MODE_PERIODIC:
282 /* unsupported */
283 WARN_ON(1);
284 break;
285
286 case CLOCK_EVT_MODE_ONESHOT:
287 case CLOCK_EVT_MODE_RESUME:
288 break;
289
290 case CLOCK_EVT_MODE_UNUSED:
291 case CLOCK_EVT_MODE_SHUTDOWN:
292 HYPERVISOR_set_timer_op(0); /* cancel timeout */
293 break;
294 }
295}
296
297static int xen_timerop_set_next_event(unsigned long delta,
298 struct clock_event_device *evt)
299{
300 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
301
302 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
303 BUG();
304
305 /* We may have missed the deadline, but there's no real way of
306 knowing for sure. If the event was in the past, then we'll
307 get an immediate interrupt. */
308
309 return 0;
310}
311
312static const struct clock_event_device xen_timerop_clockevent = {
313 .name = "xen",
314 .features = CLOCK_EVT_FEAT_ONESHOT,
315
316 .max_delta_ns = 0xffffffff,
317 .min_delta_ns = TIMER_SLOP,
318
319 .mult = 1,
320 .shift = 0,
321 .rating = 500,
322
323 .set_mode = xen_timerop_set_mode,
324 .set_next_event = xen_timerop_set_next_event,
325};
326
327
328
329static void xen_vcpuop_set_mode(enum clock_event_mode mode,
330 struct clock_event_device *evt)
331{
332 int cpu = smp_processor_id();
333
334 switch (mode) {
335 case CLOCK_EVT_MODE_PERIODIC:
336 WARN_ON(1); /* unsupported */
337 break;
338
339 case CLOCK_EVT_MODE_ONESHOT:
340 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
341 BUG();
342 break;
343
344 case CLOCK_EVT_MODE_UNUSED:
345 case CLOCK_EVT_MODE_SHUTDOWN:
346 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, cpu, NULL) ||
347 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
348 BUG();
349 break;
350 case CLOCK_EVT_MODE_RESUME:
351 break;
352 }
353}
354
355static int xen_vcpuop_set_next_event(unsigned long delta,
356 struct clock_event_device *evt)
357{
358 int cpu = smp_processor_id();
359 struct vcpu_set_singleshot_timer single;
360 int ret;
361
362 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
363
364 single.timeout_abs_ns = get_abs_timeout(delta);
365 single.flags = VCPU_SSHOTTMR_future;
366
367 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, cpu, &single);
368
369 BUG_ON(ret != 0 && ret != -ETIME);
370
371 return ret;
372}
373
374static const struct clock_event_device xen_vcpuop_clockevent = {
375 .name = "xen",
376 .features = CLOCK_EVT_FEAT_ONESHOT,
377
378 .max_delta_ns = 0xffffffff,
379 .min_delta_ns = TIMER_SLOP,
380
381 .mult = 1,
382 .shift = 0,
383 .rating = 500,
384
385 .set_mode = xen_vcpuop_set_mode,
386 .set_next_event = xen_vcpuop_set_next_event,
387};
388
389static const struct clock_event_device *xen_clockevent =
390 &xen_timerop_clockevent;
391
392struct xen_clock_event_device {
393 struct clock_event_device evt;
394 char *name;
395};
396static DEFINE_PER_CPU(struct xen_clock_event_device, xen_clock_events) = { .evt.irq = -1 };
397
398static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
399{
400 struct clock_event_device *evt = &__get_cpu_var(xen_clock_events).evt;
401 irqreturn_t ret;
402
403 ret = IRQ_NONE;
404 if (evt->event_handler) {
405 evt->event_handler(evt);
406 ret = IRQ_HANDLED;
407 }
408
409 do_stolen_accounting();
410
411 return ret;
412}
413
414void xen_teardown_timer(int cpu)
415{
416 struct clock_event_device *evt;
417 BUG_ON(cpu == 0);
418 evt = &per_cpu(xen_clock_events, cpu).evt;
419
420 if (evt->irq >= 0) {
421 unbind_from_irqhandler(evt->irq, NULL);
422 evt->irq = -1;
423 kfree(per_cpu(xen_clock_events, cpu).name);
424 per_cpu(xen_clock_events, cpu).name = NULL;
425 }
426}
427
428void xen_setup_timer(int cpu)
429{
430 char *name;
431 struct clock_event_device *evt;
432 int irq;
433
434 evt = &per_cpu(xen_clock_events, cpu).evt;
435 WARN(evt->irq >= 0, "IRQ%d for CPU%d is already allocated\n", evt->irq, cpu);
436 if (evt->irq >= 0)
437 xen_teardown_timer(cpu);
438
439 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
440
441 name = kasprintf(GFP_KERNEL, "timer%d", cpu);
442 if (!name)
443 name = "<timer kasprintf failed>";
444
445 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
446 IRQF_PERCPU|IRQF_NOBALANCING|IRQF_TIMER|
447 IRQF_FORCE_RESUME,
448 name, NULL);
449 (void)xen_set_irq_priority(irq, XEN_IRQ_PRIORITY_MAX);
450
451 memcpy(evt, xen_clockevent, sizeof(*evt));
452
453 evt->cpumask = cpumask_of(cpu);
454 evt->irq = irq;
455 per_cpu(xen_clock_events, cpu).name = name;
456}
457
458
459void xen_setup_cpu_clockevents(void)
460{
461 BUG_ON(preemptible());
462
463 clockevents_register_device(&__get_cpu_var(xen_clock_events).evt);
464}
465
466void xen_timer_resume(void)
467{
468 int cpu;
469
470 pvclock_resume();
471
472 if (xen_clockevent != &xen_vcpuop_clockevent)
473 return;
474
475 for_each_online_cpu(cpu) {
476 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
477 BUG();
478 }
479}
480
481static const struct pv_time_ops xen_time_ops __initconst = {
482 .sched_clock = xen_clocksource_read,
483};
484
485static void __init xen_time_init(void)
486{
487 int cpu = smp_processor_id();
488 struct timespec tp;
489
490 clocksource_register_hz(&xen_clocksource, NSEC_PER_SEC);
491
492 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL) == 0) {
493 /* Successfully turned off 100Hz tick, so we have the
494 vcpuop-based timer interface */
495 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
496 xen_clockevent = &xen_vcpuop_clockevent;
497 }
498
499 /* Set initial system time with full resolution */
500 xen_read_wallclock(&tp);
501 do_settimeofday(&tp);
502
503 setup_force_cpu_cap(X86_FEATURE_TSC);
504
505 xen_setup_runstate_info(cpu);
506 xen_setup_timer(cpu);
507 xen_setup_cpu_clockevents();
508
509 if (xen_initial_domain())
510 pvclock_gtod_register_notifier(&xen_pvclock_gtod_notifier);
511}
512
513void __init xen_init_time_ops(void)
514{
515 pv_time_ops = xen_time_ops;
516
517 x86_init.timers.timer_init = xen_time_init;
518 x86_init.timers.setup_percpu_clockev = x86_init_noop;
519 x86_cpuinit.setup_percpu_clockev = x86_init_noop;
520
521 x86_platform.calibrate_tsc = xen_tsc_khz;
522 x86_platform.get_wallclock = xen_get_wallclock;
523 /* Dom0 uses the native method to set the hardware RTC. */
524 if (!xen_initial_domain())
525 x86_platform.set_wallclock = xen_set_wallclock;
526}
527
528#ifdef CONFIG_XEN_PVHVM
529static void xen_hvm_setup_cpu_clockevents(void)
530{
531 int cpu = smp_processor_id();
532 xen_setup_runstate_info(cpu);
533 /*
534 * xen_setup_timer(cpu) - snprintf is bad in atomic context. Hence
535 * doing it xen_hvm_cpu_notify (which gets called by smp_init during
536 * early bootup and also during CPU hotplug events).
537 */
538 xen_setup_cpu_clockevents();
539}
540
541void __init xen_hvm_init_time_ops(void)
542{
543 /* vector callback is needed otherwise we cannot receive interrupts
544 * on cpu > 0 and at this point we don't know how many cpus are
545 * available */
546 if (!xen_have_vector_callback)
547 return;
548 if (!xen_feature(XENFEAT_hvm_safe_pvclock)) {
549 printk(KERN_INFO "Xen doesn't support pvclock on HVM,"
550 "disable pv timer\n");
551 return;
552 }
553
554 pv_time_ops = xen_time_ops;
555 x86_init.timers.setup_percpu_clockev = xen_time_init;
556 x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents;
557
558 x86_platform.calibrate_tsc = xen_tsc_khz;
559 x86_platform.get_wallclock = xen_get_wallclock;
560 x86_platform.set_wallclock = xen_set_wallclock;
561}
562#endif
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Xen time implementation.
4 *
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.
8 *
9 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
10 */
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>
19
20#include <asm/pvclock.h>
21#include <asm/xen/hypervisor.h>
22#include <asm/xen/hypercall.h>
23
24#include <xen/events.h>
25#include <xen/features.h>
26#include <xen/interface/xen.h>
27#include <xen/interface/vcpu.h>
28
29#include "xen-ops.h"
30
31/* Xen may fire a timer up to this many ns early */
32#define TIMER_SLOP 100000
33
34/* Get the TSC speed from Xen */
35static unsigned long xen_tsc_khz(void)
36{
37 struct pvclock_vcpu_time_info *info =
38 &HYPERVISOR_shared_info->vcpu_info[0].time;
39
40 return pvclock_tsc_khz(info);
41}
42
43u64 xen_clocksource_read(void)
44{
45 struct pvclock_vcpu_time_info *src;
46 u64 ret;
47
48 preempt_disable_notrace();
49 src = &__this_cpu_read(xen_vcpu)->time;
50 ret = pvclock_clocksource_read(src);
51 preempt_enable_notrace();
52 return ret;
53}
54
55static u64 xen_clocksource_get_cycles(struct clocksource *cs)
56{
57 return xen_clocksource_read();
58}
59
60static void xen_read_wallclock(struct timespec *ts)
61{
62 struct shared_info *s = HYPERVISOR_shared_info;
63 struct pvclock_wall_clock *wall_clock = &(s->wc);
64 struct pvclock_vcpu_time_info *vcpu_time;
65
66 vcpu_time = &get_cpu_var(xen_vcpu)->time;
67 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
68 put_cpu_var(xen_vcpu);
69}
70
71static void xen_get_wallclock(struct timespec *now)
72{
73 xen_read_wallclock(now);
74}
75
76static int xen_set_wallclock(const struct timespec *now)
77{
78 return -ENODEV;
79}
80
81static int xen_pvclock_gtod_notify(struct notifier_block *nb,
82 unsigned long was_set, void *priv)
83{
84 /* Protected by the calling core code serialization */
85 static struct timespec64 next_sync;
86
87 struct xen_platform_op op;
88 struct timespec64 now;
89 struct timekeeper *tk = priv;
90 static bool settime64_supported = true;
91 int ret;
92
93 now.tv_sec = tk->xtime_sec;
94 now.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
95
96 /*
97 * We only take the expensive HV call when the clock was set
98 * or when the 11 minutes RTC synchronization time elapsed.
99 */
100 if (!was_set && timespec64_compare(&now, &next_sync) < 0)
101 return NOTIFY_OK;
102
103again:
104 if (settime64_supported) {
105 op.cmd = XENPF_settime64;
106 op.u.settime64.mbz = 0;
107 op.u.settime64.secs = now.tv_sec;
108 op.u.settime64.nsecs = now.tv_nsec;
109 op.u.settime64.system_time = xen_clocksource_read();
110 } else {
111 op.cmd = XENPF_settime32;
112 op.u.settime32.secs = now.tv_sec;
113 op.u.settime32.nsecs = now.tv_nsec;
114 op.u.settime32.system_time = xen_clocksource_read();
115 }
116
117 ret = HYPERVISOR_platform_op(&op);
118
119 if (ret == -ENOSYS && settime64_supported) {
120 settime64_supported = false;
121 goto again;
122 }
123 if (ret < 0)
124 return NOTIFY_BAD;
125
126 /*
127 * Move the next drift compensation time 11 minutes
128 * ahead. That's emulating the sync_cmos_clock() update for
129 * the hardware RTC.
130 */
131 next_sync = now;
132 next_sync.tv_sec += 11 * 60;
133
134 return NOTIFY_OK;
135}
136
137static struct notifier_block xen_pvclock_gtod_notifier = {
138 .notifier_call = xen_pvclock_gtod_notify,
139};
140
141static struct clocksource xen_clocksource __read_mostly = {
142 .name = "xen",
143 .rating = 400,
144 .read = xen_clocksource_get_cycles,
145 .mask = ~0,
146 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
147};
148
149/*
150 Xen clockevent implementation
151
152 Xen has two clockevent implementations:
153
154 The old timer_op one works with all released versions of Xen prior
155 to version 3.0.4. This version of the hypervisor provides a
156 single-shot timer with nanosecond resolution. However, sharing the
157 same event channel is a 100Hz tick which is delivered while the
158 vcpu is running. We don't care about or use this tick, but it will
159 cause the core time code to think the timer fired too soon, and
160 will end up resetting it each time. It could be filtered, but
161 doing so has complications when the ktime clocksource is not yet
162 the xen clocksource (ie, at boot time).
163
164 The new vcpu_op-based timer interface allows the tick timer period
165 to be changed or turned off. The tick timer is not useful as a
166 periodic timer because events are only delivered to running vcpus.
167 The one-shot timer can report when a timeout is in the past, so
168 set_next_event is capable of returning -ETIME when appropriate.
169 This interface is used when available.
170*/
171
172
173/*
174 Get a hypervisor absolute time. In theory we could maintain an
175 offset between the kernel's time and the hypervisor's time, and
176 apply that to a kernel's absolute timeout. Unfortunately the
177 hypervisor and kernel times can drift even if the kernel is using
178 the Xen clocksource, because ntp can warp the kernel's clocksource.
179*/
180static s64 get_abs_timeout(unsigned long delta)
181{
182 return xen_clocksource_read() + delta;
183}
184
185static int xen_timerop_shutdown(struct clock_event_device *evt)
186{
187 /* cancel timeout */
188 HYPERVISOR_set_timer_op(0);
189
190 return 0;
191}
192
193static int xen_timerop_set_next_event(unsigned long delta,
194 struct clock_event_device *evt)
195{
196 WARN_ON(!clockevent_state_oneshot(evt));
197
198 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
199 BUG();
200
201 /* We may have missed the deadline, but there's no real way of
202 knowing for sure. If the event was in the past, then we'll
203 get an immediate interrupt. */
204
205 return 0;
206}
207
208static const struct clock_event_device xen_timerop_clockevent = {
209 .name = "xen",
210 .features = CLOCK_EVT_FEAT_ONESHOT,
211
212 .max_delta_ns = 0xffffffff,
213 .max_delta_ticks = 0xffffffff,
214 .min_delta_ns = TIMER_SLOP,
215 .min_delta_ticks = TIMER_SLOP,
216
217 .mult = 1,
218 .shift = 0,
219 .rating = 500,
220
221 .set_state_shutdown = xen_timerop_shutdown,
222 .set_next_event = xen_timerop_set_next_event,
223};
224
225static int xen_vcpuop_shutdown(struct clock_event_device *evt)
226{
227 int cpu = smp_processor_id();
228
229 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, xen_vcpu_nr(cpu),
230 NULL) ||
231 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
232 NULL))
233 BUG();
234
235 return 0;
236}
237
238static int xen_vcpuop_set_oneshot(struct clock_event_device *evt)
239{
240 int cpu = smp_processor_id();
241
242 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
243 NULL))
244 BUG();
245
246 return 0;
247}
248
249static int xen_vcpuop_set_next_event(unsigned long delta,
250 struct clock_event_device *evt)
251{
252 int cpu = smp_processor_id();
253 struct vcpu_set_singleshot_timer single;
254 int ret;
255
256 WARN_ON(!clockevent_state_oneshot(evt));
257
258 single.timeout_abs_ns = get_abs_timeout(delta);
259 /* Get an event anyway, even if the timeout is already expired */
260 single.flags = 0;
261
262 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, xen_vcpu_nr(cpu),
263 &single);
264 BUG_ON(ret != 0);
265
266 return ret;
267}
268
269static const struct clock_event_device xen_vcpuop_clockevent = {
270 .name = "xen",
271 .features = CLOCK_EVT_FEAT_ONESHOT,
272
273 .max_delta_ns = 0xffffffff,
274 .max_delta_ticks = 0xffffffff,
275 .min_delta_ns = TIMER_SLOP,
276 .min_delta_ticks = TIMER_SLOP,
277
278 .mult = 1,
279 .shift = 0,
280 .rating = 500,
281
282 .set_state_shutdown = xen_vcpuop_shutdown,
283 .set_state_oneshot = xen_vcpuop_set_oneshot,
284 .set_next_event = xen_vcpuop_set_next_event,
285};
286
287static const struct clock_event_device *xen_clockevent =
288 &xen_timerop_clockevent;
289
290struct xen_clock_event_device {
291 struct clock_event_device evt;
292 char name[16];
293};
294static DEFINE_PER_CPU(struct xen_clock_event_device, xen_clock_events) = { .evt.irq = -1 };
295
296static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
297{
298 struct clock_event_device *evt = this_cpu_ptr(&xen_clock_events.evt);
299 irqreturn_t ret;
300
301 ret = IRQ_NONE;
302 if (evt->event_handler) {
303 evt->event_handler(evt);
304 ret = IRQ_HANDLED;
305 }
306
307 return ret;
308}
309
310void xen_teardown_timer(int cpu)
311{
312 struct clock_event_device *evt;
313 evt = &per_cpu(xen_clock_events, cpu).evt;
314
315 if (evt->irq >= 0) {
316 unbind_from_irqhandler(evt->irq, NULL);
317 evt->irq = -1;
318 }
319}
320
321void xen_setup_timer(int cpu)
322{
323 struct xen_clock_event_device *xevt = &per_cpu(xen_clock_events, cpu);
324 struct clock_event_device *evt = &xevt->evt;
325 int irq;
326
327 WARN(evt->irq >= 0, "IRQ%d for CPU%d is already allocated\n", evt->irq, cpu);
328 if (evt->irq >= 0)
329 xen_teardown_timer(cpu);
330
331 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
332
333 snprintf(xevt->name, sizeof(xevt->name), "timer%d", cpu);
334
335 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
336 IRQF_PERCPU|IRQF_NOBALANCING|IRQF_TIMER|
337 IRQF_FORCE_RESUME|IRQF_EARLY_RESUME,
338 xevt->name, NULL);
339 (void)xen_set_irq_priority(irq, XEN_IRQ_PRIORITY_MAX);
340
341 memcpy(evt, xen_clockevent, sizeof(*evt));
342
343 evt->cpumask = cpumask_of(cpu);
344 evt->irq = irq;
345}
346
347
348void xen_setup_cpu_clockevents(void)
349{
350 clockevents_register_device(this_cpu_ptr(&xen_clock_events.evt));
351}
352
353void xen_timer_resume(void)
354{
355 int cpu;
356
357 pvclock_resume();
358
359 if (xen_clockevent != &xen_vcpuop_clockevent)
360 return;
361
362 for_each_online_cpu(cpu) {
363 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer,
364 xen_vcpu_nr(cpu), NULL))
365 BUG();
366 }
367}
368
369static const struct pv_time_ops xen_time_ops __initconst = {
370 .sched_clock = xen_clocksource_read,
371 .steal_clock = xen_steal_clock,
372};
373
374static struct pvclock_vsyscall_time_info *xen_clock __read_mostly;
375
376void xen_save_time_memory_area(void)
377{
378 struct vcpu_register_time_memory_area t;
379 int ret;
380
381 if (!xen_clock)
382 return;
383
384 t.addr.v = NULL;
385
386 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
387 if (ret != 0)
388 pr_notice("Cannot save secondary vcpu_time_info (err %d)",
389 ret);
390 else
391 clear_page(xen_clock);
392}
393
394void xen_restore_time_memory_area(void)
395{
396 struct vcpu_register_time_memory_area t;
397 int ret;
398
399 if (!xen_clock)
400 return;
401
402 t.addr.v = &xen_clock->pvti;
403
404 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
405
406 /*
407 * We don't disable VCLOCK_PVCLOCK entirely if it fails to register the
408 * secondary time info with Xen or if we migrated to a host without the
409 * necessary flags. On both of these cases what happens is either
410 * process seeing a zeroed out pvti or seeing no PVCLOCK_TSC_STABLE_BIT
411 * bit set. Userspace checks the latter and if 0, it discards the data
412 * in pvti and fallbacks to a system call for a reliable timestamp.
413 */
414 if (ret != 0)
415 pr_notice("Cannot restore secondary vcpu_time_info (err %d)",
416 ret);
417}
418
419static void xen_setup_vsyscall_time_info(void)
420{
421 struct vcpu_register_time_memory_area t;
422 struct pvclock_vsyscall_time_info *ti;
423 int ret;
424
425 ti = (struct pvclock_vsyscall_time_info *)get_zeroed_page(GFP_KERNEL);
426 if (!ti)
427 return;
428
429 t.addr.v = &ti->pvti;
430
431 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
432 if (ret) {
433 pr_notice("xen: VCLOCK_PVCLOCK not supported (err %d)\n", ret);
434 free_page((unsigned long)ti);
435 return;
436 }
437
438 /*
439 * If primary time info had this bit set, secondary should too since
440 * it's the same data on both just different memory regions. But we
441 * still check it in case hypervisor is buggy.
442 */
443 if (!(ti->pvti.flags & PVCLOCK_TSC_STABLE_BIT)) {
444 t.addr.v = NULL;
445 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area,
446 0, &t);
447 if (!ret)
448 free_page((unsigned long)ti);
449
450 pr_notice("xen: VCLOCK_PVCLOCK not supported (tsc unstable)\n");
451 return;
452 }
453
454 xen_clock = ti;
455 pvclock_set_pvti_cpu0_va(xen_clock);
456
457 xen_clocksource.archdata.vclock_mode = VCLOCK_PVCLOCK;
458}
459
460static void __init xen_time_init(void)
461{
462 struct pvclock_vcpu_time_info *pvti;
463 int cpu = smp_processor_id();
464 struct timespec tp;
465
466 /* As Dom0 is never moved, no penalty on using TSC there */
467 if (xen_initial_domain())
468 xen_clocksource.rating = 275;
469
470 clocksource_register_hz(&xen_clocksource, NSEC_PER_SEC);
471
472 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
473 NULL) == 0) {
474 /* Successfully turned off 100Hz tick, so we have the
475 vcpuop-based timer interface */
476 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
477 xen_clockevent = &xen_vcpuop_clockevent;
478 }
479
480 /* Set initial system time with full resolution */
481 xen_read_wallclock(&tp);
482 do_settimeofday(&tp);
483
484 setup_force_cpu_cap(X86_FEATURE_TSC);
485
486 /*
487 * We check ahead on the primary time info if this
488 * bit is supported hence speeding up Xen clocksource.
489 */
490 pvti = &__this_cpu_read(xen_vcpu)->time;
491 if (pvti->flags & PVCLOCK_TSC_STABLE_BIT) {
492 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
493 xen_setup_vsyscall_time_info();
494 }
495
496 xen_setup_runstate_info(cpu);
497 xen_setup_timer(cpu);
498 xen_setup_cpu_clockevents();
499
500 xen_time_setup_guest();
501
502 if (xen_initial_domain())
503 pvclock_gtod_register_notifier(&xen_pvclock_gtod_notifier);
504}
505
506void __ref xen_init_time_ops(void)
507{
508 pv_time_ops = xen_time_ops;
509
510 x86_init.timers.timer_init = xen_time_init;
511 x86_init.timers.setup_percpu_clockev = x86_init_noop;
512 x86_cpuinit.setup_percpu_clockev = x86_init_noop;
513
514 x86_platform.calibrate_tsc = xen_tsc_khz;
515 x86_platform.get_wallclock = xen_get_wallclock;
516 /* Dom0 uses the native method to set the hardware RTC. */
517 if (!xen_initial_domain())
518 x86_platform.set_wallclock = xen_set_wallclock;
519}
520
521#ifdef CONFIG_XEN_PVHVM
522static void xen_hvm_setup_cpu_clockevents(void)
523{
524 int cpu = smp_processor_id();
525 xen_setup_runstate_info(cpu);
526 /*
527 * xen_setup_timer(cpu) - snprintf is bad in atomic context. Hence
528 * doing it xen_hvm_cpu_notify (which gets called by smp_init during
529 * early bootup and also during CPU hotplug events).
530 */
531 xen_setup_cpu_clockevents();
532}
533
534void __init xen_hvm_init_time_ops(void)
535{
536 /*
537 * vector callback is needed otherwise we cannot receive interrupts
538 * on cpu > 0 and at this point we don't know how many cpus are
539 * available.
540 */
541 if (!xen_have_vector_callback)
542 return;
543
544 if (!xen_feature(XENFEAT_hvm_safe_pvclock)) {
545 printk(KERN_INFO "Xen doesn't support pvclock on HVM,"
546 "disable pv timer\n");
547 return;
548 }
549
550 pv_time_ops = xen_time_ops;
551 x86_init.timers.setup_percpu_clockev = xen_time_init;
552 x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents;
553
554 x86_platform.calibrate_tsc = xen_tsc_khz;
555 x86_platform.get_wallclock = xen_get_wallclock;
556 x86_platform.set_wallclock = xen_set_wallclock;
557}
558#endif