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/*
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
18#include <asm/pvclock.h>
19#include <asm/xen/hypervisor.h>
20#include <asm/xen/hypercall.h>
21
22#include <xen/events.h>
23#include <xen/features.h>
24#include <xen/interface/xen.h>
25#include <xen/interface/vcpu.h>
26
27#include "xen-ops.h"
28
29/* Xen may fire a timer up to this many ns early */
30#define TIMER_SLOP 100000
31#define NS_PER_TICK (1000000000LL / HZ)
32
33/* runstate info updated by Xen */
34static DEFINE_PER_CPU(struct vcpu_runstate_info, xen_runstate);
35
36/* snapshots of runstate info */
37static DEFINE_PER_CPU(struct vcpu_runstate_info, xen_runstate_snapshot);
38
39/* unused ns of stolen and blocked time */
40static DEFINE_PER_CPU(u64, xen_residual_stolen);
41static DEFINE_PER_CPU(u64, xen_residual_blocked);
42
43/* return an consistent snapshot of 64-bit time/counter value */
44static u64 get64(const u64 *p)
45{
46 u64 ret;
47
48 if (BITS_PER_LONG < 64) {
49 u32 *p32 = (u32 *)p;
50 u32 h, l;
51
52 /*
53 * Read high then low, and then make sure high is
54 * still the same; this will only loop if low wraps
55 * and carries into high.
56 * XXX some clean way to make this endian-proof?
57 */
58 do {
59 h = p32[1];
60 barrier();
61 l = p32[0];
62 barrier();
63 } while (p32[1] != h);
64
65 ret = (((u64)h) << 32) | l;
66 } else
67 ret = *p;
68
69 return ret;
70}
71
72/*
73 * Runstate accounting
74 */
75static void get_runstate_snapshot(struct vcpu_runstate_info *res)
76{
77 u64 state_time;
78 struct vcpu_runstate_info *state;
79
80 BUG_ON(preemptible());
81
82 state = &__get_cpu_var(xen_runstate);
83
84 /*
85 * The runstate info is always updated by the hypervisor on
86 * the current CPU, so there's no need to use anything
87 * stronger than a compiler barrier when fetching it.
88 */
89 do {
90 state_time = get64(&state->state_entry_time);
91 barrier();
92 *res = *state;
93 barrier();
94 } while (get64(&state->state_entry_time) != state_time);
95}
96
97/* return true when a vcpu could run but has no real cpu to run on */
98bool xen_vcpu_stolen(int vcpu)
99{
100 return per_cpu(xen_runstate, vcpu).state == RUNSTATE_runnable;
101}
102
103void xen_setup_runstate_info(int cpu)
104{
105 struct vcpu_register_runstate_memory_area area;
106
107 area.addr.v = &per_cpu(xen_runstate, cpu);
108
109 if (HYPERVISOR_vcpu_op(VCPUOP_register_runstate_memory_area,
110 cpu, &area))
111 BUG();
112}
113
114static void do_stolen_accounting(void)
115{
116 struct vcpu_runstate_info state;
117 struct vcpu_runstate_info *snap;
118 s64 blocked, runnable, offline, stolen;
119 cputime_t ticks;
120
121 get_runstate_snapshot(&state);
122
123 WARN_ON(state.state != RUNSTATE_running);
124
125 snap = &__get_cpu_var(xen_runstate_snapshot);
126
127 /* work out how much time the VCPU has not been runn*ing* */
128 blocked = state.time[RUNSTATE_blocked] - snap->time[RUNSTATE_blocked];
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 /* Add the appropriate number of ticks of blocked time,
146 including any left-overs from last time. */
147 blocked += __this_cpu_read(xen_residual_blocked);
148
149 if (blocked < 0)
150 blocked = 0;
151
152 ticks = iter_div_u64_rem(blocked, NS_PER_TICK, &blocked);
153 __this_cpu_write(xen_residual_blocked, blocked);
154 account_idle_ticks(ticks);
155}
156
157/* Get the TSC speed from Xen */
158static unsigned long xen_tsc_khz(void)
159{
160 struct pvclock_vcpu_time_info *info =
161 &HYPERVISOR_shared_info->vcpu_info[0].time;
162
163 return pvclock_tsc_khz(info);
164}
165
166cycle_t xen_clocksource_read(void)
167{
168 struct pvclock_vcpu_time_info *src;
169 cycle_t ret;
170
171 preempt_disable_notrace();
172 src = &__get_cpu_var(xen_vcpu)->time;
173 ret = pvclock_clocksource_read(src);
174 preempt_enable_notrace();
175 return ret;
176}
177
178static cycle_t xen_clocksource_get_cycles(struct clocksource *cs)
179{
180 return xen_clocksource_read();
181}
182
183static void xen_read_wallclock(struct timespec *ts)
184{
185 struct shared_info *s = HYPERVISOR_shared_info;
186 struct pvclock_wall_clock *wall_clock = &(s->wc);
187 struct pvclock_vcpu_time_info *vcpu_time;
188
189 vcpu_time = &get_cpu_var(xen_vcpu)->time;
190 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
191 put_cpu_var(xen_vcpu);
192}
193
194static unsigned long xen_get_wallclock(void)
195{
196 struct timespec ts;
197
198 xen_read_wallclock(&ts);
199 return ts.tv_sec;
200}
201
202static int xen_set_wallclock(unsigned long now)
203{
204 /* do nothing for domU */
205 return -1;
206}
207
208static struct clocksource xen_clocksource __read_mostly = {
209 .name = "xen",
210 .rating = 400,
211 .read = xen_clocksource_get_cycles,
212 .mask = ~0,
213 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
214};
215
216/*
217 Xen clockevent implementation
218
219 Xen has two clockevent implementations:
220
221 The old timer_op one works with all released versions of Xen prior
222 to version 3.0.4. This version of the hypervisor provides a
223 single-shot timer with nanosecond resolution. However, sharing the
224 same event channel is a 100Hz tick which is delivered while the
225 vcpu is running. We don't care about or use this tick, but it will
226 cause the core time code to think the timer fired too soon, and
227 will end up resetting it each time. It could be filtered, but
228 doing so has complications when the ktime clocksource is not yet
229 the xen clocksource (ie, at boot time).
230
231 The new vcpu_op-based timer interface allows the tick timer period
232 to be changed or turned off. The tick timer is not useful as a
233 periodic timer because events are only delivered to running vcpus.
234 The one-shot timer can report when a timeout is in the past, so
235 set_next_event is capable of returning -ETIME when appropriate.
236 This interface is used when available.
237*/
238
239
240/*
241 Get a hypervisor absolute time. In theory we could maintain an
242 offset between the kernel's time and the hypervisor's time, and
243 apply that to a kernel's absolute timeout. Unfortunately the
244 hypervisor and kernel times can drift even if the kernel is using
245 the Xen clocksource, because ntp can warp the kernel's clocksource.
246*/
247static s64 get_abs_timeout(unsigned long delta)
248{
249 return xen_clocksource_read() + delta;
250}
251
252static void xen_timerop_set_mode(enum clock_event_mode mode,
253 struct clock_event_device *evt)
254{
255 switch (mode) {
256 case CLOCK_EVT_MODE_PERIODIC:
257 /* unsupported */
258 WARN_ON(1);
259 break;
260
261 case CLOCK_EVT_MODE_ONESHOT:
262 case CLOCK_EVT_MODE_RESUME:
263 break;
264
265 case CLOCK_EVT_MODE_UNUSED:
266 case CLOCK_EVT_MODE_SHUTDOWN:
267 HYPERVISOR_set_timer_op(0); /* cancel timeout */
268 break;
269 }
270}
271
272static int xen_timerop_set_next_event(unsigned long delta,
273 struct clock_event_device *evt)
274{
275 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
276
277 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
278 BUG();
279
280 /* We may have missed the deadline, but there's no real way of
281 knowing for sure. If the event was in the past, then we'll
282 get an immediate interrupt. */
283
284 return 0;
285}
286
287static const struct clock_event_device xen_timerop_clockevent = {
288 .name = "xen",
289 .features = CLOCK_EVT_FEAT_ONESHOT,
290
291 .max_delta_ns = 0xffffffff,
292 .min_delta_ns = TIMER_SLOP,
293
294 .mult = 1,
295 .shift = 0,
296 .rating = 500,
297
298 .set_mode = xen_timerop_set_mode,
299 .set_next_event = xen_timerop_set_next_event,
300};
301
302
303
304static void xen_vcpuop_set_mode(enum clock_event_mode mode,
305 struct clock_event_device *evt)
306{
307 int cpu = smp_processor_id();
308
309 switch (mode) {
310 case CLOCK_EVT_MODE_PERIODIC:
311 WARN_ON(1); /* unsupported */
312 break;
313
314 case CLOCK_EVT_MODE_ONESHOT:
315 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
316 BUG();
317 break;
318
319 case CLOCK_EVT_MODE_UNUSED:
320 case CLOCK_EVT_MODE_SHUTDOWN:
321 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, cpu, NULL) ||
322 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
323 BUG();
324 break;
325 case CLOCK_EVT_MODE_RESUME:
326 break;
327 }
328}
329
330static int xen_vcpuop_set_next_event(unsigned long delta,
331 struct clock_event_device *evt)
332{
333 int cpu = smp_processor_id();
334 struct vcpu_set_singleshot_timer single;
335 int ret;
336
337 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
338
339 single.timeout_abs_ns = get_abs_timeout(delta);
340 single.flags = VCPU_SSHOTTMR_future;
341
342 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, cpu, &single);
343
344 BUG_ON(ret != 0 && ret != -ETIME);
345
346 return ret;
347}
348
349static const struct clock_event_device xen_vcpuop_clockevent = {
350 .name = "xen",
351 .features = CLOCK_EVT_FEAT_ONESHOT,
352
353 .max_delta_ns = 0xffffffff,
354 .min_delta_ns = TIMER_SLOP,
355
356 .mult = 1,
357 .shift = 0,
358 .rating = 500,
359
360 .set_mode = xen_vcpuop_set_mode,
361 .set_next_event = xen_vcpuop_set_next_event,
362};
363
364static const struct clock_event_device *xen_clockevent =
365 &xen_timerop_clockevent;
366static DEFINE_PER_CPU(struct clock_event_device, xen_clock_events);
367
368static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
369{
370 struct clock_event_device *evt = &__get_cpu_var(xen_clock_events);
371 irqreturn_t ret;
372
373 ret = IRQ_NONE;
374 if (evt->event_handler) {
375 evt->event_handler(evt);
376 ret = IRQ_HANDLED;
377 }
378
379 do_stolen_accounting();
380
381 return ret;
382}
383
384void xen_setup_timer(int cpu)
385{
386 const char *name;
387 struct clock_event_device *evt;
388 int irq;
389
390 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
391
392 name = kasprintf(GFP_KERNEL, "timer%d", cpu);
393 if (!name)
394 name = "<timer kasprintf failed>";
395
396 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
397 IRQF_DISABLED|IRQF_PERCPU|
398 IRQF_NOBALANCING|IRQF_TIMER|
399 IRQF_FORCE_RESUME,
400 name, NULL);
401
402 evt = &per_cpu(xen_clock_events, cpu);
403 memcpy(evt, xen_clockevent, sizeof(*evt));
404
405 evt->cpumask = cpumask_of(cpu);
406 evt->irq = irq;
407}
408
409void xen_teardown_timer(int cpu)
410{
411 struct clock_event_device *evt;
412 BUG_ON(cpu == 0);
413 evt = &per_cpu(xen_clock_events, cpu);
414 unbind_from_irqhandler(evt->irq, NULL);
415}
416
417void xen_setup_cpu_clockevents(void)
418{
419 BUG_ON(preemptible());
420
421 clockevents_register_device(&__get_cpu_var(xen_clock_events));
422}
423
424void xen_timer_resume(void)
425{
426 int cpu;
427
428 pvclock_resume();
429
430 if (xen_clockevent != &xen_vcpuop_clockevent)
431 return;
432
433 for_each_online_cpu(cpu) {
434 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
435 BUG();
436 }
437}
438
439static const struct pv_time_ops xen_time_ops __initconst = {
440 .sched_clock = xen_clocksource_read,
441};
442
443static void __init xen_time_init(void)
444{
445 int cpu = smp_processor_id();
446 struct timespec tp;
447
448 clocksource_register_hz(&xen_clocksource, NSEC_PER_SEC);
449
450 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL) == 0) {
451 /* Successfully turned off 100Hz tick, so we have the
452 vcpuop-based timer interface */
453 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
454 xen_clockevent = &xen_vcpuop_clockevent;
455 }
456
457 /* Set initial system time with full resolution */
458 xen_read_wallclock(&tp);
459 do_settimeofday(&tp);
460
461 setup_force_cpu_cap(X86_FEATURE_TSC);
462
463 xen_setup_runstate_info(cpu);
464 xen_setup_timer(cpu);
465 xen_setup_cpu_clockevents();
466}
467
468void __init xen_init_time_ops(void)
469{
470 pv_time_ops = xen_time_ops;
471
472 x86_init.timers.timer_init = xen_time_init;
473 x86_init.timers.setup_percpu_clockev = x86_init_noop;
474 x86_cpuinit.setup_percpu_clockev = x86_init_noop;
475
476 x86_platform.calibrate_tsc = xen_tsc_khz;
477 x86_platform.get_wallclock = xen_get_wallclock;
478 x86_platform.set_wallclock = xen_set_wallclock;
479}
480
481#ifdef CONFIG_XEN_PVHVM
482static void xen_hvm_setup_cpu_clockevents(void)
483{
484 int cpu = smp_processor_id();
485 xen_setup_runstate_info(cpu);
486 xen_setup_timer(cpu);
487 xen_setup_cpu_clockevents();
488}
489
490void __init xen_hvm_init_time_ops(void)
491{
492 /* vector callback is needed otherwise we cannot receive interrupts
493 * on cpu > 0 and at this point we don't know how many cpus are
494 * available */
495 if (!xen_have_vector_callback)
496 return;
497 if (!xen_feature(XENFEAT_hvm_safe_pvclock)) {
498 printk(KERN_INFO "Xen doesn't support pvclock on HVM,"
499 "disable pv timer\n");
500 return;
501 }
502
503 pv_time_ops = xen_time_ops;
504 x86_init.timers.setup_percpu_clockev = xen_time_init;
505 x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents;
506
507 x86_platform.calibrate_tsc = xen_tsc_khz;
508 x86_platform.get_wallclock = xen_get_wallclock;
509 x86_platform.set_wallclock = xen_set_wallclock;
510}
511#endif