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#include <asm/xen/cpuid.h>
24
25#include <xen/events.h>
26#include <xen/features.h>
27#include <xen/interface/xen.h>
28#include <xen/interface/vcpu.h>
29
30#include "xen-ops.h"
31
32/* Minimum amount of time until next clock event fires */
33#define TIMER_SLOP 100000
34
35static u64 xen_sched_clock_offset __read_mostly;
36
37/* Get the TSC speed from Xen */
38static unsigned long xen_tsc_khz(void)
39{
40 struct pvclock_vcpu_time_info *info =
41 &HYPERVISOR_shared_info->vcpu_info[0].time;
42
43 setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
44 return pvclock_tsc_khz(info);
45}
46
47static u64 xen_clocksource_read(void)
48{
49 struct pvclock_vcpu_time_info *src;
50 u64 ret;
51
52 preempt_disable_notrace();
53 src = &__this_cpu_read(xen_vcpu)->time;
54 ret = pvclock_clocksource_read(src);
55 preempt_enable_notrace();
56 return ret;
57}
58
59static u64 xen_clocksource_get_cycles(struct clocksource *cs)
60{
61 return xen_clocksource_read();
62}
63
64static noinstr u64 xen_sched_clock(void)
65{
66 struct pvclock_vcpu_time_info *src;
67 u64 ret;
68
69 src = &__this_cpu_read(xen_vcpu)->time;
70 ret = pvclock_clocksource_read_nowd(src);
71 ret -= xen_sched_clock_offset;
72
73 return ret;
74}
75
76static void xen_read_wallclock(struct timespec64 *ts)
77{
78 struct shared_info *s = HYPERVISOR_shared_info;
79 struct pvclock_wall_clock *wall_clock = &(s->wc);
80 struct pvclock_vcpu_time_info *vcpu_time;
81
82 vcpu_time = &get_cpu_var(xen_vcpu)->time;
83 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
84 put_cpu_var(xen_vcpu);
85}
86
87static void xen_get_wallclock(struct timespec64 *now)
88{
89 xen_read_wallclock(now);
90}
91
92static int xen_set_wallclock(const struct timespec64 *now)
93{
94 return -ENODEV;
95}
96
97static int xen_pvclock_gtod_notify(struct notifier_block *nb,
98 unsigned long was_set, void *priv)
99{
100 /* Protected by the calling core code serialization */
101 static struct timespec64 next_sync;
102
103 struct xen_platform_op op;
104 struct timespec64 now;
105 struct timekeeper *tk = priv;
106 static bool settime64_supported = true;
107 int ret;
108
109 now.tv_sec = tk->xtime_sec;
110 now.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
111
112 /*
113 * We only take the expensive HV call when the clock was set
114 * or when the 11 minutes RTC synchronization time elapsed.
115 */
116 if (!was_set && timespec64_compare(&now, &next_sync) < 0)
117 return NOTIFY_OK;
118
119again:
120 if (settime64_supported) {
121 op.cmd = XENPF_settime64;
122 op.u.settime64.mbz = 0;
123 op.u.settime64.secs = now.tv_sec;
124 op.u.settime64.nsecs = now.tv_nsec;
125 op.u.settime64.system_time = xen_clocksource_read();
126 } else {
127 op.cmd = XENPF_settime32;
128 op.u.settime32.secs = now.tv_sec;
129 op.u.settime32.nsecs = now.tv_nsec;
130 op.u.settime32.system_time = xen_clocksource_read();
131 }
132
133 ret = HYPERVISOR_platform_op(&op);
134
135 if (ret == -ENOSYS && settime64_supported) {
136 settime64_supported = false;
137 goto again;
138 }
139 if (ret < 0)
140 return NOTIFY_BAD;
141
142 /*
143 * Move the next drift compensation time 11 minutes
144 * ahead. That's emulating the sync_cmos_clock() update for
145 * the hardware RTC.
146 */
147 next_sync = now;
148 next_sync.tv_sec += 11 * 60;
149
150 return NOTIFY_OK;
151}
152
153static struct notifier_block xen_pvclock_gtod_notifier = {
154 .notifier_call = xen_pvclock_gtod_notify,
155};
156
157static int xen_cs_enable(struct clocksource *cs)
158{
159 vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK);
160 return 0;
161}
162
163static struct clocksource xen_clocksource __read_mostly = {
164 .name = "xen",
165 .rating = 400,
166 .read = xen_clocksource_get_cycles,
167 .mask = CLOCKSOURCE_MASK(64),
168 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
169 .enable = xen_cs_enable,
170};
171
172/*
173 Xen clockevent implementation
174
175 Xen has two clockevent implementations:
176
177 The old timer_op one works with all released versions of Xen prior
178 to version 3.0.4. This version of the hypervisor provides a
179 single-shot timer with nanosecond resolution. However, sharing the
180 same event channel is a 100Hz tick which is delivered while the
181 vcpu is running. We don't care about or use this tick, but it will
182 cause the core time code to think the timer fired too soon, and
183 will end up resetting it each time. It could be filtered, but
184 doing so has complications when the ktime clocksource is not yet
185 the xen clocksource (ie, at boot time).
186
187 The new vcpu_op-based timer interface allows the tick timer period
188 to be changed or turned off. The tick timer is not useful as a
189 periodic timer because events are only delivered to running vcpus.
190 The one-shot timer can report when a timeout is in the past, so
191 set_next_event is capable of returning -ETIME when appropriate.
192 This interface is used when available.
193*/
194
195
196/*
197 Get a hypervisor absolute time. In theory we could maintain an
198 offset between the kernel's time and the hypervisor's time, and
199 apply that to a kernel's absolute timeout. Unfortunately the
200 hypervisor and kernel times can drift even if the kernel is using
201 the Xen clocksource, because ntp can warp the kernel's clocksource.
202*/
203static s64 get_abs_timeout(unsigned long delta)
204{
205 return xen_clocksource_read() + delta;
206}
207
208static int xen_timerop_shutdown(struct clock_event_device *evt)
209{
210 /* cancel timeout */
211 HYPERVISOR_set_timer_op(0);
212
213 return 0;
214}
215
216static int xen_timerop_set_next_event(unsigned long delta,
217 struct clock_event_device *evt)
218{
219 WARN_ON(!clockevent_state_oneshot(evt));
220
221 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
222 BUG();
223
224 /* We may have missed the deadline, but there's no real way of
225 knowing for sure. If the event was in the past, then we'll
226 get an immediate interrupt. */
227
228 return 0;
229}
230
231static struct clock_event_device xen_timerop_clockevent __ro_after_init = {
232 .name = "xen",
233 .features = CLOCK_EVT_FEAT_ONESHOT,
234
235 .max_delta_ns = 0xffffffff,
236 .max_delta_ticks = 0xffffffff,
237 .min_delta_ns = TIMER_SLOP,
238 .min_delta_ticks = TIMER_SLOP,
239
240 .mult = 1,
241 .shift = 0,
242 .rating = 500,
243
244 .set_state_shutdown = xen_timerop_shutdown,
245 .set_next_event = xen_timerop_set_next_event,
246};
247
248static int xen_vcpuop_shutdown(struct clock_event_device *evt)
249{
250 int cpu = smp_processor_id();
251
252 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, xen_vcpu_nr(cpu),
253 NULL) ||
254 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
255 NULL))
256 BUG();
257
258 return 0;
259}
260
261static int xen_vcpuop_set_oneshot(struct clock_event_device *evt)
262{
263 int cpu = smp_processor_id();
264
265 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
266 NULL))
267 BUG();
268
269 return 0;
270}
271
272static int xen_vcpuop_set_next_event(unsigned long delta,
273 struct clock_event_device *evt)
274{
275 int cpu = smp_processor_id();
276 struct vcpu_set_singleshot_timer single;
277 int ret;
278
279 WARN_ON(!clockevent_state_oneshot(evt));
280
281 single.timeout_abs_ns = get_abs_timeout(delta);
282 /* Get an event anyway, even if the timeout is already expired */
283 single.flags = 0;
284
285 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, xen_vcpu_nr(cpu),
286 &single);
287 BUG_ON(ret != 0);
288
289 return ret;
290}
291
292static struct clock_event_device xen_vcpuop_clockevent __ro_after_init = {
293 .name = "xen",
294 .features = CLOCK_EVT_FEAT_ONESHOT,
295
296 .max_delta_ns = 0xffffffff,
297 .max_delta_ticks = 0xffffffff,
298 .min_delta_ns = TIMER_SLOP,
299 .min_delta_ticks = TIMER_SLOP,
300
301 .mult = 1,
302 .shift = 0,
303 .rating = 500,
304
305 .set_state_shutdown = xen_vcpuop_shutdown,
306 .set_state_oneshot = xen_vcpuop_set_oneshot,
307 .set_next_event = xen_vcpuop_set_next_event,
308};
309
310static const struct clock_event_device *xen_clockevent =
311 &xen_timerop_clockevent;
312
313struct xen_clock_event_device {
314 struct clock_event_device evt;
315 char name[16];
316};
317static DEFINE_PER_CPU(struct xen_clock_event_device, xen_clock_events) = { .evt.irq = -1 };
318
319static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
320{
321 struct clock_event_device *evt = this_cpu_ptr(&xen_clock_events.evt);
322 irqreturn_t ret;
323
324 ret = IRQ_NONE;
325 if (evt->event_handler) {
326 evt->event_handler(evt);
327 ret = IRQ_HANDLED;
328 }
329
330 return ret;
331}
332
333void xen_teardown_timer(int cpu)
334{
335 struct clock_event_device *evt;
336 evt = &per_cpu(xen_clock_events, cpu).evt;
337
338 if (evt->irq >= 0) {
339 unbind_from_irqhandler(evt->irq, NULL);
340 evt->irq = -1;
341 }
342}
343
344void xen_setup_timer(int cpu)
345{
346 struct xen_clock_event_device *xevt = &per_cpu(xen_clock_events, cpu);
347 struct clock_event_device *evt = &xevt->evt;
348 int irq;
349
350 WARN(evt->irq >= 0, "IRQ%d for CPU%d is already allocated\n", evt->irq, cpu);
351 if (evt->irq >= 0)
352 xen_teardown_timer(cpu);
353
354 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
355
356 snprintf(xevt->name, sizeof(xevt->name), "timer%d", cpu);
357
358 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
359 IRQF_PERCPU|IRQF_NOBALANCING|IRQF_TIMER|
360 IRQF_FORCE_RESUME|IRQF_EARLY_RESUME,
361 xevt->name, NULL);
362 (void)xen_set_irq_priority(irq, XEN_IRQ_PRIORITY_MAX);
363
364 memcpy(evt, xen_clockevent, sizeof(*evt));
365
366 evt->cpumask = cpumask_of(cpu);
367 evt->irq = irq;
368}
369
370
371void xen_setup_cpu_clockevents(void)
372{
373 clockevents_register_device(this_cpu_ptr(&xen_clock_events.evt));
374}
375
376void xen_timer_resume(void)
377{
378 int cpu;
379
380 if (xen_clockevent != &xen_vcpuop_clockevent)
381 return;
382
383 for_each_online_cpu(cpu) {
384 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer,
385 xen_vcpu_nr(cpu), NULL))
386 BUG();
387 }
388}
389
390static struct pvclock_vsyscall_time_info *xen_clock __read_mostly;
391static u64 xen_clock_value_saved;
392
393void xen_save_time_memory_area(void)
394{
395 struct vcpu_register_time_memory_area t;
396 int ret;
397
398 xen_clock_value_saved = xen_clocksource_read() - xen_sched_clock_offset;
399
400 if (!xen_clock)
401 return;
402
403 t.addr.v = NULL;
404
405 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
406 if (ret != 0)
407 pr_notice("Cannot save secondary vcpu_time_info (err %d)",
408 ret);
409 else
410 clear_page(xen_clock);
411}
412
413void xen_restore_time_memory_area(void)
414{
415 struct vcpu_register_time_memory_area t;
416 int ret;
417
418 if (!xen_clock)
419 goto out;
420
421 t.addr.v = &xen_clock->pvti;
422
423 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
424
425 /*
426 * We don't disable VDSO_CLOCKMODE_PVCLOCK entirely if it fails to
427 * register the secondary time info with Xen or if we migrated to a
428 * host without the necessary flags. On both of these cases what
429 * happens is either process seeing a zeroed out pvti or seeing no
430 * PVCLOCK_TSC_STABLE_BIT bit set. Userspace checks the latter and
431 * if 0, it discards the data in pvti and fallbacks to a system
432 * call for a reliable timestamp.
433 */
434 if (ret != 0)
435 pr_notice("Cannot restore secondary vcpu_time_info (err %d)",
436 ret);
437
438out:
439 /* Need pvclock_resume() before using xen_clocksource_read(). */
440 pvclock_resume();
441 xen_sched_clock_offset = xen_clocksource_read() - xen_clock_value_saved;
442}
443
444static void xen_setup_vsyscall_time_info(void)
445{
446 struct vcpu_register_time_memory_area t;
447 struct pvclock_vsyscall_time_info *ti;
448 int ret;
449
450 ti = (struct pvclock_vsyscall_time_info *)get_zeroed_page(GFP_KERNEL);
451 if (!ti)
452 return;
453
454 t.addr.v = &ti->pvti;
455
456 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
457 if (ret) {
458 pr_notice("xen: VDSO_CLOCKMODE_PVCLOCK not supported (err %d)\n", ret);
459 free_page((unsigned long)ti);
460 return;
461 }
462
463 /*
464 * If primary time info had this bit set, secondary should too since
465 * it's the same data on both just different memory regions. But we
466 * still check it in case hypervisor is buggy.
467 */
468 if (!(ti->pvti.flags & PVCLOCK_TSC_STABLE_BIT)) {
469 t.addr.v = NULL;
470 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area,
471 0, &t);
472 if (!ret)
473 free_page((unsigned long)ti);
474
475 pr_notice("xen: VDSO_CLOCKMODE_PVCLOCK not supported (tsc unstable)\n");
476 return;
477 }
478
479 xen_clock = ti;
480 pvclock_set_pvti_cpu0_va(xen_clock);
481
482 xen_clocksource.vdso_clock_mode = VDSO_CLOCKMODE_PVCLOCK;
483}
484
485/*
486 * Check if it is possible to safely use the tsc as a clocksource. This is
487 * only true if the hypervisor notifies the guest that its tsc is invariant,
488 * the tsc is stable, and the tsc instruction will never be emulated.
489 */
490static int __init xen_tsc_safe_clocksource(void)
491{
492 u32 eax, ebx, ecx, edx;
493
494 if (!(boot_cpu_has(X86_FEATURE_CONSTANT_TSC)))
495 return 0;
496
497 if (!(boot_cpu_has(X86_FEATURE_NONSTOP_TSC)))
498 return 0;
499
500 if (check_tsc_unstable())
501 return 0;
502
503 /* Leaf 4, sub-leaf 0 (0x40000x03) */
504 cpuid_count(xen_cpuid_base() + 3, 0, &eax, &ebx, &ecx, &edx);
505
506 return ebx == XEN_CPUID_TSC_MODE_NEVER_EMULATE;
507}
508
509static void __init xen_time_init(void)
510{
511 struct pvclock_vcpu_time_info *pvti;
512 int cpu = smp_processor_id();
513 struct timespec64 tp;
514
515 /*
516 * As Dom0 is never moved, no penalty on using TSC there.
517 *
518 * If it is possible for the guest to determine that the tsc is a safe
519 * clocksource, then set xen_clocksource rating below that of the tsc
520 * so that the system prefers tsc instead.
521 */
522 if (xen_initial_domain())
523 xen_clocksource.rating = 275;
524 else if (xen_tsc_safe_clocksource())
525 xen_clocksource.rating = 299;
526
527 clocksource_register_hz(&xen_clocksource, NSEC_PER_SEC);
528
529 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
530 NULL) == 0) {
531 /* Successfully turned off 100Hz tick, so we have the
532 vcpuop-based timer interface */
533 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
534 xen_clockevent = &xen_vcpuop_clockevent;
535 }
536
537 /* Set initial system time with full resolution */
538 xen_read_wallclock(&tp);
539 do_settimeofday64(&tp);
540
541 setup_force_cpu_cap(X86_FEATURE_TSC);
542
543 /*
544 * We check ahead on the primary time info if this
545 * bit is supported hence speeding up Xen clocksource.
546 */
547 pvti = &__this_cpu_read(xen_vcpu)->time;
548 if (pvti->flags & PVCLOCK_TSC_STABLE_BIT) {
549 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
550 xen_setup_vsyscall_time_info();
551 }
552
553 xen_setup_runstate_info(cpu);
554 xen_setup_timer(cpu);
555 xen_setup_cpu_clockevents();
556
557 xen_time_setup_guest();
558
559 if (xen_initial_domain())
560 pvclock_gtod_register_notifier(&xen_pvclock_gtod_notifier);
561}
562
563static void __init xen_init_time_common(void)
564{
565 xen_sched_clock_offset = xen_clocksource_read();
566 static_call_update(pv_steal_clock, xen_steal_clock);
567 paravirt_set_sched_clock(xen_sched_clock);
568
569 x86_platform.calibrate_tsc = xen_tsc_khz;
570 x86_platform.get_wallclock = xen_get_wallclock;
571}
572
573void __init xen_init_time_ops(void)
574{
575 xen_init_time_common();
576
577 x86_init.timers.timer_init = xen_time_init;
578 x86_init.timers.setup_percpu_clockev = x86_init_noop;
579 x86_cpuinit.setup_percpu_clockev = x86_init_noop;
580
581 /* Dom0 uses the native method to set the hardware RTC. */
582 if (!xen_initial_domain())
583 x86_platform.set_wallclock = xen_set_wallclock;
584}
585
586#ifdef CONFIG_XEN_PVHVM
587static void xen_hvm_setup_cpu_clockevents(void)
588{
589 int cpu = smp_processor_id();
590 xen_setup_runstate_info(cpu);
591 /*
592 * xen_setup_timer(cpu) - snprintf is bad in atomic context. Hence
593 * doing it xen_hvm_cpu_notify (which gets called by smp_init during
594 * early bootup and also during CPU hotplug events).
595 */
596 xen_setup_cpu_clockevents();
597}
598
599void __init xen_hvm_init_time_ops(void)
600{
601 static bool hvm_time_initialized;
602
603 if (hvm_time_initialized)
604 return;
605
606 /*
607 * vector callback is needed otherwise we cannot receive interrupts
608 * on cpu > 0 and at this point we don't know how many cpus are
609 * available.
610 */
611 if (!xen_have_vector_callback)
612 return;
613
614 if (!xen_feature(XENFEAT_hvm_safe_pvclock)) {
615 pr_info_once("Xen doesn't support pvclock on HVM, disable pv timer");
616 return;
617 }
618
619 /*
620 * Only MAX_VIRT_CPUS 'vcpu_info' are embedded inside 'shared_info'.
621 * The __this_cpu_read(xen_vcpu) is still NULL when Xen HVM guest
622 * boots on vcpu >= MAX_VIRT_CPUS (e.g., kexec), To access
623 * __this_cpu_read(xen_vcpu) via xen_clocksource_read() will panic.
624 *
625 * The xen_hvm_init_time_ops() should be called again later after
626 * __this_cpu_read(xen_vcpu) is available.
627 */
628 if (!__this_cpu_read(xen_vcpu)) {
629 pr_info("Delay xen_init_time_common() as kernel is running on vcpu=%d\n",
630 xen_vcpu_nr(0));
631 return;
632 }
633
634 xen_init_time_common();
635
636 x86_init.timers.setup_percpu_clockev = xen_time_init;
637 x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents;
638
639 x86_platform.set_wallclock = xen_set_wallclock;
640
641 hvm_time_initialized = true;
642}
643#endif
644
645/* Kernel parameter to specify Xen timer slop */
646static int __init parse_xen_timer_slop(char *ptr)
647{
648 unsigned long slop = memparse(ptr, NULL);
649
650 xen_timerop_clockevent.min_delta_ns = slop;
651 xen_timerop_clockevent.min_delta_ticks = slop;
652 xen_vcpuop_clockevent.min_delta_ns = slop;
653 xen_vcpuop_clockevent.min_delta_ticks = slop;
654
655 return 0;
656}
657early_param("xen_timer_slop", parse_xen_timer_slop);