Commit | Line | Data |
---|---|---|
15c84731 JF |
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> | |
5a0e3ad6 | 14 | #include <linux/gfp.h> |
c9d76a24 | 15 | #include <linux/slab.h> |
5584880e | 16 | #include <linux/pvclock_gtod.h> |
76096863 | 17 | #include <linux/timekeeper_internal.h> |
15c84731 | 18 | |
1c7b67f7 | 19 | #include <asm/pvclock.h> |
15c84731 JF |
20 | #include <asm/xen/hypervisor.h> |
21 | #include <asm/xen/hypercall.h> | |
22 | ||
23 | #include <xen/events.h> | |
409771d2 | 24 | #include <xen/features.h> |
15c84731 JF |
25 | #include <xen/interface/xen.h> |
26 | #include <xen/interface/vcpu.h> | |
27 | ||
28 | #include "xen-ops.h" | |
29 | ||
15c84731 JF |
30 | /* Xen may fire a timer up to this many ns early */ |
31 | #define TIMER_SLOP 100000 | |
f91a8b44 | 32 | |
e93ef949 | 33 | /* Get the TSC speed from Xen */ |
409771d2 | 34 | static unsigned long xen_tsc_khz(void) |
15c84731 | 35 | { |
3807f345 | 36 | struct pvclock_vcpu_time_info *info = |
15c84731 JF |
37 | &HYPERVISOR_shared_info->vcpu_info[0].time; |
38 | ||
3807f345 | 39 | return pvclock_tsc_khz(info); |
15c84731 JF |
40 | } |
41 | ||
a5a1d1c2 | 42 | u64 xen_clocksource_read(void) |
15c84731 | 43 | { |
1c7b67f7 | 44 | struct pvclock_vcpu_time_info *src; |
a5a1d1c2 | 45 | u64 ret; |
15c84731 | 46 | |
f1c39625 | 47 | preempt_disable_notrace(); |
3251f20b | 48 | src = &__this_cpu_read(xen_vcpu)->time; |
1c7b67f7 | 49 | ret = pvclock_clocksource_read(src); |
f1c39625 | 50 | preempt_enable_notrace(); |
15c84731 JF |
51 | return ret; |
52 | } | |
53 | ||
a5a1d1c2 | 54 | static u64 xen_clocksource_get_cycles(struct clocksource *cs) |
8e19608e MD |
55 | { |
56 | return xen_clocksource_read(); | |
57 | } | |
58 | ||
15c84731 JF |
59 | static void xen_read_wallclock(struct timespec *ts) |
60 | { | |
1c7b67f7 GH |
61 | struct shared_info *s = HYPERVISOR_shared_info; |
62 | struct pvclock_wall_clock *wall_clock = &(s->wc); | |
63 | struct pvclock_vcpu_time_info *vcpu_time; | |
15c84731 | 64 | |
1c7b67f7 GH |
65 | vcpu_time = &get_cpu_var(xen_vcpu)->time; |
66 | pvclock_read_wallclock(wall_clock, vcpu_time, ts); | |
67 | put_cpu_var(xen_vcpu); | |
15c84731 JF |
68 | } |
69 | ||
3565184e | 70 | static void xen_get_wallclock(struct timespec *now) |
15c84731 | 71 | { |
3565184e | 72 | xen_read_wallclock(now); |
15c84731 | 73 | } |
15c84731 | 74 | |
3565184e | 75 | static int xen_set_wallclock(const struct timespec *now) |
15c84731 | 76 | { |
47433b8c | 77 | return -1; |
15c84731 JF |
78 | } |
79 | ||
47433b8c DV |
80 | static int xen_pvclock_gtod_notify(struct notifier_block *nb, |
81 | unsigned long was_set, void *priv) | |
15c84731 | 82 | { |
47433b8c | 83 | /* Protected by the calling core code serialization */ |
187b26a9 | 84 | static struct timespec64 next_sync; |
5584880e | 85 | |
fdb9eb9f | 86 | struct xen_platform_op op; |
76096863 SS |
87 | struct timespec64 now; |
88 | struct timekeeper *tk = priv; | |
89 | static bool settime64_supported = true; | |
90 | int ret; | |
fdb9eb9f | 91 | |
76096863 SS |
92 | now.tv_sec = tk->xtime_sec; |
93 | now.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift); | |
5584880e | 94 | |
47433b8c DV |
95 | /* |
96 | * We only take the expensive HV call when the clock was set | |
97 | * or when the 11 minutes RTC synchronization time elapsed. | |
98 | */ | |
187b26a9 | 99 | if (!was_set && timespec64_compare(&now, &next_sync) < 0) |
47433b8c | 100 | return NOTIFY_OK; |
fdb9eb9f | 101 | |
76096863 SS |
102 | again: |
103 | if (settime64_supported) { | |
104 | op.cmd = XENPF_settime64; | |
105 | op.u.settime64.mbz = 0; | |
106 | op.u.settime64.secs = now.tv_sec; | |
107 | op.u.settime64.nsecs = now.tv_nsec; | |
108 | op.u.settime64.system_time = xen_clocksource_read(); | |
109 | } else { | |
110 | op.cmd = XENPF_settime32; | |
111 | op.u.settime32.secs = now.tv_sec; | |
112 | op.u.settime32.nsecs = now.tv_nsec; | |
113 | op.u.settime32.system_time = xen_clocksource_read(); | |
114 | } | |
115 | ||
116 | ret = HYPERVISOR_platform_op(&op); | |
117 | ||
118 | if (ret == -ENOSYS && settime64_supported) { | |
119 | settime64_supported = false; | |
120 | goto again; | |
121 | } | |
122 | if (ret < 0) | |
123 | return NOTIFY_BAD; | |
fdb9eb9f | 124 | |
47433b8c DV |
125 | /* |
126 | * Move the next drift compensation time 11 minutes | |
127 | * ahead. That's emulating the sync_cmos_clock() update for | |
128 | * the hardware RTC. | |
129 | */ | |
130 | next_sync = now; | |
131 | next_sync.tv_sec += 11 * 60; | |
132 | ||
5584880e | 133 | return NOTIFY_OK; |
15c84731 JF |
134 | } |
135 | ||
5584880e DV |
136 | static struct notifier_block xen_pvclock_gtod_notifier = { |
137 | .notifier_call = xen_pvclock_gtod_notify, | |
138 | }; | |
139 | ||
15c84731 JF |
140 | static struct clocksource xen_clocksource __read_mostly = { |
141 | .name = "xen", | |
142 | .rating = 400, | |
8e19608e | 143 | .read = xen_clocksource_get_cycles, |
15c84731 | 144 | .mask = ~0, |
15c84731 JF |
145 | .flags = CLOCK_SOURCE_IS_CONTINUOUS, |
146 | }; | |
147 | ||
148 | /* | |
149 | Xen clockevent implementation | |
150 | ||
151 | Xen has two clockevent implementations: | |
152 | ||
153 | The old timer_op one works with all released versions of Xen prior | |
154 | to version 3.0.4. This version of the hypervisor provides a | |
155 | single-shot timer with nanosecond resolution. However, sharing the | |
156 | same event channel is a 100Hz tick which is delivered while the | |
157 | vcpu is running. We don't care about or use this tick, but it will | |
158 | cause the core time code to think the timer fired too soon, and | |
159 | will end up resetting it each time. It could be filtered, but | |
160 | doing so has complications when the ktime clocksource is not yet | |
161 | the xen clocksource (ie, at boot time). | |
162 | ||
163 | The new vcpu_op-based timer interface allows the tick timer period | |
164 | to be changed or turned off. The tick timer is not useful as a | |
165 | periodic timer because events are only delivered to running vcpus. | |
166 | The one-shot timer can report when a timeout is in the past, so | |
167 | set_next_event is capable of returning -ETIME when appropriate. | |
168 | This interface is used when available. | |
169 | */ | |
170 | ||
171 | ||
172 | /* | |
173 | Get a hypervisor absolute time. In theory we could maintain an | |
174 | offset between the kernel's time and the hypervisor's time, and | |
175 | apply that to a kernel's absolute timeout. Unfortunately the | |
176 | hypervisor and kernel times can drift even if the kernel is using | |
177 | the Xen clocksource, because ntp can warp the kernel's clocksource. | |
178 | */ | |
179 | static s64 get_abs_timeout(unsigned long delta) | |
180 | { | |
181 | return xen_clocksource_read() + delta; | |
182 | } | |
183 | ||
955381dd | 184 | static int xen_timerop_shutdown(struct clock_event_device *evt) |
15c84731 | 185 | { |
955381dd VK |
186 | /* cancel timeout */ |
187 | HYPERVISOR_set_timer_op(0); | |
188 | ||
189 | return 0; | |
15c84731 JF |
190 | } |
191 | ||
192 | static int xen_timerop_set_next_event(unsigned long delta, | |
193 | struct clock_event_device *evt) | |
194 | { | |
955381dd | 195 | WARN_ON(!clockevent_state_oneshot(evt)); |
15c84731 JF |
196 | |
197 | if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0) | |
198 | BUG(); | |
199 | ||
200 | /* We may have missed the deadline, but there's no real way of | |
201 | knowing for sure. If the event was in the past, then we'll | |
202 | get an immediate interrupt. */ | |
203 | ||
204 | return 0; | |
205 | } | |
206 | ||
207 | static const struct clock_event_device xen_timerop_clockevent = { | |
955381dd VK |
208 | .name = "xen", |
209 | .features = CLOCK_EVT_FEAT_ONESHOT, | |
15c84731 | 210 | |
955381dd | 211 | .max_delta_ns = 0xffffffff, |
3d18d661 | 212 | .max_delta_ticks = 0xffffffff, |
955381dd | 213 | .min_delta_ns = TIMER_SLOP, |
3d18d661 | 214 | .min_delta_ticks = TIMER_SLOP, |
15c84731 | 215 | |
955381dd VK |
216 | .mult = 1, |
217 | .shift = 0, | |
218 | .rating = 500, | |
15c84731 | 219 | |
955381dd VK |
220 | .set_state_shutdown = xen_timerop_shutdown, |
221 | .set_next_event = xen_timerop_set_next_event, | |
15c84731 JF |
222 | }; |
223 | ||
955381dd VK |
224 | static int xen_vcpuop_shutdown(struct clock_event_device *evt) |
225 | { | |
226 | int cpu = smp_processor_id(); | |
15c84731 | 227 | |
ad5475f9 VK |
228 | if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, xen_vcpu_nr(cpu), |
229 | NULL) || | |
230 | HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu), | |
231 | NULL)) | |
955381dd | 232 | BUG(); |
15c84731 | 233 | |
955381dd VK |
234 | return 0; |
235 | } | |
236 | ||
237 | static int xen_vcpuop_set_oneshot(struct clock_event_device *evt) | |
15c84731 JF |
238 | { |
239 | int cpu = smp_processor_id(); | |
240 | ||
ad5475f9 VK |
241 | if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu), |
242 | NULL)) | |
955381dd VK |
243 | BUG(); |
244 | ||
245 | return 0; | |
15c84731 JF |
246 | } |
247 | ||
248 | static int xen_vcpuop_set_next_event(unsigned long delta, | |
249 | struct clock_event_device *evt) | |
250 | { | |
251 | int cpu = smp_processor_id(); | |
252 | struct vcpu_set_singleshot_timer single; | |
253 | int ret; | |
254 | ||
955381dd | 255 | WARN_ON(!clockevent_state_oneshot(evt)); |
15c84731 JF |
256 | |
257 | single.timeout_abs_ns = get_abs_timeout(delta); | |
c06b6d70 SS |
258 | /* Get an event anyway, even if the timeout is already expired */ |
259 | single.flags = 0; | |
15c84731 | 260 | |
ad5475f9 VK |
261 | ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, xen_vcpu_nr(cpu), |
262 | &single); | |
c06b6d70 | 263 | BUG_ON(ret != 0); |
15c84731 JF |
264 | |
265 | return ret; | |
266 | } | |
267 | ||
268 | static const struct clock_event_device xen_vcpuop_clockevent = { | |
269 | .name = "xen", | |
270 | .features = CLOCK_EVT_FEAT_ONESHOT, | |
271 | ||
272 | .max_delta_ns = 0xffffffff, | |
3d18d661 | 273 | .max_delta_ticks = 0xffffffff, |
15c84731 | 274 | .min_delta_ns = TIMER_SLOP, |
3d18d661 | 275 | .min_delta_ticks = TIMER_SLOP, |
15c84731 JF |
276 | |
277 | .mult = 1, | |
278 | .shift = 0, | |
279 | .rating = 500, | |
280 | ||
955381dd VK |
281 | .set_state_shutdown = xen_vcpuop_shutdown, |
282 | .set_state_oneshot = xen_vcpuop_set_oneshot, | |
15c84731 JF |
283 | .set_next_event = xen_vcpuop_set_next_event, |
284 | }; | |
285 | ||
286 | static const struct clock_event_device *xen_clockevent = | |
287 | &xen_timerop_clockevent; | |
31620a19 KRW |
288 | |
289 | struct xen_clock_event_device { | |
290 | struct clock_event_device evt; | |
7be0772d | 291 | char name[16]; |
31620a19 KRW |
292 | }; |
293 | static DEFINE_PER_CPU(struct xen_clock_event_device, xen_clock_events) = { .evt.irq = -1 }; | |
15c84731 JF |
294 | |
295 | static irqreturn_t xen_timer_interrupt(int irq, void *dev_id) | |
296 | { | |
89cbc767 | 297 | struct clock_event_device *evt = this_cpu_ptr(&xen_clock_events.evt); |
15c84731 JF |
298 | irqreturn_t ret; |
299 | ||
300 | ret = IRQ_NONE; | |
301 | if (evt->event_handler) { | |
302 | evt->event_handler(evt); | |
303 | ret = IRQ_HANDLED; | |
304 | } | |
305 | ||
306 | return ret; | |
307 | } | |
308 | ||
09e99da7 KRW |
309 | void xen_teardown_timer(int cpu) |
310 | { | |
311 | struct clock_event_device *evt; | |
09e99da7 KRW |
312 | evt = &per_cpu(xen_clock_events, cpu).evt; |
313 | ||
314 | if (evt->irq >= 0) { | |
315 | unbind_from_irqhandler(evt->irq, NULL); | |
316 | evt->irq = -1; | |
09e99da7 KRW |
317 | } |
318 | } | |
319 | ||
f87e4cac | 320 | void xen_setup_timer(int cpu) |
15c84731 | 321 | { |
7be0772d VK |
322 | struct xen_clock_event_device *xevt = &per_cpu(xen_clock_events, cpu); |
323 | struct clock_event_device *evt = &xevt->evt; | |
15c84731 JF |
324 | int irq; |
325 | ||
ef35a4e6 | 326 | WARN(evt->irq >= 0, "IRQ%d for CPU%d is already allocated\n", evt->irq, cpu); |
09e99da7 KRW |
327 | if (evt->irq >= 0) |
328 | xen_teardown_timer(cpu); | |
ef35a4e6 | 329 | |
15c84731 JF |
330 | printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu); |
331 | ||
7be0772d | 332 | snprintf(xevt->name, sizeof(xevt->name), "timer%d", cpu); |
15c84731 JF |
333 | |
334 | irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt, | |
9d71cee6 | 335 | IRQF_PERCPU|IRQF_NOBALANCING|IRQF_TIMER| |
8d5999df | 336 | IRQF_FORCE_RESUME|IRQF_EARLY_RESUME, |
7be0772d | 337 | xevt->name, NULL); |
8785c676 | 338 | (void)xen_set_irq_priority(irq, XEN_IRQ_PRIORITY_MAX); |
15c84731 | 339 | |
15c84731 JF |
340 | memcpy(evt, xen_clockevent, sizeof(*evt)); |
341 | ||
320ab2b0 | 342 | evt->cpumask = cpumask_of(cpu); |
15c84731 | 343 | evt->irq = irq; |
f87e4cac JF |
344 | } |
345 | ||
d68d82af | 346 | |
f87e4cac JF |
347 | void xen_setup_cpu_clockevents(void) |
348 | { | |
89cbc767 | 349 | clockevents_register_device(this_cpu_ptr(&xen_clock_events.evt)); |
15c84731 JF |
350 | } |
351 | ||
d07af1f0 JF |
352 | void xen_timer_resume(void) |
353 | { | |
354 | int cpu; | |
355 | ||
e7a3481c JF |
356 | pvclock_resume(); |
357 | ||
d07af1f0 JF |
358 | if (xen_clockevent != &xen_vcpuop_clockevent) |
359 | return; | |
360 | ||
361 | for_each_online_cpu(cpu) { | |
ad5475f9 VK |
362 | if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, |
363 | xen_vcpu_nr(cpu), NULL)) | |
d07af1f0 JF |
364 | BUG(); |
365 | } | |
366 | } | |
367 | ||
fb6ce5de | 368 | static const struct pv_time_ops xen_time_ops __initconst = { |
ca50a5f3 | 369 | .sched_clock = xen_clocksource_read, |
d34c30cc | 370 | .steal_clock = xen_steal_clock, |
409771d2 SS |
371 | }; |
372 | ||
fb6ce5de | 373 | static void __init xen_time_init(void) |
15c84731 JF |
374 | { |
375 | int cpu = smp_processor_id(); | |
c4507257 | 376 | struct timespec tp; |
15c84731 | 377 | |
94dd85f6 PI |
378 | /* As Dom0 is never moved, no penalty on using TSC there */ |
379 | if (xen_initial_domain()) | |
380 | xen_clocksource.rating = 275; | |
381 | ||
b01cc1b0 | 382 | clocksource_register_hz(&xen_clocksource, NSEC_PER_SEC); |
15c84731 | 383 | |
ad5475f9 VK |
384 | if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu), |
385 | NULL) == 0) { | |
f91a8b44 | 386 | /* Successfully turned off 100Hz tick, so we have the |
15c84731 JF |
387 | vcpuop-based timer interface */ |
388 | printk(KERN_DEBUG "Xen: using vcpuop timer interface\n"); | |
389 | xen_clockevent = &xen_vcpuop_clockevent; | |
390 | } | |
391 | ||
392 | /* Set initial system time with full resolution */ | |
c4507257 JS |
393 | xen_read_wallclock(&tp); |
394 | do_settimeofday(&tp); | |
15c84731 | 395 | |
404ee5b1 | 396 | setup_force_cpu_cap(X86_FEATURE_TSC); |
15c84731 | 397 | |
be012920 | 398 | xen_setup_runstate_info(cpu); |
15c84731 | 399 | xen_setup_timer(cpu); |
f87e4cac | 400 | xen_setup_cpu_clockevents(); |
5584880e | 401 | |
ecb23dc6 JG |
402 | xen_time_setup_guest(); |
403 | ||
5584880e DV |
404 | if (xen_initial_domain()) |
405 | pvclock_gtod_register_notifier(&xen_pvclock_gtod_notifier); | |
15c84731 | 406 | } |
409771d2 | 407 | |
d162809f | 408 | void __ref xen_init_time_ops(void) |
409771d2 SS |
409 | { |
410 | pv_time_ops = xen_time_ops; | |
411 | ||
412 | x86_init.timers.timer_init = xen_time_init; | |
413 | x86_init.timers.setup_percpu_clockev = x86_init_noop; | |
414 | x86_cpuinit.setup_percpu_clockev = x86_init_noop; | |
415 | ||
416 | x86_platform.calibrate_tsc = xen_tsc_khz; | |
417 | x86_platform.get_wallclock = xen_get_wallclock; | |
47433b8c DV |
418 | /* Dom0 uses the native method to set the hardware RTC. */ |
419 | if (!xen_initial_domain()) | |
420 | x86_platform.set_wallclock = xen_set_wallclock; | |
409771d2 SS |
421 | } |
422 | ||
ca65f9fc | 423 | #ifdef CONFIG_XEN_PVHVM |
409771d2 SS |
424 | static void xen_hvm_setup_cpu_clockevents(void) |
425 | { | |
426 | int cpu = smp_processor_id(); | |
427 | xen_setup_runstate_info(cpu); | |
7918c92a KRW |
428 | /* |
429 | * xen_setup_timer(cpu) - snprintf is bad in atomic context. Hence | |
430 | * doing it xen_hvm_cpu_notify (which gets called by smp_init during | |
431 | * early bootup and also during CPU hotplug events). | |
432 | */ | |
409771d2 SS |
433 | xen_setup_cpu_clockevents(); |
434 | } | |
435 | ||
fb6ce5de | 436 | void __init xen_hvm_init_time_ops(void) |
409771d2 | 437 | { |
84d582d2 BO |
438 | /* |
439 | * vector callback is needed otherwise we cannot receive interrupts | |
440 | * on cpu > 0 and at this point we don't know how many cpus are | |
441 | * available. | |
442 | */ | |
443 | if (!xen_have_vector_callback) | |
444 | return; | |
445 | ||
409771d2 SS |
446 | if (!xen_feature(XENFEAT_hvm_safe_pvclock)) { |
447 | printk(KERN_INFO "Xen doesn't support pvclock on HVM," | |
448 | "disable pv timer\n"); | |
449 | return; | |
450 | } | |
451 | ||
452 | pv_time_ops = xen_time_ops; | |
453 | x86_init.timers.setup_percpu_clockev = xen_time_init; | |
454 | x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents; | |
455 | ||
456 | x86_platform.calibrate_tsc = xen_tsc_khz; | |
457 | x86_platform.get_wallclock = xen_get_wallclock; | |
458 | x86_platform.set_wallclock = xen_set_wallclock; | |
459 | } | |
ca65f9fc | 460 | #endif |