1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2009 Adrian Chadd
5 * Copyright (c) 2012 Spectra Logic Corporation
6 * All rights reserved.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 *
29 */
30
31 /**
32 * \file dev/xen/timer/timer.c
33 * \brief A timer driver for the Xen hypervisor's PV clock.
34 */
35
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/bus.h>
42 #include <sys/kernel.h>
43 #include <sys/module.h>
44 #include <sys/time.h>
45 #include <sys/timetc.h>
46 #include <sys/timeet.h>
47 #include <sys/smp.h>
48 #include <sys/limits.h>
49 #include <sys/clock.h>
50 #include <sys/proc.h>
51
52 #include <xen/xen-os.h>
53 #include <xen/features.h>
54 #include <xen/xen_intr.h>
55 #include <xen/hypervisor.h>
56 #include <xen/interface/io/xenbus.h>
57 #include <xen/interface/vcpu.h>
58 #include <xen/error.h>
59
60 #include <machine/cpu.h>
61 #include <machine/cpufunc.h>
62 #include <machine/clock.h>
63 #include <machine/_inttypes.h>
64 #include <machine/smp.h>
65 #include <machine/pvclock.h>
66
67 #include <dev/xen/timer/timer.h>
68
69 #include "clock_if.h"
70
71 static devclass_t xentimer_devclass;
72
73 #define NSEC_IN_SEC 1000000000ULL
74 #define NSEC_IN_USEC 1000ULL
75 /* 18446744073 = int(2^64 / NSEC_IN_SC) = 1 ns in 64-bit fractions */
76 #define FRAC_IN_NSEC 18446744073LL
77
78 /* Xen timers may fire up to 100us off */
79 #define XENTIMER_MIN_PERIOD_IN_NSEC 100*NSEC_IN_USEC
80
81 /*
82 * The real resolution of the PV clock is 1ns, but the highest
83 * resolution that FreeBSD supports is 1us, so just use that.
84 */
85 #define XENCLOCK_RESOLUTION 1
86
87 #define XENTIMER_QUALITY 950
88
89 struct xentimer_pcpu_data {
90 uint64_t timer;
91 uint64_t last_processed;
92 void *irq_handle;
93 };
94
95 DPCPU_DEFINE(struct xentimer_pcpu_data, xentimer_pcpu);
96
97 DPCPU_DECLARE(struct vcpu_info *, vcpu_info);
98
99 struct xentimer_softc {
100 device_t dev;
101 struct timecounter tc;
102 struct eventtimer et;
103 };
104
105 static void
xentimer_identify(driver_t * driver,device_t parent)106 xentimer_identify(driver_t *driver, device_t parent)
107 {
108 if (!xen_domain())
109 return;
110
111 /* Handle all Xen PV timers in one device instance. */
112 if (devclass_get_device(xentimer_devclass, 0))
113 return;
114
115 BUS_ADD_CHILD(parent, 0, "xen_et", 0);
116 }
117
118 static int
xentimer_probe(device_t dev)119 xentimer_probe(device_t dev)
120 {
121 KASSERT((xen_domain()), ("Trying to use Xen timer on bare metal"));
122 /*
123 * In order to attach, this driver requires the following:
124 * - Vector callback support by the hypervisor, in order to deliver
125 * timer interrupts to the correct CPU for CPUs other than 0.
126 * - Access to the hypervisor shared info page, in order to look up
127 * each VCPU's timer information and the Xen wallclock time.
128 * - The hypervisor must say its PV clock is "safe" to use.
129 * - The hypervisor must support VCPUOP hypercalls.
130 * - The maximum number of CPUs supported by FreeBSD must not exceed
131 * the number of VCPUs supported by the hypervisor.
132 */
133 #define XTREQUIRES(condition, reason...) \
134 if (!(condition)) { \
135 device_printf(dev, ## reason); \
136 device_detach(dev); \
137 return (ENXIO); \
138 }
139
140 if (xen_hvm_domain()) {
141 XTREQUIRES(xen_vector_callback_enabled,
142 "vector callbacks unavailable\n");
143 XTREQUIRES(xen_feature(XENFEAT_hvm_safe_pvclock),
144 "HVM safe pvclock unavailable\n");
145 }
146 XTREQUIRES(HYPERVISOR_shared_info != NULL,
147 "shared info page unavailable\n");
148 XTREQUIRES(HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, 0, NULL) == 0,
149 "VCPUOPs interface unavailable\n");
150 #undef XTREQUIRES
151 device_set_desc(dev, "Xen PV Clock");
152 return (BUS_PROBE_NOWILDCARD);
153 }
154
155 /**
156 * \brief Get the current time, in nanoseconds, since the hypervisor booted.
157 *
158 * \param vcpu vcpu_info structure to fetch the time from.
159 *
160 */
161 static uint64_t
xen_fetch_vcpu_time(struct vcpu_info * vcpu)162 xen_fetch_vcpu_time(struct vcpu_info *vcpu)
163 {
164 struct pvclock_vcpu_time_info *time;
165
166 time = (struct pvclock_vcpu_time_info *) &vcpu->time;
167
168 return (pvclock_get_timecount(time));
169 }
170
171 static uint32_t
xentimer_get_timecount(struct timecounter * tc)172 xentimer_get_timecount(struct timecounter *tc)
173 {
174 uint64_t vcpu_time;
175
176 /*
177 * We don't disable preemption here because the worst that can
178 * happen is reading the vcpu_info area of a different CPU than
179 * the one we are currently running on, but that would also
180 * return a valid tc (and we avoid the overhead of
181 * critical_{enter/exit} calls).
182 */
183 vcpu_time = xen_fetch_vcpu_time(DPCPU_GET(vcpu_info));
184
185 return (vcpu_time & UINT32_MAX);
186 }
187
188 /**
189 * \brief Fetch the hypervisor boot time, known as the "Xen wallclock".
190 *
191 * \param ts Timespec to store the current stable value.
192 * \param version Pointer to store the corresponding wallclock version.
193 *
194 * \note This value is updated when Domain-0 shifts its clock to follow
195 * clock drift, e.g. as detected by NTP.
196 */
197 static void
xen_fetch_wallclock(struct timespec * ts)198 xen_fetch_wallclock(struct timespec *ts)
199 {
200 shared_info_t *src = HYPERVISOR_shared_info;
201 struct pvclock_wall_clock *wc;
202
203 wc = (struct pvclock_wall_clock *) &src->wc_version;
204
205 pvclock_get_wallclock(wc, ts);
206 }
207
208 static void
xen_fetch_uptime(struct timespec * ts)209 xen_fetch_uptime(struct timespec *ts)
210 {
211 uint64_t uptime;
212
213 uptime = xen_fetch_vcpu_time(DPCPU_GET(vcpu_info));
214
215 ts->tv_sec = uptime / NSEC_IN_SEC;
216 ts->tv_nsec = uptime % NSEC_IN_SEC;
217 }
218
219 static int
xentimer_settime(device_t dev __unused,struct timespec * ts)220 xentimer_settime(device_t dev __unused, struct timespec *ts)
221 {
222 struct xen_platform_op settime;
223 int ret;
224
225 /*
226 * Don't return EINVAL here; just silently fail if the domain isn't
227 * privileged enough to set the TOD.
228 */
229 if (!xen_initial_domain())
230 return (0);
231
232 settime.cmd = XENPF_settime64;
233 settime.u.settime64.mbz = 0;
234 settime.u.settime64.secs = ts->tv_sec;
235 settime.u.settime64.nsecs = ts->tv_nsec;
236 settime.u.settime64.system_time =
237 xen_fetch_vcpu_time(DPCPU_GET(vcpu_info));
238
239 ret = HYPERVISOR_platform_op(&settime);
240 ret = ret != 0 ? xen_translate_error(ret) : 0;
241 if (ret != 0 && bootverbose)
242 device_printf(dev, "failed to set Xen PV clock: %d\n", ret);
243
244 return (ret);
245 }
246
247 /**
248 * \brief Return current time according to the Xen Hypervisor wallclock.
249 *
250 * \param dev Xentimer device.
251 * \param ts Pointer to store the wallclock time.
252 *
253 * \note The Xen time structures document the hypervisor start time and the
254 * uptime-since-hypervisor-start (in nsec.) They need to be combined
255 * in order to calculate a TOD clock.
256 */
257 static int
xentimer_gettime(device_t dev,struct timespec * ts)258 xentimer_gettime(device_t dev, struct timespec *ts)
259 {
260 struct timespec u_ts;
261
262 timespecclear(ts);
263 xen_fetch_wallclock(ts);
264 xen_fetch_uptime(&u_ts);
265 timespecadd(ts, &u_ts, ts);
266
267 return (0);
268 }
269
270 /**
271 * \brief Handle a timer interrupt for the Xen PV timer driver.
272 *
273 * \param arg Xen timer driver softc that is expecting the interrupt.
274 */
275 static int
xentimer_intr(void * arg)276 xentimer_intr(void *arg)
277 {
278 struct xentimer_softc *sc = (struct xentimer_softc *)arg;
279 struct xentimer_pcpu_data *pcpu = DPCPU_PTR(xentimer_pcpu);
280
281 pcpu->last_processed = xen_fetch_vcpu_time(DPCPU_GET(vcpu_info));
282 if (pcpu->timer != 0 && sc->et.et_active)
283 sc->et.et_event_cb(&sc->et, sc->et.et_arg);
284
285 return (FILTER_HANDLED);
286 }
287
288 static int
xentimer_vcpu_start_timer(int vcpu,uint64_t next_time)289 xentimer_vcpu_start_timer(int vcpu, uint64_t next_time)
290 {
291 struct vcpu_set_singleshot_timer single;
292
293 single.timeout_abs_ns = next_time;
294 /* Get an event anyway, even if the timeout is already expired */
295 single.flags = 0;
296 return (HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, vcpu, &single));
297 }
298
299 static int
xentimer_vcpu_stop_timer(int vcpu)300 xentimer_vcpu_stop_timer(int vcpu)
301 {
302
303 return (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, vcpu, NULL));
304 }
305
306 /**
307 * \brief Set the next oneshot time for the current CPU.
308 *
309 * \param et Xen timer driver event timer to schedule on.
310 * \param first Delta to the next time to schedule the interrupt for.
311 * \param period Not used.
312 *
313 * \note See eventtimers(9) for more information.
314 * \note
315 *
316 * \returns 0
317 */
318 static int
xentimer_et_start(struct eventtimer * et,sbintime_t first,sbintime_t period)319 xentimer_et_start(struct eventtimer *et,
320 sbintime_t first, sbintime_t period)
321 {
322 int error;
323 struct xentimer_softc *sc = et->et_priv;
324 int cpu = PCPU_GET(vcpu_id);
325 struct xentimer_pcpu_data *pcpu = DPCPU_PTR(xentimer_pcpu);
326 struct vcpu_info *vcpu = DPCPU_GET(vcpu_info);
327 uint64_t first_in_ns, next_time;
328 #ifdef INVARIANTS
329 struct thread *td = curthread;
330 #endif
331
332 KASSERT(td->td_critnest != 0,
333 ("xentimer_et_start called without preemption disabled"));
334
335 /* See sbttots() for this formula. */
336 first_in_ns = (((first >> 32) * NSEC_IN_SEC) +
337 (((uint64_t)NSEC_IN_SEC * (uint32_t)first) >> 32));
338
339 next_time = xen_fetch_vcpu_time(vcpu) + first_in_ns;
340 error = xentimer_vcpu_start_timer(cpu, next_time);
341 if (error)
342 panic("%s: Error %d setting singleshot timer to %"PRIu64"\n",
343 device_get_nameunit(sc->dev), error, next_time);
344
345 pcpu->timer = next_time;
346 return (error);
347 }
348
349 /**
350 * \brief Cancel the event timer's currently running timer, if any.
351 */
352 static int
xentimer_et_stop(struct eventtimer * et)353 xentimer_et_stop(struct eventtimer *et)
354 {
355 int cpu = PCPU_GET(vcpu_id);
356 struct xentimer_pcpu_data *pcpu = DPCPU_PTR(xentimer_pcpu);
357
358 pcpu->timer = 0;
359 return (xentimer_vcpu_stop_timer(cpu));
360 }
361
362 /**
363 * \brief Attach a Xen PV timer driver instance.
364 *
365 * \param dev Bus device object to attach.
366 *
367 * \note
368 * \returns EINVAL
369 */
370 static int
xentimer_attach(device_t dev)371 xentimer_attach(device_t dev)
372 {
373 struct xentimer_softc *sc = device_get_softc(dev);
374 int error, i;
375
376 sc->dev = dev;
377
378 /* Bind an event channel to a VIRQ on each VCPU. */
379 CPU_FOREACH(i) {
380 struct xentimer_pcpu_data *pcpu;
381
382 pcpu = DPCPU_ID_PTR(i, xentimer_pcpu);
383 error = HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, i, NULL);
384 if (error) {
385 device_printf(dev, "Error disabling Xen periodic timer "
386 "on CPU %d\n", i);
387 return (error);
388 }
389
390 error = xen_intr_bind_virq(dev, VIRQ_TIMER, i, xentimer_intr,
391 NULL, sc, INTR_TYPE_CLK, &pcpu->irq_handle);
392 if (error) {
393 device_printf(dev, "Error %d binding VIRQ_TIMER "
394 "to VCPU %d\n", error, i);
395 return (error);
396 }
397 xen_intr_describe(pcpu->irq_handle, "c%d", i);
398 }
399
400 /* Register the event timer. */
401 sc->et.et_name = "XENTIMER";
402 sc->et.et_quality = XENTIMER_QUALITY;
403 sc->et.et_flags = ET_FLAGS_ONESHOT | ET_FLAGS_PERCPU;
404 sc->et.et_frequency = NSEC_IN_SEC;
405 /* See tstosbt() for this formula */
406 sc->et.et_min_period = (XENTIMER_MIN_PERIOD_IN_NSEC *
407 (((uint64_t)1 << 63) / 500000000) >> 32);
408 sc->et.et_max_period = ((sbintime_t)4 << 32);
409 sc->et.et_start = xentimer_et_start;
410 sc->et.et_stop = xentimer_et_stop;
411 sc->et.et_priv = sc;
412 et_register(&sc->et);
413
414 /* Register the timecounter. */
415 sc->tc.tc_name = "XENTIMER";
416 sc->tc.tc_quality = XENTIMER_QUALITY;
417 /*
418 * FIXME: due to the lack of ordering during resume, FreeBSD cannot
419 * guarantee that the Xen PV timer is resumed before any other device
420 * attempts to make use of it, so mark it as not safe for suspension
421 * (ie: remove the TC_FLAGS_SUSPEND_SAFE flag).
422 *
423 * NB: This was not a problem in previous FreeBSD versions because the
424 * timer was directly attached to the nexus, but it is an issue now
425 * that the timer is attached to the xenpv bus, and thus resumed
426 * later.
427 *
428 * sc->tc.tc_flags = TC_FLAGS_SUSPEND_SAFE;
429 */
430 /*
431 * The underlying resolution is in nanoseconds, since the timer info
432 * scales TSC frequencies using a fraction that represents time in
433 * terms of nanoseconds.
434 */
435 sc->tc.tc_frequency = NSEC_IN_SEC;
436 sc->tc.tc_counter_mask = ~0u;
437 sc->tc.tc_get_timecount = xentimer_get_timecount;
438 sc->tc.tc_priv = sc;
439 tc_init(&sc->tc);
440
441 /* Register the Hypervisor wall clock */
442 clock_register(dev, XENCLOCK_RESOLUTION);
443
444 return (0);
445 }
446
447 static int
xentimer_detach(device_t dev)448 xentimer_detach(device_t dev)
449 {
450
451 /* Implement Xen PV clock teardown - XXX see hpet_detach ? */
452 /* If possible:
453 * 1. need to deregister timecounter
454 * 2. need to deregister event timer
455 * 3. need to deregister virtual IRQ event channels
456 */
457 return (EBUSY);
458 }
459
460 static void
xentimer_percpu_resume(void * arg)461 xentimer_percpu_resume(void *arg)
462 {
463 device_t dev = (device_t) arg;
464 struct xentimer_softc *sc = device_get_softc(dev);
465
466 xentimer_et_start(&sc->et, sc->et.et_min_period, 0);
467 }
468
469 static int
xentimer_resume(device_t dev)470 xentimer_resume(device_t dev)
471 {
472 int error;
473 int i;
474
475 /* Disable the periodic timer */
476 CPU_FOREACH(i) {
477 error = HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, i, NULL);
478 if (error != 0) {
479 device_printf(dev,
480 "Error disabling Xen periodic timer on CPU %d\n",
481 i);
482 return (error);
483 }
484 }
485
486 /* Reset the last uptime value */
487 pvclock_resume();
488
489 /* Reset the RTC clock */
490 inittodr(time_second);
491
492 /* Kick the timers on all CPUs */
493 smp_rendezvous(NULL, xentimer_percpu_resume, NULL, dev);
494
495 if (bootverbose)
496 device_printf(dev, "resumed operation after suspension\n");
497
498 return (0);
499 }
500
501 static int
xentimer_suspend(device_t dev)502 xentimer_suspend(device_t dev)
503 {
504 return (0);
505 }
506
507 /*
508 * Xen early clock init
509 */
510 void
xen_clock_init(void)511 xen_clock_init(void)
512 {
513 }
514
515 /*
516 * Xen PV DELAY function
517 *
518 * When running on PVH mode we don't have an emulated i8524, so
519 * make use of the Xen time info in order to code a simple DELAY
520 * function that can be used during early boot.
521 */
522 void
xen_delay(int n)523 xen_delay(int n)
524 {
525 struct vcpu_info *vcpu = &HYPERVISOR_shared_info->vcpu_info[0];
526 uint64_t end_ns;
527 uint64_t current;
528
529 end_ns = xen_fetch_vcpu_time(vcpu);
530 end_ns += n * NSEC_IN_USEC;
531
532 for (;;) {
533 current = xen_fetch_vcpu_time(vcpu);
534 if (current >= end_ns)
535 break;
536 }
537 }
538
539 static device_method_t xentimer_methods[] = {
540 DEVMETHOD(device_identify, xentimer_identify),
541 DEVMETHOD(device_probe, xentimer_probe),
542 DEVMETHOD(device_attach, xentimer_attach),
543 DEVMETHOD(device_detach, xentimer_detach),
544 DEVMETHOD(device_suspend, xentimer_suspend),
545 DEVMETHOD(device_resume, xentimer_resume),
546 /* clock interface */
547 DEVMETHOD(clock_gettime, xentimer_gettime),
548 DEVMETHOD(clock_settime, xentimer_settime),
549 DEVMETHOD_END
550 };
551
552 static driver_t xentimer_driver = {
553 "xen_et",
554 xentimer_methods,
555 sizeof(struct xentimer_softc),
556 };
557
558 DRIVER_MODULE(xentimer, xenpv, xentimer_driver, xentimer_devclass, 0, 0);
559 MODULE_DEPEND(xentimer, xenpv, 1, 1, 1);
560