arch/x86/xen/time.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Xen time implementation.
   4  *
   5  * This is implemented in terms of a clocksource driver which uses
   6  * the hypervisor clock as a nanosecond timebase, and a clockevent
   7  * driver which uses the hypervisor's timer mechanism.
   8  *
   9  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  10  */
  11 #include <linux/kernel.h>
  12 #include <linux/interrupt.h>
  13 #include <linux/clocksource.h>
  14 #include <linux/clockchips.h>
  15 #include <linux/gfp.h>
  16 #include <linux/slab.h>
  17 #include <linux/pvclock_gtod.h>
  18 #include <linux/timekeeper_internal.h>
  19
  20 #include <asm/pvclock.h>
  21 #include <asm/xen/hypervisor.h>
  22 #include <asm/xen/hypercall.h>
  23
  24 #include <xen/events.h>
  25 #include <xen/features.h>
  26 #include <xen/interface/xen.h>
  27 #include <xen/interface/vcpu.h>
  28
  29 #include "xen-ops.h"
  30
  31 /* Xen may fire a timer up to this many ns early */
  32 #define TIMER_SLOP      100000
  33
  34 static u64 xen_sched_clock_offset __read_mostly;
  35
  36 /* Get the TSC speed from Xen */
  37 static unsigned long xen_tsc_khz(void)
  38 {
  39         struct pvclock_vcpu_time_info *info =
  40                 &HYPERVISOR_shared_info->vcpu_info[0].time;
  41
  42         return pvclock_tsc_khz(info);
  43 }
  44
  45 static u64 xen_clocksource_read(void)
  46 {
  47         struct pvclock_vcpu_time_info *src;
  48         u64 ret;
  49
  50         preempt_disable_notrace();
  51         src = &__this_cpu_read(xen_vcpu)->time;
  52         ret = pvclock_clocksource_read(src);
  53         preempt_enable_notrace();
  54         return ret;
  55 }
  56
  57 static u64 xen_clocksource_get_cycles(struct clocksource *cs)
  58 {
  59         return xen_clocksource_read();
  60 }
  61
  62 static u64 xen_sched_clock(void)
  63 {
  64         return xen_clocksource_read() - xen_sched_clock_offset;
  65 }
  66
  67 static void xen_read_wallclock(struct timespec64 *ts)
  68 {
  69         struct shared_info *s = HYPERVISOR_shared_info;
  70         struct pvclock_wall_clock *wall_clock = &(s->wc);
  71         struct pvclock_vcpu_time_info *vcpu_time;
  72
  73         vcpu_time = &get_cpu_var(xen_vcpu)->time;
  74         pvclock_read_wallclock(wall_clock, vcpu_time, ts);
  75         put_cpu_var(xen_vcpu);
  76 }
  77
  78 static void xen_get_wallclock(struct timespec64 *now)
  79 {
  80         xen_read_wallclock(now);
  81 }
  82
  83 static int xen_set_wallclock(const struct timespec64 *now)
  84 {
  85         return -ENODEV;
  86 }
  87
  88 static int xen_pvclock_gtod_notify(struct notifier_block *nb,
  89                                    unsigned long was_set, void *priv)
  90 {
  91         /* Protected by the calling core code serialization */
  92         static struct timespec64 next_sync;
  93
  94         struct xen_platform_op op;
  95         struct timespec64 now;
  96         struct timekeeper *tk = priv;
  97         static bool settime64_supported = true;
  98         int ret;
  99
 100         now.tv_sec = tk->xtime_sec;
 101         now.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
 102
 103         /*
 104          * We only take the expensive HV call when the clock was set
 105          * or when the 11 minutes RTC synchronization time elapsed.
 106          */
 107         if (!was_set && timespec64_compare(&now, &next_sync) < 0)
 108                 return NOTIFY_OK;
 109
 110 again:
 111         if (settime64_supported) {
 112                 op.cmd = XENPF_settime64;
 113                 op.u.settime64.mbz = 0;
 114                 op.u.settime64.secs = now.tv_sec;
 115                 op.u.settime64.nsecs = now.tv_nsec;
 116                 op.u.settime64.system_time = xen_clocksource_read();
 117         } else {
 118                 op.cmd = XENPF_settime32;
 119                 op.u.settime32.secs = now.tv_sec;
 120                 op.u.settime32.nsecs = now.tv_nsec;
 121                 op.u.settime32.system_time = xen_clocksource_read();
 122         }
 123
 124         ret = HYPERVISOR_platform_op(&op);
 125
 126         if (ret == -ENOSYS && settime64_supported) {
 127                 settime64_supported = false;
 128                 goto again;
 129         }
 130         if (ret < 0)
 131                 return NOTIFY_BAD;
 132
 133         /*
 134          * Move the next drift compensation time 11 minutes
 135          * ahead. That's emulating the sync_cmos_clock() update for
 136          * the hardware RTC.
 137          */
 138         next_sync = now;
 139         next_sync.tv_sec += 11 * 60;
 140
 141         return NOTIFY_OK;
 142 }
 143
 144 static struct notifier_block xen_pvclock_gtod_notifier = {
 145         .notifier_call = xen_pvclock_gtod_notify,
 146 };
 147
 148 static struct clocksource xen_clocksource __read_mostly = {
 149         .name = "xen",
 150         .rating = 400,
 151         .read = xen_clocksource_get_cycles,
 152         .mask = ~0,
 153         .flags = CLOCK_SOURCE_IS_CONTINUOUS,
 154 };
 155
 156 /*
 157    Xen clockevent implementation
 158
 159    Xen has two clockevent implementations:
 160
 161    The old timer_op one works with all released versions of Xen prior
 162    to version 3.0.4.  This version of the hypervisor provides a
 163    single-shot timer with nanosecond resolution.  However, sharing the
 164    same event channel is a 100Hz tick which is delivered while the
 165    vcpu is running.  We don't care about or use this tick, but it will
 166    cause the core time code to think the timer fired too soon, and
 167    will end up resetting it each time.  It could be filtered, but
 168    doing so has complications when the ktime clocksource is not yet
 169    the xen clocksource (ie, at boot time).
 170
 171    The new vcpu_op-based timer interface allows the tick timer period
 172    to be changed or turned off.  The tick timer is not useful as a
 173    periodic timer because events are only delivered to running vcpus.
 174    The one-shot timer can report when a timeout is in the past, so
 175    set_next_event is capable of returning -ETIME when appropriate.
 176    This interface is used when available.
 177 */
 178
 179
 180 /*
 181   Get a hypervisor absolute time.  In theory we could maintain an
 182   offset between the kernel's time and the hypervisor's time, and
 183   apply that to a kernel's absolute timeout.  Unfortunately the
 184   hypervisor and kernel times can drift even if the kernel is using
 185   the Xen clocksource, because ntp can warp the kernel's clocksource.
 186 */
 187 static s64 get_abs_timeout(unsigned long delta)
 188 {
 189         return xen_clocksource_read() + delta;
 190 }
 191
 192 static int xen_timerop_shutdown(struct clock_event_device *evt)
 193 {
 194         /* cancel timeout */
 195         HYPERVISOR_set_timer_op(0);
 196
 197         return 0;
 198 }
 199
 200 static int xen_timerop_set_next_event(unsigned long delta,
 201                                       struct clock_event_device *evt)
 202 {
 203         WARN_ON(!clockevent_state_oneshot(evt));
 204
 205         if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
 206                 BUG();
 207
 208         /* We may have missed the deadline, but there's no real way of
 209            knowing for sure.  If the event was in the past, then we'll
 210            get an immediate interrupt. */
 211
 212         return 0;
 213 }
 214
 215 static const struct clock_event_device xen_timerop_clockevent = {
 216         .name                   = "xen",
 217         .features               = CLOCK_EVT_FEAT_ONESHOT,
 218
 219         .max_delta_ns           = 0xffffffff,
 220         .max_delta_ticks        = 0xffffffff,
 221         .min_delta_ns           = TIMER_SLOP,
 222         .min_delta_ticks        = TIMER_SLOP,
 223
 224         .mult                   = 1,
 225         .shift                  = 0,
 226         .rating                 = 500,
 227
 228         .set_state_shutdown     = xen_timerop_shutdown,
 229         .set_next_event         = xen_timerop_set_next_event,
 230 };
 231
 232 static int xen_vcpuop_shutdown(struct clock_event_device *evt)
 233 {
 234         int cpu = smp_processor_id();
 235
 236         if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, xen_vcpu_nr(cpu),
 237                                NULL) ||
 238             HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
 239                                NULL))
 240                 BUG();
 241
 242         return 0;
 243 }
 244
 245 static int xen_vcpuop_set_oneshot(struct clock_event_device *evt)
 246 {
 247         int cpu = smp_processor_id();
 248
 249         if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
 250                                NULL))
 251                 BUG();
 252
 253         return 0;
 254 }
 255
 256 static int xen_vcpuop_set_next_event(unsigned long delta,
 257                                      struct clock_event_device *evt)
 258 {
 259         int cpu = smp_processor_id();
 260         struct vcpu_set_singleshot_timer single;
 261         int ret;
 262
 263         WARN_ON(!clockevent_state_oneshot(evt));
 264
 265         single.timeout_abs_ns = get_abs_timeout(delta);
 266         /* Get an event anyway, even if the timeout is already expired */
 267         single.flags = 0;
 268
 269         ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, xen_vcpu_nr(cpu),
 270                                  &single);
 271         BUG_ON(ret != 0);
 272
 273         return ret;
 274 }
 275
 276 static const struct clock_event_device xen_vcpuop_clockevent = {
 277         .name = "xen",
 278         .features = CLOCK_EVT_FEAT_ONESHOT,
 279
 280         .max_delta_ns = 0xffffffff,
 281         .max_delta_ticks = 0xffffffff,
 282         .min_delta_ns = TIMER_SLOP,
 283         .min_delta_ticks = TIMER_SLOP,
 284
 285         .mult = 1,
 286         .shift = 0,
 287         .rating = 500,
 288
 289         .set_state_shutdown = xen_vcpuop_shutdown,
 290         .set_state_oneshot = xen_vcpuop_set_oneshot,
 291         .set_next_event = xen_vcpuop_set_next_event,
 292 };
 293
 294 static const struct clock_event_device *xen_clockevent =
 295         &xen_timerop_clockevent;
 296
 297 struct xen_clock_event_device {
 298         struct clock_event_device evt;
 299         char name[16];
 300 };
 301 static DEFINE_PER_CPU(struct xen_clock_event_device, xen_clock_events) = { .evt.irq = -1 };
 302
 303 static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
 304 {
 305         struct clock_event_device *evt = this_cpu_ptr(&xen_clock_events.evt);
 306         irqreturn_t ret;
 307
 308         ret = IRQ_NONE;
 309         if (evt->event_handler) {
 310                 evt->event_handler(evt);
 311                 ret = IRQ_HANDLED;
 312         }
 313
 314         return ret;
 315 }
 316
 317 void xen_teardown_timer(int cpu)
 318 {
 319         struct clock_event_device *evt;
 320         evt = &per_cpu(xen_clock_events, cpu).evt;
 321
 322         if (evt->irq >= 0) {
 323                 unbind_from_irqhandler(evt->irq, NULL);
 324                 evt->irq = -1;
 325         }
 326 }
 327
 328 void xen_setup_timer(int cpu)
 329 {
 330         struct xen_clock_event_device *xevt = &per_cpu(xen_clock_events, cpu);
 331         struct clock_event_device *evt = &xevt->evt;
 332         int irq;
 333
 334         WARN(evt->irq >= 0, "IRQ%d for CPU%d is already allocated\n", evt->irq, cpu);
 335         if (evt->irq >= 0)
 336                 xen_teardown_timer(cpu);
 337
 338         printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
 339
 340         snprintf(xevt->name, sizeof(xevt->name), "timer%d", cpu);
 341
 342         irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
 343                                       IRQF_PERCPU|IRQF_NOBALANCING|IRQF_TIMER|
 344                                       IRQF_FORCE_RESUME|IRQF_EARLY_RESUME,
 345                                       xevt->name, NULL);
 346         (void)xen_set_irq_priority(irq, XEN_IRQ_PRIORITY_MAX);
 347
 348         memcpy(evt, xen_clockevent, sizeof(*evt));
 349
 350         evt->cpumask = cpumask_of(cpu);
 351         evt->irq = irq;
 352 }
 353
 354
 355 void xen_setup_cpu_clockevents(void)
 356 {
 357         clockevents_register_device(this_cpu_ptr(&xen_clock_events.evt));
 358 }
 359
 360 void xen_timer_resume(void)
 361 {
 362         int cpu;
 363
 364         if (xen_clockevent != &xen_vcpuop_clockevent)
 365                 return;
 366
 367         for_each_online_cpu(cpu) {
 368                 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer,
 369                                        xen_vcpu_nr(cpu), NULL))
 370                         BUG();
 371         }
 372 }
 373
 374 static const struct pv_time_ops xen_time_ops __initconst = {
 375         .sched_clock = xen_sched_clock,
 376         .steal_clock = xen_steal_clock,
 377 };
 378
 379 static struct pvclock_vsyscall_time_info *xen_clock __read_mostly;
 380 static u64 xen_clock_value_saved;
 381
 382 void xen_save_time_memory_area(void)
 383 {
 384         struct vcpu_register_time_memory_area t;
 385         int ret;
 386
 387         xen_clock_value_saved = xen_clocksource_read() - xen_sched_clock_offset;
 388
 389         if (!xen_clock)
 390                 return;
 391
 392         t.addr.v = NULL;
 393
 394         ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
 395         if (ret != 0)
 396                 pr_notice("Cannot save secondary vcpu_time_info (err %d)",
 397                           ret);
 398         else
 399                 clear_page(xen_clock);
 400 }
 401
 402 void xen_restore_time_memory_area(void)
 403 {
 404         struct vcpu_register_time_memory_area t;
 405         int ret;
 406
 407         if (!xen_clock)
 408                 goto out;
 409
 410         t.addr.v = &xen_clock->pvti;
 411
 412         ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
 413
 414         /*
 415          * We don't disable VCLOCK_PVCLOCK entirely if it fails to register the
 416          * secondary time info with Xen or if we migrated to a host without the
 417          * necessary flags. On both of these cases what happens is either
 418          * process seeing a zeroed out pvti or seeing no PVCLOCK_TSC_STABLE_BIT
 419          * bit set. Userspace checks the latter and if 0, it discards the data
 420          * in pvti and fallbacks to a system call for a reliable timestamp.
 421          */
 422         if (ret != 0)
 423                 pr_notice("Cannot restore secondary vcpu_time_info (err %d)",
 424                           ret);
 425
 426 out:
 427         /* Need pvclock_resume() before using xen_clocksource_read(). */
 428         pvclock_resume();
 429         xen_sched_clock_offset = xen_clocksource_read() - xen_clock_value_saved;
 430 }
 431
 432 static void xen_setup_vsyscall_time_info(void)
 433 {
 434         struct vcpu_register_time_memory_area t;
 435         struct pvclock_vsyscall_time_info *ti;
 436         int ret;
 437
 438         ti = (struct pvclock_vsyscall_time_info *)get_zeroed_page(GFP_KERNEL);
 439         if (!ti)
 440                 return;
 441
 442         t.addr.v = &ti->pvti;
 443
 444         ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area, 0, &t);
 445         if (ret) {
 446                 pr_notice("xen: VCLOCK_PVCLOCK not supported (err %d)\n", ret);
 447                 free_page((unsigned long)ti);
 448                 return;
 449         }
 450
 451         /*
 452          * If primary time info had this bit set, secondary should too since
 453          * it's the same data on both just different memory regions. But we
 454          * still check it in case hypervisor is buggy.
 455          */
 456         if (!(ti->pvti.flags & PVCLOCK_TSC_STABLE_BIT)) {
 457                 t.addr.v = NULL;
 458                 ret = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_time_memory_area,
 459                                          0, &t);
 460                 if (!ret)
 461                         free_page((unsigned long)ti);
 462
 463                 pr_notice("xen: VCLOCK_PVCLOCK not supported (tsc unstable)\n");
 464                 return;
 465         }
 466
 467         xen_clock = ti;
 468         pvclock_set_pvti_cpu0_va(xen_clock);
 469
 470         xen_clocksource.archdata.vclock_mode = VCLOCK_PVCLOCK;
 471 }
 472
 473 static void __init xen_time_init(void)
 474 {
 475         struct pvclock_vcpu_time_info *pvti;
 476         int cpu = smp_processor_id();
 477         struct timespec64 tp;
 478
 479         /* As Dom0 is never moved, no penalty on using TSC there */
 480         if (xen_initial_domain())
 481                 xen_clocksource.rating = 275;
 482
 483         clocksource_register_hz(&xen_clocksource, NSEC_PER_SEC);
 484
 485         if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, xen_vcpu_nr(cpu),
 486                                NULL) == 0) {
 487                 /* Successfully turned off 100Hz tick, so we have the
 488                    vcpuop-based timer interface */
 489                 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
 490                 xen_clockevent = &xen_vcpuop_clockevent;
 491         }
 492
 493         /* Set initial system time with full resolution */
 494         xen_read_wallclock(&tp);
 495         do_settimeofday64(&tp);
 496
 497         setup_force_cpu_cap(X86_FEATURE_TSC);
 498
 499         /*
 500          * We check ahead on the primary time info if this
 501          * bit is supported hence speeding up Xen clocksource.
 502          */
 503         pvti = &__this_cpu_read(xen_vcpu)->time;
 504         if (pvti->flags & PVCLOCK_TSC_STABLE_BIT) {
 505                 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
 506                 xen_setup_vsyscall_time_info();
 507         }
 508
 509         xen_setup_runstate_info(cpu);
 510         xen_setup_timer(cpu);
 511         xen_setup_cpu_clockevents();
 512
 513         xen_time_setup_guest();
 514
 515         if (xen_initial_domain())
 516                 pvclock_gtod_register_notifier(&xen_pvclock_gtod_notifier);
 517 }
 518
 519 void __init xen_init_time_ops(void)
 520 {
 521         xen_sched_clock_offset = xen_clocksource_read();
 522         pv_ops.time = xen_time_ops;
 523
 524         x86_init.timers.timer_init = xen_time_init;
 525         x86_init.timers.setup_percpu_clockev = x86_init_noop;
 526         x86_cpuinit.setup_percpu_clockev = x86_init_noop;
 527
 528         x86_platform.calibrate_tsc = xen_tsc_khz;
 529         x86_platform.get_wallclock = xen_get_wallclock;
 530         /* Dom0 uses the native method to set the hardware RTC. */
 531         if (!xen_initial_domain())
 532                 x86_platform.set_wallclock = xen_set_wallclock;
 533 }
 534
 535 #ifdef CONFIG_XEN_PVHVM
 536 static void xen_hvm_setup_cpu_clockevents(void)
 537 {
 538         int cpu = smp_processor_id();
 539         xen_setup_runstate_info(cpu);
 540         /*
 541          * xen_setup_timer(cpu) - snprintf is bad in atomic context. Hence
 542          * doing it xen_hvm_cpu_notify (which gets called by smp_init during
 543          * early bootup and also during CPU hotplug events).
 544          */
 545         xen_setup_cpu_clockevents();
 546 }
 547
 548 void __init xen_hvm_init_time_ops(void)
 549 {
 550         /*
 551          * vector callback is needed otherwise we cannot receive interrupts
 552          * on cpu > 0 and at this point we don't know how many cpus are
 553          * available.
 554          */
 555         if (!xen_have_vector_callback)
 556                 return;
 557
 558         if (!xen_feature(XENFEAT_hvm_safe_pvclock)) {
 559                 pr_info("Xen doesn't support pvclock on HVM, disable pv timer");
 560                 return;
 561         }
 562
 563         xen_sched_clock_offset = xen_clocksource_read();
 564         pv_ops.time = xen_time_ops;
 565         x86_init.timers.setup_percpu_clockev = xen_time_init;
 566         x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents;
 567
 568         x86_platform.calibrate_tsc = xen_tsc_khz;
 569         x86_platform.get_wallclock = xen_get_wallclock;
 570         x86_platform.set_wallclock = xen_set_wallclock;
 571 }
 572 #endif