arch/x86/kernel/pvclock.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*  paravirtual clock -- common code used by kvm/xen
   3
   4 */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/percpu.h>
   8 #include <linux/notifier.h>
   9 #include <linux/sched.h>
  10 #include <linux/gfp.h>
  11 #include <linux/memblock.h>
  12 #include <linux/nmi.h>
  13
  14 #include <asm/fixmap.h>
  15 #include <asm/pvclock.h>
  16 #include <asm/vgtod.h>
  17
  18 static u8 valid_flags __read_mostly = 0;
  19 static struct pvclock_vsyscall_time_info *pvti_cpu0_va __read_mostly;
  20
  21 void pvclock_set_flags(u8 flags)
  22 {
  23         valid_flags = flags;
  24 }
  25
  26 unsigned long pvclock_tsc_khz(struct pvclock_vcpu_time_info *src)
  27 {
  28         u64 pv_tsc_khz = 1000000ULL << 32;
  29
  30         do_div(pv_tsc_khz, src->tsc_to_system_mul);
  31         if (src->tsc_shift < 0)
  32                 pv_tsc_khz <<= -src->tsc_shift;
  33         else
  34                 pv_tsc_khz >>= src->tsc_shift;
  35         return pv_tsc_khz;
  36 }
  37
  38 void pvclock_touch_watchdogs(void)
  39 {
  40         touch_softlockup_watchdog_sync();
  41         clocksource_touch_watchdog();
  42         rcu_cpu_stall_reset();
  43         reset_hung_task_detector();
  44 }
  45
  46 static atomic64_t last_value = ATOMIC64_INIT(0);
  47
  48 void pvclock_resume(void)
  49 {
  50         atomic64_set(&last_value, 0);
  51 }
  52
  53 u8 pvclock_read_flags(struct pvclock_vcpu_time_info *src)
  54 {
  55         unsigned version;
  56         u8 flags;
  57
  58         do {
  59                 version = pvclock_read_begin(src);
  60                 flags = src->flags;
  61         } while (pvclock_read_retry(src, version));
  62
  63         return flags & valid_flags;
  64 }
  65
  66 u64 pvclock_clocksource_read(struct pvclock_vcpu_time_info *src)
  67 {
  68         unsigned version;
  69         u64 ret;
  70         u64 last;
  71         u8 flags;
  72
  73         do {
  74                 version = pvclock_read_begin(src);
  75                 ret = __pvclock_read_cycles(src, rdtsc_ordered());
  76                 flags = src->flags;
  77         } while (pvclock_read_retry(src, version));
  78
  79         if (unlikely((flags & PVCLOCK_GUEST_STOPPED) != 0)) {
  80                 src->flags &= ~PVCLOCK_GUEST_STOPPED;
  81                 pvclock_touch_watchdogs();
  82         }
  83
  84         if ((valid_flags & PVCLOCK_TSC_STABLE_BIT) &&
  85                 (flags & PVCLOCK_TSC_STABLE_BIT))
  86                 return ret;
  87
  88         /*
  89          * Assumption here is that last_value, a global accumulator, always goes
  90          * forward. If we are less than that, we should not be much smaller.
  91          * We assume there is an error marging we're inside, and then the correction
  92          * does not sacrifice accuracy.
  93          *
  94          * For reads: global may have changed between test and return,
  95          * but this means someone else updated poked the clock at a later time.
  96          * We just need to make sure we are not seeing a backwards event.
  97          *
  98          * For updates: last_value = ret is not enough, since two vcpus could be
  99          * updating at the same time, and one of them could be slightly behind,
 100          * making the assumption that last_value always go forward fail to hold.
 101          */
 102         last = atomic64_read(&last_value);
 103         do {
 104                 if (ret < last)
 105                         return last;
 106                 last = atomic64_cmpxchg(&last_value, last, ret);
 107         } while (unlikely(last != ret));
 108
 109         return ret;
 110 }
 111
 112 void pvclock_read_wallclock(struct pvclock_wall_clock *wall_clock,
 113                             struct pvclock_vcpu_time_info *vcpu_time,
 114                             struct timespec64 *ts)
 115 {
 116         u32 version;
 117         u64 delta;
 118         struct timespec64 now;
 119
 120         /* get wallclock at system boot */
 121         do {
 122                 version = wall_clock->version;
 123                 rmb();          /* fetch version before time */
 124                 /*
 125                  * Note: wall_clock->sec is a u32 value, so it can
 126                  * only store dates between 1970 and 2106. To allow
 127                  * times beyond that, we need to create a new hypercall
 128                  * interface with an extended pvclock_wall_clock structure
 129                  * like ARM has.
 130                  */
 131                 now.tv_sec  = wall_clock->sec;
 132                 now.tv_nsec = wall_clock->nsec;
 133                 rmb();          /* fetch time before checking version */
 134         } while ((wall_clock->version & 1) || (version != wall_clock->version));
 135
 136         delta = pvclock_clocksource_read(vcpu_time);    /* time since system boot */
 137         delta += now.tv_sec * NSEC_PER_SEC + now.tv_nsec;
 138
 139         now.tv_nsec = do_div(delta, NSEC_PER_SEC);
 140         now.tv_sec = delta;
 141
 142         set_normalized_timespec64(ts, now.tv_sec, now.tv_nsec);
 143 }
 144
 145 void pvclock_set_pvti_cpu0_va(struct pvclock_vsyscall_time_info *pvti)
 146 {
 147         WARN_ON(vclock_was_used(VCLOCK_PVCLOCK));
 148         pvti_cpu0_va = pvti;
 149 }
 150
 151 struct pvclock_vsyscall_time_info *pvclock_get_pvti_cpu0_va(void)
 152 {
 153         return pvti_cpu0_va;
 154 }
 155 EXPORT_SYMBOL_GPL(pvclock_get_pvti_cpu0_va);