arch/x86/vdso/vclock_gettime.c

   1 /*
   2  * Copyright 2006 Andi Kleen, SUSE Labs.
   3  * Subject to the GNU Public License, v.2
   4  *
   5  * Fast user context implementation of clock_gettime, gettimeofday, and time.
   6  *
   7  * The code should have no internal unresolved relocations.
   8  * Check with readelf after changing.
   9  */
  10
  11 /* Disable profiling for userspace code: */
  12 #define DISABLE_BRANCH_PROFILING
  13
  14 #include <linux/kernel.h>
  15 #include <linux/posix-timers.h>
  16 #include <linux/time.h>
  17 #include <linux/string.h>
  18 #include <asm/vsyscall.h>
  19 #include <asm/fixmap.h>
  20 #include <asm/vgtod.h>
  21 #include <asm/timex.h>
  22 #include <asm/hpet.h>
  23 #include <asm/unistd.h>
  24 #include <asm/io.h>
  25
  26 #define gtod (&VVAR(vsyscall_gtod_data))
  27
  28 notrace static cycle_t vread_tsc(void)
  29 {
  30         cycle_t ret;
  31         u64 last;
  32
  33         /*
  34          * Empirically, a fence (of type that depends on the CPU)
  35          * before rdtsc is enough to ensure that rdtsc is ordered
  36          * with respect to loads.  The various CPU manuals are unclear
  37          * as to whether rdtsc can be reordered with later loads,
  38          * but no one has ever seen it happen.
  39          */
  40         rdtsc_barrier();
  41         ret = (cycle_t)vget_cycles();
  42
  43         last = VVAR(vsyscall_gtod_data).clock.cycle_last;
  44
  45         if (likely(ret >= last))
  46                 return ret;
  47
  48         /*
  49          * GCC likes to generate cmov here, but this branch is extremely
  50          * predictable (it's just a funciton of time and the likely is
  51          * very likely) and there's a data dependence, so force GCC
  52          * to generate a branch instead.  I don't barrier() because
  53          * we don't actually need a barrier, and if this function
  54          * ever gets inlined it will generate worse code.
  55          */
  56         asm volatile ("");
  57         return last;
  58 }
  59
  60 static notrace cycle_t vread_hpet(void)
  61 {
  62         return readl((const void __iomem *)fix_to_virt(VSYSCALL_HPET) + 0xf0);
  63 }
  64
  65 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
  66 {
  67         long ret;
  68         asm("syscall" : "=a" (ret) :
  69             "0" (__NR_clock_gettime),"D" (clock), "S" (ts) : "memory");
  70         return ret;
  71 }
  72
  73 notrace static inline long vgetns(void)
  74 {
  75         long v;
  76         cycles_t cycles;
  77         if (gtod->clock.vclock_mode == VCLOCK_TSC)
  78                 cycles = vread_tsc();
  79         else
  80                 cycles = vread_hpet();
  81         v = (cycles - gtod->clock.cycle_last) & gtod->clock.mask;
  82         return (v * gtod->clock.mult) >> gtod->clock.shift;
  83 }
  84
  85 notrace static noinline int do_realtime(struct timespec *ts)
  86 {
  87         unsigned long seq, ns;
  88         do {
  89                 seq = read_seqbegin(&gtod->lock);
  90                 ts->tv_sec = gtod->wall_time_sec;
  91                 ts->tv_nsec = gtod->wall_time_nsec;
  92                 ns = vgetns();
  93         } while (unlikely(read_seqretry(&gtod->lock, seq)));
  94         timespec_add_ns(ts, ns);
  95         return 0;
  96 }
  97
  98 notrace static noinline int do_monotonic(struct timespec *ts)
  99 {
 100         unsigned long seq, ns, secs;
 101         do {
 102                 seq = read_seqbegin(&gtod->lock);
 103                 secs = gtod->wall_time_sec;
 104                 ns = gtod->wall_time_nsec + vgetns();
 105                 secs += gtod->wall_to_monotonic.tv_sec;
 106                 ns += gtod->wall_to_monotonic.tv_nsec;
 107         } while (unlikely(read_seqretry(&gtod->lock, seq)));
 108
 109         /* wall_time_nsec, vgetns(), and wall_to_monotonic.tv_nsec
 110          * are all guaranteed to be nonnegative.
 111          */
 112         while (ns >= NSEC_PER_SEC) {
 113                 ns -= NSEC_PER_SEC;
 114                 ++secs;
 115         }
 116         ts->tv_sec = secs;
 117         ts->tv_nsec = ns;
 118
 119         return 0;
 120 }
 121
 122 notrace static noinline int do_realtime_coarse(struct timespec *ts)
 123 {
 124         unsigned long seq;
 125         do {
 126                 seq = read_seqbegin(&gtod->lock);
 127                 ts->tv_sec = gtod->wall_time_coarse.tv_sec;
 128                 ts->tv_nsec = gtod->wall_time_coarse.tv_nsec;
 129         } while (unlikely(read_seqretry(&gtod->lock, seq)));
 130         return 0;
 131 }
 132
 133 notrace static noinline int do_monotonic_coarse(struct timespec *ts)
 134 {
 135         unsigned long seq, ns, secs;
 136         do {
 137                 seq = read_seqbegin(&gtod->lock);
 138                 secs = gtod->wall_time_coarse.tv_sec;
 139                 ns = gtod->wall_time_coarse.tv_nsec;
 140                 secs += gtod->wall_to_monotonic.tv_sec;
 141                 ns += gtod->wall_to_monotonic.tv_nsec;
 142         } while (unlikely(read_seqretry(&gtod->lock, seq)));
 143
 144         /* wall_time_nsec and wall_to_monotonic.tv_nsec are
 145          * guaranteed to be between 0 and NSEC_PER_SEC.
 146          */
 147         if (ns >= NSEC_PER_SEC) {
 148                 ns -= NSEC_PER_SEC;
 149                 ++secs;
 150         }
 151         ts->tv_sec = secs;
 152         ts->tv_nsec = ns;
 153
 154         return 0;
 155 }
 156
 157 notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
 158 {
 159         switch (clock) {
 160         case CLOCK_REALTIME:
 161                 if (likely(gtod->clock.vclock_mode != VCLOCK_NONE))
 162                         return do_realtime(ts);
 163                 break;
 164         case CLOCK_MONOTONIC:
 165                 if (likely(gtod->clock.vclock_mode != VCLOCK_NONE))
 166                         return do_monotonic(ts);
 167                 break;
 168         case CLOCK_REALTIME_COARSE:
 169                 return do_realtime_coarse(ts);
 170         case CLOCK_MONOTONIC_COARSE:
 171                 return do_monotonic_coarse(ts);
 172         }
 173
 174         return vdso_fallback_gettime(clock, ts);
 175 }
 176 int clock_gettime(clockid_t, struct timespec *)
 177         __attribute__((weak, alias("__vdso_clock_gettime")));
 178
 179 notrace int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz)
 180 {
 181         long ret;
 182         if (likely(gtod->clock.vclock_mode != VCLOCK_NONE)) {
 183                 if (likely(tv != NULL)) {
 184                         BUILD_BUG_ON(offsetof(struct timeval, tv_usec) !=
 185                                      offsetof(struct timespec, tv_nsec) ||
 186                                      sizeof(*tv) != sizeof(struct timespec));
 187                         do_realtime((struct timespec *)tv);
 188                         tv->tv_usec /= 1000;
 189                 }
 190                 if (unlikely(tz != NULL)) {
 191                         /* Avoid memcpy. Some old compilers fail to inline it */
 192                         tz->tz_minuteswest = gtod->sys_tz.tz_minuteswest;
 193                         tz->tz_dsttime = gtod->sys_tz.tz_dsttime;
 194                 }
 195                 return 0;
 196         }
 197         asm("syscall" : "=a" (ret) :
 198             "0" (__NR_gettimeofday), "D" (tv), "S" (tz) : "memory");
 199         return ret;
 200 }
 201 int gettimeofday(struct timeval *, struct timezone *)
 202         __attribute__((weak, alias("__vdso_gettimeofday")));
 203
 204 /*
 205  * This will break when the xtime seconds get inaccurate, but that is
 206  * unlikely
 207  */
 208 notrace time_t __vdso_time(time_t *t)
 209 {
 210         /* This is atomic on x86_64 so we don't need any locks. */
 211         time_t result = ACCESS_ONCE(VVAR(vsyscall_gtod_data).wall_time_sec);
 212
 213         if (t)
 214                 *t = result;
 215         return result;
 216 }
 217 int time(time_t *t)
 218         __attribute__((weak, alias("__vdso_time")));