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 #include <asm/pvclock.h>
  26
  27 #define gtod (&VVAR(vsyscall_gtod_data))
  28
  29 notrace static cycle_t vread_tsc(void)
  30 {
  31         cycle_t ret;
  32         u64 last;
  33
  34         /*
  35          * Empirically, a fence (of type that depends on the CPU)
  36          * before rdtsc is enough to ensure that rdtsc is ordered
  37          * with respect to loads.  The various CPU manuals are unclear
  38          * as to whether rdtsc can be reordered with later loads,
  39          * but no one has ever seen it happen.
  40          */
  41         rdtsc_barrier();
  42         ret = (cycle_t)vget_cycles();
  43
  44         last = VVAR(vsyscall_gtod_data).clock.cycle_last;
  45
  46         if (likely(ret >= last))
  47                 return ret;
  48
  49         /*
  50          * GCC likes to generate cmov here, but this branch is extremely
  51          * predictable (it's just a funciton of time and the likely is
  52          * very likely) and there's a data dependence, so force GCC
  53          * to generate a branch instead.  I don't barrier() because
  54          * we don't actually need a barrier, and if this function
  55          * ever gets inlined it will generate worse code.
  56          */
  57         asm volatile ("");
  58         return last;
  59 }
  60
  61 static notrace cycle_t vread_hpet(void)
  62 {
  63         return readl((const void __iomem *)fix_to_virt(VSYSCALL_HPET) + HPET_COUNTER);
  64 }
  65
  66 #ifdef CONFIG_PARAVIRT_CLOCK
  67
  68 static notrace const struct pvclock_vsyscall_time_info *get_pvti(int cpu)
  69 {
  70         const struct pvclock_vsyscall_time_info *pvti_base;
  71         int idx = cpu / (PAGE_SIZE/PVTI_SIZE);
  72         int offset = cpu % (PAGE_SIZE/PVTI_SIZE);
  73
  74         BUG_ON(PVCLOCK_FIXMAP_BEGIN + idx > PVCLOCK_FIXMAP_END);
  75
  76         pvti_base = (struct pvclock_vsyscall_time_info *)
  77                     __fix_to_virt(PVCLOCK_FIXMAP_BEGIN+idx);
  78
  79         return &pvti_base[offset];
  80 }
  81
  82 static notrace cycle_t vread_pvclock(int *mode)
  83 {
  84         const struct pvclock_vsyscall_time_info *pvti;
  85         cycle_t ret;
  86         u64 last;
  87         u32 version;
  88         u8 flags;
  89         unsigned cpu, cpu1;
  90
  91
  92         /*
  93          * Note: hypervisor must guarantee that:
  94          * 1. cpu ID number maps 1:1 to per-CPU pvclock time info.
  95          * 2. that per-CPU pvclock time info is updated if the
  96          *    underlying CPU changes.
  97          * 3. that version is increased whenever underlying CPU
  98          *    changes.
  99          *
 100          */
 101         do {
 102                 cpu = __getcpu() & VGETCPU_CPU_MASK;
 103                 /* TODO: We can put vcpu id into higher bits of pvti.version.
 104                  * This will save a couple of cycles by getting rid of
 105                  * __getcpu() calls (Gleb).
 106                  */
 107
 108                 pvti = get_pvti(cpu);
 109
 110                 version = __pvclock_read_cycles(&pvti->pvti, &ret, &flags);
 111
 112                 /*
 113                  * Test we're still on the cpu as well as the version.
 114                  * We could have been migrated just after the first
 115                  * vgetcpu but before fetching the version, so we
 116                  * wouldn't notice a version change.
 117                  */
 118                 cpu1 = __getcpu() & VGETCPU_CPU_MASK;
 119         } while (unlikely(cpu != cpu1 ||
 120                           (pvti->pvti.version & 1) ||
 121                           pvti->pvti.version != version));
 122
 123         if (unlikely(!(flags & PVCLOCK_TSC_STABLE_BIT)))
 124                 *mode = VCLOCK_NONE;
 125
 126         /* refer to tsc.c read_tsc() comment for rationale */
 127         last = VVAR(vsyscall_gtod_data).clock.cycle_last;
 128
 129         if (likely(ret >= last))
 130                 return ret;
 131
 132         return last;
 133 }
 134 #endif
 135
 136 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
 137 {
 138         long ret;
 139         asm("syscall" : "=a" (ret) :
 140             "0" (__NR_clock_gettime),"D" (clock), "S" (ts) : "memory");
 141         return ret;
 142 }
 143
 144 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
 145 {
 146         long ret;
 147
 148         asm("syscall" : "=a" (ret) :
 149             "0" (__NR_gettimeofday), "D" (tv), "S" (tz) : "memory");
 150         return ret;
 151 }
 152
 153
 154 notrace static inline u64 vgetsns(int *mode)
 155 {
 156         long v;
 157         cycles_t cycles;
 158         if (gtod->clock.vclock_mode == VCLOCK_TSC)
 159                 cycles = vread_tsc();
 160         else if (gtod->clock.vclock_mode == VCLOCK_HPET)
 161                 cycles = vread_hpet();
 162 #ifdef CONFIG_PARAVIRT_CLOCK
 163         else if (gtod->clock.vclock_mode == VCLOCK_PVCLOCK)
 164                 cycles = vread_pvclock(mode);
 165 #endif
 166         else
 167                 return 0;
 168         v = (cycles - gtod->clock.cycle_last) & gtod->clock.mask;
 169         return v * gtod->clock.mult;
 170 }
 171
 172 /* Code size doesn't matter (vdso is 4k anyway) and this is faster. */
 173 notrace static int __always_inline do_realtime(struct timespec *ts)
 174 {
 175         unsigned long seq;
 176         u64 ns;
 177         int mode;
 178
 179         ts->tv_nsec = 0;
 180         do {
 181                 seq = raw_read_seqcount_begin(&gtod->seq);
 182                 mode = gtod->clock.vclock_mode;
 183                 ts->tv_sec = gtod->wall_time_sec;
 184                 ns = gtod->wall_time_snsec;
 185                 ns += vgetsns(&mode);
 186                 ns >>= gtod->clock.shift;
 187         } while (unlikely(read_seqcount_retry(&gtod->seq, seq)));
 188
 189         timespec_add_ns(ts, ns);
 190         return mode;
 191 }
 192
 193 notrace static int do_monotonic(struct timespec *ts)
 194 {
 195         unsigned long seq;
 196         u64 ns;
 197         int mode;
 198
 199         ts->tv_nsec = 0;
 200         do {
 201                 seq = raw_read_seqcount_begin(&gtod->seq);
 202                 mode = gtod->clock.vclock_mode;
 203                 ts->tv_sec = gtod->monotonic_time_sec;
 204                 ns = gtod->monotonic_time_snsec;
 205                 ns += vgetsns(&mode);
 206                 ns >>= gtod->clock.shift;
 207         } while (unlikely(read_seqcount_retry(&gtod->seq, seq)));
 208         timespec_add_ns(ts, ns);
 209
 210         return mode;
 211 }
 212
 213 notrace static int do_realtime_coarse(struct timespec *ts)
 214 {
 215         unsigned long seq;
 216         do {
 217                 seq = raw_read_seqcount_begin(&gtod->seq);
 218                 ts->tv_sec = gtod->wall_time_coarse.tv_sec;
 219                 ts->tv_nsec = gtod->wall_time_coarse.tv_nsec;
 220         } while (unlikely(read_seqcount_retry(&gtod->seq, seq)));
 221         return 0;
 222 }
 223
 224 notrace static int do_monotonic_coarse(struct timespec *ts)
 225 {
 226         unsigned long seq;
 227         do {
 228                 seq = raw_read_seqcount_begin(&gtod->seq);
 229                 ts->tv_sec = gtod->monotonic_time_coarse.tv_sec;
 230                 ts->tv_nsec = gtod->monotonic_time_coarse.tv_nsec;
 231         } while (unlikely(read_seqcount_retry(&gtod->seq, seq)));
 232
 233         return 0;
 234 }
 235
 236 notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
 237 {
 238         int ret = VCLOCK_NONE;
 239
 240         switch (clock) {
 241         case CLOCK_REALTIME:
 242                 ret = do_realtime(ts);
 243                 break;
 244         case CLOCK_MONOTONIC:
 245                 ret = do_monotonic(ts);
 246                 break;
 247         case CLOCK_REALTIME_COARSE:
 248                 return do_realtime_coarse(ts);
 249         case CLOCK_MONOTONIC_COARSE:
 250                 return do_monotonic_coarse(ts);
 251         }
 252
 253         if (ret == VCLOCK_NONE)
 254                 return vdso_fallback_gettime(clock, ts);
 255         return 0;
 256 }
 257 int clock_gettime(clockid_t, struct timespec *)
 258         __attribute__((weak, alias("__vdso_clock_gettime")));
 259
 260 notrace int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz)
 261 {
 262         long ret = VCLOCK_NONE;
 263
 264         if (likely(tv != NULL)) {
 265                 BUILD_BUG_ON(offsetof(struct timeval, tv_usec) !=
 266                              offsetof(struct timespec, tv_nsec) ||
 267                              sizeof(*tv) != sizeof(struct timespec));
 268                 ret = do_realtime((struct timespec *)tv);
 269                 tv->tv_usec /= 1000;
 270         }
 271         if (unlikely(tz != NULL)) {
 272                 /* Avoid memcpy. Some old compilers fail to inline it */
 273                 tz->tz_minuteswest = gtod->sys_tz.tz_minuteswest;
 274                 tz->tz_dsttime = gtod->sys_tz.tz_dsttime;
 275         }
 276
 277         if (ret == VCLOCK_NONE)
 278                 return vdso_fallback_gtod(tv, tz);
 279         return 0;
 280 }
 281 int gettimeofday(struct timeval *, struct timezone *)
 282         __attribute__((weak, alias("__vdso_gettimeofday")));
 283
 284 /*
 285  * This will break when the xtime seconds get inaccurate, but that is
 286  * unlikely
 287  */
 288 notrace time_t __vdso_time(time_t *t)
 289 {
 290         /* This is atomic on x86_64 so we don't need any locks. */
 291         time_t result = ACCESS_ONCE(VVAR(vsyscall_gtod_data).wall_time_sec);
 292
 293         if (t)
 294                 *t = result;
 295         return result;
 296 }
 297 int time(time_t *t)
 298         __attribute__((weak, alias("__vdso_time")));