arch/x86/entry/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  * 32 Bit compat layer by Stefani Seibold <stefani@seibold.net>
   8  *  sponsored by Rohde & Schwarz GmbH & Co. KG Munich/Germany
   9  *
  10  * The code should have no internal unresolved relocations.
  11  * Check with readelf after changing.
  12  */
  13
  14 #include <uapi/linux/time.h>
  15 #include <asm/vgtod.h>
  16 #include <asm/vvar.h>
  17 #include <asm/unistd.h>
  18 #include <asm/msr.h>
  19 #include <asm/pvclock.h>
  20 #include <linux/math64.h>
  21 #include <linux/time.h>
  22 #include <linux/kernel.h>
  23
  24 #define gtod (&VVAR(vsyscall_gtod_data))
  25
  26 extern int __vdso_clock_gettime(clockid_t clock, struct timespec *ts);
  27 extern int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz);
  28 extern time_t __vdso_time(time_t *t);
  29
  30 #ifdef CONFIG_PARAVIRT_CLOCK
  31 extern u8 pvclock_page
  32         __attribute__((visibility("hidden")));
  33 #endif
  34
  35 #ifndef BUILD_VDSO32
  36
  37 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
  38 {
  39         long ret;
  40         asm("syscall" : "=a" (ret) :
  41             "0" (__NR_clock_gettime), "D" (clock), "S" (ts) : "memory");
  42         return ret;
  43 }
  44
  45 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
  46 {
  47         long ret;
  48
  49         asm("syscall" : "=a" (ret) :
  50             "0" (__NR_gettimeofday), "D" (tv), "S" (tz) : "memory");
  51         return ret;
  52 }
  53
  54
  55 #else
  56
  57 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
  58 {
  59         long ret;
  60
  61         asm(
  62                 "mov %%ebx, %%edx \n"
  63                 "mov %2, %%ebx \n"
  64                 "call __kernel_vsyscall \n"
  65                 "mov %%edx, %%ebx \n"
  66                 : "=a" (ret)
  67                 : "0" (__NR_clock_gettime), "g" (clock), "c" (ts)
  68                 : "memory", "edx");
  69         return ret;
  70 }
  71
  72 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
  73 {
  74         long ret;
  75
  76         asm(
  77                 "mov %%ebx, %%edx \n"
  78                 "mov %2, %%ebx \n"
  79                 "call __kernel_vsyscall \n"
  80                 "mov %%edx, %%ebx \n"
  81                 : "=a" (ret)
  82                 : "0" (__NR_gettimeofday), "g" (tv), "c" (tz)
  83                 : "memory", "edx");
  84         return ret;
  85 }
  86
  87 #endif
  88
  89 #ifdef CONFIG_PARAVIRT_CLOCK
  90 static notrace const struct pvclock_vsyscall_time_info *get_pvti0(void)
  91 {
  92         return (const struct pvclock_vsyscall_time_info *)&pvclock_page;
  93 }
  94
  95 static notrace cycle_t vread_pvclock(int *mode)
  96 {
  97         const struct pvclock_vcpu_time_info *pvti = &get_pvti0()->pvti;
  98         cycle_t ret;
  99         u64 tsc, pvti_tsc;
 100         u64 last, delta, pvti_system_time;
 101         u32 version, pvti_tsc_to_system_mul, pvti_tsc_shift;
 102
 103         /*
 104          * Note: The kernel and hypervisor must guarantee that cpu ID
 105          * number maps 1:1 to per-CPU pvclock time info.
 106          *
 107          * Because the hypervisor is entirely unaware of guest userspace
 108          * preemption, it cannot guarantee that per-CPU pvclock time
 109          * info is updated if the underlying CPU changes or that that
 110          * version is increased whenever underlying CPU changes.
 111          *
 112          * On KVM, we are guaranteed that pvti updates for any vCPU are
 113          * atomic as seen by *all* vCPUs.  This is an even stronger
 114          * guarantee than we get with a normal seqlock.
 115          *
 116          * On Xen, we don't appear to have that guarantee, but Xen still
 117          * supplies a valid seqlock using the version field.
 118          *
 119          * We only do pvclock vdso timing at all if
 120          * PVCLOCK_TSC_STABLE_BIT is set, and we interpret that bit to
 121          * mean that all vCPUs have matching pvti and that the TSC is
 122          * synced, so we can just look at vCPU 0's pvti.
 123          */
 124
 125         do {
 126                 version = pvti->version;
 127
 128                 smp_rmb();
 129
 130                 if (unlikely(!(pvti->flags & PVCLOCK_TSC_STABLE_BIT))) {
 131                         *mode = VCLOCK_NONE;
 132                         return 0;
 133                 }
 134
 135                 tsc = rdtsc_ordered();
 136                 pvti_tsc_to_system_mul = pvti->tsc_to_system_mul;
 137                 pvti_tsc_shift = pvti->tsc_shift;
 138                 pvti_system_time = pvti->system_time;
 139                 pvti_tsc = pvti->tsc_timestamp;
 140
 141                 /* Make sure that the version double-check is last. */
 142                 smp_rmb();
 143         } while (unlikely((version & 1) || version != pvti->version));
 144
 145         delta = tsc - pvti_tsc;
 146         ret = pvti_system_time +
 147                 pvclock_scale_delta(delta, pvti_tsc_to_system_mul,
 148                                     pvti_tsc_shift);
 149
 150         /* refer to vread_tsc() comment for rationale */
 151         last = gtod->cycle_last;
 152
 153         if (likely(ret >= last))
 154                 return ret;
 155
 156         return last;
 157 }
 158 #endif
 159
 160 notrace static cycle_t vread_tsc(void)
 161 {
 162         cycle_t ret = (cycle_t)rdtsc_ordered();
 163         u64 last = gtod->cycle_last;
 164
 165         if (likely(ret >= last))
 166                 return ret;
 167
 168         /*
 169          * GCC likes to generate cmov here, but this branch is extremely
 170          * predictable (it's just a function of time and the likely is
 171          * very likely) and there's a data dependence, so force GCC
 172          * to generate a branch instead.  I don't barrier() because
 173          * we don't actually need a barrier, and if this function
 174          * ever gets inlined it will generate worse code.
 175          */
 176         asm volatile ("");
 177         return last;
 178 }
 179
 180 notrace static inline u64 vgetsns(int *mode)
 181 {
 182         u64 v;
 183         cycles_t cycles;
 184
 185         if (gtod->vclock_mode == VCLOCK_TSC)
 186                 cycles = vread_tsc();
 187 #ifdef CONFIG_PARAVIRT_CLOCK
 188         else if (gtod->vclock_mode == VCLOCK_PVCLOCK)
 189                 cycles = vread_pvclock(mode);
 190 #endif
 191         else
 192                 return 0;
 193         v = (cycles - gtod->cycle_last) & gtod->mask;
 194         return v * gtod->mult;
 195 }
 196
 197 /* Code size doesn't matter (vdso is 4k anyway) and this is faster. */
 198 notrace static int __always_inline do_realtime(struct timespec *ts)
 199 {
 200         unsigned long seq;
 201         u64 ns;
 202         int mode;
 203
 204         do {
 205                 seq = gtod_read_begin(gtod);
 206                 mode = gtod->vclock_mode;
 207                 ts->tv_sec = gtod->wall_time_sec;
 208                 ns = gtod->wall_time_snsec;
 209                 ns += vgetsns(&mode);
 210                 ns >>= gtod->shift;
 211         } while (unlikely(gtod_read_retry(gtod, seq)));
 212
 213         ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
 214         ts->tv_nsec = ns;
 215
 216         return mode;
 217 }
 218
 219 notrace static int __always_inline do_monotonic(struct timespec *ts)
 220 {
 221         unsigned long seq;
 222         u64 ns;
 223         int mode;
 224
 225         do {
 226                 seq = gtod_read_begin(gtod);
 227                 mode = gtod->vclock_mode;
 228                 ts->tv_sec = gtod->monotonic_time_sec;
 229                 ns = gtod->monotonic_time_snsec;
 230                 ns += vgetsns(&mode);
 231                 ns >>= gtod->shift;
 232         } while (unlikely(gtod_read_retry(gtod, seq)));
 233
 234         ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
 235         ts->tv_nsec = ns;
 236
 237         return mode;
 238 }
 239
 240 notrace static void do_realtime_coarse(struct timespec *ts)
 241 {
 242         unsigned long seq;
 243         do {
 244                 seq = gtod_read_begin(gtod);
 245                 ts->tv_sec = gtod->wall_time_coarse_sec;
 246                 ts->tv_nsec = gtod->wall_time_coarse_nsec;
 247         } while (unlikely(gtod_read_retry(gtod, seq)));
 248 }
 249
 250 notrace static void do_monotonic_coarse(struct timespec *ts)
 251 {
 252         unsigned long seq;
 253         do {
 254                 seq = gtod_read_begin(gtod);
 255                 ts->tv_sec = gtod->monotonic_time_coarse_sec;
 256                 ts->tv_nsec = gtod->monotonic_time_coarse_nsec;
 257         } while (unlikely(gtod_read_retry(gtod, seq)));
 258 }
 259
 260 notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
 261 {
 262         switch (clock) {
 263         case CLOCK_REALTIME:
 264                 if (do_realtime(ts) == VCLOCK_NONE)
 265                         goto fallback;
 266                 break;
 267         case CLOCK_MONOTONIC:
 268                 if (do_monotonic(ts) == VCLOCK_NONE)
 269                         goto fallback;
 270                 break;
 271         case CLOCK_REALTIME_COARSE:
 272                 do_realtime_coarse(ts);
 273                 break;
 274         case CLOCK_MONOTONIC_COARSE:
 275                 do_monotonic_coarse(ts);
 276                 break;
 277         default:
 278                 goto fallback;
 279         }
 280
 281         return 0;
 282 fallback:
 283         return vdso_fallback_gettime(clock, ts);
 284 }
 285 int clock_gettime(clockid_t, struct timespec *)
 286         __attribute__((weak, alias("__vdso_clock_gettime")));
 287
 288 notrace int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz)
 289 {
 290         if (likely(tv != NULL)) {
 291                 if (unlikely(do_realtime((struct timespec *)tv) == VCLOCK_NONE))
 292                         return vdso_fallback_gtod(tv, tz);
 293                 tv->tv_usec /= 1000;
 294         }
 295         if (unlikely(tz != NULL)) {
 296                 tz->tz_minuteswest = gtod->tz_minuteswest;
 297                 tz->tz_dsttime = gtod->tz_dsttime;
 298         }
 299
 300         return 0;
 301 }
 302 int gettimeofday(struct timeval *, struct timezone *)
 303         __attribute__((weak, alias("__vdso_gettimeofday")));
 304
 305 /*
 306  * This will break when the xtime seconds get inaccurate, but that is
 307  * unlikely
 308  */
 309 notrace time_t __vdso_time(time_t *t)
 310 {
 311         /* This is atomic on x86 so we don't need any locks. */
 312         time_t result = ACCESS_ONCE(gtod->wall_time_sec);
 313
 314         if (t)
 315                 *t = result;
 316         return result;
 317 }
 318 int time(time_t *t)
 319         __attribute__((weak, alias("__vdso_time")));