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  * 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 /* Disable profiling for userspace code: */
  15 #define DISABLE_BRANCH_PROFILING
  16
  17 #include <uapi/linux/time.h>
  18 #include <asm/vgtod.h>
  19 #include <asm/hpet.h>
  20 #include <asm/vvar.h>
  21 #include <asm/unistd.h>
  22 #include <asm/msr.h>
  23 #include <linux/math64.h>
  24 #include <linux/time.h>
  25
  26 #define gtod (&VVAR(vsyscall_gtod_data))
  27
  28 extern int __vdso_clock_gettime(clockid_t clock, struct timespec *ts);
  29 extern int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz);
  30 extern time_t __vdso_time(time_t *t);
  31
  32 #ifdef CONFIG_HPET_TIMER
  33 extern u8 hpet_page
  34         __attribute__((visibility("hidden")));
  35
  36 static notrace cycle_t vread_hpet(void)
  37 {
  38         return *(const volatile u32 *)(&hpet_page + HPET_COUNTER);
  39 }
  40 #endif
  41
  42 #ifndef BUILD_VDSO32
  43
  44 #include <linux/kernel.h>
  45 #include <asm/vsyscall.h>
  46 #include <asm/fixmap.h>
  47 #include <asm/pvclock.h>
  48
  49 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
  50 {
  51         long ret;
  52         asm("syscall" : "=a" (ret) :
  53             "0" (__NR_clock_gettime), "D" (clock), "S" (ts) : "memory");
  54         return ret;
  55 }
  56
  57 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
  58 {
  59         long ret;
  60
  61         asm("syscall" : "=a" (ret) :
  62             "0" (__NR_gettimeofday), "D" (tv), "S" (tz) : "memory");
  63         return ret;
  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 = gtod->cycle_last;
 128
 129         if (likely(ret >= last))
 130                 return ret;
 131
 132         return last;
 133 }
 134 #endif
 135
 136 #else
 137
 138 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
 139 {
 140         long ret;
 141
 142         asm(
 143                 "mov %%ebx, %%edx \n"
 144                 "mov %2, %%ebx \n"
 145                 "call __kernel_vsyscall \n"
 146                 "mov %%edx, %%ebx \n"
 147                 : "=a" (ret)
 148                 : "0" (__NR_clock_gettime), "g" (clock), "c" (ts)
 149                 : "memory", "edx");
 150         return ret;
 151 }
 152
 153 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
 154 {
 155         long ret;
 156
 157         asm(
 158                 "mov %%ebx, %%edx \n"
 159                 "mov %2, %%ebx \n"
 160                 "call __kernel_vsyscall \n"
 161                 "mov %%edx, %%ebx \n"
 162                 : "=a" (ret)
 163                 : "0" (__NR_gettimeofday), "g" (tv), "c" (tz)
 164                 : "memory", "edx");
 165         return ret;
 166 }
 167
 168 #ifdef CONFIG_PARAVIRT_CLOCK
 169
 170 static notrace cycle_t vread_pvclock(int *mode)
 171 {
 172         *mode = VCLOCK_NONE;
 173         return 0;
 174 }
 175 #endif
 176
 177 #endif
 178
 179 notrace static cycle_t vread_tsc(void)
 180 {
 181         cycle_t ret;
 182         u64 last;
 183
 184         /*
 185          * Empirically, a fence (of type that depends on the CPU)
 186          * before rdtsc is enough to ensure that rdtsc is ordered
 187          * with respect to loads.  The various CPU manuals are unclear
 188          * as to whether rdtsc can be reordered with later loads,
 189          * but no one has ever seen it happen.
 190          */
 191         rdtsc_barrier();
 192         ret = (cycle_t)__native_read_tsc();
 193
 194         last = gtod->cycle_last;
 195
 196         if (likely(ret >= last))
 197                 return ret;
 198
 199         /*
 200          * GCC likes to generate cmov here, but this branch is extremely
 201          * predictable (it's just a funciton of time and the likely is
 202          * very likely) and there's a data dependence, so force GCC
 203          * to generate a branch instead.  I don't barrier() because
 204          * we don't actually need a barrier, and if this function
 205          * ever gets inlined it will generate worse code.
 206          */
 207         asm volatile ("");
 208         return last;
 209 }
 210
 211 notrace static inline u64 vgetsns(int *mode)
 212 {
 213         u64 v;
 214         cycles_t cycles;
 215
 216         if (gtod->vclock_mode == VCLOCK_TSC)
 217                 cycles = vread_tsc();
 218 #ifdef CONFIG_HPET_TIMER
 219         else if (gtod->vclock_mode == VCLOCK_HPET)
 220                 cycles = vread_hpet();
 221 #endif
 222 #ifdef CONFIG_PARAVIRT_CLOCK
 223         else if (gtod->vclock_mode == VCLOCK_PVCLOCK)
 224                 cycles = vread_pvclock(mode);
 225 #endif
 226         else
 227                 return 0;
 228         v = (cycles - gtod->cycle_last) & gtod->mask;
 229         return v * gtod->mult;
 230 }
 231
 232 /* Code size doesn't matter (vdso is 4k anyway) and this is faster. */
 233 notrace static int __always_inline do_realtime(struct timespec *ts)
 234 {
 235         unsigned long seq;
 236         u64 ns;
 237         int mode;
 238
 239         do {
 240                 seq = gtod_read_begin(gtod);
 241                 mode = gtod->vclock_mode;
 242                 ts->tv_sec = gtod->wall_time_sec;
 243                 ns = gtod->wall_time_snsec;
 244                 ns += vgetsns(&mode);
 245                 ns >>= gtod->shift;
 246         } while (unlikely(gtod_read_retry(gtod, seq)));
 247
 248         ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
 249         ts->tv_nsec = ns;
 250
 251         return mode;
 252 }
 253
 254 notrace static int __always_inline do_monotonic(struct timespec *ts)
 255 {
 256         unsigned long seq;
 257         u64 ns;
 258         int mode;
 259
 260         do {
 261                 seq = gtod_read_begin(gtod);
 262                 mode = gtod->vclock_mode;
 263                 ts->tv_sec = gtod->monotonic_time_sec;
 264                 ns = gtod->monotonic_time_snsec;
 265                 ns += vgetsns(&mode);
 266                 ns >>= gtod->shift;
 267         } while (unlikely(gtod_read_retry(gtod, seq)));
 268
 269         ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
 270         ts->tv_nsec = ns;
 271
 272         return mode;
 273 }
 274
 275 notrace static void do_realtime_coarse(struct timespec *ts)
 276 {
 277         unsigned long seq;
 278         do {
 279                 seq = gtod_read_begin(gtod);
 280                 ts->tv_sec = gtod->wall_time_coarse_sec;
 281                 ts->tv_nsec = gtod->wall_time_coarse_nsec;
 282         } while (unlikely(gtod_read_retry(gtod, seq)));
 283 }
 284
 285 notrace static void do_monotonic_coarse(struct timespec *ts)
 286 {
 287         unsigned long seq;
 288         do {
 289                 seq = gtod_read_begin(gtod);
 290                 ts->tv_sec = gtod->monotonic_time_coarse_sec;
 291                 ts->tv_nsec = gtod->monotonic_time_coarse_nsec;
 292         } while (unlikely(gtod_read_retry(gtod, seq)));
 293 }
 294
 295 notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
 296 {
 297         switch (clock) {
 298         case CLOCK_REALTIME:
 299                 if (do_realtime(ts) == VCLOCK_NONE)
 300                         goto fallback;
 301                 break;
 302         case CLOCK_MONOTONIC:
 303                 if (do_monotonic(ts) == VCLOCK_NONE)
 304                         goto fallback;
 305                 break;
 306         case CLOCK_REALTIME_COARSE:
 307                 do_realtime_coarse(ts);
 308                 break;
 309         case CLOCK_MONOTONIC_COARSE:
 310                 do_monotonic_coarse(ts);
 311                 break;
 312         default:
 313                 goto fallback;
 314         }
 315
 316         return 0;
 317 fallback:
 318         return vdso_fallback_gettime(clock, ts);
 319 }
 320 int clock_gettime(clockid_t, struct timespec *)
 321         __attribute__((weak, alias("__vdso_clock_gettime")));
 322
 323 notrace int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz)
 324 {
 325         if (likely(tv != NULL)) {
 326                 if (unlikely(do_realtime((struct timespec *)tv) == VCLOCK_NONE))
 327                         return vdso_fallback_gtod(tv, tz);
 328                 tv->tv_usec /= 1000;
 329         }
 330         if (unlikely(tz != NULL)) {
 331                 tz->tz_minuteswest = gtod->tz_minuteswest;
 332                 tz->tz_dsttime = gtod->tz_dsttime;
 333         }
 334
 335         return 0;
 336 }
 337 int gettimeofday(struct timeval *, struct timezone *)
 338         __attribute__((weak, alias("__vdso_gettimeofday")));
 339
 340 /*
 341  * This will break when the xtime seconds get inaccurate, but that is
 342  * unlikely
 343  */
 344 notrace time_t __vdso_time(time_t *t)
 345 {
 346         /* This is atomic on x86 so we don't need any locks. */
 347         time_t result = ACCESS_ONCE(gtod->wall_time_sec);
 348
 349         if (t)
 350                 *t = result;
 351         return result;
 352 }
 353 int time(time_t *t)
 354         __attribute__((weak, alias("__vdso_time")));