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 #include <uapi/linux/time.h>
  15 #include <asm/vgtod.h>
  16 #include <asm/hpet.h>
  17 #include <asm/vvar.h>
  18 #include <asm/unistd.h>
  19 #include <asm/msr.h>
  20 #include <linux/math64.h>
  21 #include <linux/time.h>
  22
  23 #define gtod (&VVAR(vsyscall_gtod_data))
  24
  25 extern int __vdso_clock_gettime(clockid_t clock, struct timespec *ts);
  26 extern int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz);
  27 extern time_t __vdso_time(time_t *t);
  28
  29 #ifdef CONFIG_HPET_TIMER
  30 extern u8 hpet_page
  31         __attribute__((visibility("hidden")));
  32
  33 static notrace cycle_t vread_hpet(void)
  34 {
  35         return *(const volatile u32 *)(&hpet_page + HPET_COUNTER);
  36 }
  37 #endif
  38
  39 #ifndef BUILD_VDSO32
  40
  41 #include <linux/kernel.h>
  42 #include <asm/vsyscall.h>
  43 #include <asm/fixmap.h>
  44 #include <asm/pvclock.h>
  45
  46 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
  47 {
  48         long ret;
  49         asm("syscall" : "=a" (ret) :
  50             "0" (__NR_clock_gettime), "D" (clock), "S" (ts) : "memory");
  51         return ret;
  52 }
  53
  54 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
  55 {
  56         long ret;
  57
  58         asm("syscall" : "=a" (ret) :
  59             "0" (__NR_gettimeofday), "D" (tv), "S" (tz) : "memory");
  60         return ret;
  61 }
  62
  63 #ifdef CONFIG_PARAVIRT_CLOCK
  64
  65 static notrace const struct pvclock_vsyscall_time_info *get_pvti(int cpu)
  66 {
  67         const struct pvclock_vsyscall_time_info *pvti_base;
  68         int idx = cpu / (PAGE_SIZE/PVTI_SIZE);
  69         int offset = cpu % (PAGE_SIZE/PVTI_SIZE);
  70
  71         BUG_ON(PVCLOCK_FIXMAP_BEGIN + idx > PVCLOCK_FIXMAP_END);
  72
  73         pvti_base = (struct pvclock_vsyscall_time_info *)
  74                     __fix_to_virt(PVCLOCK_FIXMAP_BEGIN+idx);
  75
  76         return &pvti_base[offset];
  77 }
  78
  79 static notrace cycle_t vread_pvclock(int *mode)
  80 {
  81         const struct pvclock_vsyscall_time_info *pvti;
  82         cycle_t ret;
  83         u64 last;
  84         u32 version;
  85         u8 flags;
  86         unsigned cpu, cpu1;
  87
  88
  89         /*
  90          * Note: hypervisor must guarantee that:
  91          * 1. cpu ID number maps 1:1 to per-CPU pvclock time info.
  92          * 2. that per-CPU pvclock time info is updated if the
  93          *    underlying CPU changes.
  94          * 3. that version is increased whenever underlying CPU
  95          *    changes.
  96          *
  97          */
  98         do {
  99                 cpu = __getcpu() & VGETCPU_CPU_MASK;
 100                 /* TODO: We can put vcpu id into higher bits of pvti.version.
 101                  * This will save a couple of cycles by getting rid of
 102                  * __getcpu() calls (Gleb).
 103                  */
 104
 105                 pvti = get_pvti(cpu);
 106
 107                 version = __pvclock_read_cycles(&pvti->pvti, &ret, &flags);
 108
 109                 /*
 110                  * Test we're still on the cpu as well as the version.
 111                  * We could have been migrated just after the first
 112                  * vgetcpu but before fetching the version, so we
 113                  * wouldn't notice a version change.
 114                  */
 115                 cpu1 = __getcpu() & VGETCPU_CPU_MASK;
 116         } while (unlikely(cpu != cpu1 ||
 117                           (pvti->pvti.version & 1) ||
 118                           pvti->pvti.version != version));
 119
 120         if (unlikely(!(flags & PVCLOCK_TSC_STABLE_BIT)))
 121                 *mode = VCLOCK_NONE;
 122
 123         /* refer to tsc.c read_tsc() comment for rationale */
 124         last = gtod->cycle_last;
 125
 126         if (likely(ret >= last))
 127                 return ret;
 128
 129         return last;
 130 }
 131 #endif
 132
 133 #else
 134
 135 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
 136 {
 137         long ret;
 138
 139         asm(
 140                 "mov %%ebx, %%edx \n"
 141                 "mov %2, %%ebx \n"
 142                 "call __kernel_vsyscall \n"
 143                 "mov %%edx, %%ebx \n"
 144                 : "=a" (ret)
 145                 : "0" (__NR_clock_gettime), "g" (clock), "c" (ts)
 146                 : "memory", "edx");
 147         return ret;
 148 }
 149
 150 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
 151 {
 152         long ret;
 153
 154         asm(
 155                 "mov %%ebx, %%edx \n"
 156                 "mov %2, %%ebx \n"
 157                 "call __kernel_vsyscall \n"
 158                 "mov %%edx, %%ebx \n"
 159                 : "=a" (ret)
 160                 : "0" (__NR_gettimeofday), "g" (tv), "c" (tz)
 161                 : "memory", "edx");
 162         return ret;
 163 }
 164
 165 #ifdef CONFIG_PARAVIRT_CLOCK
 166
 167 static notrace cycle_t vread_pvclock(int *mode)
 168 {
 169         *mode = VCLOCK_NONE;
 170         return 0;
 171 }
 172 #endif
 173
 174 #endif
 175
 176 notrace static cycle_t vread_tsc(void)
 177 {
 178         cycle_t ret;
 179         u64 last;
 180
 181         /*
 182          * Empirically, a fence (of type that depends on the CPU)
 183          * before rdtsc is enough to ensure that rdtsc is ordered
 184          * with respect to loads.  The various CPU manuals are unclear
 185          * as to whether rdtsc can be reordered with later loads,
 186          * but no one has ever seen it happen.
 187          */
 188         rdtsc_barrier();
 189         ret = (cycle_t)__native_read_tsc();
 190
 191         last = gtod->cycle_last;
 192
 193         if (likely(ret >= last))
 194                 return ret;
 195
 196         /*
 197          * GCC likes to generate cmov here, but this branch is extremely
 198          * predictable (it's just a funciton of time and the likely is
 199          * very likely) and there's a data dependence, so force GCC
 200          * to generate a branch instead.  I don't barrier() because
 201          * we don't actually need a barrier, and if this function
 202          * ever gets inlined it will generate worse code.
 203          */
 204         asm volatile ("");
 205         return last;
 206 }
 207
 208 notrace static inline u64 vgetsns(int *mode)
 209 {
 210         u64 v;
 211         cycles_t cycles;
 212
 213         if (gtod->vclock_mode == VCLOCK_TSC)
 214                 cycles = vread_tsc();
 215 #ifdef CONFIG_HPET_TIMER
 216         else if (gtod->vclock_mode == VCLOCK_HPET)
 217                 cycles = vread_hpet();
 218 #endif
 219 #ifdef CONFIG_PARAVIRT_CLOCK
 220         else if (gtod->vclock_mode == VCLOCK_PVCLOCK)
 221                 cycles = vread_pvclock(mode);
 222 #endif
 223         else
 224                 return 0;
 225         v = (cycles - gtod->cycle_last) & gtod->mask;
 226         return v * gtod->mult;
 227 }
 228
 229 /* Code size doesn't matter (vdso is 4k anyway) and this is faster. */
 230 notrace static int __always_inline do_realtime(struct timespec *ts)
 231 {
 232         unsigned long seq;
 233         u64 ns;
 234         int mode;
 235
 236         do {
 237                 seq = gtod_read_begin(gtod);
 238                 mode = gtod->vclock_mode;
 239                 ts->tv_sec = gtod->wall_time_sec;
 240                 ns = gtod->wall_time_snsec;
 241                 ns += vgetsns(&mode);
 242                 ns >>= gtod->shift;
 243         } while (unlikely(gtod_read_retry(gtod, seq)));
 244
 245         ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
 246         ts->tv_nsec = ns;
 247
 248         return mode;
 249 }
 250
 251 notrace static int __always_inline do_monotonic(struct timespec *ts)
 252 {
 253         unsigned long seq;
 254         u64 ns;
 255         int mode;
 256
 257         do {
 258                 seq = gtod_read_begin(gtod);
 259                 mode = gtod->vclock_mode;
 260                 ts->tv_sec = gtod->monotonic_time_sec;
 261                 ns = gtod->monotonic_time_snsec;
 262                 ns += vgetsns(&mode);
 263                 ns >>= gtod->shift;
 264         } while (unlikely(gtod_read_retry(gtod, seq)));
 265
 266         ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
 267         ts->tv_nsec = ns;
 268
 269         return mode;
 270 }
 271
 272 notrace static void do_realtime_coarse(struct timespec *ts)
 273 {
 274         unsigned long seq;
 275         do {
 276                 seq = gtod_read_begin(gtod);
 277                 ts->tv_sec = gtod->wall_time_coarse_sec;
 278                 ts->tv_nsec = gtod->wall_time_coarse_nsec;
 279         } while (unlikely(gtod_read_retry(gtod, seq)));
 280 }
 281
 282 notrace static void do_monotonic_coarse(struct timespec *ts)
 283 {
 284         unsigned long seq;
 285         do {
 286                 seq = gtod_read_begin(gtod);
 287                 ts->tv_sec = gtod->monotonic_time_coarse_sec;
 288                 ts->tv_nsec = gtod->monotonic_time_coarse_nsec;
 289         } while (unlikely(gtod_read_retry(gtod, seq)));
 290 }
 291
 292 notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
 293 {
 294         switch (clock) {
 295         case CLOCK_REALTIME:
 296                 if (do_realtime(ts) == VCLOCK_NONE)
 297                         goto fallback;
 298                 break;
 299         case CLOCK_MONOTONIC:
 300                 if (do_monotonic(ts) == VCLOCK_NONE)
 301                         goto fallback;
 302                 break;
 303         case CLOCK_REALTIME_COARSE:
 304                 do_realtime_coarse(ts);
 305                 break;
 306         case CLOCK_MONOTONIC_COARSE:
 307                 do_monotonic_coarse(ts);
 308                 break;
 309         default:
 310                 goto fallback;
 311         }
 312
 313         return 0;
 314 fallback:
 315         return vdso_fallback_gettime(clock, ts);
 316 }
 317 int clock_gettime(clockid_t, struct timespec *)
 318         __attribute__((weak, alias("__vdso_clock_gettime")));
 319
 320 notrace int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz)
 321 {
 322         if (likely(tv != NULL)) {
 323                 if (unlikely(do_realtime((struct timespec *)tv) == VCLOCK_NONE))
 324                         return vdso_fallback_gtod(tv, tz);
 325                 tv->tv_usec /= 1000;
 326         }
 327         if (unlikely(tz != NULL)) {
 328                 tz->tz_minuteswest = gtod->tz_minuteswest;
 329                 tz->tz_dsttime = gtod->tz_dsttime;
 330         }
 331
 332         return 0;
 333 }
 334 int gettimeofday(struct timeval *, struct timezone *)
 335         __attribute__((weak, alias("__vdso_gettimeofday")));
 336
 337 /*
 338  * This will break when the xtime seconds get inaccurate, but that is
 339  * unlikely
 340  */
 341 notrace time_t __vdso_time(time_t *t)
 342 {
 343         /* This is atomic on x86 so we don't need any locks. */
 344         time_t result = ACCESS_ONCE(gtod->wall_time_sec);
 345
 346         if (t)
 347                 *t = result;
 348         return result;
 349 }
 350 int time(time_t *t)
 351         __attribute__((weak, alias("__vdso_time")));