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 <asm/mshyperv.h>
  21 #include <linux/math64.h>
  22 #include <linux/time.h>
  23 #include <linux/kernel.h>
  24
  25 #define gtod (&VVAR(vsyscall_gtod_data))
  26
  27 extern int __vdso_clock_gettime(clockid_t clock, struct timespec *ts);
  28 extern int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz);
  29 extern time_t __vdso_time(time_t *t);
  30
  31 #ifdef CONFIG_PARAVIRT_CLOCK
  32 extern u8 pvclock_page[PAGE_SIZE]
  33         __attribute__((visibility("hidden")));
  34 #endif
  35
  36 #ifdef CONFIG_HYPERV_TSCPAGE
  37 extern u8 hvclock_page[PAGE_SIZE]
  38         __attribute__((visibility("hidden")));
  39 #endif
  40
  41 #ifndef BUILD_VDSO32
  42
  43 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
  44 {
  45         long ret;
  46         asm ("syscall" : "=a" (ret), "=m" (*ts) :
  47              "0" (__NR_clock_gettime), "D" (clock), "S" (ts) :
  48              "memory", "rcx", "r11");
  49         return ret;
  50 }
  51
  52 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
  53 {
  54         long ret;
  55
  56         asm ("syscall" : "=a" (ret), "=m" (*tv), "=m" (*tz) :
  57              "0" (__NR_gettimeofday), "D" (tv), "S" (tz) :
  58              "memory", "rcx", "r11");
  59         return ret;
  60 }
  61
  62
  63 #else
  64
  65 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
  66 {
  67         long ret;
  68
  69         asm (
  70                 "mov %%ebx, %%edx \n"
  71                 "mov %[clock], %%ebx \n"
  72                 "call __kernel_vsyscall \n"
  73                 "mov %%edx, %%ebx \n"
  74                 : "=a" (ret), "=m" (*ts)
  75                 : "0" (__NR_clock_gettime), [clock] "g" (clock), "c" (ts)
  76                 : "memory", "edx");
  77         return ret;
  78 }
  79
  80 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
  81 {
  82         long ret;
  83
  84         asm (
  85                 "mov %%ebx, %%edx \n"
  86                 "mov %[tv], %%ebx \n"
  87                 "call __kernel_vsyscall \n"
  88                 "mov %%edx, %%ebx \n"
  89                 : "=a" (ret), "=m" (*tv), "=m" (*tz)
  90                 : "0" (__NR_gettimeofday), [tv] "g" (tv), "c" (tz)
  91                 : "memory", "edx");
  92         return ret;
  93 }
  94
  95 #endif
  96
  97 #ifdef CONFIG_PARAVIRT_CLOCK
  98 static notrace const struct pvclock_vsyscall_time_info *get_pvti0(void)
  99 {
 100         return (const struct pvclock_vsyscall_time_info *)&pvclock_page;
 101 }
 102
 103 static notrace u64 vread_pvclock(int *mode)
 104 {
 105         const struct pvclock_vcpu_time_info *pvti = &get_pvti0()->pvti;
 106         u64 ret;
 107         u64 last;
 108         u32 version;
 109
 110         /*
 111          * Note: The kernel and hypervisor must guarantee that cpu ID
 112          * number maps 1:1 to per-CPU pvclock time info.
 113          *
 114          * Because the hypervisor is entirely unaware of guest userspace
 115          * preemption, it cannot guarantee that per-CPU pvclock time
 116          * info is updated if the underlying CPU changes or that that
 117          * version is increased whenever underlying CPU changes.
 118          *
 119          * On KVM, we are guaranteed that pvti updates for any vCPU are
 120          * atomic as seen by *all* vCPUs.  This is an even stronger
 121          * guarantee than we get with a normal seqlock.
 122          *
 123          * On Xen, we don't appear to have that guarantee, but Xen still
 124          * supplies a valid seqlock using the version field.
 125          *
 126          * We only do pvclock vdso timing at all if
 127          * PVCLOCK_TSC_STABLE_BIT is set, and we interpret that bit to
 128          * mean that all vCPUs have matching pvti and that the TSC is
 129          * synced, so we can just look at vCPU 0's pvti.
 130          */
 131
 132         do {
 133                 version = pvclock_read_begin(pvti);
 134
 135                 if (unlikely(!(pvti->flags & PVCLOCK_TSC_STABLE_BIT))) {
 136                         *mode = VCLOCK_NONE;
 137                         return 0;
 138                 }
 139
 140                 ret = __pvclock_read_cycles(pvti, rdtsc_ordered());
 141         } while (pvclock_read_retry(pvti, version));
 142
 143         /* refer to vread_tsc() comment for rationale */
 144         last = gtod->cycle_last;
 145
 146         if (likely(ret >= last))
 147                 return ret;
 148
 149         return last;
 150 }
 151 #endif
 152 #ifdef CONFIG_HYPERV_TSCPAGE
 153 static notrace u64 vread_hvclock(int *mode)
 154 {
 155         const struct ms_hyperv_tsc_page *tsc_pg =
 156                 (const struct ms_hyperv_tsc_page *)&hvclock_page;
 157         u64 current_tick = hv_read_tsc_page(tsc_pg);
 158
 159         if (current_tick != U64_MAX)
 160                 return current_tick;
 161
 162         *mode = VCLOCK_NONE;
 163         return 0;
 164 }
 165 #endif
 166
 167 notrace static u64 vread_tsc(void)
 168 {
 169         u64 ret = (u64)rdtsc_ordered();
 170         u64 last = gtod->cycle_last;
 171
 172         if (likely(ret >= last))
 173                 return ret;
 174
 175         /*
 176          * GCC likes to generate cmov here, but this branch is extremely
 177          * predictable (it's just a function of time and the likely is
 178          * very likely) and there's a data dependence, so force GCC
 179          * to generate a branch instead.  I don't barrier() because
 180          * we don't actually need a barrier, and if this function
 181          * ever gets inlined it will generate worse code.
 182          */
 183         asm volatile ("");
 184         return last;
 185 }
 186
 187 notrace static inline u64 vgetsns(int *mode)
 188 {
 189         u64 v;
 190         cycles_t cycles;
 191
 192         if (gtod->vclock_mode == VCLOCK_TSC)
 193                 cycles = vread_tsc();
 194
 195         /*
 196          * For any memory-mapped vclock type, we need to make sure that gcc
 197          * doesn't cleverly hoist a load before the mode check.  Otherwise we
 198          * might end up touching the memory-mapped page even if the vclock in
 199          * question isn't enabled, which will segfault.  Hence the barriers.
 200          */
 201 #ifdef CONFIG_PARAVIRT_CLOCK
 202         else if (gtod->vclock_mode == VCLOCK_PVCLOCK) {
 203                 barrier();
 204                 cycles = vread_pvclock(mode);
 205         }
 206 #endif
 207 #ifdef CONFIG_HYPERV_TSCPAGE
 208         else if (gtod->vclock_mode == VCLOCK_HVCLOCK) {
 209                 barrier();
 210                 cycles = vread_hvclock(mode);
 211         }
 212 #endif
 213         else
 214                 return 0;
 215         v = (cycles - gtod->cycle_last) & gtod->mask;
 216         return v * gtod->mult;
 217 }
 218
 219 /* Code size doesn't matter (vdso is 4k anyway) and this is faster. */
 220 notrace static int __always_inline do_realtime(struct timespec *ts)
 221 {
 222         unsigned long seq;
 223         u64 ns;
 224         int mode;
 225
 226         do {
 227                 seq = gtod_read_begin(gtod);
 228                 mode = gtod->vclock_mode;
 229                 ts->tv_sec = gtod->wall_time_sec;
 230                 ns = gtod->wall_time_snsec;
 231                 ns += vgetsns(&mode);
 232                 ns >>= gtod->shift;
 233         } while (unlikely(gtod_read_retry(gtod, seq)));
 234
 235         ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
 236         ts->tv_nsec = ns;
 237
 238         return mode;
 239 }
 240
 241 notrace static int __always_inline do_monotonic(struct timespec *ts)
 242 {
 243         unsigned long seq;
 244         u64 ns;
 245         int mode;
 246
 247         do {
 248                 seq = gtod_read_begin(gtod);
 249                 mode = gtod->vclock_mode;
 250                 ts->tv_sec = gtod->monotonic_time_sec;
 251                 ns = gtod->monotonic_time_snsec;
 252                 ns += vgetsns(&mode);
 253                 ns >>= gtod->shift;
 254         } while (unlikely(gtod_read_retry(gtod, seq)));
 255
 256         ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
 257         ts->tv_nsec = ns;
 258
 259         return mode;
 260 }
 261
 262 notrace static void do_realtime_coarse(struct timespec *ts)
 263 {
 264         unsigned long seq;
 265         do {
 266                 seq = gtod_read_begin(gtod);
 267                 ts->tv_sec = gtod->wall_time_coarse_sec;
 268                 ts->tv_nsec = gtod->wall_time_coarse_nsec;
 269         } while (unlikely(gtod_read_retry(gtod, seq)));
 270 }
 271
 272 notrace static void do_monotonic_coarse(struct timespec *ts)
 273 {
 274         unsigned long seq;
 275         do {
 276                 seq = gtod_read_begin(gtod);
 277                 ts->tv_sec = gtod->monotonic_time_coarse_sec;
 278                 ts->tv_nsec = gtod->monotonic_time_coarse_nsec;
 279         } while (unlikely(gtod_read_retry(gtod, seq)));
 280 }
 281
 282 notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
 283 {
 284         switch (clock) {
 285         case CLOCK_REALTIME:
 286                 if (do_realtime(ts) == VCLOCK_NONE)
 287                         goto fallback;
 288                 break;
 289         case CLOCK_MONOTONIC:
 290                 if (do_monotonic(ts) == VCLOCK_NONE)
 291                         goto fallback;
 292                 break;
 293         case CLOCK_REALTIME_COARSE:
 294                 do_realtime_coarse(ts);
 295                 break;
 296         case CLOCK_MONOTONIC_COARSE:
 297                 do_monotonic_coarse(ts);
 298                 break;
 299         default:
 300                 goto fallback;
 301         }
 302
 303         return 0;
 304 fallback:
 305         return vdso_fallback_gettime(clock, ts);
 306 }
 307 int clock_gettime(clockid_t, struct timespec *)
 308         __attribute__((weak, alias("__vdso_clock_gettime")));
 309
 310 notrace int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz)
 311 {
 312         if (likely(tv != NULL)) {
 313                 if (unlikely(do_realtime((struct timespec *)tv) == VCLOCK_NONE))
 314                         return vdso_fallback_gtod(tv, tz);
 315                 tv->tv_usec /= 1000;
 316         }
 317         if (unlikely(tz != NULL)) {
 318                 tz->tz_minuteswest = gtod->tz_minuteswest;
 319                 tz->tz_dsttime = gtod->tz_dsttime;
 320         }
 321
 322         return 0;
 323 }
 324 int gettimeofday(struct timeval *, struct timezone *)
 325         __attribute__((weak, alias("__vdso_gettimeofday")));
 326
 327 /*
 328  * This will break when the xtime seconds get inaccurate, but that is
 329  * unlikely
 330  */
 331 notrace time_t __vdso_time(time_t *t)
 332 {
 333         /* This is atomic on x86 so we don't need any locks. */
 334         time_t result = ACCESS_ONCE(gtod->wall_time_sec);
 335
 336         if (t)
 337                 *t = result;
 338         return result;
 339 }
 340 int time(time_t *t)
 341         __attribute__((weak, alias("__vdso_time")));