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
  33         __attribute__((visibility("hidden")));
  34 #endif
  35
  36 #ifdef CONFIG_HYPERV_TSCPAGE
  37 extern u8 hvclock_page
  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) :
  47             "0" (__NR_clock_gettime), "D" (clock), "S" (ts) : "memory");
  48         return ret;
  49 }
  50
  51 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
  52 {
  53         long ret;
  54
  55         asm("syscall" : "=a" (ret) :
  56             "0" (__NR_gettimeofday), "D" (tv), "S" (tz) : "memory");
  57         return ret;
  58 }
  59
  60
  61 #else
  62
  63 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
  64 {
  65         long ret;
  66
  67         asm(
  68                 "mov %%ebx, %%edx \n"
  69                 "mov %2, %%ebx \n"
  70                 "call __kernel_vsyscall \n"
  71                 "mov %%edx, %%ebx \n"
  72                 : "=a" (ret)
  73                 : "0" (__NR_clock_gettime), "g" (clock), "c" (ts)
  74                 : "memory", "edx");
  75         return ret;
  76 }
  77
  78 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
  79 {
  80         long ret;
  81
  82         asm(
  83                 "mov %%ebx, %%edx \n"
  84                 "mov %2, %%ebx \n"
  85                 "call __kernel_vsyscall \n"
  86                 "mov %%edx, %%ebx \n"
  87                 : "=a" (ret)
  88                 : "0" (__NR_gettimeofday), "g" (tv), "c" (tz)
  89                 : "memory", "edx");
  90         return ret;
  91 }
  92
  93 #endif
  94
  95 #ifdef CONFIG_PARAVIRT_CLOCK
  96 static notrace const struct pvclock_vsyscall_time_info *get_pvti0(void)
  97 {
  98         return (const struct pvclock_vsyscall_time_info *)&pvclock_page;
  99 }
 100
 101 static notrace u64 vread_pvclock(int *mode)
 102 {
 103         const struct pvclock_vcpu_time_info *pvti = &get_pvti0()->pvti;
 104         u64 ret;
 105         u64 last;
 106         u32 version;
 107
 108         /*
 109          * Note: The kernel and hypervisor must guarantee that cpu ID
 110          * number maps 1:1 to per-CPU pvclock time info.
 111          *
 112          * Because the hypervisor is entirely unaware of guest userspace
 113          * preemption, it cannot guarantee that per-CPU pvclock time
 114          * info is updated if the underlying CPU changes or that that
 115          * version is increased whenever underlying CPU changes.
 116          *
 117          * On KVM, we are guaranteed that pvti updates for any vCPU are
 118          * atomic as seen by *all* vCPUs.  This is an even stronger
 119          * guarantee than we get with a normal seqlock.
 120          *
 121          * On Xen, we don't appear to have that guarantee, but Xen still
 122          * supplies a valid seqlock using the version field.
 123          *
 124          * We only do pvclock vdso timing at all if
 125          * PVCLOCK_TSC_STABLE_BIT is set, and we interpret that bit to
 126          * mean that all vCPUs have matching pvti and that the TSC is
 127          * synced, so we can just look at vCPU 0's pvti.
 128          */
 129
 130         do {
 131                 version = pvclock_read_begin(pvti);
 132
 133                 if (unlikely(!(pvti->flags & PVCLOCK_TSC_STABLE_BIT))) {
 134                         *mode = VCLOCK_NONE;
 135                         return 0;
 136                 }
 137
 138                 ret = __pvclock_read_cycles(pvti, rdtsc_ordered());
 139         } while (pvclock_read_retry(pvti, version));
 140
 141         /* refer to vread_tsc() comment for rationale */
 142         last = gtod->cycle_last;
 143
 144         if (likely(ret >= last))
 145                 return ret;
 146
 147         return last;
 148 }
 149 #endif
 150 #ifdef CONFIG_HYPERV_TSCPAGE
 151 static notrace u64 vread_hvclock(int *mode)
 152 {
 153         const struct ms_hyperv_tsc_page *tsc_pg =
 154                 (const struct ms_hyperv_tsc_page *)&hvclock_page;
 155         u64 current_tick = hv_read_tsc_page(tsc_pg);
 156
 157         if (current_tick != U64_MAX)
 158                 return current_tick;
 159
 160         *mode = VCLOCK_NONE;
 161         return 0;
 162 }
 163 #endif
 164
 165 notrace static u64 vread_tsc(void)
 166 {
 167         u64 ret = (u64)rdtsc_ordered();
 168         u64 last = gtod->cycle_last;
 169
 170         if (likely(ret >= last))
 171                 return ret;
 172
 173         /*
 174          * GCC likes to generate cmov here, but this branch is extremely
 175          * predictable (it's just a function of time and the likely is
 176          * very likely) and there's a data dependence, so force GCC
 177          * to generate a branch instead.  I don't barrier() because
 178          * we don't actually need a barrier, and if this function
 179          * ever gets inlined it will generate worse code.
 180          */
 181         asm volatile ("");
 182         return last;
 183 }
 184
 185 notrace static inline u64 vgetsns(int *mode)
 186 {
 187         u64 v;
 188         cycles_t cycles;
 189
 190         if (gtod->vclock_mode == VCLOCK_TSC)
 191                 cycles = vread_tsc();
 192 #ifdef CONFIG_PARAVIRT_CLOCK
 193         else if (gtod->vclock_mode == VCLOCK_PVCLOCK)
 194                 cycles = vread_pvclock(mode);
 195 #endif
 196 #ifdef CONFIG_HYPERV_TSCPAGE
 197         else if (gtod->vclock_mode == VCLOCK_HVCLOCK)
 198                 cycles = vread_hvclock(mode);
 199 #endif
 200         else
 201                 return 0;
 202         v = (cycles - gtod->cycle_last) & gtod->mask;
 203         return v * gtod->mult;
 204 }
 205
 206 /* Code size doesn't matter (vdso is 4k anyway) and this is faster. */
 207 notrace static int __always_inline do_realtime(struct timespec *ts)
 208 {
 209         unsigned long seq;
 210         u64 ns;
 211         int mode;
 212
 213         do {
 214                 seq = gtod_read_begin(gtod);
 215                 mode = gtod->vclock_mode;
 216                 ts->tv_sec = gtod->wall_time_sec;
 217                 ns = gtod->wall_time_snsec;
 218                 ns += vgetsns(&mode);
 219                 ns >>= gtod->shift;
 220         } while (unlikely(gtod_read_retry(gtod, seq)));
 221
 222         ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
 223         ts->tv_nsec = ns;
 224
 225         return mode;
 226 }
 227
 228 notrace static int __always_inline do_monotonic(struct timespec *ts)
 229 {
 230         unsigned long seq;
 231         u64 ns;
 232         int mode;
 233
 234         do {
 235                 seq = gtod_read_begin(gtod);
 236                 mode = gtod->vclock_mode;
 237                 ts->tv_sec = gtod->monotonic_time_sec;
 238                 ns = gtod->monotonic_time_snsec;
 239                 ns += vgetsns(&mode);
 240                 ns >>= gtod->shift;
 241         } while (unlikely(gtod_read_retry(gtod, seq)));
 242
 243         ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
 244         ts->tv_nsec = ns;
 245
 246         return mode;
 247 }
 248
 249 notrace static void do_realtime_coarse(struct timespec *ts)
 250 {
 251         unsigned long seq;
 252         do {
 253                 seq = gtod_read_begin(gtod);
 254                 ts->tv_sec = gtod->wall_time_coarse_sec;
 255                 ts->tv_nsec = gtod->wall_time_coarse_nsec;
 256         } while (unlikely(gtod_read_retry(gtod, seq)));
 257 }
 258
 259 notrace static void do_monotonic_coarse(struct timespec *ts)
 260 {
 261         unsigned long seq;
 262         do {
 263                 seq = gtod_read_begin(gtod);
 264                 ts->tv_sec = gtod->monotonic_time_coarse_sec;
 265                 ts->tv_nsec = gtod->monotonic_time_coarse_nsec;
 266         } while (unlikely(gtod_read_retry(gtod, seq)));
 267 }
 268
 269 notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
 270 {
 271         switch (clock) {
 272         case CLOCK_REALTIME:
 273                 if (do_realtime(ts) == VCLOCK_NONE)
 274                         goto fallback;
 275                 break;
 276         case CLOCK_MONOTONIC:
 277                 if (do_monotonic(ts) == VCLOCK_NONE)
 278                         goto fallback;
 279                 break;
 280         case CLOCK_REALTIME_COARSE:
 281                 do_realtime_coarse(ts);
 282                 break;
 283         case CLOCK_MONOTONIC_COARSE:
 284                 do_monotonic_coarse(ts);
 285                 break;
 286         default:
 287                 goto fallback;
 288         }
 289
 290         return 0;
 291 fallback:
 292         return vdso_fallback_gettime(clock, ts);
 293 }
 294 int clock_gettime(clockid_t, struct timespec *)
 295         __attribute__((weak, alias("__vdso_clock_gettime")));
 296
 297 notrace int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz)
 298 {
 299         if (likely(tv != NULL)) {
 300                 if (unlikely(do_realtime((struct timespec *)tv) == VCLOCK_NONE))
 301                         return vdso_fallback_gtod(tv, tz);
 302                 tv->tv_usec /= 1000;
 303         }
 304         if (unlikely(tz != NULL)) {
 305                 tz->tz_minuteswest = gtod->tz_minuteswest;
 306                 tz->tz_dsttime = gtod->tz_dsttime;
 307         }
 308
 309         return 0;
 310 }
 311 int gettimeofday(struct timeval *, struct timezone *)
 312         __attribute__((weak, alias("__vdso_gettimeofday")));
 313
 314 /*
 315  * This will break when the xtime seconds get inaccurate, but that is
 316  * unlikely
 317  */
 318 notrace time_t __vdso_time(time_t *t)
 319 {
 320         /* This is atomic on x86 so we don't need any locks. */
 321         time_t result = READ_ONCE(gtod->wall_time_sec);
 322
 323         if (t)
 324                 *t = result;
 325         return result;
 326 }
 327 int time(time_t *t)
 328         __attribute__((weak, alias("__vdso_time")));