2 * Copyright 2006 Andi Kleen, SUSE Labs.
3 * Subject to the GNU Public License, v.2
5 * Fast user context implementation of clock_gettime, gettimeofday, and time.
7 * 32 Bit compat layer by Stefani Seibold <stefani@seibold.net>
8 * sponsored by Rohde & Schwarz GmbH & Co. KG Munich/Germany
10 * The code should have no internal unresolved relocations.
11 * Check with readelf after changing.
14 #include <uapi/linux/time.h>
15 #include <asm/vgtod.h>
17 #include <asm/unistd.h>
19 #include <asm/pvclock.h>
20 #include <linux/math64.h>
21 #include <linux/time.h>
22 #include <linux/kernel.h>
24 #define gtod (&VVAR(vsyscall_gtod_data))
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);
30 #ifdef CONFIG_PARAVIRT_CLOCK
31 extern u8 pvclock_page
32 __attribute__((visibility("hidden")));
37 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
40 asm ("syscall" : "=a" (ret), "=m" (*ts) :
41 "0" (__NR_clock_gettime), "D" (clock), "S" (ts) :
42 "memory", "rcx", "r11");
46 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
50 asm ("syscall" : "=a" (ret), "=m" (*tv), "=m" (*tz) :
51 "0" (__NR_gettimeofday), "D" (tv), "S" (tz) :
52 "memory", "rcx", "r11");
59 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
65 "mov %[clock], %%ebx \n"
66 "call __kernel_vsyscall \n"
68 : "=a" (ret), "=m" (*ts)
69 : "0" (__NR_clock_gettime), [clock] "g" (clock), "c" (ts)
74 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
81 "call __kernel_vsyscall \n"
83 : "=a" (ret), "=m" (*tv), "=m" (*tz)
84 : "0" (__NR_gettimeofday), [tv] "g" (tv), "c" (tz)
91 #ifdef CONFIG_PARAVIRT_CLOCK
92 static notrace const struct pvclock_vsyscall_time_info *get_pvti0(void)
94 return (const struct pvclock_vsyscall_time_info *)&pvclock_page;
97 static notrace cycle_t vread_pvclock(int *mode)
99 const struct pvclock_vcpu_time_info *pvti = &get_pvti0()->pvti;
105 * Note: The kernel and hypervisor must guarantee that cpu ID
106 * number maps 1:1 to per-CPU pvclock time info.
108 * Because the hypervisor is entirely unaware of guest userspace
109 * preemption, it cannot guarantee that per-CPU pvclock time
110 * info is updated if the underlying CPU changes or that that
111 * version is increased whenever underlying CPU changes.
113 * On KVM, we are guaranteed that pvti updates for any vCPU are
114 * atomic as seen by *all* vCPUs. This is an even stronger
115 * guarantee than we get with a normal seqlock.
117 * On Xen, we don't appear to have that guarantee, but Xen still
118 * supplies a valid seqlock using the version field.
120 * We only do pvclock vdso timing at all if
121 * PVCLOCK_TSC_STABLE_BIT is set, and we interpret that bit to
122 * mean that all vCPUs have matching pvti and that the TSC is
123 * synced, so we can just look at vCPU 0's pvti.
127 version = pvclock_read_begin(pvti);
129 if (unlikely(!(pvti->flags & PVCLOCK_TSC_STABLE_BIT))) {
134 ret = __pvclock_read_cycles(pvti, rdtsc_ordered());
135 } while (pvclock_read_retry(pvti, version));
137 /* refer to vread_tsc() comment for rationale */
138 last = gtod->cycle_last;
140 if (likely(ret >= last))
147 notrace static cycle_t vread_tsc(void)
149 cycle_t ret = (cycle_t)rdtsc_ordered();
150 u64 last = gtod->cycle_last;
152 if (likely(ret >= last))
156 * GCC likes to generate cmov here, but this branch is extremely
157 * predictable (it's just a function of time and the likely is
158 * very likely) and there's a data dependence, so force GCC
159 * to generate a branch instead. I don't barrier() because
160 * we don't actually need a barrier, and if this function
161 * ever gets inlined it will generate worse code.
167 notrace static inline u64 vgetsns(int *mode)
172 if (gtod->vclock_mode == VCLOCK_TSC)
173 cycles = vread_tsc();
174 #ifdef CONFIG_PARAVIRT_CLOCK
175 else if (gtod->vclock_mode == VCLOCK_PVCLOCK)
176 cycles = vread_pvclock(mode);
180 v = (cycles - gtod->cycle_last) & gtod->mask;
181 return v * gtod->mult;
184 /* Code size doesn't matter (vdso is 4k anyway) and this is faster. */
185 notrace static int __always_inline do_realtime(struct timespec *ts)
192 seq = gtod_read_begin(gtod);
193 mode = gtod->vclock_mode;
194 ts->tv_sec = gtod->wall_time_sec;
195 ns = gtod->wall_time_snsec;
196 ns += vgetsns(&mode);
198 } while (unlikely(gtod_read_retry(gtod, seq)));
200 ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
206 notrace static int __always_inline do_monotonic(struct timespec *ts)
213 seq = gtod_read_begin(gtod);
214 mode = gtod->vclock_mode;
215 ts->tv_sec = gtod->monotonic_time_sec;
216 ns = gtod->monotonic_time_snsec;
217 ns += vgetsns(&mode);
219 } while (unlikely(gtod_read_retry(gtod, seq)));
221 ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
227 notrace static void do_realtime_coarse(struct timespec *ts)
231 seq = gtod_read_begin(gtod);
232 ts->tv_sec = gtod->wall_time_coarse_sec;
233 ts->tv_nsec = gtod->wall_time_coarse_nsec;
234 } while (unlikely(gtod_read_retry(gtod, seq)));
237 notrace static void do_monotonic_coarse(struct timespec *ts)
241 seq = gtod_read_begin(gtod);
242 ts->tv_sec = gtod->monotonic_time_coarse_sec;
243 ts->tv_nsec = gtod->monotonic_time_coarse_nsec;
244 } while (unlikely(gtod_read_retry(gtod, seq)));
247 notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
251 if (do_realtime(ts) == VCLOCK_NONE)
254 case CLOCK_MONOTONIC:
255 if (do_monotonic(ts) == VCLOCK_NONE)
258 case CLOCK_REALTIME_COARSE:
259 do_realtime_coarse(ts);
261 case CLOCK_MONOTONIC_COARSE:
262 do_monotonic_coarse(ts);
270 return vdso_fallback_gettime(clock, ts);
272 int clock_gettime(clockid_t, struct timespec *)
273 __attribute__((weak, alias("__vdso_clock_gettime")));
275 notrace int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz)
277 if (likely(tv != NULL)) {
278 if (unlikely(do_realtime((struct timespec *)tv) == VCLOCK_NONE))
279 return vdso_fallback_gtod(tv, tz);
282 if (unlikely(tz != NULL)) {
283 tz->tz_minuteswest = gtod->tz_minuteswest;
284 tz->tz_dsttime = gtod->tz_dsttime;
289 int gettimeofday(struct timeval *, struct timezone *)
290 __attribute__((weak, alias("__vdso_gettimeofday")));
293 * This will break when the xtime seconds get inaccurate, but that is
296 notrace time_t __vdso_time(time_t *t)
298 /* This is atomic on x86 so we don't need any locks. */
299 time_t result = ACCESS_ONCE(gtod->wall_time_sec);
306 __attribute__((weak, alias("__vdso_time")));