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 * The code should have no internal unresolved relocations.
8 * Check with readelf after changing.
11 /* Disable profiling for userspace code: */
12 #define DISABLE_BRANCH_PROFILING
14 #include <linux/kernel.h>
15 #include <linux/posix-timers.h>
16 #include <linux/time.h>
17 #include <linux/string.h>
18 #include <asm/vsyscall.h>
19 #include <asm/fixmap.h>
20 #include <asm/vgtod.h>
21 #include <asm/timex.h>
23 #include <asm/unistd.h>
25 #include <asm/pvclock.h>
27 #define gtod (&VVAR(vsyscall_gtod_data))
29 notrace static cycle_t vread_tsc(void)
35 * Empirically, a fence (of type that depends on the CPU)
36 * before rdtsc is enough to ensure that rdtsc is ordered
37 * with respect to loads. The various CPU manuals are unclear
38 * as to whether rdtsc can be reordered with later loads,
39 * but no one has ever seen it happen.
42 ret = (cycle_t)vget_cycles();
44 last = VVAR(vsyscall_gtod_data).clock.cycle_last;
46 if (likely(ret >= last))
50 * GCC likes to generate cmov here, but this branch is extremely
51 * predictable (it's just a funciton of time and the likely is
52 * very likely) and there's a data dependence, so force GCC
53 * to generate a branch instead. I don't barrier() because
54 * we don't actually need a barrier, and if this function
55 * ever gets inlined it will generate worse code.
61 static notrace cycle_t vread_hpet(void)
63 return readl((const void __iomem *)fix_to_virt(VSYSCALL_HPET) + HPET_COUNTER);
66 #ifdef CONFIG_PARAVIRT_CLOCK
68 static notrace const struct pvclock_vsyscall_time_info *get_pvti(int cpu)
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);
74 BUG_ON(PVCLOCK_FIXMAP_BEGIN + idx > PVCLOCK_FIXMAP_END);
76 pvti_base = (struct pvclock_vsyscall_time_info *)
77 __fix_to_virt(PVCLOCK_FIXMAP_BEGIN+idx);
79 return &pvti_base[offset];
82 static notrace cycle_t vread_pvclock(int *mode)
84 const struct pvclock_vsyscall_time_info *pvti;
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
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).
108 pvti = get_pvti(cpu);
110 version = __pvclock_read_cycles(&pvti->pvti, &ret, &flags);
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.
118 cpu1 = __getcpu() & VGETCPU_CPU_MASK;
119 } while (unlikely(cpu != cpu1 ||
120 (pvti->pvti.version & 1) ||
121 pvti->pvti.version != version));
123 if (unlikely(!(flags & PVCLOCK_TSC_STABLE_BIT)))
126 /* refer to tsc.c read_tsc() comment for rationale */
127 last = VVAR(vsyscall_gtod_data).clock.cycle_last;
129 if (likely(ret >= last))
136 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
139 asm("syscall" : "=a" (ret) :
140 "0" (__NR_clock_gettime),"D" (clock), "S" (ts) : "memory");
144 notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
148 asm("syscall" : "=a" (ret) :
149 "0" (__NR_gettimeofday), "D" (tv), "S" (tz) : "memory");
154 notrace static inline u64 vgetsns(int *mode)
158 if (gtod->clock.vclock_mode == VCLOCK_TSC)
159 cycles = vread_tsc();
160 else if (gtod->clock.vclock_mode == VCLOCK_HPET)
161 cycles = vread_hpet();
162 #ifdef CONFIG_PARAVIRT_CLOCK
163 else if (gtod->clock.vclock_mode == VCLOCK_PVCLOCK)
164 cycles = vread_pvclock(mode);
168 v = (cycles - gtod->clock.cycle_last) & gtod->clock.mask;
169 return v * gtod->clock.mult;
172 /* Code size doesn't matter (vdso is 4k anyway) and this is faster. */
173 notrace static int __always_inline do_realtime(struct timespec *ts)
181 seq = read_seqcount_begin(>od->seq);
182 mode = gtod->clock.vclock_mode;
183 ts->tv_sec = gtod->wall_time_sec;
184 ns = gtod->wall_time_snsec;
185 ns += vgetsns(&mode);
186 ns >>= gtod->clock.shift;
187 } while (unlikely(read_seqcount_retry(>od->seq, seq)));
189 timespec_add_ns(ts, ns);
193 notrace static int do_monotonic(struct timespec *ts)
201 seq = read_seqcount_begin(>od->seq);
202 mode = gtod->clock.vclock_mode;
203 ts->tv_sec = gtod->monotonic_time_sec;
204 ns = gtod->monotonic_time_snsec;
205 ns += vgetsns(&mode);
206 ns >>= gtod->clock.shift;
207 } while (unlikely(read_seqcount_retry(>od->seq, seq)));
208 timespec_add_ns(ts, ns);
213 notrace static int do_realtime_coarse(struct timespec *ts)
217 seq = read_seqcount_begin(>od->seq);
218 ts->tv_sec = gtod->wall_time_coarse.tv_sec;
219 ts->tv_nsec = gtod->wall_time_coarse.tv_nsec;
220 } while (unlikely(read_seqcount_retry(>od->seq, seq)));
224 notrace static int do_monotonic_coarse(struct timespec *ts)
228 seq = read_seqcount_begin(>od->seq);
229 ts->tv_sec = gtod->monotonic_time_coarse.tv_sec;
230 ts->tv_nsec = gtod->monotonic_time_coarse.tv_nsec;
231 } while (unlikely(read_seqcount_retry(>od->seq, seq)));
236 notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
238 int ret = VCLOCK_NONE;
242 ret = do_realtime(ts);
244 case CLOCK_MONOTONIC:
245 ret = do_monotonic(ts);
247 case CLOCK_REALTIME_COARSE:
248 return do_realtime_coarse(ts);
249 case CLOCK_MONOTONIC_COARSE:
250 return do_monotonic_coarse(ts);
253 if (ret == VCLOCK_NONE)
254 return vdso_fallback_gettime(clock, ts);
257 int clock_gettime(clockid_t, struct timespec *)
258 __attribute__((weak, alias("__vdso_clock_gettime")));
260 notrace int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz)
262 long ret = VCLOCK_NONE;
264 if (likely(tv != NULL)) {
265 BUILD_BUG_ON(offsetof(struct timeval, tv_usec) !=
266 offsetof(struct timespec, tv_nsec) ||
267 sizeof(*tv) != sizeof(struct timespec));
268 ret = do_realtime((struct timespec *)tv);
271 if (unlikely(tz != NULL)) {
272 /* Avoid memcpy. Some old compilers fail to inline it */
273 tz->tz_minuteswest = gtod->sys_tz.tz_minuteswest;
274 tz->tz_dsttime = gtod->sys_tz.tz_dsttime;
277 if (ret == VCLOCK_NONE)
278 return vdso_fallback_gtod(tv, tz);
281 int gettimeofday(struct timeval *, struct timezone *)
282 __attribute__((weak, alias("__vdso_gettimeofday")));
285 * This will break when the xtime seconds get inaccurate, but that is
288 notrace time_t __vdso_time(time_t *t)
290 /* This is atomic on x86_64 so we don't need any locks. */
291 time_t result = ACCESS_ONCE(VVAR(vsyscall_gtod_data).wall_time_sec);
298 __attribute__((weak, alias("__vdso_time")));