Merge branches 'acpi-resources', 'acpi-battery', 'acpi-doc' and 'acpi-pnp'
[linux-2.6-block.git] / arch / x86 / vdso / vclock_gettime.c
CommitLineData
2aae950b
AK
1/*
2 * Copyright 2006 Andi Kleen, SUSE Labs.
3 * Subject to the GNU Public License, v.2
4 *
f144a6b4 5 * Fast user context implementation of clock_gettime, gettimeofday, and time.
2aae950b 6 *
7a59ed41
SS
7 * 32 Bit compat layer by Stefani Seibold <stefani@seibold.net>
8 * sponsored by Rohde & Schwarz GmbH & Co. KG Munich/Germany
9 *
2aae950b
AK
10 * The code should have no internal unresolved relocations.
11 * Check with readelf after changing.
2aae950b
AK
12 */
13
7a59ed41 14#include <uapi/linux/time.h>
2aae950b 15#include <asm/vgtod.h>
2aae950b 16#include <asm/hpet.h>
7c03156f 17#include <asm/vvar.h>
2aae950b 18#include <asm/unistd.h>
7c03156f
SS
19#include <asm/msr.h>
20#include <linux/math64.h>
21#include <linux/time.h>
2aae950b 22
8c49d9a7 23#define gtod (&VVAR(vsyscall_gtod_data))
2aae950b 24
7a59ed41
SS
25extern int __vdso_clock_gettime(clockid_t clock, struct timespec *ts);
26extern int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz);
27extern time_t __vdso_time(time_t *t);
28
7c03156f 29#ifdef CONFIG_HPET_TIMER
f40c3300
AL
30extern u8 hpet_page
31 __attribute__((visibility("hidden")));
32
33static notrace cycle_t vread_hpet(void)
7c03156f 34{
f40c3300 35 return *(const volatile u32 *)(&hpet_page + HPET_COUNTER);
7c03156f
SS
36}
37#endif
38
7a59ed41
SS
39#ifndef BUILD_VDSO32
40
7c03156f
SS
41#include <linux/kernel.h>
42#include <asm/vsyscall.h>
43#include <asm/fixmap.h>
44#include <asm/pvclock.h>
45
411f790c
SS
46notrace 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;
98d0ac38
AL
52}
53
411f790c 54notrace static long vdso_fallback_gtod(struct timeval *tv, struct timezone *tz)
98d0ac38 55{
411f790c
SS
56 long ret;
57
58 asm("syscall" : "=a" (ret) :
59 "0" (__NR_gettimeofday), "D" (tv), "S" (tz) : "memory");
60 return ret;
98d0ac38
AL
61}
62
51c19b4f
MT
63#ifdef CONFIG_PARAVIRT_CLOCK
64
65static 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
79static 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;
51c19b4f
MT
85 u8 flags;
86 unsigned cpu, cpu1;
87
88
89 /*
73459e2a
PB
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 *
51c19b4f
MT
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
73459e2a 105 pvti = get_pvti(cpu);
0a4e6be9 106
51c19b4f
MT
107 version = __pvclock_read_cycles(&pvti->pvti, &ret, &flags);
108
109 /*
110 * Test we're still on the cpu as well as the version.
73459e2a
PB
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.
51c19b4f 114 */
73459e2a
PB
115 cpu1 = __getcpu() & VGETCPU_CPU_MASK;
116 } while (unlikely(cpu != cpu1 ||
117 (pvti->pvti.version & 1) ||
118 pvti->pvti.version != version));
51c19b4f
MT
119
120 if (unlikely(!(flags & PVCLOCK_TSC_STABLE_BIT)))
121 *mode = VCLOCK_NONE;
122
123 /* refer to tsc.c read_tsc() comment for rationale */
7c03156f 124 last = gtod->cycle_last;
51c19b4f
MT
125
126 if (likely(ret >= last))
127 return ret;
128
129 return last;
130}
131#endif
132
7a59ed41
SS
133#else
134
7a59ed41
SS
135notrace 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"
6f121e54 142 "call __kernel_vsyscall \n"
7a59ed41
SS
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
150notrace 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"
6f121e54 157 "call __kernel_vsyscall \n"
7a59ed41
SS
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
167static notrace cycle_t vread_pvclock(int *mode)
168{
169 *mode = VCLOCK_NONE;
170 return 0;
171}
172#endif
173
174#endif
175
411f790c 176notrace static cycle_t vread_tsc(void)
2aae950b 177{
411f790c
SS
178 cycle_t ret;
179 u64 last;
2aae950b 180
411f790c
SS
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();
7a59ed41 189 ret = (cycle_t)__native_read_tsc();
a939e817 190
7c03156f 191 last = gtod->cycle_last;
a939e817 192
411f790c
SS
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}
a939e817 207
51c19b4f 208notrace static inline u64 vgetsns(int *mode)
2aae950b 209{
7a59ed41 210 u64 v;
98d0ac38 211 cycles_t cycles;
7c03156f
SS
212
213 if (gtod->vclock_mode == VCLOCK_TSC)
98d0ac38 214 cycles = vread_tsc();
7a59ed41 215#ifdef CONFIG_HPET_TIMER
7c03156f 216 else if (gtod->vclock_mode == VCLOCK_HPET)
98d0ac38 217 cycles = vread_hpet();
7a59ed41 218#endif
51c19b4f 219#ifdef CONFIG_PARAVIRT_CLOCK
7c03156f 220 else if (gtod->vclock_mode == VCLOCK_PVCLOCK)
51c19b4f
MT
221 cycles = vread_pvclock(mode);
222#endif
a939e817
JS
223 else
224 return 0;
7c03156f
SS
225 v = (cycles - gtod->cycle_last) & gtod->mask;
226 return v * gtod->mult;
2aae950b
AK
227}
228
5f293474
AL
229/* Code size doesn't matter (vdso is 4k anyway) and this is faster. */
230notrace static int __always_inline do_realtime(struct timespec *ts)
2aae950b 231{
650ea024
JS
232 unsigned long seq;
233 u64 ns;
a939e817
JS
234 int mode;
235
2aae950b 236 do {
7c03156f
SS
237 seq = gtod_read_begin(gtod);
238 mode = gtod->vclock_mode;
2aae950b 239 ts->tv_sec = gtod->wall_time_sec;
650ea024 240 ns = gtod->wall_time_snsec;
51c19b4f 241 ns += vgetsns(&mode);
7c03156f
SS
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;
a939e817 247
a939e817 248 return mode;
2aae950b
AK
249}
250
7a59ed41 251notrace static int __always_inline do_monotonic(struct timespec *ts)
2aae950b 252{
650ea024
JS
253 unsigned long seq;
254 u64 ns;
a939e817
JS
255 int mode;
256
2aae950b 257 do {
7c03156f
SS
258 seq = gtod_read_begin(gtod);
259 mode = gtod->vclock_mode;
91ec87d5 260 ts->tv_sec = gtod->monotonic_time_sec;
650ea024 261 ns = gtod->monotonic_time_snsec;
51c19b4f 262 ns += vgetsns(&mode);
7c03156f
SS
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;
0f51f285 268
a939e817 269 return mode;
2aae950b
AK
270}
271
ce39c640 272notrace static void do_realtime_coarse(struct timespec *ts)
da15cfda 273{
274 unsigned long seq;
275 do {
7c03156f
SS
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)));
da15cfda 280}
281
ce39c640 282notrace static void do_monotonic_coarse(struct timespec *ts)
da15cfda 283{
91ec87d5 284 unsigned long seq;
da15cfda 285 do {
7c03156f
SS
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)));
da15cfda 290}
291
23adec55 292notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
2aae950b 293{
0d7b8547
AL
294 switch (clock) {
295 case CLOCK_REALTIME:
ce39c640
SS
296 if (do_realtime(ts) == VCLOCK_NONE)
297 goto fallback;
0d7b8547
AL
298 break;
299 case CLOCK_MONOTONIC:
ce39c640
SS
300 if (do_monotonic(ts) == VCLOCK_NONE)
301 goto fallback;
0d7b8547
AL
302 break;
303 case CLOCK_REALTIME_COARSE:
ce39c640
SS
304 do_realtime_coarse(ts);
305 break;
0d7b8547 306 case CLOCK_MONOTONIC_COARSE:
ce39c640
SS
307 do_monotonic_coarse(ts);
308 break;
309 default:
310 goto fallback;
0d7b8547
AL
311 }
312
a939e817 313 return 0;
ce39c640
SS
314fallback:
315 return vdso_fallback_gettime(clock, ts);
2aae950b
AK
316}
317int clock_gettime(clockid_t, struct timespec *)
318 __attribute__((weak, alias("__vdso_clock_gettime")));
319
23adec55 320notrace int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz)
2aae950b 321{
a939e817 322 if (likely(tv != NULL)) {
0df1ea2b
SS
323 if (unlikely(do_realtime((struct timespec *)tv) == VCLOCK_NONE))
324 return vdso_fallback_gtod(tv, tz);
a939e817 325 tv->tv_usec /= 1000;
2aae950b 326 }
a939e817 327 if (unlikely(tz != NULL)) {
7c03156f
SS
328 tz->tz_minuteswest = gtod->tz_minuteswest;
329 tz->tz_dsttime = gtod->tz_dsttime;
a939e817
JS
330 }
331
a939e817 332 return 0;
2aae950b
AK
333}
334int gettimeofday(struct timeval *, struct timezone *)
335 __attribute__((weak, alias("__vdso_gettimeofday")));
f144a6b4 336
0d7b8547
AL
337/*
338 * This will break when the xtime seconds get inaccurate, but that is
339 * unlikely
340 */
f144a6b4
AL
341notrace time_t __vdso_time(time_t *t)
342{
7a59ed41 343 /* This is atomic on x86 so we don't need any locks. */
af8c93d8 344 time_t result = ACCESS_ONCE(gtod->wall_time_sec);
f144a6b4
AL
345
346 if (t)
347 *t = result;
348 return result;
349}
350int time(time_t *t)
351 __attribute__((weak, alias("__vdso_time")));