aboutsummaryrefslogtreecommitdiffstats
path: root/kernel
diff options
context:
space:
mode:
authorIngo Molnar <mingo@elte.hu>2008-08-11 18:07:02 -0400
committerIngo Molnar <mingo@elte.hu>2008-08-11 18:07:02 -0400
commite26b33e9552c29c1d3fe67dc602c6264c29f5dc7 (patch)
tree98f99cedc86db77665d1a3b582444b9049aacc58 /kernel
parent279ef6bbb8308488398c8f33b04c760148428378 (diff)
parentcf206bffbb7542df54043fad9898113172af99d8 (diff)
Merge branch 'sched/clock' into sched/urgent
Diffstat (limited to 'kernel')
-rw-r--r--kernel/sched_clock.c163
-rw-r--r--kernel/time/tick-sched.c2
2 files changed, 49 insertions, 116 deletions
diff --git a/kernel/sched_clock.c b/kernel/sched_clock.c
index 5a2dc7d8fd98..204991a0bfa7 100644
--- a/kernel/sched_clock.c
+++ b/kernel/sched_clock.c
@@ -42,12 +42,9 @@ unsigned long long __attribute__((weak)) sched_clock(void)
42 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ); 42 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
43} 43}
44 44
45#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK 45static __read_mostly int sched_clock_running;
46 46
47#define MULTI_SHIFT 15 47#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
48/* Max is double, Min is 1/2 */
49#define MAX_MULTI (2LL << MULTI_SHIFT)
50#define MIN_MULTI (1LL << (MULTI_SHIFT-1))
51 48
52struct sched_clock_data { 49struct sched_clock_data {
53 /* 50 /*
@@ -58,14 +55,9 @@ struct sched_clock_data {
58 raw_spinlock_t lock; 55 raw_spinlock_t lock;
59 56
60 unsigned long tick_jiffies; 57 unsigned long tick_jiffies;
61 u64 prev_raw;
62 u64 tick_raw; 58 u64 tick_raw;
63 u64 tick_gtod; 59 u64 tick_gtod;
64 u64 clock; 60 u64 clock;
65 s64 multi;
66#ifdef CONFIG_NO_HZ
67 int check_max;
68#endif
69}; 61};
70 62
71static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data); 63static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
@@ -80,8 +72,6 @@ static inline struct sched_clock_data *cpu_sdc(int cpu)
80 return &per_cpu(sched_clock_data, cpu); 72 return &per_cpu(sched_clock_data, cpu);
81} 73}
82 74
83static __read_mostly int sched_clock_running;
84
85void sched_clock_init(void) 75void sched_clock_init(void)
86{ 76{
87 u64 ktime_now = ktime_to_ns(ktime_get()); 77 u64 ktime_now = ktime_to_ns(ktime_get());
@@ -93,90 +83,39 @@ void sched_clock_init(void)
93 83
94 scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED; 84 scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
95 scd->tick_jiffies = now_jiffies; 85 scd->tick_jiffies = now_jiffies;
96 scd->prev_raw = 0;
97 scd->tick_raw = 0; 86 scd->tick_raw = 0;
98 scd->tick_gtod = ktime_now; 87 scd->tick_gtod = ktime_now;
99 scd->clock = ktime_now; 88 scd->clock = ktime_now;
100 scd->multi = 1 << MULTI_SHIFT;
101#ifdef CONFIG_NO_HZ
102 scd->check_max = 1;
103#endif
104 } 89 }
105 90
106 sched_clock_running = 1; 91 sched_clock_running = 1;
107} 92}
108 93
109#ifdef CONFIG_NO_HZ
110/*
111 * The dynamic ticks makes the delta jiffies inaccurate. This
112 * prevents us from checking the maximum time update.
113 * Disable the maximum check during stopped ticks.
114 */
115void sched_clock_tick_stop(int cpu)
116{
117 struct sched_clock_data *scd = cpu_sdc(cpu);
118
119 scd->check_max = 0;
120}
121
122void sched_clock_tick_start(int cpu)
123{
124 struct sched_clock_data *scd = cpu_sdc(cpu);
125
126 scd->check_max = 1;
127}
128
129static int check_max(struct sched_clock_data *scd)
130{
131 return scd->check_max;
132}
133#else
134static int check_max(struct sched_clock_data *scd)
135{
136 return 1;
137}
138#endif /* CONFIG_NO_HZ */
139
140/* 94/*
141 * update the percpu scd from the raw @now value 95 * update the percpu scd from the raw @now value
142 * 96 *
143 * - filter out backward motion 97 * - filter out backward motion
144 * - use jiffies to generate a min,max window to clip the raw values 98 * - use jiffies to generate a min,max window to clip the raw values
145 */ 99 */
146static void __update_sched_clock(struct sched_clock_data *scd, u64 now, u64 *time) 100static u64 __update_sched_clock(struct sched_clock_data *scd, u64 now)
147{ 101{
148 unsigned long now_jiffies = jiffies; 102 unsigned long now_jiffies = jiffies;
149 long delta_jiffies = now_jiffies - scd->tick_jiffies; 103 long delta_jiffies = now_jiffies - scd->tick_jiffies;
150 u64 clock = scd->clock; 104 u64 clock = scd->clock;
151 u64 min_clock, max_clock; 105 u64 min_clock, max_clock;
152 s64 delta = now - scd->prev_raw; 106 s64 delta = now - scd->tick_raw;
153 107
154 WARN_ON_ONCE(!irqs_disabled()); 108 WARN_ON_ONCE(!irqs_disabled());
155 109 min_clock = scd->tick_gtod + delta_jiffies * TICK_NSEC;
156 /*
157 * At schedule tick the clock can be just under the gtod. We don't
158 * want to push it too prematurely.
159 */
160 min_clock = scd->tick_gtod + (delta_jiffies * TICK_NSEC);
161 if (min_clock > TICK_NSEC)
162 min_clock -= TICK_NSEC / 2;
163 110
164 if (unlikely(delta < 0)) { 111 if (unlikely(delta < 0)) {
165 clock++; 112 clock++;
166 goto out; 113 goto out;
167 } 114 }
168 115
169 /* 116 max_clock = min_clock + TICK_NSEC;
170 * The clock must stay within a jiffie of the gtod.
171 * But since we may be at the start of a jiffy or the end of one
172 * we add another jiffy buffer.
173 */
174 max_clock = scd->tick_gtod + (2 + delta_jiffies) * TICK_NSEC;
175
176 delta *= scd->multi;
177 delta >>= MULTI_SHIFT;
178 117
179 if (unlikely(clock + delta > max_clock) && check_max(scd)) { 118 if (unlikely(clock + delta > max_clock)) {
180 if (clock < max_clock) 119 if (clock < max_clock)
181 clock = max_clock; 120 clock = max_clock;
182 else 121 else
@@ -189,12 +128,10 @@ static void __update_sched_clock(struct sched_clock_data *scd, u64 now, u64 *tim
189 if (unlikely(clock < min_clock)) 128 if (unlikely(clock < min_clock))
190 clock = min_clock; 129 clock = min_clock;
191 130
192 if (time) 131 scd->tick_jiffies = now_jiffies;
193 *time = clock; 132 scd->clock = clock;
194 else { 133
195 scd->prev_raw = now; 134 return clock;
196 scd->clock = clock;
197 }
198} 135}
199 136
200static void lock_double_clock(struct sched_clock_data *data1, 137static void lock_double_clock(struct sched_clock_data *data1,
@@ -212,7 +149,7 @@ static void lock_double_clock(struct sched_clock_data *data1,
212u64 sched_clock_cpu(int cpu) 149u64 sched_clock_cpu(int cpu)
213{ 150{
214 struct sched_clock_data *scd = cpu_sdc(cpu); 151 struct sched_clock_data *scd = cpu_sdc(cpu);
215 u64 now, clock; 152 u64 now, clock, this_clock, remote_clock;
216 153
217 if (unlikely(!sched_clock_running)) 154 if (unlikely(!sched_clock_running))
218 return 0ull; 155 return 0ull;
@@ -221,43 +158,44 @@ u64 sched_clock_cpu(int cpu)
221 now = sched_clock(); 158 now = sched_clock();
222 159
223 if (cpu != raw_smp_processor_id()) { 160 if (cpu != raw_smp_processor_id()) {
224 /*
225 * in order to update a remote cpu's clock based on our
226 * unstable raw time rebase it against:
227 * tick_raw (offset between raw counters)
228 * tick_gotd (tick offset between cpus)
229 */
230 struct sched_clock_data *my_scd = this_scd(); 161 struct sched_clock_data *my_scd = this_scd();
231 162
232 lock_double_clock(scd, my_scd); 163 lock_double_clock(scd, my_scd);
233 164
234 now -= my_scd->tick_raw; 165 this_clock = __update_sched_clock(my_scd, now);
235 now += scd->tick_raw; 166 remote_clock = scd->clock;
236 167
237 now += my_scd->tick_gtod; 168 /*
238 now -= scd->tick_gtod; 169 * Use the opportunity that we have both locks
170 * taken to couple the two clocks: we take the
171 * larger time as the latest time for both
172 * runqueues. (this creates monotonic movement)
173 */
174 if (likely(remote_clock < this_clock)) {
175 clock = this_clock;
176 scd->clock = clock;
177 } else {
178 /*
179 * Should be rare, but possible:
180 */
181 clock = remote_clock;
182 my_scd->clock = remote_clock;
183 }
239 184
240 __raw_spin_unlock(&my_scd->lock); 185 __raw_spin_unlock(&my_scd->lock);
241
242 __update_sched_clock(scd, now, &clock);
243
244 __raw_spin_unlock(&scd->lock);
245
246 } else { 186 } else {
247 __raw_spin_lock(&scd->lock); 187 __raw_spin_lock(&scd->lock);
248 __update_sched_clock(scd, now, NULL); 188 clock = __update_sched_clock(scd, now);
249 clock = scd->clock;
250 __raw_spin_unlock(&scd->lock);
251 } 189 }
252 190
191 __raw_spin_unlock(&scd->lock);
192
253 return clock; 193 return clock;
254} 194}
255 195
256void sched_clock_tick(void) 196void sched_clock_tick(void)
257{ 197{
258 struct sched_clock_data *scd = this_scd(); 198 struct sched_clock_data *scd = this_scd();
259 unsigned long now_jiffies = jiffies;
260 s64 mult, delta_gtod, delta_raw;
261 u64 now, now_gtod; 199 u64 now, now_gtod;
262 200
263 if (unlikely(!sched_clock_running)) 201 if (unlikely(!sched_clock_running))
@@ -269,29 +207,14 @@ void sched_clock_tick(void)
269 now = sched_clock(); 207 now = sched_clock();
270 208
271 __raw_spin_lock(&scd->lock); 209 __raw_spin_lock(&scd->lock);
272 __update_sched_clock(scd, now, NULL); 210 __update_sched_clock(scd, now);
273 /* 211 /*
274 * update tick_gtod after __update_sched_clock() because that will 212 * update tick_gtod after __update_sched_clock() because that will
275 * already observe 1 new jiffy; adding a new tick_gtod to that would 213 * already observe 1 new jiffy; adding a new tick_gtod to that would
276 * increase the clock 2 jiffies. 214 * increase the clock 2 jiffies.
277 */ 215 */
278 delta_gtod = now_gtod - scd->tick_gtod;
279 delta_raw = now - scd->tick_raw;
280
281 if ((long)delta_raw > 0) {
282 mult = delta_gtod << MULTI_SHIFT;
283 do_div(mult, delta_raw);
284 scd->multi = mult;
285 if (scd->multi > MAX_MULTI)
286 scd->multi = MAX_MULTI;
287 else if (scd->multi < MIN_MULTI)
288 scd->multi = MIN_MULTI;
289 } else
290 scd->multi = 1 << MULTI_SHIFT;
291
292 scd->tick_raw = now; 216 scd->tick_raw = now;
293 scd->tick_gtod = now_gtod; 217 scd->tick_gtod = now_gtod;
294 scd->tick_jiffies = now_jiffies;
295 __raw_spin_unlock(&scd->lock); 218 __raw_spin_unlock(&scd->lock);
296} 219}
297 220
@@ -310,7 +233,6 @@ EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
310void sched_clock_idle_wakeup_event(u64 delta_ns) 233void sched_clock_idle_wakeup_event(u64 delta_ns)
311{ 234{
312 struct sched_clock_data *scd = this_scd(); 235 struct sched_clock_data *scd = this_scd();
313 u64 now = sched_clock();
314 236
315 /* 237 /*
316 * Override the previous timestamp and ignore all 238 * Override the previous timestamp and ignore all
@@ -319,15 +241,28 @@ void sched_clock_idle_wakeup_event(u64 delta_ns)
319 * rq clock: 241 * rq clock:
320 */ 242 */
321 __raw_spin_lock(&scd->lock); 243 __raw_spin_lock(&scd->lock);
322 scd->prev_raw = now;
323 scd->clock += delta_ns; 244 scd->clock += delta_ns;
324 scd->multi = 1 << MULTI_SHIFT;
325 __raw_spin_unlock(&scd->lock); 245 __raw_spin_unlock(&scd->lock);
326 246
327 touch_softlockup_watchdog(); 247 touch_softlockup_watchdog();
328} 248}
329EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event); 249EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
330 250
251#else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
252
253void sched_clock_init(void)
254{
255 sched_clock_running = 1;
256}
257
258u64 sched_clock_cpu(int cpu)
259{
260 if (unlikely(!sched_clock_running))
261 return 0;
262
263 return sched_clock();
264}
265
331#endif 266#endif
332 267
333unsigned long long cpu_clock(int cpu) 268unsigned long long cpu_clock(int cpu)
diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
index 825b4c00fe44..f5da526424a9 100644
--- a/kernel/time/tick-sched.c
+++ b/kernel/time/tick-sched.c
@@ -289,7 +289,6 @@ void tick_nohz_stop_sched_tick(int inidle)
289 ts->tick_stopped = 1; 289 ts->tick_stopped = 1;
290 ts->idle_jiffies = last_jiffies; 290 ts->idle_jiffies = last_jiffies;
291 rcu_enter_nohz(); 291 rcu_enter_nohz();
292 sched_clock_tick_stop(cpu);
293 } 292 }
294 293
295 /* 294 /*
@@ -392,7 +391,6 @@ void tick_nohz_restart_sched_tick(void)
392 select_nohz_load_balancer(0); 391 select_nohz_load_balancer(0);
393 now = ktime_get(); 392 now = ktime_get();
394 tick_do_update_jiffies64(now); 393 tick_do_update_jiffies64(now);
395 sched_clock_tick_start(cpu);
396 cpu_clear(cpu, nohz_cpu_mask); 394 cpu_clear(cpu, nohz_cpu_mask);
397 395
398 /* 396 /*