From c2bc11113c50449f23c40b724fe410fc2380a8e9 Mon Sep 17 00:00:00 2001 From: John Stultz Date: Thu, 27 Oct 2011 18:12:42 -0700 Subject: [PATCH 1/4] time: Improve documentation of timekeeeping_adjust() After getting a number of questions in private emails about the math around admittedly very complex timekeeping_adjust() and timekeeping_big_adjust(), I figure the code needs some better comments. Hopefully the explanations are clear enough and don't muddy the water any worse. Still needs documentation for ntp_error, but I couldn't recall exactly the full explanation behind the code that's there (although I do recall once working it out when Roman first proposed it). Given a bit more time I can probably work it out, but I don't want to hold back this documentation until then. Signed-off-by: John Stultz Cc: Chen Jie Cc: Steven Rostedt Link: http://lkml.kernel.org/r/1319764362-32367-1-git-send-email-john.stultz@linaro.org Signed-off-by: Ingo Molnar --- kernel/time/timekeeping.c | 81 ++++++++++++++++++++++++++++++++++++++- 1 file changed, 80 insertions(+), 1 deletion(-) diff --git a/kernel/time/timekeeping.c b/kernel/time/timekeeping.c index 2b021b0e8507..025e136f3881 100644 --- a/kernel/time/timekeeping.c +++ b/kernel/time/timekeeping.c @@ -802,14 +802,44 @@ static void timekeeping_adjust(s64 offset) s64 error, interval = timekeeper.cycle_interval; int adj; + /* + * The point of this is to check if the error is greater then half + * an interval. + * + * First we shift it down from NTP_SHIFT to clocksource->shifted nsecs. + * + * Note we subtract one in the shift, so that error is really error*2. + * This "saves" dividing(shifting) intererval twice, but keeps the + * (error > interval) comparision as still measuring if error is + * larger then half an interval. + * + * Note: It does not "save" on aggrivation when reading the code. + */ error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1); if (error > interval) { + /* + * We now divide error by 4(via shift), which checks if + * the error is greater then twice the interval. + * If it is greater, we need a bigadjust, if its smaller, + * we can adjust by 1. + */ error >>= 2; + /* + * XXX - In update_wall_time, we round up to the next + * nanosecond, and store the amount rounded up into + * the error. This causes the likely below to be unlikely. + * + * The properfix is to avoid rounding up by using + * the high precision timekeeper.xtime_nsec instead of + * xtime.tv_nsec everywhere. Fixing this will take some + * time. + */ if (likely(error <= interval)) adj = 1; else adj = timekeeping_bigadjust(error, &interval, &offset); } else if (error < -interval) { + /* See comment above, this is just switched for the negative */ error >>= 2; if (likely(error >= -interval)) { adj = -1; @@ -817,9 +847,58 @@ static void timekeeping_adjust(s64 offset) offset = -offset; } else adj = timekeeping_bigadjust(error, &interval, &offset); - } else + } else /* No adjustment needed */ return; + /* + * So the following can be confusing. + * + * To keep things simple, lets assume adj == 1 for now. + * + * When adj != 1, remember that the interval and offset values + * have been appropriately scaled so the math is the same. + * + * The basic idea here is that we're increasing the multiplier + * by one, this causes the xtime_interval to be incremented by + * one cycle_interval. This is because: + * xtime_interval = cycle_interval * mult + * So if mult is being incremented by one: + * xtime_interval = cycle_interval * (mult + 1) + * Its the same as: + * xtime_interval = (cycle_interval * mult) + cycle_interval + * Which can be shortened to: + * xtime_interval += cycle_interval + * + * So offset stores the non-accumulated cycles. Thus the current + * time (in shifted nanoseconds) is: + * now = (offset * adj) + xtime_nsec + * Now, even though we're adjusting the clock frequency, we have + * to keep time consistent. In other words, we can't jump back + * in time, and we also want to avoid jumping forward in time. + * + * So given the same offset value, we need the time to be the same + * both before and after the freq adjustment. + * now = (offset * adj_1) + xtime_nsec_1 + * now = (offset * adj_2) + xtime_nsec_2 + * So: + * (offset * adj_1) + xtime_nsec_1 = + * (offset * adj_2) + xtime_nsec_2 + * And we know: + * adj_2 = adj_1 + 1 + * So: + * (offset * adj_1) + xtime_nsec_1 = + * (offset * (adj_1+1)) + xtime_nsec_2 + * (offset * adj_1) + xtime_nsec_1 = + * (offset * adj_1) + offset + xtime_nsec_2 + * Canceling the sides: + * xtime_nsec_1 = offset + xtime_nsec_2 + * Which gives us: + * xtime_nsec_2 = xtime_nsec_1 - offset + * Which simplfies to: + * xtime_nsec -= offset + * + * XXX - TODO: Doc ntp_error calculation. + */ timekeeper.mult += adj; timekeeper.xtime_interval += interval; timekeeper.xtime_nsec -= offset; From d65670a78cdbfae94f20a9e05ec705871d7cdf2b Mon Sep 17 00:00:00 2001 From: John Stultz Date: Mon, 31 Oct 2011 17:06:35 -0400 Subject: [PATCH 2/4] clocksource: Avoid selecting mult values that might overflow when adjusted For some frequencies, the clocks_calc_mult_shift() function will unfortunately select mult values very close to 0xffffffff. This has the potential to overflow when NTP adjusts the clock, adding to the mult value. This patch adds a clocksource.maxadj value, which provides an approximation of an 11% adjustment(NTP limits adjustments to 500ppm and the tick adjustment is limited to 10%), which could be made to the clocksource.mult value. This is then used to both check that the current mult value won't overflow/underflow, as well as warning us if the timekeeping_adjust() code pushes over that 11% boundary. v2: Fix max_adjustment calculation, and improve WARN_ONCE messages. v3: Don't warn before maxadj has actually been set CC: Yong Zhang CC: David Daney CC: Thomas Gleixner CC: Chen Jie CC: zhangfx CC: stable@kernel.org Reported-by: Chen Jie Reported-by: zhangfx Tested-by: Yong Zhang Signed-off-by: John Stultz --- include/linux/clocksource.h | 3 +- kernel/time/clocksource.c | 58 ++++++++++++++++++++++++++++++------- kernel/time/timekeeping.c | 7 +++++ 3 files changed, 57 insertions(+), 11 deletions(-) diff --git a/include/linux/clocksource.h b/include/linux/clocksource.h index 139c4db55f17..c86c940d1de3 100644 --- a/include/linux/clocksource.h +++ b/include/linux/clocksource.h @@ -156,6 +156,7 @@ extern u64 timecounter_cyc2time(struct timecounter *tc, * @mult: cycle to nanosecond multiplier * @shift: cycle to nanosecond divisor (power of two) * @max_idle_ns: max idle time permitted by the clocksource (nsecs) + * @maxadj maximum adjustment value to mult (~11%) * @flags: flags describing special properties * @archdata: arch-specific data * @suspend: suspend function for the clocksource, if necessary @@ -172,7 +173,7 @@ struct clocksource { u32 mult; u32 shift; u64 max_idle_ns; - + u32 maxadj; #ifdef CONFIG_ARCH_CLOCKSOURCE_DATA struct arch_clocksource_data archdata; #endif diff --git a/kernel/time/clocksource.c b/kernel/time/clocksource.c index cf52fda2e096..cfc65e1eb9fb 100644 --- a/kernel/time/clocksource.c +++ b/kernel/time/clocksource.c @@ -491,6 +491,22 @@ void clocksource_touch_watchdog(void) clocksource_resume_watchdog(); } +/** + * clocksource_max_adjustment- Returns max adjustment amount + * @cs: Pointer to clocksource + * + */ +static u32 clocksource_max_adjustment(struct clocksource *cs) +{ + u64 ret; + /* + * We won't try to correct for more then 11% adjustments (110,000 ppm), + */ + ret = (u64)cs->mult * 11; + do_div(ret,100); + return (u32)ret; +} + /** * clocksource_max_deferment - Returns max time the clocksource can be deferred * @cs: Pointer to clocksource @@ -503,25 +519,28 @@ static u64 clocksource_max_deferment(struct clocksource *cs) /* * Calculate the maximum number of cycles that we can pass to the * cyc2ns function without overflowing a 64-bit signed result. The - * maximum number of cycles is equal to ULLONG_MAX/cs->mult which - * is equivalent to the below. - * max_cycles < (2^63)/cs->mult - * max_cycles < 2^(log2((2^63)/cs->mult)) - * max_cycles < 2^(log2(2^63) - log2(cs->mult)) - * max_cycles < 2^(63 - log2(cs->mult)) - * max_cycles < 1 << (63 - log2(cs->mult)) + * maximum number of cycles is equal to ULLONG_MAX/(cs->mult+cs->maxadj) + * which is equivalent to the below. + * max_cycles < (2^63)/(cs->mult + cs->maxadj) + * max_cycles < 2^(log2((2^63)/(cs->mult + cs->maxadj))) + * max_cycles < 2^(log2(2^63) - log2(cs->mult + cs->maxadj)) + * max_cycles < 2^(63 - log2(cs->mult + cs->maxadj)) + * max_cycles < 1 << (63 - log2(cs->mult + cs->maxadj)) * Please note that we add 1 to the result of the log2 to account for * any rounding errors, ensure the above inequality is satisfied and * no overflow will occur. */ - max_cycles = 1ULL << (63 - (ilog2(cs->mult) + 1)); + max_cycles = 1ULL << (63 - (ilog2(cs->mult + cs->maxadj) + 1)); /* * The actual maximum number of cycles we can defer the clocksource is * determined by the minimum of max_cycles and cs->mask. + * Note: Here we subtract the maxadj to make sure we don't sleep for + * too long if there's a large negative adjustment. */ max_cycles = min_t(u64, max_cycles, (u64) cs->mask); - max_nsecs = clocksource_cyc2ns(max_cycles, cs->mult, cs->shift); + max_nsecs = clocksource_cyc2ns(max_cycles, cs->mult - cs->maxadj, + cs->shift); /* * To ensure that the clocksource does not wrap whilst we are idle, @@ -640,7 +659,6 @@ static void clocksource_enqueue(struct clocksource *cs) void __clocksource_updatefreq_scale(struct clocksource *cs, u32 scale, u32 freq) { u64 sec; - /* * Calc the maximum number of seconds which we can run before * wrapping around. For clocksources which have a mask > 32bit @@ -661,6 +679,20 @@ void __clocksource_updatefreq_scale(struct clocksource *cs, u32 scale, u32 freq) clocks_calc_mult_shift(&cs->mult, &cs->shift, freq, NSEC_PER_SEC / scale, sec * scale); + + /* + * for clocksources that have large mults, to avoid overflow. + * Since mult may be adjusted by ntp, add an safety extra margin + * + */ + cs->maxadj = clocksource_max_adjustment(cs); + while ((cs->mult + cs->maxadj < cs->mult) + || (cs->mult - cs->maxadj > cs->mult)) { + cs->mult >>= 1; + cs->shift--; + cs->maxadj = clocksource_max_adjustment(cs); + } + cs->max_idle_ns = clocksource_max_deferment(cs); } EXPORT_SYMBOL_GPL(__clocksource_updatefreq_scale); @@ -701,6 +733,12 @@ EXPORT_SYMBOL_GPL(__clocksource_register_scale); */ int clocksource_register(struct clocksource *cs) { + /* calculate max adjustment for given mult/shift */ + cs->maxadj = clocksource_max_adjustment(cs); + WARN_ONCE(cs->mult + cs->maxadj < cs->mult, + "Clocksource %s might overflow on 11%% adjustment\n", + cs->name); + /* calculate max idle time permitted for this clocksource */ cs->max_idle_ns = clocksource_max_deferment(cs); diff --git a/kernel/time/timekeeping.c b/kernel/time/timekeeping.c index 2b021b0e8507..e65ff3171102 100644 --- a/kernel/time/timekeeping.c +++ b/kernel/time/timekeeping.c @@ -820,6 +820,13 @@ static void timekeeping_adjust(s64 offset) } else return; + WARN_ONCE(timekeeper.clock->maxadj && + (timekeeper.mult + adj > timekeeper.clock->mult + + timekeeper.clock->maxadj), + "Adjusting %s more then 11%% (%ld vs %ld)\n", + timekeeper.clock->name, (long)timekeeper.mult + adj, + (long)timekeeper.clock->mult + + timekeeper.clock->maxadj); timekeeper.mult += adj; timekeeper.xtime_interval += interval; timekeeper.xtime_nsec -= offset; From d004e024058a0eaca097513ce62cbcf978913e0a Mon Sep 17 00:00:00 2001 From: Hector Palacios Date: Mon, 14 Nov 2011 11:15:25 +0100 Subject: [PATCH 3/4] timekeeping: add arch_offset hook to ktime_get functions ktime_get and ktime_get_ts were calling timekeeping_get_ns() but later they were not calling arch_gettimeoffset() so architectures using this mechanism returned 0 ns when calling these functions. This happened for example when running Busybox's ping which calls syscall(__NR_clock_gettime, CLOCK_MONOTONIC, ts) which eventually calls ktime_get. As a result the returned ping travel time was zero. CC: stable@kernel.org Signed-off-by: Hector Palacios Signed-off-by: John Stultz --- kernel/time/timekeeping.c | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/kernel/time/timekeeping.c b/kernel/time/timekeeping.c index e9f60d311436..237841378c03 100644 --- a/kernel/time/timekeeping.c +++ b/kernel/time/timekeeping.c @@ -249,6 +249,8 @@ ktime_t ktime_get(void) secs = xtime.tv_sec + wall_to_monotonic.tv_sec; nsecs = xtime.tv_nsec + wall_to_monotonic.tv_nsec; nsecs += timekeeping_get_ns(); + /* If arch requires, add in gettimeoffset() */ + nsecs += arch_gettimeoffset(); } while (read_seqretry(&xtime_lock, seq)); /* @@ -280,6 +282,8 @@ void ktime_get_ts(struct timespec *ts) *ts = xtime; tomono = wall_to_monotonic; nsecs = timekeeping_get_ns(); + /* If arch requires, add in gettimeoffset() */ + nsecs += arch_gettimeoffset(); } while (read_seqretry(&xtime_lock, seq)); From 27c9cd7e601632b3794e1c3344d37b86917ffb43 Mon Sep 17 00:00:00 2001 From: Jeff Ohlstein Date: Fri, 18 Nov 2011 15:47:10 -0800 Subject: [PATCH 4/4] hrtimer: Fix extra wakeups from __remove_hrtimer() __remove_hrtimer() attempts to reprogram the clockevent device when the timer being removed is the next to expire. However, __remove_hrtimer() reprograms the clockevent *before* removing the timer from the timerqueue and thus when hrtimer_force_reprogram() finds the next timer to expire it finds the timer we're trying to remove. This is especially noticeable when the system switches to NOHz mode and the system tick is removed. The timer tick is removed from the system but the clockevent is programmed to wakeup in another HZ anyway. Silence the extra wakeup by removing the timer from the timerqueue before calling hrtimer_force_reprogram() so that we actually program the clockevent for the next timer to expire. This was broken by 998adc3 "hrtimers: Convert hrtimers to use timerlist infrastructure". Signed-off-by: Jeff Ohlstein Cc: stable@vger.kernel.org Link: http://lkml.kernel.org/r/1321660030-8520-1-git-send-email-johlstei@codeaurora.org Signed-off-by: Thomas Gleixner --- kernel/hrtimer.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/kernel/hrtimer.c b/kernel/hrtimer.c index a9205e32a059..2043c08d36c8 100644 --- a/kernel/hrtimer.c +++ b/kernel/hrtimer.c @@ -885,10 +885,13 @@ static void __remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base, unsigned long newstate, int reprogram) { + struct timerqueue_node *next_timer; if (!(timer->state & HRTIMER_STATE_ENQUEUED)) goto out; - if (&timer->node == timerqueue_getnext(&base->active)) { + next_timer = timerqueue_getnext(&base->active); + timerqueue_del(&base->active, &timer->node); + if (&timer->node == next_timer) { #ifdef CONFIG_HIGH_RES_TIMERS /* Reprogram the clock event device. if enabled */ if (reprogram && hrtimer_hres_active()) { @@ -901,7 +904,6 @@ static void __remove_hrtimer(struct hrtimer *timer, } #endif } - timerqueue_del(&base->active, &timer->node); if (!timerqueue_getnext(&base->active)) base->cpu_base->active_bases &= ~(1 << base->index); out: