From c2bc11113c50449f23c40b724fe410fc2380a8e9 Mon Sep 17 00:00:00 2001 From: John Stultz Date: Thu, 27 Oct 2011 18:12:42 -0700 Subject: 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(-) (limited to 'kernel/time') 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; -- cgit v1.2.1 From d65670a78cdbfae94f20a9e05ec705871d7cdf2b Mon Sep 17 00:00:00 2001 From: John Stultz Date: Mon, 31 Oct 2011 17:06:35 -0400 Subject: 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 --- kernel/time/clocksource.c | 58 +++++++++++++++++++++++++++++++++++++++-------- kernel/time/timekeeping.c | 7 ++++++ 2 files changed, 55 insertions(+), 10 deletions(-) (limited to 'kernel/time') 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; -- cgit v1.2.1 From d004e024058a0eaca097513ce62cbcf978913e0a Mon Sep 17 00:00:00 2001 From: Hector Palacios Date: Mon, 14 Nov 2011 11:15:25 +0100 Subject: 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(+) (limited to 'kernel/time') 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)); -- cgit v1.2.1 From b1f919664d04a8d0ba29cb76673c7ca3325a2006 Mon Sep 17 00:00:00 2001 From: "Yang Honggang (Joseph)" Date: Thu, 1 Dec 2011 22:22:41 -0500 Subject: clocksource: Fix bug with max_deferment margin calculation In order to leave a margin of 12.5% we should >> 3 not >> 5. CC: stable@kernel.org Signed-off-by: Yang Honggang (Joseph) [jstultz: Modified commit subject] Signed-off-by: John Stultz --- kernel/time/clocksource.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'kernel/time') diff --git a/kernel/time/clocksource.c b/kernel/time/clocksource.c index cfc65e1eb9fb..da2f760e780c 100644 --- a/kernel/time/clocksource.c +++ b/kernel/time/clocksource.c @@ -548,7 +548,7 @@ static u64 clocksource_max_deferment(struct clocksource *cs) * note a margin of 12.5% is used because this can be computed with * a shift, versus say 10% which would require division. */ - return max_nsecs - (max_nsecs >> 5); + return max_nsecs - (max_nsecs >> 3); } #ifndef CONFIG_ARCH_USES_GETTIMEOFFSET @@ -669,7 +669,7 @@ void __clocksource_updatefreq_scale(struct clocksource *cs, u32 scale, u32 freq) * ~ 0.06ppm granularity for NTP. We apply the same 12.5% * margin as we do in clocksource_max_deferment() */ - sec = (cs->mask - (cs->mask >> 5)); + sec = (cs->mask - (cs->mask >> 3)); do_div(sec, freq); do_div(sec, scale); if (!sec) -- cgit v1.2.1 From c1be84309c58b1e7c6d626e28fba41a22b364c3d Mon Sep 17 00:00:00 2001 From: Thomas Gleixner Date: Fri, 2 Dec 2011 12:34:16 +0100 Subject: tick-broadcast: Stop active broadcast device when replacing it When a better rated broadcast device is installed, then the current active device is not disabled, which results in two running broadcast devices. Signed-off-by: Thomas Gleixner Cc: stable@vger.kernel.org --- kernel/time/tick-broadcast.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'kernel/time') diff --git a/kernel/time/tick-broadcast.c b/kernel/time/tick-broadcast.c index f954282d9a82..fd4a7b1625a2 100644 --- a/kernel/time/tick-broadcast.c +++ b/kernel/time/tick-broadcast.c @@ -71,7 +71,7 @@ int tick_check_broadcast_device(struct clock_event_device *dev) (dev->features & CLOCK_EVT_FEAT_C3STOP)) return 0; - clockevents_exchange_device(NULL, dev); + clockevents_exchange_device(tick_broadcast_device.evtdev, dev); tick_broadcast_device.evtdev = dev; if (!cpumask_empty(tick_get_broadcast_mask())) tick_broadcast_start_periodic(dev); -- cgit v1.2.1 From de28f25e8244c7353abed8de0c7792f5f883588c Mon Sep 17 00:00:00 2001 From: Thomas Gleixner Date: Fri, 2 Dec 2011 16:02:45 +0100 Subject: clockevents: Set noop handler in clockevents_exchange_device() If a device is shutdown, then there might be a pending interrupt, which will be processed after we reenable interrupts, which causes the original handler to be run. If the old handler is the (broadcast) periodic handler the shutdown state might hang the kernel completely. Signed-off-by: Thomas Gleixner Cc: stable@vger.kernel.org --- kernel/time/clockevents.c | 1 + 1 file changed, 1 insertion(+) (limited to 'kernel/time') diff --git a/kernel/time/clockevents.c b/kernel/time/clockevents.c index 1ecd6ba36d6c..c4eb71c8b2ea 100644 --- a/kernel/time/clockevents.c +++ b/kernel/time/clockevents.c @@ -387,6 +387,7 @@ void clockevents_exchange_device(struct clock_event_device *old, * released list and do a notify add later. */ if (old) { + old->event_handler = clockevents_handle_noop; clockevents_set_mode(old, CLOCK_EVT_MODE_UNUSED); list_del(&old->list); list_add(&old->list, &clockevents_released); -- cgit v1.2.1 From c9c024b3f3e07d087974db4c0dc46217fff3a6c0 Mon Sep 17 00:00:00 2001 From: Thomas Gleixner Date: Mon, 5 Dec 2011 21:20:23 +0100 Subject: alarmtimers: Fix time comparison The expiry function compares the timer against current time and does not expire the timer when the expiry time is >= now. That's wrong. If the timer is set for now, then it must expire. Make the condition expiry > now for breaking out the loop. Signed-off-by: Thomas Gleixner Acked-by: John Stultz Cc: stable@kernel.org --- kernel/time/alarmtimer.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'kernel/time') diff --git a/kernel/time/alarmtimer.c b/kernel/time/alarmtimer.c index c436e790b21b..8a46f5d64504 100644 --- a/kernel/time/alarmtimer.c +++ b/kernel/time/alarmtimer.c @@ -195,7 +195,7 @@ static enum hrtimer_restart alarmtimer_fired(struct hrtimer *timer) struct alarm *alarm; ktime_t expired = next->expires; - if (expired.tv64 >= now.tv64) + if (expired.tv64 > now.tv64) break; alarm = container_of(next, struct alarm, node); -- cgit v1.2.1