linux/kernel/time/timekeeping.c, branch v3.17

timekeeping: Update timekeeper before updating vsyscall and pvclock

2014-09-06T10:58:18Z

The update_walltime() code works on the shadow timekeeper to make the seqcount protected region as short as possible. But that update to the shadow timekeeper does not update all timekeeper fields because it's sufficient to do that once before it becomes life. One of these fields is tkr.base_mono. That stays stale in the shadow timekeeper unless an operation happens which copies the real timekeeper to the shadow. The update function is called after the update calls to vsyscall and pvclock. While not correct, it did not cause any problems because none of the invoked update functions used base_mono. commit cbcf2dd3b3d4 (x86: kvm: Make kvm_get_time_and_clockread() nanoseconds based) changed that in the kvm pvclock update function, so the stale mono_base value got used and caused kvm-clock to malfunction. Put the update where it belongs and fix the issue. Reported-by: Chris J Arges Reported-by: Paolo Bonzini Cc: Gleb Natapov Cc: John Stultz Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1409050000570.3333@nanos Signed-off-by: Thomas Gleixner

timekeeping: Another fix to the VSYSCALL_OLD update_vsyscall

2014-08-14T17:04:11Z

Benjamin Herrenschmidt pointed out that I further missed modifying update_vsyscall after the wall_to_mono value was changed to a timespec64. This causes issues on powerpc32, which expects a 32bit timespec. This patch fixes the problem by properly converting from a timespec64 to a timespec before passing the value on to the arch-specific vsyscall logic. [ Thomas is currently on vacation, but reviewed it and wanted me to send this fix on to you directly. ] Cc: LKML Cc: Thomas Gleixner Cc: Ingo Molnar Cc: Benjamin Herrenschmidt Reported-by: Benjamin Herrenschmidt Reviewed-by: Thomas Gleixner Signed-off-by: John Stultz Signed-off-by: Linus Torvalds

timekeeping: Use cached ntp_tick_length when accumulating error

2014-07-23T22:01:57Z

By caching the ntp_tick_length() when we correct the frequency error, and then using that cached value to accumulate error, we avoid large initial errors when the tick length is changed. This makes convergence happen much faster in the simulator, since the initial error doesn't have to be slowly whittled away. This initially seems like an accounting error, but Miroslav pointed out that ntp_tick_length() can change mid-tick, so when we apply it in the error accumulation, we are applying any recent change to the entire tick. This approach chooses to apply changes in the ntp_tick_length() only to the next tick, which allows us to calculate the freq correction before using the new tick length, which avoids accummulating error. Credit to Miroslav for pointing this out and providing the original patch this functionality has been pulled out from, along with the rational. Cc: Miroslav Lichvar Cc: Richard Cochran Cc: Prarit Bhargava Reported-by: Miroslav Lichvar Signed-off-by: John Stultz

timekeeping: Rework frequency adjustments to work better w/ nohz

2014-07-23T22:01:56Z

The existing timekeeping_adjust logic has always been complicated to understand. Further, since it was developed prior to NOHZ becoming common, its not surprising it performs poorly when NOHZ is enabled. Since Miroslav pointed out the problematic nature of the existing code in the NOHZ case, I've tried to refactor the code to perform better. The problem with the previous approach was that it tried to adjust for the total cumulative error using a scaled dampening factor. This resulted in large errors to be corrected slowly, while small errors were corrected quickly. With NOHZ the timekeeping code doesn't know how far out the next tick will be, so this results in bad over-correction to small errors, and insufficient correction to large errors. Inspired by Miroslav's patch, I've refactored the code to try to address the correction in two steps. 1) Check the future freq error for the next tick, and if the frequency error is large, try to make sure we correct it so it doesn't cause much accumulated error. 2) Then make a small single unit adjustment to correct any cumulative error that has collected over time. This method performs fairly well in the simulator Miroslav created. Major credit to Miroslav for pointing out the issue, providing the original patch to resolve this, a simulator for testing, as well as helping debug and resolve issues in my implementation so that it performed closer to his original implementation. Cc: Miroslav Lichvar Cc: Richard Cochran Cc: Prarit Bhargava Reported-by: Miroslav Lichvar Signed-off-by: John Stultz

timekeeping: Minor fixup for timespec64->timespec assignment

2014-07-23T22:01:56Z

In the GENERIC_TIME_VSYSCALL_OLD update_vsyscall implementation, we take the tk_xtime() value, which returns a timespec64, and store it in a timespec. This luckily is ok, since the only architectures that use GENERIC_TIME_VSYSCALL_OLD are ia64 and ppc64, which are both 64 bit systems where timespec64 is the same as a timespec. Even so, for cleanliness reasons, use the conversion function to assign the proper type. Signed-off-by: John Stultz

timekeeping: Provide fast and NMI safe access to CLOCK_MONOTONIC

2014-07-23T22:01:55Z

Tracers want a correlated time between the kernel instrumentation and user space. We really do not want to export sched_clock() to user space, so we need to provide something sensible for this. Using separate data structures with an non blocking sequence count based update mechanism allows us to do that. The data structure required for the readout has a sequence counter and two copies of the timekeeping data. On the update side: smp_wmb(); tkf->seq++; smp_wmb(); update(tkf->base[0], tk); smp_wmb(); tkf->seq++; smp_wmb(); update(tkf->base[1], tk); On the reader side: do { seq = tkf->seq; smp_rmb(); idx = seq & 0x01; now = now(tkf->base[idx]); smp_rmb(); } while (seq != tkf->seq) So if a NMI hits the update of base[0] it will use base[1] which is still consistent, but this timestamp is not guaranteed to be monotonic across an update. The timestamp is calculated by: now = base_mono + clock_delta * slope So if the update lowers the slope, readers who are forced to the not yet updated second array are still using the old steeper slope. tmono ^ | o n | o n | u | o |o |12345678---> reader order o = old slope u = update n = new slope So reader 6 will observe time going backwards versus reader 5. While other CPUs are likely to be able observe that, the only way for a CPU local observation is when an NMI hits in the middle of the update. Timestamps taken from that NMI context might be ahead of the following timestamps. Callers need to be aware of that and deal with it. V2: Got rid of clock monotonic raw and reorganized the data structures. Folded in the barrier fix from Mathieu. Signed-off-by: Thomas Gleixner Cc: Peter Zijlstra Cc: Steven Rostedt Cc: Mathieu Desnoyers Signed-off-by: John Stultz

timekeeping: Use tk_read_base as argument for timekeeping_get_ns()

2014-07-23T22:01:53Z

All the function needs is in the tk_read_base struct. No functional change for the current code, just a preparatory patch for the NMI safe accessor to clock monotonic which will use struct tk_read_base as well. Signed-off-by: Thomas Gleixner Cc: Steven Rostedt Cc: Peter Zijlstra Cc: Mathieu Desnoyers Signed-off-by: John Stultz

timekeeping: Create struct tk_read_base and use it in struct timekeeper

2014-07-23T22:01:53Z

The members of the new struct are the required ones for the new NMI safe accessor to clcok monotonic. In order to reuse the existing timekeeping code and to make the update of the fast NMI safe timekeepers a simple memcpy use the struct for the timekeeper as well and convert all users. Signed-off-by: Thomas Gleixner Cc: Peter Zijlstra Cc: Ingo Molnar Cc: Mathieu Desnoyers Signed-off-by: John Stultz

timekeeping: Restructure the timekeeper some more

2014-07-23T22:01:52Z

Access to time requires to touch two cachelines at minimum 1) The timekeeper data structure 2) The clocksource data structure The access to the clocksource data structure can be avoided as almost all clocksource implementations ignore the argument to the read callback, which is a pointer to the clocksource. But the core needs to touch it to access the members @read and @mask. So we are better off by copying the @read function pointer and the @mask from the clocksource to the core data structure itself. For the most used ktime_get() access all required data including the @read and @mask copies fits together with the sequence counter into a single 64 byte cacheline. For the other time access functions we touch in the current code three cache lines in the worst case. But with the clocksource data copies we can reduce that to two adjacent cachelines, which is more efficient than disjunct cache lines. Signed-off-by: Thomas Gleixner Signed-off-by: John Stultz

clocksource: Get rid of cycle_last

2014-07-23T22:01:52Z

cycle_last was added to the clocksource to support the TSC validation. We moved that to the core code, so we can get rid of the extra copy. Signed-off-by: Thomas Gleixner Signed-off-by: John Stultz