[PATCH v2 0/4] Stop the idle tick on polling idle on nohz_full CPUs

[Marco Crivellari]

Hi,

On nohz_full CPUs booted with idle=poll, the tick interrupts the task once
after every wakeup from idle. do_idle() restarts the tick before polling,
and on wakeup tick_nohz_idle_exit() only re-evaluates the tick if it is
stopped, so it keeps running and fires in userspace before
tick_nohz_irq_exit() stops it. Without idle=poll this doesn't happen:
since commit a5183862e76f ("tick/nohz: Conditionally restart tick on idle
exit") a nohz_full CPU leaves idle with the tick still stopped.

Marcelo Tosatti proposed stopping the tick on polling idle entry:
  https://lore.kernel.org/all/aQiWfnnSzxsnwa2o@tpad/

Following Frederic's review, this series does it only on nohz_full CPUs,
and handles on the target CPU the cases where stopping the tick in polling
idle would leave a timer unprogrammed.

[ P1 ] tick/nohz: Re-evaluate the tick on return to user space
  On nohz_full CPUs the tick is re-evaluated on IRQ exit and on idle exit.

  This is not enough on return to user space. If the tick is left running in
  the kernel, for example because a tick dependency was cleared
  (tick_nohz_dep_clear*() don't kick the CPU) or a timer that kept it
  running was cancelled, it fires at least once in userspace before
  tick_nohz_irq_exit() stops it.

  Stop it from context tracking right before resuming userspace. When the
  tick is already stopped, the common case, this costs a few flag tests on
  each return to userspace.

[ P2 ] Don't handle polling idle as dynticks idle

  TS_FLAG_INIDLE tells nohz whether the tick is managed by the idle loop or
  on IRQ exit. With idle=poll it is set while polling, but cpu_idle_poll()
  doesn't run the idle loop again on interrupts: a nohz_full kick or the
  expiry of a stopped tick isn't re-evaluated until something sets
  TIF_NEED_RESCHED. The tick is therefore restarted before polling, and on
  nohz_full CPUs it then keeps running after the wake-up and interrupts
  userspace.

  Enter dynticks idle only on the cpuidle path, from cpuidle_idle_call(),
  and leave it before cpu_idle_poll(). While polling, the tick is then
  managed on IRQ exit as for a running task: on nohz_full CPUs it stays
  stopped if possible, and nothing is left to do on wake-up, so the task
  doesn't pay for a tick stop after waking up. On other CPUs, leaving
  dynticks idle restarts the tick as before.

[ P3 ] tick/nohz: Remove tick_nohz_idle_restart_tick()

  Polling idle now leaves dynticks idle through tick_nohz_idle_exit()
  before cpu_idle_poll(), which restarts the tick where needed.
  tick_nohz_idle_restart_tick() has no users left: remove it.

[ P4 ] Documentation: no_hz: Add tick stop in polling idle

  Document the above in the idle=poll item of "Known Issues".

With big effort on the Claude side, I made a test suite to help testing
the work:

  https://github.com/DispatchCode/nohz-full-polling-idle-tick-test-suite

The following is a short summary of the results.

On x86_64 guest, HZ=1000, nohz_full=2-3 idle=poll), a task alternating
100us sleeps with 10ms of userspace work, 200 times:
  - ticks in userspace:              200/200 -> 0/200
  - ticks while polling:             35 -> 0
  - housekeeping CPUs:               unchanged (tick runs while polling)
  - P2 without the kick change:      15 of 20 timers enqueued from an
                                     interrupt on the polling CPU lost
  - P2 without the stopped tick fix: the second of two queued timers up
                                     to 1.9s late
Tick dependencies (POSIX CPU timers, two tasks on the CPU) and CPU hotplug
were also checked.

Thanks,

  Marco

---
Changes in v2:
 - P1: The only change part of this patch is re-evaluate the tick on
   return to user space.
 
 - P2: Changed as suggested by Frederic: TS_FLAG_INIDLE is only set when
   the CPU goes idle and not in polling idle; tick_nohz_idle_enter() is
   called on every idle loop and runs with interrupt disabled. The tick
   is not restarted before polling idle, so
   tick_nohz_idle_restart_tick() has been removed.

 - P3: Remove tick_nohz_idle_restart_tick(), now unused.

Marco Crivellari (4):
  tick/nohz: Re-evaluate the tick on return to user space
  sched/idle: Don't handle polling idle as dynticks idle
  tick/nohz: Remove tick_nohz_idle_restart_tick()
  Documentation: no_hz: Add tick stop in polling idle

 Documentation/timers/no_hz.rst |  2 +
 include/linux/tick.h           | 10 ++++-
 kernel/context_tracking.c      |  8 ++++
 kernel/sched/idle.c            | 13 ++++++-
 kernel/time/tick-sched.c       | 68 ++++++++++++++++++++--------------
 5 files changed, 69 insertions(+), 32 deletions(-)

-- 
2.56.0