When a tick broadcast clockevent device is initialized for one shot mode
then tick_broadcast_setup_oneshot() OR's the periodic broadcast mode
cpumask into the oneshot broadcast cpumask.

This is required when switching from periodic broadcast mode to oneshot
broadcast mode to ensure that CPUs which are waiting for periodic
broadcast are woken up on the next tick.

But it is subtly broken, when an active broadcast device is replaced and
the system is already in oneshot (NOHZ/HIGHRES) mode. Victor observed
this and debugged the issue.

Then the OR of the periodic broadcast CPU mask is wrong as the periodic
cpumask bits are sticky after tick_broadcast_enable() set it for a CPU
unless explicitly cleared via tick_broadcast_disable().

That means that this sets all other CPUs which have tick broadcasting
enabled at that point unconditionally in the oneshot broadcast mask.

If the affected CPUs were already idle and had their bits set in the
oneshot broadcast mask then this does no harm. But for non idle CPUs
which were not set this corrupts their state.

On their next invocation of tick_broadcast_enable() they observe the bit
set, which indicates that the broadcast for the CPU is already set up.
As a consequence they fail to update the broadcast event even if their
earliest expiring timer is before the actually programmed broadcast
event.

If the programmed broadcast event is far in the future, then this can
cause stalls or trigger the hung task detector.

Avoid this by telling tick_broadcast_setup_oneshot() explicitly whether
this is the initial switch over from periodic to oneshot broadcast which
must take the periodic broadcast mask into account. In the case of
initialization of a replacement device this prevents that the broadcast
oneshot mask is modified.

There is a second problem with broadcast device replacement in this
function. The broadcast device is only armed when the previous state of
the device was periodic.

That is correct for the switch from periodic broadcast mode to oneshot
broadcast mode as the underlying broadcast device could operate in
oneshot state already due to lack of periodic state in hardware. In that
case it is already armed to expire at the next tick.

For the replacement case this is wrong as the device is in shutdown
state. That means that any already pending broadcast event will not be
armed.

This went unnoticed because any CPU which goes idle will observe that
the broadcast device has an expiry time of KTIME_MAX and therefore any
CPUs next timer event will be earlier and cause a reprogramming of the
broadcast device. But that does not guarantee that the events of the
CPUs which were already in idle are delivered on time.

Fix this by arming the newly installed device for an immediate event
which will reevaluate the per CPU expiry times and reprogram the
broadcast device accordingly. This is simpler than caching the last
expiry time in yet another place or saving it before the device exchange
and handing it down to the setup function. Replacement of broadcast
devices is not a frequent operation and usually happens once somewhere
late in the boot process.

Fixes: 9c336c9935cf ("tick/broadcast: Allow late registered device to enter oneshot mode")
Reported-by: Victor Hassan <victor@allwinnertech.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Link: https://lore.kernel.org/r/87pm7d2z1i.ffs@tglx

Pull more timer updates from Thomas Gleixner:
 "Timekeeping and clocksource/event driver updates the second batch:

   - A trivial documentation fix in the timekeeping core

   - A really boring set of small fixes, enhancements and cleanups in
     the drivers code. No new clocksource/clockevent drivers for a
     change"

* tag 'timers-core-2023-04-28' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  timekeeping: Fix references to nonexistent ktime_get_fast_ns()
  dt-bindings: timer: rockchip: Add rk3588 compatible
  dt-bindings: timer: rockchip: Drop superfluous rk3288 compatible
  clocksource/drivers/ti: Use of_property_read_bool() for boolean properties
  clocksource/drivers/timer-ti-dm: Fix finding alwon timer
  clocksource/drivers/davinci: Fix memory leak in davinci_timer_register when init fails
  clocksource/drivers/stm32-lp: Drop of_match_ptr for ID table
  clocksource/drivers/timer-ti-dm: Convert to platform remove callback returning void
  clocksource/drivers/timer-tegra186: Convert to platform remove callback returning void
  clocksource/drivers/timer-ti-dm: Improve error message in .remove
  clocksource/drivers/timer-stm32-lp: Mark driver as non-removable
  clocksource/drivers/sh_mtu2: Mark driver as non-removable
  clocksource/drivers/timer-ti-dm: Use of_address_to_resource()
  clocksource/drivers/timer-imx-gpt: Remove non-DT function
  clocksource/drivers/timer-mediatek: Split out CPUXGPT timers
  clocksource/drivers/exynos_mct: Explicitly return 0 for shared timer

Pull driver core updates from Greg KH:
 "Here is the large set of driver core changes for 6.4-rc1.

  Once again, a busy development cycle, with lots of changes happening
  in the driver core in the quest to be able to move "struct bus" and
  "struct class" into read-only memory, a task now complete with these
  changes.

  This will make the future rust interactions with the driver core more
  "provably correct" as well as providing more obvious lifetime rules
  for all busses and classes in the kernel.

  The changes required for this did touch many individual classes and
  busses as many callbacks were changed to take const * parameters
  instead. All of these changes have been submitted to the various
  subsystem maintainers, giving them plenty of time to review, and most
  of them actually did so.

  Other than those changes, included in here are a small set of other
  things:

   - kobject logging improvements

   - cacheinfo improvements and updates

   - obligatory fw_devlink updates and fixes

   - documentation updates

   - device property cleanups and const * changes

   - firwmare loader dependency fixes.

  All of these have been in linux-next for a while with no reported
  problems"

* tag 'driver-core-6.4-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/driver-core: (120 commits)
  device property: make device_property functions take const device *
  driver core: update comments in device_rename()
  driver core: Don't require dynamic_debug for initcall_debug probe timing
  firmware_loader: rework crypto dependencies
  firmware_loader: Strip off \n from customized path
  zram: fix up permission for the hot_add sysfs file
  cacheinfo: Add use_arch[|_cache]_info field/function
  arch_topology: Remove early cacheinfo error message if -ENOENT
  cacheinfo: Check cache properties are present in DT
  cacheinfo: Check sib_leaf in cache_leaves_are_shared()
  cacheinfo: Allow early level detection when DT/ACPI info is missing/broken
  cacheinfo: Add arm64 early level initializer implementation
  cacheinfo: Add arch specific early level initializer
  tty: make tty_class a static const structure
  driver core: class: remove struct class_interface * from callbacks
  driver core: class: mark the struct class in struct class_interface constant
  driver core: class: make class_register() take a const *
  driver core: class: mark class_release() as taking a const *
  driver core: remove incorrect comment for device_create*
  MIPS: vpe-cmp: remove module owner pointer from struct class usage.
  ...

There was never a function named ktime_get_fast_ns().
Presumably these should refer to ktime_get_mono_fast_ns() instead.

Fixes: c1ce406e80fb15fa ("timekeeping: Fix up function documentation for the NMI safe accessors")
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: John Stultz <jstultz@google.com>
Link: https://lore.kernel.org/r/06df7b3cbd94f016403bbf6cd2b38e4368e7468f.1682516546.git.geert+renesas@glider.be

Pull timers and timekeeping updates from Thomas Gleixner:

 - Improve the VDSO build time checks to cover all dynamic relocations

   VDSO does not allow dynamic relocations, but the build time check is
   incomplete and fragile.

   It's based on architectures specifying the relocation types to search
   for and does not handle R_*_NONE relocation entries correctly.
   R_*_NONE relocations are injected by some GNU ld variants if they
   fail to determine the exact .rel[a]/dyn_size to cover trailing zeros.
   R_*_NONE relocations must be ignored by dynamic loaders, so they
   should be ignored in the build time check too.

   Remove the architecture specific relocation types to check for and
   validate strictly that no other relocations than R_*_NONE end up in
   the VSDO .so file.

 - Prefer signal delivery to the current thread for
   CLOCK_PROCESS_CPUTIME_ID based posix-timers

   Such timers prefer to deliver the signal to the main thread of a
   process even if the context in which the timer expires is the current
   task. This has the downside that it might wake up an idle thread.

   As there is no requirement or guarantee that the signal has to be
   delivered to the main thread, avoid this by preferring the current
   task if it is part of the thread group which shares sighand.

   This not only avoids waking idle threads, it also distributes the
   signal delivery in case of multiple timers firing in the context of
   different threads close to each other better.

 - Align the tick period properly (again)

   For a long time the tick was starting at CLOCK_MONOTONIC zero, which
   allowed users space applications to either align with the tick or to
   place a periodic computation so that it does not interfere with the
   tick. The alignement of the tick period was more by chance than by
   intention as the tick is set up before a high resolution clocksource
   is installed, i.e. timekeeping is still tick based and the tick
   period advances from there.

   The early enablement of sched_clock() broke this alignement as the
   time accumulated by sched_clock() is taken into account when
   timekeeping is initialized. So the base value now(CLOCK_MONOTONIC) is
   not longer a multiple of tick periods, which breaks applications
   which relied on that behaviour.

   Cure this by aligning the tick starting point to the next multiple of
   tick periods, i.e 1000ms/CONFIG_HZ.

 - A set of NOHZ fixes and enhancements:

     * Cure the concurrent writer race for idle and IO sleeptime
       statistics

       The statitic values which are exposed via /proc/stat are updated
       from the CPU local idle exit and remotely by cpufreq, but that
       happens without any form of serialization. As a consequence
       sleeptimes can be accounted twice or worse.

       Prevent this by restricting the accumulation writeback to the CPU
       local idle exit and let the remote access compute the accumulated
       value.

     * Protect idle/iowait sleep time with a sequence count

       Reading idle/iowait sleep time, e.g. from /proc/stat, can race
       with idle exit updates. As a consequence the readout may result
       in random and potentially going backwards values.

       Protect this by a sequence count, which fixes the idle time
       statistics issue, but cannot fix the iowait time problem because
       iowait time accounting races with remote wake ups decrementing
       the remote runqueues nr_iowait counter. The latter is impossible
       to fix, so the only way to deal with that is to document it
       properly and to remove the assertion in the selftest which
       triggers occasionally due to that.

     * Restructure struct tick_sched for better cache layout

     * Some small cleanups and a better cache layout for struct
       tick_sched

 - Implement the missing timer_wait_running() callback for POSIX CPU
   timers

   For unknown reason the introduction of the timer_wait_running()
   callback missed to fixup posix CPU timers, which went unnoticed for
   almost four years.

   While initially only targeted to prevent livelocks between a timer
   deletion and the timer expiry function on PREEMPT_RT enabled kernels,
   it turned out that fixing this for mainline is not as trivial as just
   implementing a stub similar to the hrtimer/timer callbacks.

   The reason is that for CONFIG_POSIX_CPU_TIMERS_TASK_WORK enabled
   systems there is a livelock issue independent of RT.

   CONFIG_POSIX_CPU_TIMERS_TASK_WORK=y moves the expiry of POSIX CPU
   timers out from hard interrupt context to task work, which is handled
   before returning to user space or to a VM. The expiry mechanism moves
   the expired timers to a stack local list head with sighand lock held.
   Once sighand is dropped the task can be preempted and a task which
   wants to delete a timer will spin-wait until the expiry task is
   scheduled back in. In the worst case this will end up in a livelock
   when the preempting task and the expiry task are pinned on the same
   CPU.

   The timer wheel has a timer_wait_running() mechanism for RT, which
   uses a per CPU timer-base expiry lock which is held by the expiry
   code and the task waiting for the timer function to complete blocks
   on that lock.

   This does not work in the same way for posix CPU timers as there is
   no timer base and expiry for process wide timers can run on any task
   belonging to that process, but the concept of waiting on an expiry
   lock can be used too in a slightly different way.

   Add a per task mutex to struct posix_cputimers_work, let the expiry
   task hold it accross the expiry function and let the deleting task
   which waits for the expiry to complete block on the mutex.

   In the non-contended case this results in an extra
   mutex_lock()/unlock() pair on both sides.

   This avoids spin-waiting on a task which is scheduled out, prevents
   the livelock and cures the problem for RT and !RT systems

* tag 'timers-core-2023-04-24' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  posix-cpu-timers: Implement the missing timer_wait_running callback
  selftests/proc: Assert clock_gettime(CLOCK_BOOTTIME) VS /proc/uptime monotonicity
  selftests/proc: Remove idle time monotonicity assertions
  MAINTAINERS: Remove stale email address
  timers/nohz: Remove middle-function __tick_nohz_idle_stop_tick()
  timers/nohz: Add a comment about broken iowait counter update race
  timers/nohz: Protect idle/iowait sleep time under seqcount
  timers/nohz: Only ever update sleeptime from idle exit
  timers/nohz: Restructure and reshuffle struct tick_sched
  tick/common: Align tick period with the HZ tick.
  selftests/timers/posix_timers: Test delivery of signals across threads
  posix-timers: Prefer delivery of signals to the current thread
  vdso: Improve cmd_vdso_check to check all dynamic relocations

For some unknown reason the introduction of the timer_wait_running callback
missed to fixup posix CPU timers, which went unnoticed for almost four years.
Marco reported recently that the WARN_ON() in timer_wait_running()
triggers with a posix CPU timer test case.

Posix CPU timers have two execution models for expiring timers depending on
CONFIG_POSIX_CPU_TIMERS_TASK_WORK:

1) If not enabled, the expiry happens in hard interrupt context so
   spin waiting on the remote CPU is reasonably time bound.

   Implement an empty stub function for that case.

2) If enabled, the expiry happens in task work before returning to user
   space or guest mode. The expired timers are marked as firing and moved
   from the timer queue to a local list head with sighand lock held. Once
   the timers are moved, sighand lock is dropped and the expiry happens in
   fully preemptible context. That means the expiring task can be scheduled
   out, migrated, interrupted etc. So spin waiting on it is more than
   suboptimal.

   The timer wheel has a timer_wait_running() mechanism for RT, which uses
   a per CPU timer-base expiry lock which is held by the expiry code and the
   task waiting for the timer function to complete blocks on that lock.

   This does not work in the same way for posix CPU timers as there is no
   timer base and expiry for process wide timers can run on any task
   belonging to that process, but the concept of waiting on an expiry lock
   can be used too in a slightly different way:

    - Add a mutex to struct posix_cputimers_work. This struct is per task
      and used to schedule the expiry task work from the timer interrupt.

    - Add a task_struct pointer to struct cpu_timer which is used to store
      a the task which runs the expiry. That's filled in when the task
      moves the expired timers to the local expiry list. That's not
      affecting the size of the k_itimer union as there are bigger union
      members already

    - Let the task take the expiry mutex around the expiry function

    - Let the waiter acquire a task reference with rcu_read_lock() held and
      block on the expiry mutex

   This avoids spin-waiting on a task which might not even be on a CPU and
   works nicely for RT too.

Fixes: ec8f954a40da ("posix-timers: Use a callback for cancel synchronization on PREEMPT_RT")
Reported-by: Marco Elver <elver@google.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Marco Elver <elver@google.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/87zg764ojw.ffs@tglx

There is no need for the __tick_nohz_idle_stop_tick() function between
tick_nohz_idle_stop_tick() and its implementation. Remove that
unnecessary step.

Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20230222144649.624380-6-frederic@kernel.org

The per-cpu iowait task counter is incremented locally upon sleeping.
But since the task can be woken to (and by) another CPU, the counter may
then be decremented remotely. This is the source of a race involving
readers VS writer of idle/iowait sleeptime.

The following scenario shows an example where a /proc/stat reader
observes a pending sleep time as IO whereas that pending sleep time
later eventually gets accounted as non-IO.

    CPU 0                       CPU  1                    CPU 2
    -----                       -----                     ------
    //io_schedule() TASK A
    current->in_iowait = 1
    rq(0)->nr_iowait++
    //switch to idle
                        // READ /proc/stat
                        // See nr_iowait_cpu(0) == 1
                        return ts->iowait_sleeptime +
                               ktime_sub(ktime_get(), ts->idle_entrytime)

                                                          //try_to_wake_up(TASK A)
                                                          rq(0)->nr_iowait--
    //idle exit
    // See nr_iowait_cpu(0) == 0
    ts->idle_sleeptime += ktime_sub(ktime_get(), ts->idle_entrytime)

As a result subsequent reads on /proc/stat may expose backward progress.

This is unfortunately hardly fixable. Just add a comment about that
condition.

Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20230222144649.624380-5-frederic@kernel.org

Reading idle/IO sleep time (eg: from /proc/stat) can race with idle exit
updates because the state machine handling the stats is not atomic and
requires a coherent read batch.

As a result reading the sleep time may report irrelevant or backward
values.

Fix this with protecting the simple state machine within a seqcount.
This is expected to be cheap enough not to add measurable performance
impact on the idle path.

Note this only fixes reader VS writer condition partitially. A race
remains that involves remote updates of the CPU iowait task counter. It
can hardly be fixed.

Reported-by: Yu Liao <liaoyu15@huawei.com>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20230222144649.624380-4-frederic@kernel.org

The idle and IO sleeptime statistics appearing in /proc/stat can be
currently updated from two sites: locally on idle exit and remotely
by cpufreq. However there is no synchronization mechanism protecting
concurrent updates. It is therefore possible to account the sleeptime
twice, among all the other possible broken scenarios.

To prevent from breaking the sleeptime accounting source, restrict the
sleeptime updates to the local idle exit site. If there is a delta to
add since the last update, IO/Idle sleep time readers will now only
compute the delta without actually writing it back to the internal idle
statistic fields.

This fixes a writer VS writer race. Note there are still two known
reader VS writer races to handle. A subsequent patch will fix one.

Reported-by: Yu Liao <liaoyu15@huawei.com>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20230222144649.624380-3-frederic@kernel.org