Revision 6e2df0581f569038719cf2bc2b3baa3fcc83cab4 authored by Peter Zijlstra on 08 November 2019, 10:11:52 UTC, committed by Peter Zijlstra on 08 November 2019, 21:34:14 UTC
Commit 67692435c411 ("sched: Rework pick_next_task() slow-path") inadvertly introduced a race because it changed a previously unexplored dependency between dropping the rq->lock and sched_class::put_prev_task(). The comments about dropping rq->lock, in for example newidle_balance(), only mentions the task being current and ->on_cpu being set. But when we look at the 'change' pattern (in for example sched_setnuma()): queued = task_on_rq_queued(p); /* p->on_rq == TASK_ON_RQ_QUEUED */ running = task_current(rq, p); /* rq->curr == p */ if (queued) dequeue_task(...); if (running) put_prev_task(...); /* change task properties */ if (queued) enqueue_task(...); if (running) set_next_task(...); It becomes obvious that if we do this after put_prev_task() has already been called on @p, things go sideways. This is exactly what the commit in question allows to happen when it does: prev->sched_class->put_prev_task(rq, prev, rf); if (!rq->nr_running) newidle_balance(rq, rf); The newidle_balance() call will drop rq->lock after we've called put_prev_task() and that allows the above 'change' pattern to interleave and mess up the state. Furthermore, it turns out we lost the RT-pull when we put the last DL task. Fix both problems by extracting the balancing from put_prev_task() and doing a multi-class balance() pass before put_prev_task(). Fixes: 67692435c411 ("sched: Rework pick_next_task() slow-path") Reported-by: Quentin Perret <qperret@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Quentin Perret <qperret@google.com> Tested-by: Valentin Schneider <valentin.schneider@arm.com>
1 parent e3b8b6a
hwlat_detector.rst
=========================
Hardware Latency Detector
=========================
Introduction
-------------
The tracer hwlat_detector is a special purpose tracer that is used to
detect large system latencies induced by the behavior of certain underlying
hardware or firmware, independent of Linux itself. The code was developed
originally to detect SMIs (System Management Interrupts) on x86 systems,
however there is nothing x86 specific about this patchset. It was
originally written for use by the "RT" patch since the Real Time
kernel is highly latency sensitive.
SMIs are not serviced by the Linux kernel, which means that it does not
even know that they are occuring. SMIs are instead set up by BIOS code
and are serviced by BIOS code, usually for "critical" events such as
management of thermal sensors and fans. Sometimes though, SMIs are used for
other tasks and those tasks can spend an inordinate amount of time in the
handler (sometimes measured in milliseconds). Obviously this is a problem if
you are trying to keep event service latencies down in the microsecond range.
The hardware latency detector works by hogging one of the cpus for configurable
amounts of time (with interrupts disabled), polling the CPU Time Stamp Counter
for some period, then looking for gaps in the TSC data. Any gap indicates a
time when the polling was interrupted and since the interrupts are disabled,
the only thing that could do that would be an SMI or other hardware hiccup
(or an NMI, but those can be tracked).
Note that the hwlat detector should *NEVER* be used in a production environment.
It is intended to be run manually to determine if the hardware platform has a
problem with long system firmware service routines.
Usage
------
Write the ASCII text "hwlat" into the current_tracer file of the tracing system
(mounted at /sys/kernel/tracing or /sys/kernel/tracing). It is possible to
redefine the threshold in microseconds (us) above which latency spikes will
be taken into account.
Example::
# echo hwlat > /sys/kernel/tracing/current_tracer
# echo 100 > /sys/kernel/tracing/tracing_thresh
The /sys/kernel/tracing/hwlat_detector interface contains the following files:
- width - time period to sample with CPUs held (usecs)
must be less than the total window size (enforced)
- window - total period of sampling, width being inside (usecs)
By default the width is set to 500,000 and window to 1,000,000, meaning that
for every 1,000,000 usecs (1s) the hwlat detector will spin for 500,000 usecs
(0.5s). If tracing_thresh contains zero when hwlat tracer is enabled, it will
change to a default of 10 usecs. If any latencies that exceed the threshold is
observed then the data will be written to the tracing ring buffer.
The minimum sleep time between periods is 1 millisecond. Even if width
is less than 1 millisecond apart from window, to allow the system to not
be totally starved.
If tracing_thresh was zero when hwlat detector was started, it will be set
back to zero if another tracer is loaded. Note, the last value in
tracing_thresh that hwlat detector had will be saved and this value will
be restored in tracing_thresh if it is still zero when hwlat detector is
started again.
The following tracing directory files are used by the hwlat_detector:
in /sys/kernel/tracing:
- tracing_threshold - minimum latency value to be considered (usecs)
- tracing_max_latency - maximum hardware latency actually observed (usecs)
- tracing_cpumask - the CPUs to move the hwlat thread across
- hwlat_detector/width - specified amount of time to spin within window (usecs)
- hwlat_detector/window - amount of time between (width) runs (usecs)
The hwlat detector's kernel thread will migrate across each CPU specified in
tracing_cpumask between each window. To limit the migration, either modify
tracing_cpumask, or modify the hwlat kernel thread (named [hwlatd]) CPU
affinity directly, and the migration will stop.
![swh spinner](/static/img/swh-spinner.gif)
Computing file changes ...