The select_idle_sibling() (SIS) rewrite in commit:
10e2f1acd0 ("sched/core: Rewrite and improve select_idle_siblings()")
... replaced a domain iteration with a search that broadly speaking
does a wrapped walk of the scheduler domain sharing a last-level-cache.
While this had a number of improvements, one consequence is that two tasks
that share a waker/wakee relationship push each other around a socket. Even
though two tasks may be active, all cores are evenly used. This is great from
a search perspective and spreads a load across individual cores, but it has
adverse consequences for cpufreq. As each CPU has relatively low utilisation,
cpufreq may decide the utilisation is too low to used a higher P-state and
overall computation throughput suffers.
While individual cpufreq and cpuidle drivers may compensate by artifically
boosting P-state (at c0) or avoiding lower C-states (during idle), it does
not help if hardware-based cpufreq (e.g. HWP) is used.
This patch tracks a recently used CPU based on what CPU a task was running
on when it last was a waker a CPU it was recently using when a task is a
wakee. During SIS, the recently used CPU is used as a target if it's still
allowed by the task and is idle.
The benefit may be non-obvious so consider an example of two tasks
communicating back and forth. Task A may be an application doing IO where
task B is a kworker or kthread like journald. Task A may issue IO, wake
B and B wakes up A on completion. With the existing scheme this may look
like the following (potentially different IDs if SMT is in use but similar
principal applies).
A (cpu 0) wake B (wakes on cpu 1)
B (cpu 1) wake A (wakes on cpu 2)
A (cpu 2) wake B (wakes on cpu 3)
etc.
A careful reader may wonder why CPU 0 was not idle when B wakes A the
first time and it's simply due to the fact that A can be rescheduled to
another CPU and the pattern is that prev == target when B tries to wakeup A
and the information about CPU 0 has been lost.
With this patch, the pattern is more likely to be:
A (cpu 0) wake B (wakes on cpu 1)
B (cpu 1) wake A (wakes on cpu 0)
A (cpu 0) wake B (wakes on cpu 1)
etc
i.e. two communicating casts are more likely to use just two cores instead
of all available cores sharing a LLC.
The most dramatic speedup was noticed on dbench using the XFS filesystem on
UMA as clients interact heavily with workqueues in that configuration. Note
that a similar speedup is not observed on ext4 as the wakeup pattern
is different:
4.15.0-rc9 4.15.0-rc9
waprev-v1 biasancestor-v1
Hmean 1 287.54 ( 0.00%) 817.01 ( 184.14%)
Hmean 2 1268.12 ( 0.00%) 1781.24 ( 40.46%)
Hmean 4 1739.68 ( 0.00%) 1594.47 ( -8.35%)
Hmean 8 2464.12 ( 0.00%) 2479.56 ( 0.63%)
Hmean 64 1455.57 ( 0.00%) 1434.68 ( -1.44%)
The results can be less dramatic on NUMA where automatic balancing interferes
with the test. It's also known that network benchmarks running on localhost
also benefit quite a bit from this patch (roughly 10% on netperf RR for UDP
and TCP depending on the machine). Hackbench also seens small improvements
(6-11% depending on machine and thread count). The facebook schbench was also
tested but in most cases showed little or no different to wakeup latencies.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180130104555.4125-5-mgorman@techsingularity.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
wake_affine_idle() prefers to move a task to the current CPU if the
wakeup is due to an interrupt. The expectation is that the interrupt
data is cache hot and relevant to the waking task as well as avoiding
a search. However, there is no way to determine if there was cache hot
data on the previous CPU that may exceed the interrupt data. Furthermore,
round-robin delivery of interrupts can migrate tasks around a socket where
each CPU is under-utilised. This can interact badly with cpufreq which
makes decisions based on per-cpu data. It has been observed on machines
with HWP that p-states are not boosted to their maximum levels even though
the workload is latency and throughput sensitive.
This patch uses the previous CPU for the task if it's idle and cache-affine
with the current CPU even if the current CPU is idle due to the wakup
being related to the interrupt. This reduces migrations at the cost of
the interrupt data not being cache hot when the task wakes.
A variety of workloads were tested on various machines and no adverse
impact was noticed that was outside noise. dbench on ext4 on UMA showed
roughly 10% reduction in the number of CPU migrations and it is a case
where interrupts are frequent for IO competions. In most cases, the
difference in performance is quite small but variability is often
reduced. For example, this is the result for pgbench running on a UMA
machine with different numbers of clients.
4.15.0-rc9 4.15.0-rc9
baseline waprev-v1
Hmean 1 22096.28 ( 0.00%) 22734.86 ( 2.89%)
Hmean 4 74633.42 ( 0.00%) 75496.77 ( 1.16%)
Hmean 7 115017.50 ( 0.00%) 113030.81 ( -1.73%)
Hmean 12 126209.63 ( 0.00%) 126613.40 ( 0.32%)
Hmean 16 131886.91 ( 0.00%) 130844.35 ( -0.79%)
Stddev 1 636.38 ( 0.00%) 417.11 ( 34.46%)
Stddev 4 614.64 ( 0.00%) 583.24 ( 5.11%)
Stddev 7 542.46 ( 0.00%) 435.45 ( 19.73%)
Stddev 12 173.93 ( 0.00%) 171.50 ( 1.40%)
Stddev 16 671.42 ( 0.00%) 680.30 ( -1.32%)
CoeffVar 1 2.88 ( 0.00%) 1.83 ( 36.26%)
Note that the different in performance is marginal but for low utilisation,
there is less variability.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180130104555.4125-4-mgorman@techsingularity.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When issuing an IPI RT push, where an IPI is sent to each CPU that has more
than one RT task scheduled on it, it references the root domain's rto_mask,
that contains all the CPUs within the root domain that has more than one RT
task in the runable state. The problem is, after the IPIs are initiated, the
rq->lock is released. This means that the root domain that is associated to
the run queue could be freed while the IPIs are going around.
Add a sched_get_rd() and a sched_put_rd() that will increment and decrement
the root domain's ref count respectively. This way when initiating the IPIs,
the scheduler will up the root domain's ref count before releasing the
rq->lock, ensuring that the root domain does not go away until the IPI round
is complete.
Reported-by: Pavan Kondeti <pkondeti@codeaurora.org>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Fixes: 4bdced5c9a ("sched/rt: Simplify the IPI based RT balancing logic")
Link: http://lkml.kernel.org/r/CAEU1=PkiHO35Dzna8EQqNSKW1fr1y1zRQ5y66X117MG06sQtNA@mail.gmail.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When the rto_push_irq_work_func() is called, it looks at the RT overloaded
bitmask in the root domain via the runqueue (rq->rd). The problem is that
during CPU up and down, nothing here stops rq->rd from changing between
taking the rq->rd->rto_lock and releasing it. That means the lock that is
released is not the same lock that was taken.
Instead of using this_rq()->rd to get the root domain, as the irq work is
part of the root domain, we can simply get the root domain from the irq work
that is passed to the routine:
container_of(work, struct root_domain, rto_push_work)
This keeps the root domain consistent.
Reported-by: Pavan Kondeti <pkondeti@codeaurora.org>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Fixes: 4bdced5c9a ("sched/rt: Simplify the IPI based RT balancing logic")
Link: http://lkml.kernel.org/r/CAEU1=PkiHO35Dzna8EQqNSKW1fr1y1zRQ5y66X117MG06sQtNA@mail.gmail.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Allow expedited membarrier to be used for data shared between processes
through shared memory.
Processes wishing to receive the membarriers register with
MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED. Those which want to issue
membarrier invoke MEMBARRIER_CMD_GLOBAL_EXPEDITED.
This allows extremely simple kernel-level implementation: we have almost
everything we need with the PRIVATE_EXPEDITED barrier code. All we need
to do is to add a flag in the mm_struct that will be used to check
whether we need to send the IPI to the current thread of each CPU.
There is a slight downside to this approach compared to targeting
specific shared memory users: when performing a membarrier operation,
all registered "global" receivers will get the barrier, even if they
don't share a memory mapping with the sender issuing
MEMBARRIER_CMD_GLOBAL_EXPEDITED.
This registration approach seems to fit the requirement of not
disturbing processes that really deeply care about real-time: they
simply should not register with MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED.
In order to align the membarrier command names, the "MEMBARRIER_CMD_SHARED"
command is renamed to "MEMBARRIER_CMD_GLOBAL", keeping an alias of
MEMBARRIER_CMD_SHARED to MEMBARRIER_CMD_GLOBAL for UAPI header backward
compatibility.
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrea Parri <parri.andrea@gmail.com>
Cc: Andrew Hunter <ahh@google.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Avi Kivity <avi@scylladb.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Dave Watson <davejwatson@fb.com>
Cc: David Sehr <sehr@google.com>
Cc: Greg Hackmann <ghackmann@google.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Maged Michael <maged.michael@gmail.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Russell King <linux@armlinux.org.uk>
Cc: Will Deacon <will.deacon@arm.com>
Cc: linux-api@vger.kernel.org
Link: http://lkml.kernel.org/r/20180129202020.8515-5-mathieu.desnoyers@efficios.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull printk updates from Petr Mladek:
- Add a console_msg_format command line option:
The value "default" keeps the old "[time stamp] text\n" format. The
value "syslog" allows to see the syslog-like "<log
level>[timestamp] text" format.
This feature was requested by people doing regression tests, for
example, 0day robot. They want to have both filtered and full logs
at hands.
- Reduce the risk of softlockup:
Pass the console owner in a busy loop.
This is a new approach to the old problem. It was first proposed by
Steven Rostedt on Kernel Summit 2017. It marks a context in which
the console_lock owner calls console drivers and could not sleep.
On the other side, printk() callers could detect this state and use
a busy wait instead of a simple console_trylock(). Finally, the
console_lock owner checks if there is a busy waiter at the end of
the special context and eventually passes the console_lock to the
waiter.
The hand-off works surprisingly well and helps in many situations.
Well, there is still a possibility of the softlockup, for example,
when the flood of messages stops and the last owner still has too
much to flush.
There is increasing number of people having problems with
printk-related softlockups. We might eventually need to get better
solution. Anyway, this looks like a good start and promising
direction.
- Do not allow to schedule in console_unlock() called from printk():
This reverts an older controversial commit. The reschedule helped
to avoid softlockups. But it also slowed down the console output.
This patch is obsoleted by the new console waiter logic described
above. In fact, the reschedule made the hand-off less effective.
- Deprecate "%pf" and "%pF" format specifier:
It was needed on ia64, ppc64 and parisc64 to dereference function
descriptors and show the real function address. It is done
transparently by "%ps" and "pS" format specifier now.
Sergey Senozhatsky found that all the function descriptors were in
a special elf section and could be easily detected.
- Remove printk_symbol() API:
It has been obsoleted by "%pS" format specifier, and this change
helped to remove few continuous lines and a less intuitive old API.
- Remove redundant memsets:
Sergey removed unnecessary memset when processing printk.devkmsg
command line option.
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/pmladek/printk: (27 commits)
printk: drop redundant devkmsg_log_str memsets
printk: Never set console_may_schedule in console_trylock()
printk: Hide console waiter logic into helpers
printk: Add console owner and waiter logic to load balance console writes
kallsyms: remove print_symbol() function
checkpatch: add pF/pf deprecation warning
symbol lookup: introduce dereference_symbol_descriptor()
parisc64: Add .opd based function descriptor dereference
powerpc64: Add .opd based function descriptor dereference
ia64: Add .opd based function descriptor dereference
sections: split dereference_function_descriptor()
openrisc: Fix conflicting types for _exext and _stext
lib: do not use print_symbol()
irq debug: do not use print_symbol()
sysfs: do not use print_symbol()
drivers: do not use print_symbol()
x86: do not use print_symbol()
unicore32: do not use print_symbol()
sh: do not use print_symbol()
mn10300: do not use print_symbol()
...
Pull scheduler updates from Ingo Molnar:
"The main changes in this cycle were:
- Implement frequency/CPU invariance and OPP selection for
SCHED_DEADLINE (Juri Lelli)
- Tweak the task migration logic for better multi-tasking
workload scalability (Mel Gorman)
- Misc cleanups, fixes and improvements"
* 'sched-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/deadline: Make bandwidth enforcement scale-invariant
sched/cpufreq: Move arch_scale_{freq,cpu}_capacity() outside of #ifdef CONFIG_SMP
sched/cpufreq: Remove arch_scale_freq_capacity()'s 'sd' parameter
sched/cpufreq: Always consider all CPUs when deciding next freq
sched/cpufreq: Split utilization signals
sched/cpufreq: Change the worker kthread to SCHED_DEADLINE
sched/deadline: Move CPU frequency selection triggering points
sched/cpufreq: Use the DEADLINE utilization signal
sched/deadline: Implement "runtime overrun signal" support
sched/fair: Only immediately migrate tasks due to interrupts if prev and target CPUs share cache
sched/fair: Correct obsolete comment about cpufreq_update_util()
sched/fair: Remove impossible condition from find_idlest_group_cpu()
sched/cpufreq: Don't pass flags to sugov_set_iowait_boost()
sched/cpufreq: Initialize sg_cpu->flags to 0
sched/fair: Consider RT/IRQ pressure in capacity_spare_wake()
sched/fair: Use 'unsigned long' for utilization, consistently
sched/core: Rework and clarify prepare_lock_switch()
sched/fair: Remove unused 'curr' parameter from wakeup_gran
sched/headers: Constify object_is_on_stack()
Pull RCU updates from Ingo Molnar:
"The main RCU changes in this cycle were:
- Updates to use cond_resched() instead of cond_resched_rcu_qs()
where feasible (currently everywhere except in kernel/rcu and in
kernel/torture.c). Also a couple of fixes to avoid sending IPIs to
offline CPUs.
- Updates to simplify RCU's dyntick-idle handling.
- Updates to remove almost all uses of smp_read_barrier_depends() and
read_barrier_depends().
- Torture-test updates.
- Miscellaneous fixes"
* 'core-rcu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (72 commits)
torture: Save a line in stutter_wait(): while -> for
torture: Eliminate torture_runnable and perf_runnable
torture: Make stutter less vulnerable to compilers and races
locking/locktorture: Fix num reader/writer corner cases
locking/locktorture: Fix rwsem reader_delay
torture: Place all torture-test modules in one MAINTAINERS group
rcutorture/kvm-build.sh: Skip build directory check
rcutorture: Simplify functions.sh include path
rcutorture: Simplify logging
rcutorture/kvm-recheck-*: Improve result directory readability check
rcutorture/kvm.sh: Support execution from any directory
rcutorture/kvm.sh: Use consistent help text for --qemu-args
rcutorture/kvm.sh: Remove unused variable, `alldone`
rcutorture: Remove unused script, config2frag.sh
rcutorture/configinit: Fix build directory error message
rcutorture: Preempt RCU-preempt readers more vigorously
torture: Reduce #ifdefs for preempt_schedule()
rcu: Remove have_rcu_nocb_mask from tree_plugin.h
rcu: Add comment giving debug strategy for double call_rcu()
tracing, rcu: Hide trace event rcu_nocb_wake when not used
...
Before commit:
e33a9bba85 ("sched/core: move IO scheduling accounting from io_schedule_timeout() into scheduler")
delayacct_blkio_end() was called after context-switching into the task which
completed I/O.
This resulted in double counting: the task would account a delay both waiting
for I/O and for time spent in the runqueue.
With e33a9bba85, delayacct_blkio_end() is called by try_to_wake_up().
In ttwu, we have not yet context-switched. This is more correct, in that
the delay accounting ends when the I/O is complete.
But delayacct_blkio_end() relies on 'get_current()', and we have not yet
context-switched into the task whose I/O completed. This results in the
wrong task having its delay accounting statistics updated.
Instead of doing that, pass the task_struct being woken to delayacct_blkio_end(),
so that it can update the statistics of the correct task.
Signed-off-by: Josh Snyder <joshs@netflix.com>
Acked-by: Tejun Heo <tj@kernel.org>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: <stable@vger.kernel.org>
Cc: Brendan Gregg <bgregg@netflix.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-block@vger.kernel.org
Fixes: e33a9bba85 ("sched/core: move IO scheduling accounting from io_schedule_timeout() into scheduler")
Link: http://lkml.kernel.org/r/1513613712-571-1-git-send-email-joshs@netflix.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull scheduler fixes from Ingo Molnar:
"A Kconfig fix, a build fix and a membarrier bug fix"
* 'sched-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
membarrier: Disable preemption when calling smp_call_function_many()
sched/isolation: Make CONFIG_CPU_ISOLATION=y depend on SMP or COMPILE_TEST
ia64, sched/cputime: Fix build error if CONFIG_VIRT_CPU_ACCOUNTING_NATIVE=y
