Now that we've removed the rq->lock requirement from the first part of
ttwu() and can compute placement without holding any rq->lock, ensure
we execute the second half of ttwu() on the actual cpu we want the
task to run on.
This avoids having to take rq->lock and doing the task enqueue
remotely, saving lots on cacheline transfers.
As measured using: http://oss.oracle.com/~mason/sembench.c
$ for i in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor ; do echo performance > $i; done
$ echo 4096 32000 64 128 > /proc/sys/kernel/sem
$ ./sembench -t 2048 -w 1900 -o 0
unpatched: run time 30 seconds 647278 worker burns per second
patched: run time 30 seconds 816715 worker burns per second
Reviewed-by: Frank Rowand <frank.rowand@am.sony.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110405152729.515897185@chello.nl
Currently ttwu() does two rq->lock acquisitions, once on the task's
old rq, holding it over the p->state fiddling and load-balance pass.
Then it drops the old rq->lock to acquire the new rq->lock.
By having serialized ttwu(), p->sched_class, p->cpus_allowed with
p->pi_lock, we can now drop the whole first rq->lock acquisition.
The p->pi_lock serializing concurrent ttwu() calls protects p->state,
which we will set to TASK_WAKING to bridge possible p->pi_lock to
rq->lock gaps and serialize set_task_cpu() calls against
task_rq_lock().
The p->pi_lock serialization of p->sched_class allows us to call
scheduling class methods without holding the rq->lock, and the
serialization of p->cpus_allowed allows us to do the load-balancing
bits without races.
Reviewed-by: Frank Rowand <frank.rowand@am.sony.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110405152729.354401150@chello.nl
Merge reason: Pick up this upstream commit:
6631e635c6: block: don't flush plugged IO on forced preemtion scheduling
As it modifies the scheduler and we'll queue up dependent patches.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We really only want to unplug the pending IO when the process actually
goes to sleep. So move the test for flushing the plug up to the place
where we actually deactivate the task - where we have properly checked
for preemption and for the process really sleeping.
Acked-by: Jens Axboe <jaxboe@fusionio.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Make XEN_SAVE_RESTORE select HIBERNATE_CALLBACKS.
Remove XEN_SAVE_RESTORE dependency from PM_SLEEP.
Signed-off-by: Shriram Rajagopalan <rshriram@cs.ubc.ca>
Acked-by: Ian Campbell <ian.campbell@citrix.com>
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Xen save/restore is going to use hibernate device callbacks for
quiescing devices and putting them back to normal operations and it
would need to select CONFIG_HIBERNATION for this purpose. However,
that also would cause the hibernate interfaces for user space to be
enabled, which might confuse user space, because the Xen kernels
don't support hibernation. Moreover, it would be wasteful, as it
would make the Xen kernels include a substantial amount of code that
they would never use.
To address this issue introduce new power management Kconfig option
CONFIG_HIBERNATE_CALLBACKS, such that it will only select the code
that is necessary for the hibernate device callbacks to work and make
CONFIG_HIBERNATION select it. Then, Xen save/restore will be able to
select CONFIG_HIBERNATE_CALLBACKS without dragging the entire
hibernate code along with it.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Tested-by: Shriram Rajagopalan <rshriram@cs.ubc.ca>
Instead of relying on static allocations for the sched_domain and
sched_group trees, dynamically allocate and RCU free them.
Allocating this dynamically also allows for some build_sched_groups()
simplification since we can now (like with other simplifications) rely
on the sched_domain tree instead of hard-coded knowledge.
One tricky to note is that detach_destroy_domains() needs to hold
rcu_read_lock() over the entire tear-down, per-cpu is not sufficient
since that can lead to partial sched_group existance (could possibly
be solved by doing the tear-down backwards but this is much more
robust).
A concequence of the above is that we can no longer print the
sched_domain debug stuff from cpu_attach_domain() since that might now
run with preemption disabled (due to classic RCU etc.) and
sched_domain_debug() does some GFP_KERNEL allocations.
Another thing to note is that we now fully rely on normal RCU and not
RCU-sched, this is because with the new and exiting RCU flavours we
grew over the years BH doesn't necessarily hold off RCU-sched grace
periods (-rt is known to break this). This would in fact already cause
us grief since we do sched_domain/sched_group iterations from softirq
context.
This patch is somewhat larger than I would like it to be, but I didn't
find any means of shrinking/splitting this.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/20110407122942.245307941@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Again, instead of relying on knowing the possible domains and their
order, simply rely on the sched_domain tree and whatever domains are
present in there to initialize the sched_group cpu_power.
Note: we need to iterate the CPU mask backwards because of the
cpumask_first() condition for iterating up the tree. By iterating the
mask backwards we ensure all groups of a domain are set-up before
starting on the parent groups that rely on its children to be
completely done.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/20110407122942.187335414@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The NODE sched_domain is 'special' in that it allocates sched_groups
per CPU, instead of sharing the sched_groups between all CPUs.
While this might have some benefits on large NUMA and avoid remote
memory accesses when iterating the sched_groups, this does break
current code that assumes sched_groups are shared between all
sched_domains (since the dynamic cpu_power patches).
So refactor the NODE groups to behave like all other groups.
(The ALLNODES domain again shared its groups across the CPUs for some
reason).
If someone does measure a performance decrease due to this change we
need to revisit this and come up with another way to have both dynamic
cpu_power and NUMA work nice together.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/20110407122941.978111700@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The scheduler load balancer has specific code to deal with cases of
unbalanced system due to lots of unmovable tasks (for example because of
hard CPU affinity). In those situation, it excludes the busiest CPU that
has pinned tasks for load balance consideration such that it can perform
second 2nd load balance pass on the rest of the system.
This all works as designed if there is only one cgroup in the system.
However, when we have multiple cgroups, this logic has false positives and
triggers multiple load balance passes despite there are actually no pinned
tasks at all.
The reason it has false positives is that the all pinned logic is deep in
the lowest function of can_migrate_task() and is too low level:
load_balance_fair() iterates each task group and calls balance_tasks() to
migrate target load. Along the way, balance_tasks() will also set a
all_pinned variable. Given that task-groups are iterated, this all_pinned
variable is essentially the status of last group in the scanning process.
Task group can have number of reasons that no load being migrated, none
due to cpu affinity. However, this status bit is being propagated back up
to the higher level load_balance(), which incorrectly think that no tasks
were moved. It kick off the all pinned logic and start multiple passes
attempt to move load onto puller CPU.
To fix this, move the all_pinned aggregation up at the iterator level.
This ensures that the status is aggregated over all task-groups, not just
last one in the list.
Signed-off-by: Ken Chen <kenchen@google.com>
Cc: stable@kernel.org
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/BANLkTi=ernzNawaR5tJZEsV_QVnfxqXmsQ@mail.gmail.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
