Heiko Carstens <heiko.carstens@de.ibm.com> wrote:
The boot sequence on s390 sometimes takes ages and we spend a very long
time (up to one or two minutes) in calibrate_migration_costs. The time
spent there differs from boot to boot. Also the calculated costs differ
a lot. I've seen differences by up to a factor of 15 (yes, factor not
percent). Also I doubt that making these measurements make much sense on
a completely virtualized architecture where you cannot tell how much cpu
time you will get anyway.
So introduce the CONFIG_DEFAULT_MIGRATION_COST method for an architecture
to set the scheduler migration costs. This turns off automatic detection
of migration costs. Makes sense on virtual platforms, where migration
costs are hard to measure accurately.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Revert commit d7102e95b7b9c00277562c29aad421d2d521c5f6:
[PATCH] sched: filter affine wakeups
Apparently caused more than 10% performance regression for aim7 benchmark.
The setup in use is 16-cpu HP rx8620, 64Gb of memory and 12 MSA1000s with 144
disks. Each disk is 72Gb with a single ext3 filesystem (courtesy of HP, who
supplied benchmark results).
The problem is, for aim7, the wake-up pattern is random, but it still needs
load balancing action in the wake-up path to achieve best performance. With
the above commit, lack of load balancing hurts that workload.
However, for workloads like database transaction processing, the requirement
is exactly opposite. In the wake up path, best performance is achieved with
absolutely zero load balancing. We simply wake up the process on the CPU that
it was previously run. Worst performance is obtained when we do load
balancing at wake up.
There isn't an easy way to auto detect the workload characteristics. Ingo's
earlier patch that detects idle CPU and decide whether to load balance or not
doesn't perform with aim7 either since all CPUs are busy (it causes even
bigger perf. regression).
Revert commit d7102e95b7, which causes more
than 10% performance regression with aim7.
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
I don't think the code is quite ready, which is why I asked for Peter's
additions to also be merged before I acked it (although it turned out that
it still isn't quite ready with his additions either).
Basically I have had similar observations to Suresh in that it does not
play nicely with the rest of the balancing infrastructure (and raised
similar concerns in my review).
The samples (group of 4) I got for "maximum recorded imbalance" on a 2x2
SMP+HT Xeon are as follows:
| Following boot | hackbench 20 | hackbench 40
-----------+----------------+---------------------+---------------------
2.6.16-rc2 | 30,37,100,112 | 5600,5530,6020,6090 | 6390,7090,8760,8470
+nosmpnice | 3, 2, 4, 2 | 28, 150, 294, 132 | 348, 348, 294, 347
Hackbench raw performance is down around 15% with smpnice (but that in
itself isn't a huge deal because it is just a benchmark). However, the
samples show that the imbalance passed into move_tasks is increased by
about a factor of 10-30. I think this would also go some way to explaining
latency blips turning up in the balancing code (though I haven't actually
measured that).
We'll probably have to revert this in the SUSE kernel.
Cc: "Siddha, Suresh B" <suresh.b.siddha@intel.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: Peter Williams <pwil3058@bigpond.net.au>
Cc: "Martin J. Bligh" <mbligh@aracnet.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
migration_cost prints after every CPU hotplug event. Make it print only
once at boot.
Signed-off-by: Chuck Ebbert <76306.1226@compuserve.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
percpu_data blindly allocates bootmem memory to store NR_CPUS instances of
cpudata, instead of allocating memory only for possible cpus.
As a preparation for changing that, we need to convert various 0 -> NR_CPUS
loops to use for_each_cpu().
(The above only applies to users of asm-generic/percpu.h. powerpc has gone it
alone and is presently only allocating memory for present CPUs, so it's
currently corrupting memory).
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: James Bottomley <James.Bottomley@steeleye.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: Jens Axboe <axboe@suse.de>
Cc: Anton Blanchard <anton@samba.org>
Acked-by: William Irwin <wli@holomorphy.com>
Cc: Andi Kleen <ak@muc.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Change sched_getaffinity() so that it returns a bitmap that indicates the
legally schedulable cpus that a task is allowed to run on.
Without this patch, if CONFIG_HOTPLUG_CPU is enabled, sched_getaffinity()
unconditionally returns (at least on IA64) a mask with NR_CPUS bits set.
This conveys no useful infornmation except for a kernel compile option.
This fixes a breakage we obseved running recent kernels. We have MPI jobs
that use sched_getaffinity() to determine where to place their threads.
Placing them on non-existant cpus is problematic :-)
Signed-off-by: Jack Steiner <steiner@sgi.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: Nathan Lynch <ntl@pobox.com>
Cc: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This reduces the amount of time the migration cost calculations cost
during bootup. Based on numbers by Tony Luck <tony.luck@intel.com>.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Currently, a negative policy argument passed into the
'sys_sched_setscheduler()' system call, will return with success. However,
the manpage for 'sys_sched_setscheduler' says:
EINVAL The scheduling policy is not one of the recognized policies, or the
parameter p does not make sense for the policy.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Remove the "inline" keyword from a bunch of big functions in the kernel with
the goal of shrinking it by 30kb to 40kb
Signed-off-by: Arjan van de Ven <arjan@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Jeff Garzik <jgarzik@pobox.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add a new SCHED_BATCH (3) scheduling policy: such tasks are presumed
CPU-intensive, and will acquire a constant +5 priority level penalty. Such
policy is nice for workloads that are non-interactive, but which do not
want to give up their nice levels. The policy is also useful for workloads
that want a deterministic scheduling policy without interactivity causing
extra preemptions (between that workload's tasks).
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Cc: Michael Kerrisk <mtk-manpages@gmx.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
)
From: Nick Piggin <nickpiggin@yahoo.com.au>
Track the last waker CPU, and only consider wakeup-balancing if there's a
match between current waker CPU and the previous waker CPU. This ensures
that there is some correlation between two subsequent wakeup events before
we move the task. Should help random-wakeup workloads on large SMP
systems, by reducing the migration attempts by a factor of nr_cpus.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
)
From: Ingo Molnar <mingo@elte.hu>
This is the latest version of the scheduler cache-hot-auto-tune patch.
The first problem was that detection time scaled with O(N^2), which is
unacceptable on larger SMP and NUMA systems. To solve this:
- I've added a 'domain distance' function, which is used to cache
measurement results. Each distance is only measured once. This means
that e.g. on NUMA distances of 0, 1 and 2 might be measured, on HT
distances 0 and 1, and on SMP distance 0 is measured. The code walks
the domain tree to determine the distance, so it automatically follows
whatever hierarchy an architecture sets up. This cuts down on the boot
time significantly and removes the O(N^2) limit. The only assumption
is that migration costs can be expressed as a function of domain
distance - this covers the overwhelming majority of existing systems,
and is a good guess even for more assymetric systems.
[ People hacking systems that have assymetries that break this
assumption (e.g. different CPU speeds) should experiment a bit with
the cpu_distance() function. Adding a ->migration_distance factor to
the domain structure would be one possible solution - but lets first
see the problem systems, if they exist at all. Lets not overdesign. ]
Another problem was that only a single cache-size was used for measuring
the cost of migration, and most architectures didnt set that variable
up. Furthermore, a single cache-size does not fit NUMA hierarchies with
L3 caches and does not fit HT setups, where different CPUs will often
have different 'effective cache sizes'. To solve this problem:
- Instead of relying on a single cache-size provided by the platform and
sticking to it, the code now auto-detects the 'effective migration
cost' between two measured CPUs, via iterating through a wide range of
cachesizes. The code searches for the maximum migration cost, which
occurs when the working set of the test-workload falls just below the
'effective cache size'. I.e. real-life optimized search is done for
the maximum migration cost, between two real CPUs.
This, amongst other things, has the positive effect hat if e.g. two
CPUs share a L2/L3 cache, a different (and accurate) migration cost
will be found than between two CPUs on the same system that dont share
any caches.
(The reliable measurement of migration costs is tricky - see the source
for details.)
Furthermore i've added various boot-time options to override/tune
migration behavior.
Firstly, there's a blanket override for autodetection:
migration_cost=1000,2000,3000
will override the depth 0/1/2 values with 1msec/2msec/3msec values.
Secondly, there's a global factor that can be used to increase (or
decrease) the autodetected values:
migration_factor=120
will increase the autodetected values by 20%. This option is useful to
tune things in a workload-dependent way - e.g. if a workload is
cache-insensitive then CPU utilization can be maximized by specifying
migration_factor=0.
I've tested the autodetection code quite extensively on x86, on 3
P3/Xeon/2MB, and the autodetected values look pretty good:
Dual Celeron (128K L2 cache):
---------------------
migration cost matrix (max_cache_size: 131072, cpu: 467 MHz):
---------------------
[00] [01]
[00]: - 1.7(1)
[01]: 1.7(1) -
---------------------
cacheflush times [2]: 0.0 (0) 1.7 (1784008)
---------------------
Here the slow memory subsystem dominates system performance, and even
though caches are small, the migration cost is 1.7 msecs.
Dual HT P4 (512K L2 cache):
---------------------
migration cost matrix (max_cache_size: 524288, cpu: 2379 MHz):
---------------------
[00] [01] [02] [03]
[00]: - 0.4(1) 0.0(0) 0.4(1)
[01]: 0.4(1) - 0.4(1) 0.0(0)
[02]: 0.0(0) 0.4(1) - 0.4(1)
[03]: 0.4(1) 0.0(0) 0.4(1) -
---------------------
cacheflush times [2]: 0.0 (33900) 0.4 (448514)
---------------------
Here it can be seen that there is no migration cost between two HT
siblings (CPU#0/2 and CPU#1/3 are separate physical CPUs). A fast memory
system makes inter-physical-CPU migration pretty cheap: 0.4 msecs.
8-way P3/Xeon [2MB L2 cache]:
---------------------
migration cost matrix (max_cache_size: 2097152, cpu: 700 MHz):
---------------------
[00] [01] [02] [03] [04] [05] [06] [07]
[00]: - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[01]: 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[02]: 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[03]: 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[04]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1)
[05]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1)
[06]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1)
[07]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) -
---------------------
cacheflush times [2]: 0.0 (0) 19.2 (19281756)
---------------------
This one has huge caches and a relatively slow memory subsystem - so the
migration cost is 19 msecs.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Ashok Raj <ashok.raj@intel.com>
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Cc: <wilder@us.ibm.com>
Signed-off-by: John Hawkes <hawkes@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
They are referred to often so avoid potential false sharing for them.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
- Move capable() from sched.h to capability.h;
- Use <linux/capability.h> where capable() is used
(in include/, block/, ipc/, kernel/, a few drivers/,
mm/, security/, & sound/;
many more drivers/ to go)
Signed-off-by: Randy Dunlap <rdunlap@xenotime.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
more mutex debugging: check for held locks during memory freeing,
task exit, enable sysrq printouts, etc.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Arjan van de Ven <arjan@infradead.org>
RCU tasklist_lock and RCU signal handling: send signals RCU-read-locked
instead of tasklist_lock read-locked. This is a scalability improvement on
SMP and a preemption-latency improvement under PREEMPT_RCU.
Signed-off-by: Paul E. McKenney <paulmck@us.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Acked-by: William Irwin <wli@holomorphy.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Oleg Nesterov <oleg@tv-sign.ru>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
encapsulates the rest of arch-dependent operations with thread_info access.
Two new helpers - setup_thread_stack() and end_of_stack(). For normal case
the former consists of copying thread_info of parent to new thread_info and
the latter returns pointer immediately past the end of thread_info.
Signed-off-by: Al Viro <viro@parcelfarce.linux.theplanet.co.uk>
Signed-off-by: Roman Zippel <zippel@linux-m68k.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
new helper - task_thread_info(task). On platforms that have thread_info
allocated separately (i.e. in default case) it simply returns
task->thread_info. m68k wants (and for good reasons) to embed its thread_info
into task_struct. So it will (in later patch) have task_thread_info() of its
own. For now we just add a macro for generic case and convert existing
instances of its body in core kernel to uses of new macro. Obviously safe -
all normal architectures get the same preprocessor output they used to get.
Signed-off-by: Al Viro <viro@parcelfarce.linux.theplanet.co.uk>
Signed-off-by: Roman Zippel <zippel@linux-m68k.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
recalc_task_prio() is called from activate_task() to calculate dynamic
priority and interactive credit for the activating task. For real-time
scheduling process, all that dynamic calculation is thrown away at the end
because rt priority is fixed. Patch to optimize recalc_task_prio() away
for rt processes.
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: Nick Piggin <piggin@cyberone.com.au>
Cc: Con Kolivas <kernel@kolivas.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Make some changes to the NEED_RESCHED and POLLING_NRFLAG to reduce
confusion, and make their semantics rigid. Improves efficiency of
resched_task and some cpu_idle routines.
* In resched_task:
- TIF_NEED_RESCHED is only cleared with the task's runqueue lock held,
and as we hold it during resched_task, then there is no need for an
atomic test and set there. The only other time this should be set is
when the task's quantum expires, in the timer interrupt - this is
protected against because the rq lock is irq-safe.
- If TIF_NEED_RESCHED is set, then we don't need to do anything. It
won't get unset until the task get's schedule()d off.
- If we are running on the same CPU as the task we resched, then set
TIF_NEED_RESCHED and no further action is required.
- If we are running on another CPU, and TIF_POLLING_NRFLAG is *not* set
after TIF_NEED_RESCHED has been set, then we need to send an IPI.
Using these rules, we are able to remove the test and set operation in
resched_task, and make clear the previously vague semantics of
POLLING_NRFLAG.
* In idle routines:
- Enter cpu_idle with preempt disabled. When the need_resched() condition
becomes true, explicitly call schedule(). This makes things a bit clearer
(IMO), but haven't updated all architectures yet.
- Many do a test and clear of TIF_NEED_RESCHED for some reason. According
to the resched_task rules, this isn't needed (and actually breaks the
assumption that TIF_NEED_RESCHED is only cleared with the runqueue lock
held). So remove that. Generally one less locked memory op when switching
to the idle thread.
- Many idle routines clear TIF_POLLING_NRFLAG, and only set it in the inner
most polling idle loops. The above resched_task semantics allow it to be
set until before the last time need_resched() is checked before going into
a halt requiring interrupt wakeup.
Many idle routines simply never enter such a halt, and so POLLING_NRFLAG
can be always left set, completely eliminating resched IPIs when rescheduling
the idle task.
POLLING_NRFLAG width can be increased, to reduce the chance of resched IPIs.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Con Kolivas <kernel@kolivas.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The intermittent scheduling of the migration thread at ultra high priority
makes the smp nice handling see that runqueue as being heavily loaded. The
migration thread itself actually handles the balancing so its influence on
priority balancing should be ignored.
Signed-off-by: Con Kolivas <kernel@kolivas.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The priority biasing was off by mutliplying the total load by the total
priority bias and this ruins the ratio of loads between runqueues. This
patch should correct the ratios of loads between runqueues to be proportional
to overall load. -2nd attempt.
From: Dave Kleikamp <shaggy@austin.ibm.com>
This patch fixes a divide-by-zero error that I hit on a two-way i386
machine. rq->nr_running is tested to be non-zero, but may change by the
time it is used in the division. Saving the value to a local variable
ensures that the same value that is checked is used in the division.
Signed-off-by: Con Kolivas <kernel@kolivas.org>
Signed-off-by: Dave Kleikamp <shaggy@austin.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
To intensify the 'nice' support across physical cpus on SMP we can bias the
loads on idle rebalancing. To prevent idle rebalance from trying to pull tasks
from queues that appear heavily loaded we only bias the load if there is more
than one task running.
Add some minor micro-optimisations and have only one return from __source_load
and __target_load functions.
Fix the fact that target_load was not biased by priority when type == 0.
Signed-off-by: Con Kolivas <kernel@kolivas.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Real time tasks' effect on prio_bias should be based on their real time
priority level instead of their static_prio which is based on nice.
Signed-off-by: Con Kolivas <kernel@kolivas.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
prio_bias should only be adjusted in set_user_nice if p is actually currently
queued.
Signed-off-by: Con Kolivas <kernel@kolivas.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch implements 'nice' support across physical cpus on SMP.
It introduces an extra runqueue variable prio_bias which is the sum of the
(inverted) static priorities of all the tasks on the runqueue.
This is then used to bias busy rebalancing between runqueues to obtain good
distribution of tasks of different nice values. By biasing the balancing only
during busy rebalancing we can avoid having any significant loss of throughput
by not affecting the carefully tuned idle balancing already in place. If all
tasks are running at the same nice level this code should also have minimal
effect. The code is optimised out in the !CONFIG_SMP case.
Signed-off-by: Con Kolivas <kernel@kolivas.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
I didn't find any possible modular usage in the kernel.
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Replace smp_processor_id() with any_online_cpu(cpu_online_map) in order to
avoid lots of "BUG: using smp_processor_id() in preemptible [00000001]
code:..." messages in case taking a cpu online fails.
All the traces start at the last notifier_call_chain(...) in kernel/cpu.c.
Since we hold the cpu_control semaphore it shouldn't be any problem to access
cpu_online_map.
The reason why cpu_up failed is simply that the cpu that was supposed to be
taken online wasn't even there. That is because on s390 we never know when a
new cpu comes and therefore cpu_possible_map consists of only ones and doesn't
reflect reality.
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Do not transfer remaining time slice to another cpu on process exit.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Simplify the UP (1 CPU) implementatin of set_cpus_allowed.
The one CPU is hardcoded to be cpu 0 - so just test for that bit, and avoid
having to pick up the cpu_online_map.
Also, unexport cpu_online_map: it was only needed for set_cpus_allowed().
Signed-off-by: Paul Jackson <pj@sgi.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
update_mem_hiwater has attracted various criticisms, in particular from those
concerned with mm scalability. Originally it was called whenever rss or
total_vm got raised. Then many of those callsites were replaced by a timer
tick call from account_system_time. Now Frank van Maarseveen reports that to
be found inadequate. How about this? Works for Frank.
Replace update_mem_hiwater, a poor combination of two unrelated ops, by macros
update_hiwater_rss and update_hiwater_vm. Don't attempt to keep
mm->hiwater_rss up to date at timer tick, nor every time we raise rss (usually
by 1): those are hot paths. Do the opposite, update only when about to lower
rss (usually by many), or just before final accounting in do_exit. Handle
mm->hiwater_vm in the same way, though it's much less of an issue. Demand
that whoever collects these hiwater statistics do the work of taking the
maximum with rss or total_vm.
And there has been no collector of these hiwater statistics in the tree. The
new convention needs an example, so match Frank's usage by adding a VmPeak
line above VmSize to /proc/<pid>/status, and also a VmHWM line above VmRSS
(High-Water-Mark or High-Water-Memory).
There was a particular anomaly during mremap move, that hiwater_vm might be
captured too high. A fleeting such anomaly remains, but it's quickly
corrected now, whereas before it would stick.
What locking? None: if the app is racy then these statistics will be racy,
it's not worth any overhead to make them exact. But whenever it suits,
hiwater_vm is updated under exclusive mmap_sem, and hiwater_rss under
page_table_lock (for now) or with preemption disabled (later on): without
going to any trouble, minimize the time between reading current values and
updating, to minimize those occasions when a racing thread bumps a count up
and back down in between.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
With CONFIG_SMP=n:
*** Warning: "cpu_online_map" [drivers/firmware/dcdbas.ko] undefined!
due to set_cpus_allowed().
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
fix up the runqueue lock owner only if we truly did a context-switch
with the runqueue lock held. Impacts ia64, mips, sparc64 and arm.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
..and only enable them for ia64. The functions are only valid
when the whole system has been totally stopped and no scheduler
activity is ongoing on any CPU, and interrupts are globally
disabled.
In other words, they aren't useful for anything else. So make
sure that nobody can use them by mistake.
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Scheduler hooks to see/change which process is deemed to be on a cpu.
Signed-off-by: Keith Owens <kaos@sgi.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Don't pull tasks from a group if that would cause the group's total load to
drop below its total cpu_power (ie. cause the group to start going idle).
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Jack Steiner brought this issue at my OLS talk.
Take a scenario where two tasks are pinned to two HT threads in a physical
package. Idle packages in the system will keep kicking migration_thread on
the busy package with out any success.
We will run into similar scenarios in the presence of CMP/NUMA.
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
In sys_sched_yield(), we cache current->array in the "array" variable, thus
there's no need to dereference "current" again later.
Signed-Off-By: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
If an idle sibling of an HT queue encounters a busy sibling, then make
higher level load balancing of the non-idle variety.
Performance of multiprocessor HT systems with low numbers of tasks
(generally < number of virtual CPUs) can be significantly worse than the
exact same workloads when running in non-HT mode. The reason is largely
due to poor scheduling behaviour.
This patch improves the situation, making the performance gap far less
significant on one problematic test case (tbench).
Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
During periodic load balancing, don't hold this runqueue's lock while
scanning remote runqueues, which can take a non trivial amount of time
especially on very large systems.
Holding the runqueue lock will only help to stabilise ->nr_running, however
this doesn't do much to help because tasks being woken will simply get held
up on the runqueue lock, so ->nr_running would not provide a really
accurate picture of runqueue load in that case anyway.
What's more, ->nr_running (and possibly the cpu_load averages) of remote
runqueues won't be stable anyway, so load balancing is always an inexact
operation.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Similarly to the earlier change in load_balance, only lock the runqueue in
load_balance_newidle if the busiest queue found has a nr_running > 1. This
will reduce frequency of expensive remote runqueue lock aquisitions in the
schedule() path on some workloads.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
William Weston reported unusually high scheduling latencies on his x86 HT
box, on the -RT kernel. I managed to reproduce it on my HT box and the
latency tracer shows the incident in action:
_------=> CPU#
/ _-----=> irqs-off
| / _----=> need-resched
|| / _---=> hardirq/softirq
||| / _--=> preempt-depth
|||| /
||||| delay
cmd pid ||||| time | caller
\ / ||||| \ | /
du-2803 3Dnh2 0us : __trace_start_sched_wakeup (try_to_wake_up)
..............................................................
... we are running on CPU#3, PID 2778 gets woken to CPU#1: ...
..............................................................
du-2803 3Dnh2 0us : __trace_start_sched_wakeup <<...>-2778> (73 1)
du-2803 3Dnh2 0us : _raw_spin_unlock (try_to_wake_up)
................................................
... still on CPU#3, we send an IPI to CPU#1: ...
................................................
du-2803 3Dnh1 0us : resched_task (try_to_wake_up)
du-2803 3Dnh1 1us : smp_send_reschedule (try_to_wake_up)
du-2803 3Dnh1 1us : send_IPI_mask_bitmask (smp_send_reschedule)
du-2803 3Dnh1 2us : _raw_spin_unlock_irqrestore (try_to_wake_up)
...............................................
... 1 usec later, the IPI arrives on CPU#1: ...
...............................................
<idle>-0 1Dnh. 2us : smp_reschedule_interrupt (c0100c5a 0 0)
So far so good, this is the normal wakeup/preemption mechanism. But here
comes the scheduler anomaly on CPU#1:
<idle>-0 1Dnh. 2us : preempt_schedule_irq (need_resched)
<idle>-0 1Dnh. 2us : preempt_schedule_irq (need_resched)
<idle>-0 1Dnh. 3us : __schedule (preempt_schedule_irq)
<idle>-0 1Dnh. 3us : profile_hit (__schedule)
<idle>-0 1Dnh1 3us : sched_clock (__schedule)
<idle>-0 1Dnh1 4us : _raw_spin_lock_irq (__schedule)
<idle>-0 1Dnh1 4us : _raw_spin_lock_irqsave (__schedule)
<idle>-0 1Dnh2 5us : _raw_spin_unlock (__schedule)
<idle>-0 1Dnh1 5us : preempt_schedule (__schedule)
<idle>-0 1Dnh1 6us : _raw_spin_lock (__schedule)
<idle>-0 1Dnh2 6us : find_next_bit (__schedule)
<idle>-0 1Dnh2 6us : _raw_spin_lock (__schedule)
<idle>-0 1Dnh3 7us : find_next_bit (__schedule)
<idle>-0 1Dnh3 7us : find_next_bit (__schedule)
<idle>-0 1Dnh3 8us : _raw_spin_unlock (__schedule)
<idle>-0 1Dnh2 8us : preempt_schedule (__schedule)
<idle>-0 1Dnh2 8us : find_next_bit (__schedule)
<idle>-0 1Dnh2 9us : trace_stop_sched_switched (__schedule)
<idle>-0 1Dnh2 9us : _raw_spin_lock (trace_stop_sched_switched)
<idle>-0 1Dnh3 10us : trace_stop_sched_switched <<...>-2778> (73 8c)
<idle>-0 1Dnh3 10us : _raw_spin_unlock (trace_stop_sched_switched)
<idle>-0 1Dnh1 10us : _raw_spin_unlock (__schedule)
<idle>-0 1Dnh. 11us : local_irq_enable_noresched (preempt_schedule_irq)
<idle>-0 1Dnh. 11us < (0)
we didnt pick up pid 2778! It only gets scheduled much later:
<...>-2778 1Dnh2 412us : __switch_to (__schedule)
<...>-2778 1Dnh2 413us : __schedule <<idle>-0> (8c 73)
<...>-2778 1Dnh2 413us : _raw_spin_unlock (__schedule)
<...>-2778 1Dnh1 413us : trace_stop_sched_switched (__schedule)
<...>-2778 1Dnh1 414us : _raw_spin_lock (trace_stop_sched_switched)
<...>-2778 1Dnh2 414us : trace_stop_sched_switched <<...>-2778> (73 1)
<...>-2778 1Dnh2 414us : _raw_spin_unlock (trace_stop_sched_switched)
<...>-2778 1Dnh1 415us : trace_stop_sched_switched (__schedule)
the reason for this anomaly is the following code in dependent_sleeper():
/*
* If a user task with lower static priority than the
* running task on the SMT sibling is trying to schedule,
* delay it till there is proportionately less timeslice
* left of the sibling task to prevent a lower priority
* task from using an unfair proportion of the
* physical cpu's resources. -ck
*/
[...]
if (((smt_curr->time_slice * (100 - sd->per_cpu_gain) /
100) > task_timeslice(p)))
ret = 1;
Note that in contrast to the comment above, we dont actually do the check
based on static priority, we do the check based on timeslices. But
timeslices go up and down, and even highprio tasks can randomly have very
low timeslices (just before their next refill) and can thus be judged as
'lowprio' by the above piece of code. This condition is clearly buggy.
The correct test is to check for static_prio _and_ to check for the
preemption priority. Even on different static priority levels, a
higher-prio interactive task should not be delayed due to a
higher-static-prio CPU hog.
There is a symmetric bug in the 'kick SMT sibling' code of this function as
well, which can be solved in a similar way.
The patch below (against the current scheduler queue in -mm) fixes both
bugs. I have build and boot-tested this on x86 SMT, and nice +20 tasks
still get properly throttled - so the dependent-sleeper logic is still in
action.
btw., these bugs pessimised the SMT scheduler because the 'delay wakeup'
property was applied too liberally, so this fix is likely a throughput
improvement as well.
I separated out a smt_slice() function to make the code easier to read.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch implements a task state bit (TASK_NONINTERACTIVE), which can be
used by blocking points to mark the task's wait as "non-interactive". This
does not mean the task will be considered a CPU-hog - the wait will simply
not have an effect on the waiting task's priority - positive or negative
alike. Right now only pipe_wait() will make use of it, because it's a
common source of not-so-interactive waits (kernel compilation jobs, etc.).
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add relevant checks into find_idlest_group() and find_idlest_cpu() to make
them return only the groups that have allowed CPUs and allowed CPUs
respectively.
Signed-off-by: M.Baris Demiray <baris@labristeknoloji.com>
Signed-off-by: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The hyperthread aware nice handling currently puts to sleep any non real
time task when a real time task is running on its sibling cpu. This can
lead to prolonged starvation by having the non real time task pegged to the
cpu with load balancing not pulling that task away.
Currently we force lower priority hyperthread tasks to run a percentage of
time difference based on timeslice differences which is meaningless when
comparing real time tasks to SCHED_NORMAL tasks. We can allow non real
time tasks to run with real time tasks on the sibling up to per_cpu_gain%
if we use jiffies as a counter.
Cleanups and micro-optimisations to the relevant code section should make
it more understandable as well.
Signed-off-by: Con Kolivas <kernel@kolivas.org>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch (written by me and also containing many suggestions of Arjan van
de Ven) does a major cleanup of the spinlock code. It does the following
things:
- consolidates and enhances the spinlock/rwlock debugging code
- simplifies the asm/spinlock.h files
- encapsulates the raw spinlock type and moves generic spinlock
features (such as ->break_lock) into the generic code.
- cleans up the spinlock code hierarchy to get rid of the spaghetti.
Most notably there's now only a single variant of the debugging code,
located in lib/spinlock_debug.c. (previously we had one SMP debugging
variant per architecture, plus a separate generic one for UP builds)
Also, i've enhanced the rwlock debugging facility, it will now track
write-owners. There is new spinlock-owner/CPU-tracking on SMP builds too.
All locks have lockup detection now, which will work for both soft and hard
spin/rwlock lockups.
The arch-level include files now only contain the minimally necessary
subset of the spinlock code - all the rest that can be generalized now
lives in the generic headers:
include/asm-i386/spinlock_types.h | 16
include/asm-x86_64/spinlock_types.h | 16
I have also split up the various spinlock variants into separate files,
making it easier to see which does what. The new layout is:
SMP | UP
----------------------------|-----------------------------------
asm/spinlock_types_smp.h | linux/spinlock_types_up.h
linux/spinlock_types.h | linux/spinlock_types.h
asm/spinlock_smp.h | linux/spinlock_up.h
linux/spinlock_api_smp.h | linux/spinlock_api_up.h
linux/spinlock.h | linux/spinlock.h
/*
* here's the role of the various spinlock/rwlock related include files:
*
* on SMP builds:
*
* asm/spinlock_types.h: contains the raw_spinlock_t/raw_rwlock_t and the
* initializers
*
* linux/spinlock_types.h:
* defines the generic type and initializers
*
* asm/spinlock.h: contains the __raw_spin_*()/etc. lowlevel
* implementations, mostly inline assembly code
*
* (also included on UP-debug builds:)
*
* linux/spinlock_api_smp.h:
* contains the prototypes for the _spin_*() APIs.
*
* linux/spinlock.h: builds the final spin_*() APIs.
*
* on UP builds:
*
* linux/spinlock_type_up.h:
* contains the generic, simplified UP spinlock type.
* (which is an empty structure on non-debug builds)
*
* linux/spinlock_types.h:
* defines the generic type and initializers
*
* linux/spinlock_up.h:
* contains the __raw_spin_*()/etc. version of UP
* builds. (which are NOPs on non-debug, non-preempt
* builds)
*
* (included on UP-non-debug builds:)
*
* linux/spinlock_api_up.h:
* builds the _spin_*() APIs.
*
* linux/spinlock.h: builds the final spin_*() APIs.
*/
All SMP and UP architectures are converted by this patch.
arm, i386, ia64, ppc, ppc64, s390/s390x, x64 was build-tested via
crosscompilers. m32r, mips, sh, sparc, have not been tested yet, but should
be mostly fine.
From: Grant Grundler <grundler@parisc-linux.org>
Booted and lightly tested on a500-44 (64-bit, SMP kernel, dual CPU).
Builds 32-bit SMP kernel (not booted or tested). I did not try to build
non-SMP kernels. That should be trivial to fix up later if necessary.
I converted bit ops atomic_hash lock to raw_spinlock_t. Doing so avoids
some ugly nesting of linux/*.h and asm/*.h files. Those particular locks
are well tested and contained entirely inside arch specific code. I do NOT
expect any new issues to arise with them.
If someone does ever need to use debug/metrics with them, then they will
need to unravel this hairball between spinlocks, atomic ops, and bit ops
that exist only because parisc has exactly one atomic instruction: LDCW
(load and clear word).
From: "Luck, Tony" <tony.luck@intel.com>
ia64 fix
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Arjan van de Ven <arjanv@infradead.org>
Signed-off-by: Grant Grundler <grundler@parisc-linux.org>
Cc: Matthew Wilcox <willy@debian.org>
Signed-off-by: Hirokazu Takata <takata@linux-m32r.org>
Signed-off-by: Mikael Pettersson <mikpe@csd.uu.se>
Signed-off-by: Benoit Boissinot <benoit.boissinot@ens-lyon.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
For architecture like ia64, the switch stack structure is fairly large
(currently 528 bytes). For context switch intensive application, we found
that significant amount of cache misses occurs in switch_to() function.
The following patch adds a hook in the schedule() function to prefetch
switch stack structure as soon as 'next' task is determined. This allows
maximum overlap in prefetch cache lines for that structure.
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: "Luck, Tony" <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
For a NUMA system with multiple CPUs per node, declaring a cpu-exclusive
cpuset that includes only some, but not all, of the CPUs in a node will mangle
the sched domain structures.
Signed-off-by: John Hawkes <hawkes@sgi.com>
Cc; Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Ingo Molnar