From: Srivatsa Vaddagiri <vatsa@in.ibm.com>

migrate_all_tasks is currently run with rest of the machine stopped.
It iterates thr' the complete task table, turning off cpu affinity of any task
that it finds affine to the dying cpu. Depending on the task table
size this can take considerable time. All this time machine is stopped, doing
nothing.

Stopping the machine for such extended periods can be avoided if we do
task migration in CPU_DEAD notification and that's precisely what this patch
does.

The patch puts idle task to the _front_ of the dying CPU's runqueue at the 
highest priority possible. This cause idle thread to run _immediately_ after
kstopmachine thread yields. Idle thread notices that its cpu is offline and
dies quickly. Task migration can then be done at leisure in CPU_DEAD
notification, when rest of the CPUs are running.

Some advantages with this approach are:

	- More scalable. Predicatable amout of time that machine is stopped.
	- No changes to hot path/core code. We are just exploiting scheduler
	  rules which runs the next high-priority task on the runqueue. Also
	  since I put idle task to the _front_ of the runqueue, there
	  are no races when a equally high priority task is woken up
	  and added to the runqueue. It gets in at the back of the runqueue,
	  _after_ idle task!
	- cpu_is_offline check that is presenty required in try_to_wake_up,
	  idle_balance and rebalance_tick can be removed, thus speeding them
	  up a bit

From: Srivatsa Vaddagiri <vatsa@in.ibm.com>

  Rusty mentioned that the unlikely hints against cpu_is_offline is
  redundant since the macro already has that hint.  Patch below removes those
  redundant hints I added.

From: Nick Piggin <nickpiggin@yahoo.com.au>

The SMT wake_idle code really wants to look at a non-local CPU's domain in
order to check for idle siblings.

So change the domain attachment code a little bit so we continue to hold a
runqueue's lock while attaching a new domain.  This means the locking rules
have changed to: you may access your own domain without any lock, you must
hold a remote runqueue's lock in order to view its domain.

From: Nick Piggin <nickpiggin@yahoo.com.au>

This actually does produce better code, especially under the locked
section.

Turns a conditional + unconditional jump under the lock in the unlikely
case into a cmov outside the lock.

From: Nick Piggin <nickpiggin@yahoo.com.au>

It makes NEWLY_IDLE balances cause find_busiest_group return the busiest
available group even if there isn't an imbalance.  Basically - try a bit
harder to prevent schedule emptying the runqueue.

It is quite aggressive, but that isn't so bad because we don't (by default)
do NEWLY_IDLE balancing across NUMA nodes, and NEWLY_IDLE balancing is always
restricted to cache_hot tasks.

It picked up a little bit of idle time that dbt2-pgsql was seeing...

From: Ingo Molnar <mingo@elte.hu>

Implement balancing during clone().  It does the following things:

- introduces SD_BALANCE_CLONE that can serve as a tool for an
  architecture to limit the search-idlest-CPU scope on clone().
  E.g. the 512-CPU systems should rather not enable this.

- uses the highest sd for the imbalance_pct, not this_rq (which didnt
  make sense).

- unifies balance-on-exec and balance-on-clone via the find_idlest_cpu()
  function. Gets rid of sched_best_cpu() which was still a bit
  inconsistent IMO, it used 'min_load < load' as a condition for
  balancing - while a more correct approach would be to use half of the
  imbalance_pct, like passive balancing does.

- the patch also reintroduces the possibility to do SD_BALANCE_EXEC on
  SMP systems, and activates it - to get testing.

- NOTE: there's one thing in this patch that is slightly unclean: i
  introduced wake_up_forked_thread. I did this to make it easier to get
  rid of this patch later (wake_up_forked_process() has lots of
  dependencies in various architectures). If this capability remains in
  the kernel then i'll clean it up and introduce one function for
  wake_up_forked_process/thread.

- NOTE2: i added the SD_BALANCE_CLONE flag to the NUMA CPU template too.
  Some NUMA architectures probably want to disable this.

From: Ingo Molnar <mingo@elte.hu>

This does the source/target cleanup.  This is a no-functionality patch which
also adds more comments to explain these functions.

From: Nick Piggin <nickpiggin@yahoo.com.au>

This patch starts to balance woken processes when half the relevant domain's
imbalance_pct is reached.  Previously balancing would start after a small,
constant difference in waker/wakee runqueue loads was reached, which would
cause too much process movement when there are lots of processes running.

It also turns wake balancing into a domain flag while previously it was always
on.  Now sched domains can "soft partition" an SMP system without using
processor affinities.

From: Ingo Molnar <mingo@elte.hu>

This re-adds cleanups which were lost in splitups of an earlier patch.

From: Nick Piggin <nickpiggin@yahoo.com.au>

The attached patch is required to work correctly with the CPU hotplug
framework.  John Hawkes reports successful booting with this.

From: Ingo Molnar <mingo@elte.hu>

The attached patch extends sync wakeups to the process sys_exit() path too:
the chldwait wakeup can be done sync, since we know that the process is
going to exit (and thus deschedule).

The most visible effect of this change is strace's behavior on SMP systems:
it now stays on a single CPU, together with the traced child.  (previously
it would run in parallel to the child, bouncing around madly.)

From: Ingo Molnar <mingo@elte.hu>

Helper function for later patches

From: Ingo Molnar <mingo@elte.hu>

Uninline things

From: Nick Piggin <nickpiggin@yahoo.com.au>

Minor cleanups from Ingo's patch including task_hot (do it right in
try_to_wake_up too).

From: Nick Piggin <nickpiggin@yahoo.com.au>

De-racify the sched domain setup code.  This involves creating a dummy
"init" domain during sched_init (which is called early).

When topology information becomes available, the sched domains are then
built and attached.  The attach mechanism is asynchronous and uses the
migration threads, which perform the switch with interrupts off.  This is a
quiescent state, so domains can still be lockless on the read side.  It
also allows us to change the domains at runtime without much more work. 
This is something SGI is interested in to elegantly do soft partitioning of
their systems without having to use hard cpu affinities (which cause
balancing problems of their own).

The current setup code also has a race somewhere because it is unable to
boot on a 384 CPU system.



From: Anton Blanchard <anton@samba.org>

   This is basically a mindless ppc64 merge of the x86 changes to sched
   domain init code.

   Actually if I produce a sibling_map[] then the x86 code and the ppc64
   will be identical.  Maybe we can merge it.

From: Anton Blanchard <anton@samba.org>

It seems someone has been making trivial changes without using grep.

From: Anton Blanchard <anton@samba.org>

Below are the diffs between the current ppc64 sched init stuff and x86.

- Ignore the POWER5 specific stuff, I dont set up a sibling map yet.
- What should I set cache_hot_time to?

large cpumask typechecking requirements (perhaps useful on x86 as well):
- cpu->cpumask = CPU_MASK_NONE -> cpus_clear(cpu->cpumask);
- cpus_and(nodemask, node_to_cpumask(i), cpu_possible_map) doesnt work,
  need to use a temporary

From: Nick Piggin <nickpiggin@yahoo.com.au>

After the for_each_domain change, the warn here won't trigger, instead it
will oops in the if statement.  Also, make sure we don't pass an empty
cpumask to for_each_cpu.

From: Nick Piggin <nickpiggin@yahoo.com.au>

From: John Hawkes

The following brings up performance on a 64-way Altix.  This system being on
the smaller end of the scale should also be applicable to other NUMA systems.

From: Nick Piggin <nickpiggin@yahoo.com.au>

Imbalance calculations were not right.  This would cause unneeded migration.

From: Nick Piggin <nickpiggin@yahoo.com.au>

Make affine wakes and "passive load balancing" more conservative.  Aggressive
affine wakeups were causing huge regressions in dbt3-pgsql on 8-way non NUMA
systems at OSDL's STP.

From: Rusty Russell <rusty@rustcorp.com.au>

From: Srivatsa Vaddagiri <vatsa@in.ibm.com>
From: Andrew Morton <akpm@osdl.org>
From: Rusty Russell <rusty@rustcorp.com.au>

We want to get rid of lock_cpu_hotplug() in sched_migrate_task.  Found
that lockless migration of execing task is _extremely_ racy.  The
races I hit are described below, alongwith probable solutions.

Task migration done elsewhere should be safe (?) since they either
hold the lock (sys_sched_setaffinity) or are done entirely with preemption 
disabled (load_balance).

   sched_balance_exec does:

	a. disables preemption
	b. finds new_cpu for current
	c. enables preemption
	d. calls sched_migrate_task to migrate current to new_cpu

   and sched_migrate_task does:

	e. task_rq_lock(p)
	f. migrate_task(p, dest_cpu ..)
		(if we have to wait for migration thread)
		g. task_rq_unlock()
		h. wake_up_process(rq->migration_thread)
		i. wait_for_completion()

   Several things can happen here:

	1. new_cpu can go down after h and before migration thread has
	   got around to handle the request

	   ==> we need to add a cpu_is_offline check in __migrate_task

	2. new_cpu can go down between c and d or before f.

	   ===> Even though this case is automatically handled by the above 
	        change (migrate_task being called on a running task, current,
		will delegate migration to migration thread), would it be 
	 	good practice to avoid calling migrate_task in the first place
		itself when dest_cpu is offline. This means adding another
	 	cpu_is_offline check after e in sched_migrate_task

	3. The 'current' task can get preempted _immediately_ after
	   g and when it comes back, task_cpu(p) can be dead. In
	   which case, it is invalid to do wake_up on a non-existent migration 
	   thread.  (rq->migration_thread can be NULL).

	   ===> We should disable preemption thr' g and h

	4. Before migration thread gets around to handle the request, its cpu
	   goes dead. This will leave unhandled migration requests in the dead 
	   cpu. 

	   ===> We need to wakeup sleeping requestors (if any) in CPU_DEAD
	        notification.

I really wonder if we can get rid of these issues by avoiding balancing at 
exec time and instead have it balanced during load_balance ..Alternately
if this is valuable and we want to retain it, I think we still need to
consider a read/write sem, with sched_migrate_task doing down_read_trylock.
This may eliminate the deadlock I hit between cpu_up and CPU_UP_PREPARE 
notification, which had forced me away from r/w sem.

Anyway patch below addresses the above races. Its against 2.6.6-rc2-mm1
and has been tested on a 4way Intel Pentium SMP m/c.


Rusty sez:

Two other changes:
1) I grabbed a reference to the thread, rather than using
preempt_disable().  It's the more obvious way I think.

2) Why the wait_to_die code?  It might be needed if we move tasks after
stop_machine, but for nowI don't see the problem with the migration
thread running on the wrong CPU for a bit: nothing is on this runqueue
so active_load_balance is safe, and __migrate task will be a noop (due
to cpu_is_offline() check).  If there is a problem, your fix is racy,
because we could be preempted immediately afterwards.

So I just stop the kthread then wakeup any remaining...

From: Ingo Molnar <mingo@elte.hu>

The trivial fixes.

- added recent trivial bits from Nick's and my patches.
- hotplug CPU fix
- early init cleanup

From: Martin Hicks <mort@wildopensource.com>

Another optimization patch from Jack Steiner, intended to reduce TLB
flushes during process migration.

Most architextures should define tlb_migrate_prepare() to be flush_tlb_mm(),
but on i386, it would be a wasted flush, because i386 disconnects previous
cpus from the tlb flush automatically.

From: Nick Piggin <piggin@cyberone.com.au>

This patch removes the per runqueue array of NR_CPU arrays.  Each time we
want to check a remote CPU's load we check nr_running as well anyway, so
introduce a cpu_load which is the load of the local runqueue and is kept
updated in the timer tick.  Put them in the same cacheline.

This has additional benefits of having the cpu_load consistent across all
CPUs and more up to date.  It is sampled better too, being updated once per
timer tick.

This shouldn't make much difference in scheduling behaviour, but all
benchmarks are either as good or better on the 16-way NUMAQ: hackbench,
reaim, volanomark are about the same, tbench and dbench are maybe a bit
better.  kernbench is about one percent better.

John reckons it isn't a big deal, but it does save 4K per CPU or 2MB total
on his big systems, so I figure it must be a bit kinder on the caches.  I
think it is just nicer in general anyway.

From: Con Kolivas <kernel@kolivas.org>

This patch provides full per-package priority support for SMT processors
(aka pentium4 hyperthreading) when combined with CONFIG_SCHED_SMT.

It maintains cpu percentage distribution within each physical cpu package
by limiting the time a lower priority task can run on a sibling cpu
concurrently with a higher priority task.

It introduces a new flag into the scheduler domain
unsigned int per_cpu_gain;	/* CPU % gained by adding domain cpus */

This is empirically set to 15% for pentium4 at the moment and can be
modified to support different values dynamically as newer processors come
out with improved SMT performance.  It should not matter how many siblings
there are.

How it works is it compares tasks running on sibling cpus and when a lower
static priority task is running it will delay it till
high_priority_timeslice * (100 - per_cpu_gain) / 100 <= low_prio_timeslice

eg.  a nice 19 task timeslice is 10ms and nice 0 timeslice is 102ms On
vanilla the nice 0 task runs on one logical cpu while the nice 19 task runs
unabated on the other logical cpu.  With smtnice the nice 0 runs on one
logical cpu for 102ms and the nice 19 sleeps till the nice 0 task has 12ms
remaining and then will schedule.

Real time tasks and kernel threads are not altered by this code, and kernel
threads do not delay lower priority user tasks.

with lots of thanks to Zwane Mwaikambo and Nick Piggin for help with the
coding of this version.

If this is merged, it is probably best to delay pushing this upstream in
mainline till sched_domains gets tested for at least one major release.

From: Nick Piggin <piggin@cyberone.com.au>

This changes sched domains to contain all possible CPUs, and check for
online as needed.  It's in order to play nicely with CPU hotplug.

From: Nick Piggin <piggin@cyberone.com.au>

The following patch implements a cpu_power member to struct sched_group.

This allows special casing to be removed for SMT groups in the balancing
code.  It does not take CPU hotplug into account yet, but that shouldn't be
too hard.

I have tested it on the NUMAQ by pretending it has SMT.  Works as expected.
Active balances across nodes.

From: Rusty Russell <rusty@rustcorp.com.au>,
      Nick Piggin <piggin@cyberone.com.au>

The current sched_balance_exec() sets the task's cpus_allowed mask
temporarily to move it to a different CPU.  This has several issues,
including the fact that a task will see its affinity at a bogus value.

So we change the migration_req_t to explicitly specify a destination CPU,
rather than the migration thread deriving it from cpus_allowed.  If the
requested CPU is no longer valid (racing with another set_cpus_allowed,
say), it can be ignored: if the task is not allowed on this CPU, there will
be another migration request pending.

This change allows sched_balance_exec() to tell the migration thread what
to do without changing the cpus_allowed mask.

So we rename __set_cpus_allowed() to move_task(), as the cpus_allowed mask
is now set by the caller.  And move_task_away(), which the migration thread
uses to actually perform the move, is renamed __move_task().

I also ignore offline CPUs in sched_best_cpu(), so sched_migrate_task()
doesn't need to check for offline CPUs.

Ulterior motive: this approach also plays well with CPU Hotplug.
Previously that patch might have seen a task with cpus_allowed only
containing the dying CPU (temporarily due to sched_balance_exec) and
forcibly reset it to all cpus, which might be wrong.  The other approach is
to hold the cpucontrol sem around sched_balance_exec(), which is too much
of a bottleneck.

From: Nick Piggin <piggin@cyberone.com.au>

John Hawkes discribed this problem to me:

There *is* a small problem in this area, though, that SuSE avoids.
"jiffies" gets updated by cpu0.  The other CPUs may, over time, get out of
sync (and they're initialized on ia64 to start out being out of sync), so
it's no guarantee that every CPU will wake up from its timer interrupt and
see a "jiffies" value that is guaranteed to be last_jiffies+1.  Sometimes
the jiffies value may be unchanged since the last wakeup.  Sometimes the
jiffies value may have incremented by 2 (or more, especially if cpu0's
interrupts are disabled for long stretches of time).  So an algoithm that
says, "I'll call load_balance() only when jiffies is *exactly* N" is going
to fail on occasion, either by calling load_balance() too often or not
often enough.  ***

I fixed this by adding a last_balance field to struct sched_domain, and
working off that.

From: Nick Piggin <piggin@cyberone.com.au>

The following patch builds a scheduling description for the i386
architecture using cpu_sibling_map to set up SMT if CONFIG_SCHED_SMT is
set.

It could be made more fancy and collapse degenerate domains at runtime (ie.
1 sibling per CPU, or 1 NUMA node in the computer).


From: Zwane Mwaikambo <zwane@arm.linux.org.uk>

   This fixes an oops due to cpu_sibling_map being uninitialised when a
   system with no MP table (most UP boxen) boots a CONFIG_SMT kernel.  What
   also happens is that the cpu_group lists end up not being terminated
   properly, but this oops kills it first.  Patch tested on UP w/o MP table,
   2x P2 and UP Xeon w/ no siblings.

From: "Martin J. Bligh" <mbligh@aracnet.com>,
      Nick Piggin <piggin@cyberone.com.au>

   Change arch_init_sched_domains to use cpu_online_map

From: Anton Blanchard <anton@samba.org>

   Fix build with NR_CPUS > BITS_PER_LONG

From: Nick Piggin <piggin@cyberone.com.au>

This is a (somewhat) trivial patch which converts cpu_sibling_map from an
array of CPUs to an array of cpumasks.  Needed for >2 siblings per package,
but it actually can simplify code as it allows the cpu_sibling_map to be
set up even when there is 1 sibling per package.  Intel want this, I use it
in the next patch to build scheduling domains for the P4 HT.

From: Thomas Schlichter <thomas.schlichter@web.de>

   Build fix

From: "Pallipadi, Venkatesh" <venkatesh.pallipadi@intel.com>

   Fix to handle more than 2 siblings per package.

From: Nick Piggin <piggin@cyberone.com.au>

This patch gets the sched_domain scheduler working better WRT balancing.
Its been tested on the NUMAQ.  Among other things it changes to the way SMT
load calculation works so as not to active load blances when it shouldn't.

It still has a problem with SMT and NUMA: it will put a task on each
sibling in a node before moving tasks to another node.  It should probably
start moving tasks after each *physical* CPU is filled.

To fix, you need "how much CPU power in this domain?" At the moment we
approximate # runqueues == CPU power, and hack around it at the CPU
physical domain by counting all sibling runqueues as 1.

It isn't hard to correctly work the CPU power out, but once CPU hotplug is
in the equation it becomes much more hotplug events.  If anyone is actually
interested in getting this fixed, that is.

From: Nick Piggin <piggin@cyberone.com.au>

Anton was attempting to make a sched domain topology for his POWER5 and was
having some trouble.

This patch only includes code which is ifdefed out, but hopefully it will
be of some use to implementors.

From: Nick Piggin <piggin@cyberone.com.au>

This is the core sched domains patch.  It can handle any number of levels
in a scheduling heirachy, and allows architectures to easily customize how
the scheduler behaves.  It also provides progressive balancing backoff
needed by SGI on their large systems (although they have not yet tested
it).

It is built on top of (well, uses ideas from) my previous SMP/NUMA work, and
gets results very similar to them when using the default scheduling
description.

Benchmarks
==========

Martin was seeing I think 10-20% better system times in kernbench on the 32
way.  I was seeing improvements in dbench, tbench, kernbench, reaim,
hackbench on a 16-way NUMAQ.  Hackbench in fact had a non linear element
which is all but eliminated.  Large improvements in volanomark.

Cross node task migration was decreased in all above benchmarks, sometimes by
a factor of 100!!  Cross CPU migration was also generally decreased.  See
this post:
http://groups.google.com.au/groups?hl=en&lr=&ie=UTF-8&oe=UTF-8&frame=right&th=a406c910b30cbac4&seekm=UAdQ.3hj.5%40gated-at.bofh.it#link2

Results on a hyperthreading P4 are equivalent to Ingo's shared runqueues
patch (which is a big improvement).

Some examples on the 16-way NUMAQ (this is slightly older sched domain code):

 http://www.kerneltrap.org/~npiggin/w26/hbench.png
 http://www.kerneltrap.org/~npiggin/w26/vmark.html

From: Jes Sorensen <jes@wildopensource.com>

   Tiny patch to make -mm3 compile on an NUMA box with NR_CPUS >
   BITS_PER_LONG.

From: "Martin J. Bligh" <mbligh@aracnet.com>

   Fix a minor nit with the find_busiest_group code.  No functional change,
   but makes the code simpler and clearer.  This patch does two things ... 
   adds some more expansive comments, and removes this if clause:

      if (*imbalance < SCHED_LOAD_SCALE
                      && max_load - this_load > SCHED_LOAD_SCALE)
		*imbalance = SCHED_LOAD_SCALE;

   If we remove the scaling factor, we're basically conditionally doing:

	if (*imbalance < 1)
		*imbalance = 1;

   Which is pointless, as the very next thing we do is to remove the
   scaling factor, rounding up to the nearest integer as we do:

	*imbalance = (*imbalance + SCHED_LOAD_SCALE - 1) >> SCHED_LOAD_SHIFT;

   Thus the if statement is redundant, and only makes the code harder to
   read ;-)

From: Rick Lindsley <ricklind@us.ibm.com>

   In find_busiest_group(), after we exit the do/while, we select our
   imbalance.  But max_load, avg_load, and this_load are all unsigned, so
   min(x,y) will make a bad choice if max_load < avg_load < this_load (that
   is, a choice between two negative [very large] numbers).

   Unfortunately, there is a bug when max_load never gets changed from zero
   (look in the loop and think what happens if the only load on the machine is
   being created by cpu groups of which we are a member).  And you have a
   recipe for some really bogus values for imbalance.

   Even if you fix the max_load == 0 bug, there will still be times when
   avg_load - this_load will be negative (thus very large) and you'll make the
   decision to move stuff when you shouldn't have.

   This patch allows for this_load to set max_load, which if I understand
   the logic properly is correct.  With this patch applied, the algorithm is
   *much* more conservative ...  maybe *too* conservative but that's for
   another round of testing ...

From: Ingo Molnar <mingo@elte.hu>

   sched-find-busiest-fix

From: Nick Piggin <piggin@cyberone.com.au>

From: Frank Cornelis <frank.cornelis@elis.ugent.be>

In order to get the best possible resolution we need to use NR_CPUS instead
of the constant value 10.  load is an int, so no need to worry about
overflows...

From: Ingo Molnar <mingo@elte.hu>

From: Nick Piggin <piggin@cyberone.com.au> wrote:

It removes the last place where we mess with run_list open coded.

This should help some laptops where the generic PCI
code might otherwise believe that this range is unused.
The ACPI IO range is usually not visible as a standard
BAR.

Various people hit this in earlier kernels. The x86-64 kernel did not compile 
without CONFIG_IOMMU_GART in various configurations. Just add the missing symbol 
and export it. Also export iommu_merge while I am at it.

This fixes a bug in the new machine check handler on x86-64.

One nasty part was that when you got an MCE during boot up
then it would not always print it on the screen, but still
panic because it attempted to kill the idle task.

This patch does:
 - Always use KERN_EMERG when printing MCEs
 - Always panic and print on screen before killing idle loop
   or init.