Commit Graph

19 Commits

Author SHA1 Message Date
Yong Zhang
5710f15b52 sched/cpupri: Remove cpupri->pri_active
Since [sched/cpupri: Remove the vec->lock], member pri_active
of struct cpupri is not needed any more, just remove it. Also
clean stuff related to it.

Signed-off-by: Yong Zhang <yong.zhang0@gmail.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110806001004.GA2207@zhy
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-08-14 12:01:11 +02:00
Steven Rostedt
d473750b40 sched/cpupri: Fix memory barriers for vec updates to always be in order
[ This patch actually compiles. Thanks to Mike Galbraith for pointing
that out. I compiled and booted this patch with no issues. ]

Re-examining the cpupri patch, I see there's a possible race because the
update of the two priorities vec->counts are not protected by a memory
barrier.

When a RT runqueue is overloaded and wants to push an RT task to another
runqueue, it scans the RT priority vectors in a loop from lowest
priority to highest.

When we queue or dequeue an RT task that changes a runqueue's highest
priority task, we update the vectors to show that a runqueue is rated at
a different priority. To do this, we first set the new priority mask,
and increment the vec->count, and then set the old priority mask by
decrementing the vec->count.

If we are lowering the runqueue's RT priority rating, it will trigger a
RT pull, and we do not care if we miss pushing to this runqueue or not.

But if we raise the priority, but the priority is still lower than an RT
task that is looking to be pushed, we must make sure that this runqueue
is still seen by the push algorithm (the loop).

Because the loop reads from lowest to highest, and the new priority is
set before the old one is cleared, we will either see the new or old
priority set and the vector will be checked.

But! Since there's no memory barrier between the updates of the two, the
old count may be decremented first before the new count is incremented.
This means the loop may see the old count of zero and skip it, and also
the new count of zero before it was updated. A possible runqueue that
the RT task could move to could be missed.

A conditional memory barrier is placed between the vec->count updates
and is only called when both updates are done.

The smp_wmb() has also been changed to smp_mb__before_atomic_inc/dec(),
as they are not needed by archs that already synchronize
atomic_inc/dec().

The smp_rmb() has been moved to be called at every iteration of the loop
so that the race between seeing the two updates is visible by each
iteration of the loop, as an arch is free to optimize the reading of
memory of the counters in the loop.

Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/1312547269.18583.194.camel@gandalf.stny.rr.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-08-14 12:01:07 +02:00
Steven Rostedt
c92211d9b7 sched/cpupri: Remove the vec->lock
sched/cpupri: Remove the vec->lock

The cpupri vec->lock has been showing up as a top contention
lately. This is because of the RT push/pull logic takes an
agressive approach for migrating RT tasks. The cpupri logic is
in place to improve the performance of the push/pull when dealing
with large number CPU machines.

The problem though is a vec->lock is required, where a vec is a
global per RT priority structure. That is, if there are lots of
RT tasks at the same priority, every time they are added or removed
from the RT queue, this global vec->lock is taken. Now that more
kernel threads are becoming RT (RCU boost and threaded interrupts)
this is becoming much more of an issue.

There are two variables that are being synced by the vec->lock.
The cpupri bitmask, and the vec->counter. The cpupri bitmask
is one bit per priority. If a RT priority vec has a process queued,
then the vec->count is > 0 and the cpupri bitmask is set for that
RT priority.

If the cpupri bitmask gets out of sync with the vec->counter, we could
end up pushing a low proirity RT task to a high priority queue.
That RT task that could have run immediately could be queued on a
run queue with a higher priority task indefinitely.

The solution is not to use the cpupri bitmask and just look at the
vec->count directly when doing a pull. The cpupri bitmask is just
a fast way to scan the RT priorities when a pull is made. Instead
of using the bitmask, and just examine all RT priorities, and
look at the vec->counts, we could eliminate the vec->lock. The
scan of RT tasks is to find a run queue that we can push an RT task
to, and we do not push to a high priority queue, thus the scan only
needs to go from 1 to RT task->prio, and not all 100 RT priorities.

The push algorithm, which does the scan of RT priorities (and
scan of the bitmask) only happens when we have an overloaded RT run
queue (more than one RT task queued). The grabbing of the vec->lock
happens every time any RT task is queued or dequeued on the run
queue for that priority. The slowing down of the scan by not using
a bitmask is negligible by the speed up of removing the vec->lock
contention, and replacing it with an atomic counter and memory barrier.

To prove this, I wrote a patch that times both the loop and the code
that grabs the vec->locks. I passed the patches to various people
(and companies) to test and show the results. I let everyone choose
their own load to test, giving different loads on the system,
for various different setups.

Here's some of the results: (snipping to a few CPUs to not make
this change log huge, but the results were consistent across
the entire system).

System 1 (24 CPUs)

Before patch:
CPU:    Name    Count   Max     Min     Average Total
----    ----    -----   ---     ---     ------- -----
[...]
cpu 20: loop    3057    1.766   0.061   0.642   1963.170
        vec     6782949 90.469  0.089   0.414   2811760.503
cpu 21: loop    2617    1.723   0.062   0.641   1679.074
        vec     6782810 90.499  0.089   0.291   1978499.900
cpu 22: loop    2212    1.863   0.063   0.699   1547.160
        vec     6767244 85.685  0.089   0.435   2949676.898
cpu 23: loop    2320    2.013   0.062   0.594   1380.265
        vec     6781694 87.923  0.088   0.431   2928538.224

After patch:
cpu 20: loop    2078    1.579   0.061   0.533   1108.006
        vec     6164555 5.704   0.060   0.143   885185.809
cpu 21: loop    2268    1.712   0.065   0.575   1305.248
        vec     6153376 5.558   0.060   0.187   1154960.469
cpu 22: loop    1542    1.639   0.095   0.533   823.249
        vec     6156510 5.720   0.060   0.190   1172727.232
cpu 23: loop    1650    1.733   0.068   0.545   900.781
        vec     6170784 5.533   0.060   0.167   1034287.953

All times are in microseconds. The 'loop' is the amount of time spent
doing the loop across the priorities (before patch uses bitmask).
the 'vec' is the amount of time in the code that requires grabbing
the vec->lock. The second patch just does not have the vec lock, but
encompasses the same code.

Amazingly the loop code even went down on average. The vec code went
from .5 down to .18, that's more than half the time spent!

Note, more than one test was run, but they all had the same results.

System 2 (64 CPUs)

Before patch:
CPU:    Name    Count   Max     Min     Average Total
----    ----    -----   ---     ---     ------- -----
cpu 60: loop    0       0       0       0       0
        vec     5410840 277.954 0.084   0.782   4232895.727
cpu 61: loop    0       0       0       0       0
        vec     4915648 188.399 0.084   0.570   2803220.301
cpu 62: loop    0       0       0       0       0
        vec     5356076 276.417 0.085   0.786   4214544.548
cpu 63: loop    0       0       0       0       0
        vec     4891837 170.531 0.085   0.799   3910948.833

After patch:
cpu 60: loop    0       0       0       0       0
        vec     5365118 5.080   0.021   0.063   340490.267
cpu 61: loop    0       0       0       0       0
        vec     4898590 1.757   0.019   0.071   347903.615
cpu 62: loop    0       0       0       0       0
        vec     5737130 3.067   0.021   0.119   687108.734
cpu 63: loop    0       0       0       0       0
        vec     4903228 1.822   0.021   0.071   348506.477

The test run during the measurement did not have any (very few,
from other CPUs) RT tasks pushing. But this shows that it helped
out tremendously with the contention, as the contention happens
because the vec->lock is taken only on queuing at an RT priority,
and different CPUs that queue tasks at the same priority will
have contention.

I tested on my own 4 CPU machine with the following results:

Before patch:
CPU:    Name    Count   Max     Min     Average Total
----    ----    -----   ---     ---     ------- -----
cpu 0:  loop    2377    1.489   0.158   0.588   1398.395
        vec     4484    770.146 2.301   4.396   19711.755
cpu 1:  loop    2169    1.962   0.160   0.576   1250.110
        vec     4425    152.769 2.297   4.030   17834.228
cpu 2:  loop    2324    1.749   0.155   0.559   1299.799
        vec     4368    779.632 2.325   4.665   20379.268
cpu 3:  loop    2325    1.629   0.157   0.561   1306.113
        vec     4650    408.782 2.394   4.348   20222.577

After patch:
CPU:    Name    Count   Max     Min     Average Total
----    ----    -----   ---     ---     ------- -----
cpu 0:  loop    2121    1.616   0.113   0.636   1349.189
        vec     4303    1.151   0.225   0.421   1811.966
cpu 1:  loop    2130    1.638   0.178   0.644   1372.927
        vec     4627    1.379   0.235   0.428   1983.648
cpu 2:  loop    2056    1.464   0.165   0.637   1310.141
        vec     4471    1.311   0.217   0.433   1937.927
cpu 3:  loop    2154    1.481   0.162   0.601   1295.083
        vec     4236    1.253   0.230   0.425   1803.008

This was running my migrate.c code that can be found at:
http://lwn.net/Articles/425763/

The migrate code does stress the RT tasks a bit. This shows that
the loop did increase a little after the patch, but not by much.
The vec code dropped dramatically. From 4.3us down to .42us.
That's a 10x improvement!

Tested-by: Mike Galbraith <mgalbraith@suse.de>
Tested-by: Luis Claudio R. Gonçalves <lgoncalv@redhat.com>
Tested-by: Matthew Hank Sabins<msabins@linux.vnet.ibm.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Reviewed-by: Gregory Haskins <gregory.haskins@gmail.com>
Acked-by: Hillf Danton <dhillf@gmail.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Chris Mason <chris.mason@oracle.com>
Link: http://lkml.kernel.org/r/1312317372.18583.101.camel@gandalf.stny.rr.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-08-14 12:01:03 +02:00
Pekka Enberg
68c38fc3cb sched: No need for bootmem special cases
As of commit dcce284 ("mm: Extend gfp masking to the page
allocator") and commit 7e85ee0 ("slab,slub: don't enable
interrupts during early boot"), the slab allocator makes
sure we don't attempt to sleep during boot.

Therefore, remove bootmem special cases from the scheduler
and use plain GFP_KERNEL instead.

Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1279225102-2572-1-git-send-email-penberg@cs.helsinki.fi>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-07-17 12:06:22 +02:00
Tejun Heo
5a0e3ad6af include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files.  percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.

percpu.h -> slab.h dependency is about to be removed.  Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability.  As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.

  http://userweb.kernel.org/~tj/misc/slabh-sweep.py

The script does the followings.

* Scan files for gfp and slab usages and update includes such that
  only the necessary includes are there.  ie. if only gfp is used,
  gfp.h, if slab is used, slab.h.

* When the script inserts a new include, it looks at the include
  blocks and try to put the new include such that its order conforms
  to its surrounding.  It's put in the include block which contains
  core kernel includes, in the same order that the rest are ordered -
  alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
  doesn't seem to be any matching order.

* If the script can't find a place to put a new include (mostly
  because the file doesn't have fitting include block), it prints out
  an error message indicating which .h file needs to be added to the
  file.

The conversion was done in the following steps.

1. The initial automatic conversion of all .c files updated slightly
   over 4000 files, deleting around 700 includes and adding ~480 gfp.h
   and ~3000 slab.h inclusions.  The script emitted errors for ~400
   files.

2. Each error was manually checked.  Some didn't need the inclusion,
   some needed manual addition while adding it to implementation .h or
   embedding .c file was more appropriate for others.  This step added
   inclusions to around 150 files.

3. The script was run again and the output was compared to the edits
   from #2 to make sure no file was left behind.

4. Several build tests were done and a couple of problems were fixed.
   e.g. lib/decompress_*.c used malloc/free() wrappers around slab
   APIs requiring slab.h to be added manually.

5. The script was run on all .h files but without automatically
   editing them as sprinkling gfp.h and slab.h inclusions around .h
   files could easily lead to inclusion dependency hell.  Most gfp.h
   inclusion directives were ignored as stuff from gfp.h was usually
   wildly available and often used in preprocessor macros.  Each
   slab.h inclusion directive was examined and added manually as
   necessary.

6. percpu.h was updated not to include slab.h.

7. Build test were done on the following configurations and failures
   were fixed.  CONFIG_GCOV_KERNEL was turned off for all tests (as my
   distributed build env didn't work with gcov compiles) and a few
   more options had to be turned off depending on archs to make things
   build (like ipr on powerpc/64 which failed due to missing writeq).

   * x86 and x86_64 UP and SMP allmodconfig and a custom test config.
   * powerpc and powerpc64 SMP allmodconfig
   * sparc and sparc64 SMP allmodconfig
   * ia64 SMP allmodconfig
   * s390 SMP allmodconfig
   * alpha SMP allmodconfig
   * um on x86_64 SMP allmodconfig

8. percpu.h modifications were reverted so that it could be applied as
   a separate patch and serve as bisection point.

Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.

Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-30 22:02:32 +09:00
Jiri Kosina
318ae2edc3 Merge branch 'for-next' into for-linus
Conflicts:
	Documentation/filesystems/proc.txt
	arch/arm/mach-u300/include/mach/debug-macro.S
	drivers/net/qlge/qlge_ethtool.c
	drivers/net/qlge/qlge_main.c
	drivers/net/typhoon.c
2010-03-08 16:55:37 +01:00
Akinobu Mita
984b3f5746 bitops: rename for_each_bit() to for_each_set_bit()
Rename for_each_bit to for_each_set_bit in the kernel source tree.  To
permit for_each_clear_bit(), should that ever be added.

The patch includes a macro to map the old for_each_bit() onto the new
for_each_set_bit().  This is a (very) temporary thing to ease the migration.

[akpm@linux-foundation.org: add temporary for_each_bit()]
Suggested-by: Alexey Dobriyan <adobriyan@gmail.com>
Suggested-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: David Woodhouse <dwmw2@infradead.org>
Cc: Artem Bityutskiy <dedekind@infradead.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 11:26:23 -08:00
Adam Buchbinder
2a61aa4016 Fix misspellings of "invocation" in comments.
Some comments misspell "invocation"; this fixes them. No code
changes.

Signed-off-by: Adam Buchbinder <adam.buchbinder@gmail.com>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
2010-02-04 11:55:45 +01:00
Akinobu Mita
90fdbdb484 sched: Use for_each_bit
No change in functionality.

Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
LKML-Reference: <1264938810-4173-1-git-send-email-akinobu.mita@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-02-02 06:58:27 +01:00
Thomas Gleixner
fe841226bd sched: Convert cpupri lock to raw_spinlock
Convert locks which cannot be sleeping locks in preempt-rt to
raw_spinlocks.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: Ingo Molnar <mingo@elte.hu>
2009-12-14 23:55:33 +01:00
Steven Rostedt
c3a2ae3d93 sched: Add new prio to cpupri before removing old prio
We need to add the new prio to the cpupri accounting before
removing the old prio. This is because removing the old prio
first will open a race window where the cpu will be removed
from pri_active. In this case the cpu will not be visible for
RT push and pulls. This could cause a RT task to not migrate
appropriately, and create a very large latency.

This bug was found with the use of ftrace sched events and
trace_printk.

Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090729042526.438281019@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-08-02 14:26:09 +02:00
Gregory Haskins
07903af152 sched: Fix race in cpupri introduced by cpumask_var changes
Background:

Several race conditions in the scheduler have cropped up
recently, which Steven and I have tracked down using ftrace.
The most recent one turns out to be a race in how the scheduler
determines a suitable migration target for RT tasks, introduced
recently with commit:

    commit 68e74568fb
    Date:   Tue Nov 25 02:35:13 2008 +1030

        sched: convert struct cpupri_vec cpumask_var_t.

The original design of cpupri allowed lockless readers to
quickly determine a best-estimate target.  Races between the
pri_active bitmap and the vec->mask were handled in the
original code because we would detect and return "0" when this
occured.  The design was predicated on the *effective*
atomicity (*) of caching the result of cpus_and() between the
cpus_allowed and the vec->mask.

Commit 68e74568 changed the behavior such that vec->mask is
accessed multiple times.  This introduces a subtle race, the
result of which means we can have a result that returns "1",
but with an empty bitmap.

*) yes, we know cpus_and() is not a locked operator across the
   entire composite array, but it is implicitly atomic on a
   per-word basis which is all the design required to work.

Implementation:

Rather than forgoing the lockless design, or reverting to a
stack-based cpumask_t, we simply check for when the race has
been encountered and continue processing in the event that the
race is hit.  This renders the removal race as if the priority
bit had been atomically cleared as well, and allows the
algorithm to execute correctly.

Signed-off-by: Gregory Haskins <ghaskins@novell.com>
CC: Rusty Russell <rusty@rustcorp.com.au>
CC: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090730145728.25226.92769.stgit@dev.haskins.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-08-02 14:23:29 +02:00
Li Zefan
fd5e1b5dba sched: Remove unneeded __ref tag
Those two functions no longer call alloc_bootmmem_cpumask_var(),
so no need to tag them with __init_refok.

Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
LKML-Reference: <4A35DD5B.9050106@cn.fujitsu.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-17 16:08:04 +02:00
Pekka Enberg
0fb5302916 sched: use slab in cpupri_init()
Lets not use the bootmem allocator in cpupri_init() as slab is already up when
it is run.

Cc: Ingo Molnar <mingo@elte.hu>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Yinghai Lu <yinghai@kernel.org>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
2009-06-11 19:27:12 +03:00
Yinghai Lu
eaa958402e cpumask: alloc zeroed cpumask for static cpumask_var_ts
These are defined as static cpumask_var_t so if MAXSMP is not used,
they are cleared already.  Avoid surprises when MAXSMP is enabled.

Signed-off-by: Yinghai Lu <yinghai.lu@kernel.org>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2009-06-09 22:30:27 +09:30
Rusty Russell
13b8bd0a57 sched_rt: don't allocate cpumask in fastpath
Impact: cleanup

As pointed out by Steven Rostedt.  Since the arg in question is
unused, we simply change cpupri_find() to accept NULL.

Reported-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
LKML-Reference: <200903251501.22664.rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-04-01 13:24:51 +02:00
Li Zefan
db2f59c8c9 sched: fix section mismatch
init_rootdomain() calls alloc_bootmem_cpumask_var() at system boot,
so does cpupri_init().

Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-06 11:07:15 +01:00
Rusty Russell
68e74568fb sched: convert struct cpupri_vec cpumask_var_t.
Impact: stack usage reduction, (future) size reduction for large NR_CPUS.

Dynamically allocating cpumasks (when CONFIG_CPUMASK_OFFSTACK) saves
space for small nr_cpu_ids but big CONFIG_NR_CPUS.

The fact cpupro_init is called both before and after the slab is
available makes for an ugly parameter unfortunately.

We also use cpumask_any_and to get rid of a temporary in cpupri_find.

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-11-24 17:52:22 +01:00
Gregory Haskins
6e0534f278 sched: use a 2-d bitmap for searching lowest-pri CPU
The current code use a linear algorithm which causes scaling issues
on larger SMP machines.  This patch replaces that algorithm with a
2-dimensional bitmap to reduce latencies in the wake-up path.

Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Acked-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2008-06-06 15:19:28 +02:00