Skip to content

K
Kernel

openeuler / Kernel
12 个月 前同步成功

通知 5
Star 0
Fork 0
K
Kernel

体验新版 GitCode，发现更多精彩内容 >>

切换分支/标签
  1. 19 4月, 2013 1 次提交
    • W
      mutex: Move mutex spinning code from sched/core.c back to mutex.c · 41fcb9f2
      由 Waiman Long 提交于 
      As mentioned by Ingo, the SCHED_FEAT_OWNER_SPIN scheduler
feature bit was really just an early hack to make with/without
mutex-spinning testable. So it is no longer necessary.

This patch removes the SCHED_FEAT_OWNER_SPIN feature bit and
move the mutex spinning code from kernel/sched/core.c back to
kernel/mutex.c which is where they should belong.
Signed-off-by: NWaiman Long <Waiman.Long@hp.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Chandramouleeswaran Aswin <aswin@hp.com>
Cc: Davidlohr Bueso <davidlohr.bueso@hp.com>
Cc: Norton Scott J <scott.norton@hp.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Dave Jones <davej@redhat.com>
Cc: Clark Williams <williams@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1366226594-5506-2-git-send-email-Waiman.Long@hp.comSigned-off-by: NIngo Molnar <mingo@kernel.org>
      41fcb9f2
  3. 12 4月, 2013 1 次提交
    • T
      kthread: Prevent unpark race which puts threads on the wrong cpu · f2530dc7
      由 Thomas Gleixner 提交于 
      The smpboot threads rely on the park/unpark mechanism which binds per
cpu threads on a particular core. Though the functionality is racy:

CPU0	       	 	CPU1  	     	    CPU2
unpark(T)				    wake_up_process(T)
  clear(SHOULD_PARK)	T runs
			leave parkme() due to !SHOULD_PARK  
  bind_to(CPU2)		BUG_ON(wrong CPU)						    

We cannot let the tasks move themself to the target CPU as one of
those tasks is actually the migration thread itself, which requires
that it starts running on the target cpu right away.

The solution to this problem is to prevent wakeups in park mode which
are not from unpark(). That way we can guarantee that the association
of the task to the target cpu is working correctly.

Add a new task state (TASK_PARKED) which prevents other wakeups and
use this state explicitly for the unpark wakeup.

Peter noticed: Also, since the task state is visible to userspace and
all the parked tasks are still in the PID space, its a good hint in ps
and friends that these tasks aren't really there for the moment.

The migration thread has another related issue.

CPU0	      	     	 CPU1
Bring up CPU2
create_thread(T)
park(T)
 wait_for_completion()
			 parkme()
			 complete()
sched_set_stop_task()
			 schedule(TASK_PARKED)

The sched_set_stop_task() call is issued while the task is on the
runqueue of CPU1 and that confuses the hell out of the stop_task class
on that cpu. So we need the same synchronizaion before
sched_set_stop_task().
Reported-by: NDave Jones <davej@redhat.com>
Reported-and-tested-by: NDave Hansen <dave@sr71.net>
Reported-and-tested-by: NBorislav Petkov <bp@alien8.de>
Acked-by: NPeter Ziljstra <peterz@infradead.org>
Cc: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Cc: dhillf@gmail.com
Cc: Ingo Molnar <mingo@kernel.org>
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/alpine.LFD.2.02.1304091635430.21884@ionosSigned-off-by: NThomas Gleixner <tglx@linutronix.de>
      f2530dc7
  5. 20 3月, 2013 1 次提交
    • T
      sched: replace PF_THREAD_BOUND with PF_NO_SETAFFINITY · 14a40ffc
      由 Tejun Heo 提交于 
      PF_THREAD_BOUND was originally used to mark kernel threads which were
bound to a specific CPU using kthread_bind() and a task with the flag
set allows cpus_allowed modifications only to itself.  Workqueue is
currently abusing it to prevent userland from meddling with
cpus_allowed of workqueue workers.

What we need is a flag to prevent userland from messing with
cpus_allowed of certain kernel tasks.  In kernel, anyone can
(incorrectly) squash the flag, and, for worker-type usages,
restricting cpus_allowed modification to the task itself doesn't
provide meaningful extra proection as other tasks can inject work
items to the task anyway.

This patch replaces PF_THREAD_BOUND with PF_NO_SETAFFINITY.
sched_setaffinity() checks the flag and return -EINVAL if set.
set_cpus_allowed_ptr() is no longer affected by the flag.

This will allow simplifying workqueue worker CPU affinity management.
Signed-off-by: NTejun Heo <tj@kernel.org>
Acked-by: NIngo Molnar <mingo@kernel.org>
Reviewed-by: NLai Jiangshan <laijs@cn.fujitsu.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
      14a40ffc
  7. 28 2月, 2013 1 次提交
  9. 24 2月, 2013 1 次提交
    • M
      mm: teach mm by current context info to not do I/O during memory allocation · 21caf2fc
      由 Ming Lei 提交于 
      This patch introduces PF_MEMALLOC_NOIO on process flag('flags' field of
'struct task_struct'), so that the flag can be set by one task to avoid
doing I/O inside memory allocation in the task's context.

The patch trys to solve one deadlock problem caused by block device, and
the problem may happen at least in the below situations:

- during block device runtime resume, if memory allocation with
  GFP_KERNEL is called inside runtime resume callback of any one of its
  ancestors(or the block device itself), the deadlock may be triggered
  inside the memory allocation since it might not complete until the block
  device becomes active and the involed page I/O finishes.  The situation
  is pointed out first by Alan Stern.  It is not a good approach to
  convert all GFP_KERNEL[1] in the path into GFP_NOIO because several
  subsystems may be involved(for example, PCI, USB and SCSI may be
  involved for usb mass stoarage device, network devices involved too in
  the iSCSI case)

- during block device runtime suspend, because runtime resume need to
  wait for completion of concurrent runtime suspend.

- during error handling of usb mass storage deivce, USB bus reset will
  be put on the device, so there shouldn't have any memory allocation with
  GFP_KERNEL during USB bus reset, otherwise the deadlock similar with
  above may be triggered.  Unfortunately, any usb device may include one
  mass storage interface in theory, so it requires all usb interface
  drivers to handle the situation.  In fact, most usb drivers don't know
  how to handle bus reset on the device and don't provide .pre_set() and
  .post_reset() callback at all, so USB core has to unbind and bind driver
  for these devices.  So it is still not practical to resort to GFP_NOIO
  for solving the problem.

Also the introduced solution can be used by block subsystem or block
drivers too, for example, set the PF_MEMALLOC_NOIO flag before doing
actual I/O transfer.

It is not a good idea to convert all these GFP_KERNEL in the affected
path into GFP_NOIO because these functions doing that may be implemented
as library and will be called in many other contexts.

In fact, memalloc_noio_flags() can convert some of current static
GFP_NOIO allocation into GFP_KERNEL back in other non-affected contexts,
at least almost all GFP_NOIO in USB subsystem can be converted into
GFP_KERNEL after applying the approach and make allocation with GFP_NOIO
only happen in runtime resume/bus reset/block I/O transfer contexts
generally.

[1], several GFP_KERNEL allocation examples in runtime resume path

- pci subsystem
acpi_os_allocate
	<-acpi_ut_allocate
		<-ACPI_ALLOCATE_ZEROED
			<-acpi_evaluate_object
				<-__acpi_bus_set_power
					<-acpi_bus_set_power
						<-acpi_pci_set_power_state
							<-platform_pci_set_power_state
								<-pci_platform_power_transition
									<-__pci_complete_power_transition
										<-pci_set_power_state
											<-pci_restore_standard_config
												<-pci_pm_runtime_resume
- usb subsystem
usb_get_status
	<-finish_port_resume
		<-usb_port_resume
			<-generic_resume
				<-usb_resume_device
					<-usb_resume_both
						<-usb_runtime_resume

- some individual usb drivers
usblp, uvc, gspca, most of dvb-usb-v2 media drivers, cpia2, az6007, ....

That is just what I have found.  Unfortunately, this allocation can only
be found by human being now, and there should be many not found since
any function in the resume path(call tree) may allocate memory with
GFP_KERNEL.
Signed-off-by: NMing Lei <ming.lei@canonical.com>
Signed-off-by: NMinchan Kim <minchan@kernel.org>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Oliver Neukum <oneukum@suse.de>
Cc: Jiri Kosina <jiri.kosina@suse.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Rafael J. Wysocki" <rjw@sisk.pl>
Cc: Greg KH <greg@kroah.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Decotigny <david.decotigny@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      21caf2fc
  11. 20 2月, 2013 1 次提交
  13. 14 2月, 2013 1 次提交
  15. 08 2月, 2013 2 次提交
  17. 04 2月, 2013 1 次提交
  19. 28 1月, 2013 2 次提交
    • F
      cputime: Safely read cputime of full dynticks CPUs · 6a61671b
      由 Frederic Weisbecker 提交于 
      While remotely reading the cputime of a task running in a
full dynticks CPU, the values stored in utime/stime fields
of struct task_struct may be stale. Its values may be those
of the last kernel <-> user transition time snapshot and
we need to add the tickless time spent since this snapshot.

To fix this, flush the cputime of the dynticks CPUs on
kernel <-> user transition and record the time / context
where we did this. Then on top of this snapshot and the current
time, perform the fixup on the reader side from task_times()
accessors.
Signed-off-by: NFrederic Weisbecker <fweisbec@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Li Zhong <zhong@linux.vnet.ibm.com>
Cc: Namhyung Kim <namhyung.kim@lge.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Paul Gortmaker <paul.gortmaker@windriver.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
[fixed kvm module related build errors]
Signed-off-by: NSedat Dilek <sedat.dilek@gmail.com>
      6a61671b
    • F
      cputime: Use accessors to read task cputime stats · 6fac4829
      由 Frederic Weisbecker 提交于 
      This is in preparation for the full dynticks feature. While
remotely reading the cputime of a task running in a full
dynticks CPU, we'll need to do some extra-computation. This
way we can account the time it spent tickless in userspace
since its last cputime snapshot.
Signed-off-by: NFrederic Weisbecker <fweisbec@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Li Zhong <zhong@linux.vnet.ibm.com>
Cc: Namhyung Kim <namhyung.kim@lge.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Paul Gortmaker <paul.gortmaker@windriver.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
      6fac4829
  21. 25 1月, 2013 2 次提交
    • Y
      sched/rt: Avoid updating RT entry timeout twice within one tick period · 57d2aa00
      由 Ying Xue 提交于 
      The issue below was found in 2.6.34-rt rather than mainline rt
kernel, but the issue still exists upstream as well.

So please let me describe how it was noticed on 2.6.34-rt:

On this version, each softirq has its own thread, it means there
is at least one RT FIFO task per cpu. The priority of these
tasks is set to 49 by default. If user launches an RT FIFO task
with priority lower than 49 of softirq RT tasks, it's possible
there are two RT FIFO tasks enqueued one cpu runqueue at one
moment. By current strategy of balancing RT tasks, when it comes
to RT tasks, we really need to put them off to a CPU that they
can run on as soon as possible. Even if it means a bit of cache
line flushing, we want RT tasks to be run with the least latency.

When the user RT FIFO task which just launched before is
running, the sched timer tick of the current cpu happens. In this
tick period, the timeout value of the user RT task will be
updated once. Subsequently, we try to wake up one softirq RT
task on its local cpu. As the priority of current user RT task
is lower than the softirq RT task, the current task will be
preempted by the higher priority softirq RT task. Before
preemption, we check to see if current can readily move to a
different cpu. If so, we will reschedule to allow the RT push logic
to try to move current somewhere else. Whenever the woken
softirq RT task runs, it first tries to migrate the user FIFO RT
task over to a cpu that is running a task of lesser priority. If
migration is done, it will send a reschedule request to the found
cpu by IPI interrupt. Once the target cpu responds the IPI
interrupt, it will pick the migrated user RT task to preempt its
current task. When the user RT task is running on the new cpu,
the sched timer tick of the cpu fires. So it will tick the user
RT task again. This also means the RT task timeout value will be
updated again. As the migration may be done in one tick period,
it means the user RT task timeout value will be updated twice
within one tick.

If we set a limit on the amount of cpu time for the user RT task
by setrlimit(RLIMIT_RTTIME), the SIGXCPU signal should be posted
upon reaching the soft limit.

But exactly when the SIGXCPU signal should be sent depends on the
RT task timeout value. In fact the timeout mechanism of sending
the SIGXCPU signal assumes the RT task timeout is increased once
every tick.

However, currently the timeout value may be added twice per
tick. So it results in the SIGXCPU signal being sent earlier
than expected.

To solve this issue, we prevent the timeout value from increasing
twice within one tick time by remembering the jiffies value of
last updating the timeout. As long as the RT task's jiffies is
different with the global jiffies value, we allow its timeout to
be updated.
Signed-off-by: NYing Xue <ying.xue@windriver.com>
Signed-off-by: NFan Du <fan.du@windriver.com>
Reviewed-by: NYong Zhang <yong.zhang0@gmail.com>
Acked-by: NSteven Rostedt <rostedt@goodmis.org>
Cc: <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1342508623-2887-1-git-send-email-ying.xue@windriver.comSigned-off-by: NIngo Molnar <mingo@kernel.org>
      57d2aa00
    • L
      sched: split out css_online/css_offline from tg creation/destruction · ace783b9
      由 Li Zefan 提交于 
      This is a preparaton for later patches.

- What do we gain from cpu_cgroup_css_online():

After ss->css_alloc() and before ss->css_online(), there's a small
window that tg->css.cgroup is NULL. With this change, tg won't be seen
before ss->css_online(), where it's added to the global list, so we're
guaranteed we'll never see NULL tg->css.cgroup.

- What do we gain from cpu_cgroup_css_offline():

tg is freed via RCU, so is cgroup. Without this change, This is how
synchronization works:

cgroup_rmdir()
  no ss->css_offline()
diput()
  syncornize_rcu()
  ss->css_free()       <-- unregister tg, and free it via call_rcu()
  kfree_rcu(cgroup)    <-- wait possible refs to cgroup, and free cgroup

We can't just kfree(cgroup), because tg might access tg->css.cgroup.

With this change:

cgroup_rmdir()
  ss->css_offline()    <-- unregister tg
diput()
  synchronize_rcu()    <-- wait possible refs to tg and cgroup
  ss->css_free()       <-- free tg
  kfree_rcu(cgroup)    <-- free cgroup

As you see, kfree_rcu() is redundant now.
Signed-off-by: NLi Zefan <lizefan@huawei.com>
Signed-off-by: NTejun Heo <tj@kernel.org>
Acked-by: NIngo Molnar <mingo@kernel.org>
      ace783b9
  23. 23 1月, 2013 1 次提交
  25. 17 1月, 2013 1 次提交
    • T
      module, async: async_synchronize_full() on module init iff async is used · 774a1221
      由 Tejun Heo 提交于 
      If the default iosched is built as module, the kernel may deadlock
while trying to load the iosched module on device probe if the probing
was running off async.  This is because async_synchronize_full() at
the end of module init ends up waiting for the async job which
initiated the module loading.

 async A				modprobe

 1. finds a device
 2. registers the block device
 3. request_module(default iosched)
					4. modprobe in userland
					5. load and init module
					6. async_synchronize_full()

Async A waits for modprobe to finish in request_module() and modprobe
waits for async A to finish in async_synchronize_full().

Because there's no easy to track dependency once control goes out to
userland, implementing properly nested flushing is difficult.  For
now, make module init perform async_synchronize_full() iff module init
has queued async jobs as suggested by Linus.

This avoids the described deadlock because iosched module doesn't use
async and thus wouldn't invoke async_synchronize_full().  This is
hacky and incomplete.  It will deadlock if async module loading nests;
however, this works around the known problem case and seems to be the
best of bad options.

For more details, please refer to the following thread.

  http://thread.gmane.org/gmane.linux.kernel/1420814Signed-off-by: NTejun Heo <tj@kernel.org>
Reported-by: NAlex Riesen <raa.lkml@gmail.com>
Tested-by: NMing Lei <ming.lei@canonical.com>
Tested-by: NAlex Riesen <raa.lkml@gmail.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      774a1221
  27. 20 12月, 2012 1 次提交
  29. 19 12月, 2012 1 次提交
    • G
      memcg: skip memcg kmem allocations in specified code regions · 0e9d92f2
      由 Glauber Costa 提交于 
      Create a mechanism that skip memcg allocations during certain pieces of
our core code.  It basically works in the same way as
preempt_disable()/preempt_enable(): By marking a region under which all
allocations will be accounted to the root memcg.

We need this to prevent races in early cache creation, when we
allocate data using caches that are not necessarily created already.
Signed-off-by: NGlauber Costa <glommer@parallels.com>
yCc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Frederic Weisbecker <fweisbec@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: JoonSoo Kim <js1304@gmail.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Rik van Riel <riel@redhat.com>
Cc: Suleiman Souhlal <suleiman@google.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      0e9d92f2
  31. 18 12月, 2012 1 次提交
  33. 12 12月, 2012 2 次提交
    • D
      mm, oom: fix race when specifying a thread as the oom origin · e1e12d2f
      由 David Rientjes 提交于 
      test_set_oom_score_adj() and compare_swap_oom_score_adj() are used to
specify that current should be killed first if an oom condition occurs in
between the two calls.

The usage is

	short oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
	...
	compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX, oom_score_adj);

to store the thread's oom_score_adj, temporarily change it to the maximum
score possible, and then restore the old value if it is still the same.

This happens to still be racy, however, if the user writes
OOM_SCORE_ADJ_MAX to /proc/pid/oom_score_adj in between the two calls.
The compare_swap_oom_score_adj() will then incorrectly reset the old value
prior to the write of OOM_SCORE_ADJ_MAX.

To fix this, introduce a new oom_flags_t member in struct signal_struct
that will be used for per-thread oom killer flags.  KSM and swapoff can
now use a bit in this member to specify that threads should be killed
first in oom conditions without playing around with oom_score_adj.

This also allows the correct oom_score_adj to always be shown when reading
/proc/pid/oom_score.
Signed-off-by: NDavid Rientjes <rientjes@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: NMichal Hocko <mhocko@suse.cz>
Cc: Anton Vorontsov <anton.vorontsov@linaro.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      e1e12d2f
    • D
      mm, oom: change type of oom_score_adj to short · a9c58b90
      由 David Rientjes 提交于 
      The maximum oom_score_adj is 1000 and the minimum oom_score_adj is -1000,
so this range can be represented by the signed short type with no
functional change.  The extra space this frees up in struct signal_struct
will be used for per-thread oom kill flags in the next patch.
Signed-off-by: NDavid Rientjes <rientjes@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: NMichal Hocko <mhocko@suse.cz>
Cc: Anton Vorontsov <anton.vorontsov@linaro.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      a9c58b90
  35. 11 12月, 2012 5 次提交
    • M
      mm: sched: numa: Control enabling and disabling of NUMA balancing · 1a687c2e
      由 Mel Gorman 提交于 
      This patch adds Kconfig options and kernel parameters to allow the
enabling and disabling of automatic NUMA balancing. The existance
of such a switch was and is very important when debugging problems
related to transparent hugepages and we should have the same for
automatic NUMA placement.
Signed-off-by: NMel Gorman <mgorman@suse.de>
      1a687c2e
    • M
      mm: sched: Adapt the scanning rate if a NUMA hinting fault does not migrate · b8593bfd
      由 Mel Gorman 提交于 
      The PTE scanning rate and fault rates are two of the biggest sources of
system CPU overhead with automatic NUMA placement.  Ideally a proper policy
would detect if a workload was properly placed, schedule and adjust the
PTE scanning rate accordingly. We do not track the necessary information
to do that but we at least know if we migrated or not.

This patch scans slower if a page was not migrated as the result of a
NUMA hinting fault up to sysctl_numa_balancing_scan_period_max which is
now higher than the previous default. Once every minute it will reset
the scanner in case of phase changes.

This is hilariously crude and the numbers are arbitrary. Workloads will
converge quite slowly in comparison to what a proper policy should be able
to do. On the plus side, we will chew up less CPU for workloads that have
no need for automatic balancing.
Signed-off-by: NMel Gorman <mgorman@suse.de>
      b8593bfd
    • P
      mm: sched: numa: Implement slow start for working set sampling · 4b96a29b
      由 Peter Zijlstra 提交于 
      Add a 1 second delay before starting to scan the working set of
a task and starting to balance it amongst nodes.

[ note that before the constant per task WSS sampling rate patch
  the initial scan would happen much later still, in effect that
  patch caused this regression. ]

The theory is that short-run tasks benefit very little from NUMA
placement: they come and go, and they better stick to the node
they were started on. As tasks mature and rebalance to other CPUs
and nodes, so does their NUMA placement have to change and so
does it start to matter more and more.

In practice this change fixes an observable kbuild regression:

   # [ a perf stat --null --repeat 10 test of ten bzImage builds to /dev/shm ]

   !NUMA:
   45.291088843 seconds time elapsed                                          ( +-  0.40% )
   45.154231752 seconds time elapsed                                          ( +-  0.36% )

   +NUMA, no slow start:
   46.172308123 seconds time elapsed                                          ( +-  0.30% )
   46.343168745 seconds time elapsed                                          ( +-  0.25% )

   +NUMA, 1 sec slow start:
   45.224189155 seconds time elapsed                                          ( +-  0.25% )
   45.160866532 seconds time elapsed                                          ( +-  0.17% )

and it also fixes an observable perf bench (hackbench) regression:

   # perf stat --null --repeat 10 perf bench sched messaging

   -NUMA:

   -NUMA:                  0.246225691 seconds time elapsed                   ( +-  1.31% )
   +NUMA no slow start:    0.252620063 seconds time elapsed                   ( +-  1.13% )

   +NUMA 1sec delay:       0.248076230 seconds time elapsed                   ( +-  1.35% )

The implementation is simple and straightforward, most of the patch
deals with adding the /proc/sys/kernel/numa_balancing_scan_delay_ms tunable
knob.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
[ Wrote the changelog, ran measurements, tuned the default. ]
Signed-off-by: NIngo Molnar <mingo@kernel.org>
Signed-off-by: NMel Gorman <mgorman@suse.de>
Reviewed-by: NRik van Riel <riel@redhat.com>
      4b96a29b
    • P
      mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate · 6e5fb223
      由 Peter Zijlstra 提交于 
      Previously, to probe the working set of a task, we'd use
a very simple and crude method: mark all of its address
space PROT_NONE.

That method has various (obvious) disadvantages:

 - it samples the working set at dissimilar rates,
   giving some tasks a sampling quality advantage
   over others.

 - creates performance problems for tasks with very
   large working sets

 - over-samples processes with large address spaces but
   which only very rarely execute

Improve that method by keeping a rotating offset into the
address space that marks the current position of the scan,
and advance it by a constant rate (in a CPU cycles execution
proportional manner). If the offset reaches the last mapped
address of the mm then it then it starts over at the first
address.

The per-task nature of the working set sampling functionality in this tree
allows such constant rate, per task, execution-weight proportional sampling
of the working set, with an adaptive sampling interval/frequency that
goes from once per 100ms up to just once per 8 seconds.  The current
sampling volume is 256 MB per interval.

As tasks mature and converge their working set, so does the
sampling rate slow down to just a trickle, 256 MB per 8
seconds of CPU time executed.

This, beyond being adaptive, also rate-limits rarely
executing systems and does not over-sample on overloaded
systems.

[ In AutoNUMA speak, this patch deals with the effective sampling
  rate of the 'hinting page fault'. AutoNUMA's scanning is
  currently rate-limited, but it is also fundamentally
  single-threaded, executing in the knuma_scand kernel thread,
  so the limit in AutoNUMA is global and does not scale up with
  the number of CPUs, nor does it scan tasks in an execution
  proportional manner.

  So the idea of rate-limiting the scanning was first implemented
  in the AutoNUMA tree via a global rate limit. This patch goes
  beyond that by implementing an execution rate proportional
  working set sampling rate that is not implemented via a single
  global scanning daemon. ]

[ Dan Carpenter pointed out a possible NULL pointer dereference in the
  first version of this patch. ]
Based-on-idea-by: NAndrea Arcangeli <aarcange@redhat.com>
Bug-Found-By: NDan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
[ Wrote changelog and fixed bug. ]
Signed-off-by: NIngo Molnar <mingo@kernel.org>
Signed-off-by: NMel Gorman <mgorman@suse.de>
Reviewed-by: NRik van Riel <riel@redhat.com>
      6e5fb223
    • P
      mm: numa: Add fault driven placement and migration · cbee9f88
      由 Peter Zijlstra 提交于 
      NOTE: This patch is based on "sched, numa, mm: Add fault driven
	placement and migration policy" but as it throws away all the policy
	to just leave a basic foundation I had to drop the signed-offs-by.

This patch creates a bare-bones method for setting PTEs pte_numa in the
context of the scheduler that when faulted later will be faulted onto the
node the CPU is running on.  In itself this does nothing useful but any
placement policy will fundamentally depend on receiving hints on placement
from fault context and doing something intelligent about it.
Signed-off-by: NMel Gorman <mgorman@suse.de>
Acked-by: NRik van Riel <riel@redhat.com>
      cbee9f88
  37. 29 11月, 2012 6 次提交
  39. 28 11月, 2012 1 次提交
  41. 24 10月, 2012 7 次提交