commit a10d206ef1 upstream.
Each grace period is supposed to have at least one callback waiting
for that grace period to complete. However, if CONFIG_NO_HZ=n, an
extra callback-free grace period is no big problem -- it will chew up
a tiny bit of CPU time, but it will complete normally. In contrast,
CONFIG_NO_HZ=y kernels have the potential for all the CPUs to go to
sleep indefinitely, in turn indefinitely delaying completion of the
callback-free grace period. Given that nothing is waiting on this grace
period, this is also not a problem.
That is, unless RCU CPU stall warnings are also enabled, as they are
in recent kernels. In this case, if a CPU wakes up after at least one
minute of inactivity, an RCU CPU stall warning will result. The reason
that no one noticed until quite recently is that most systems have enough
OS noise that they will never remain absolutely idle for a full minute.
But there are some embedded systems with cut-down userspace configurations
that consistently get into this situation.
All this begs the question of exactly how a callback-free grace period
gets started in the first place. This can happen due to the fact that
CPUs do not necessarily agree on which grace period is in progress.
If a CPU still believes that the grace period that just completed is
still ongoing, it will believe that it has callbacks that need to wait for
another grace period, never mind the fact that the grace period that they
were waiting for just completed. This CPU can therefore erroneously
decide to start a new grace period. Note that this can happen in
TREE_RCU and TREE_PREEMPT_RCU even on a single-CPU system: Deadlock
considerations mean that the CPU that detected the end of the grace
period is not necessarily officially informed of this fact for some time.
Once this CPU notices that the earlier grace period completed, it will
invoke its callbacks. It then won't have any callbacks left. If no
other CPU has any callbacks, we now have a callback-free grace period.
This commit therefore makes CPUs check more carefully before starting a
new grace period. This new check relies on an array of tail pointers
into each CPU's list of callbacks. If the CPU is up to date on which
grace periods have completed, it checks to see if any callbacks follow
the RCU_DONE_TAIL segment, otherwise it checks to see if any callbacks
follow the RCU_WAIT_TAIL segment. The reason that this works is that
the RCU_WAIT_TAIL segment will be promoted to the RCU_DONE_TAIL segment
as soon as the CPU is officially notified that the old grace period
has ended.
This change is to cpu_needs_another_gp(), which is called in a number
of places. The only one that really matters is in rcu_start_gp(), where
the root rcu_node structure's ->lock is held, which prevents any
other CPU from starting or completing a grace period, so that the
comparison that determines whether the CPU is missing the completion
of a grace period is stable.
Reported-by: Becky Bruce <bgillbruce@gmail.com>
Reported-by: Subodh Nijsure <snijsure@grid-net.com>
Reported-by: Paul Walmsley <paul@pwsan.com>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Paul Walmsley <paul@pwsan.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 959d1af8cf upstream.
WORK_STRUCT_PENDING is used to claim ownership of a work item and
process_one_work() releases it before starting execution. When
someone else grabs PENDING, all pre-release updates to the work item
should be visible and all updates made by the new owner should happen
afterwards.
Grabbing PENDING uses test_and_set_bit() and thus has a full barrier;
however, clearing doesn't have a matching wmb. Given the preceding
spin_unlock and use of clear_bit, I don't believe this can be a
problem on an actual machine and there hasn't been any related report
but it still is theretically possible for clear_pending to permeate
upwards and happen before work->entry update.
Add an explicit smp_wmb() before work_clear_pending().
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit f96972f2dc upstream.
As kernel_power_off() calls disable_nonboot_cpus(), we may also want to
have kernel_restart() call disable_nonboot_cpus(). Doing so can help
machines that require boot cpu be the last alive cpu during reboot to
survive with kernel restart.
This fixes one reboot issue seen on imx6q (Cortex-A9 Quad). The machine
requires that the restart routine be run on the primary cpu rather than
secondary ones. Otherwise, the secondary core running the restart
routine will fail to come to online after reboot.
Signed-off-by: Shawn Guo <shawn.guo@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit b5740f4b2c upstream.
try_to_wake_up() has a problem which may change status from TASK_DEAD to
TASK_RUNNING in race condition with SMI or guest environment of virtual
machine. As a result, exited task is scheduled() again and panic occurs.
Here is the sequence how it occurs:
----------------------------------+-----------------------------
|
CPU A | CPU B
----------------------------------+-----------------------------
TASK A calls exit()....
do_exit()
exit_mm()
down_read(mm->mmap_sem);
rwsem_down_failed_common()
set TASK_UNINTERRUPTIBLE
set waiter.task <= task A
list_add to sem->wait_list
:
raw_spin_unlock_irq()
(I/O interruption occured)
__rwsem_do_wake(mmap_sem)
list_del(&waiter->list);
waiter->task = NULL
wake_up_process(task A)
try_to_wake_up()
(task is still
TASK_UNINTERRUPTIBLE)
p->on_rq is still 1.)
ttwu_do_wakeup()
(*A)
:
(I/O interruption handler finished)
if (!waiter.task)
schedule() is not called
due to waiter.task is NULL.
tsk->state = TASK_RUNNING
:
check_preempt_curr();
:
task->state = TASK_DEAD
(*B)
<--- set TASK_RUNNING (*C)
schedule()
(exit task is running again)
BUG_ON() is called!
--------------------------------------------------------
The execution time between (*A) and (*B) is usually very short,
because the interruption is disabled, and setting TASK_RUNNING at (*C)
must be executed before setting TASK_DEAD.
HOWEVER, if SMI is interrupted between (*A) and (*B),
(*C) is able to execute AFTER setting TASK_DEAD!
Then, exited task is scheduled again, and BUG_ON() is called....
If the system works on guest system of virtual machine, the time
between (*A) and (*B) may be also long due to scheduling of hypervisor,
and same phenomenon can occur.
By this patch, do_exit() waits for releasing task->pi_lock which is used
in try_to_wake_up(). It guarantees the task becomes TASK_DEAD after
waking up.
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Acked-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/20120117174031.3118.E1E9C6FF@jp.fujitsu.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Cc: Michal Hocko <mhocko@suse.cz>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit cee58483cf upstream
Andreas Bombe reported that the added ktime_t overflow checking added to
timespec_valid in commit 4e8b14526c ("time: Improve sanity checking of
timekeeping inputs") was causing problems with X.org because it caused
timeouts larger then KTIME_T to be invalid.
Previously, these large timeouts would be clamped to KTIME_MAX and would
never expire, which is valid.
This patch splits the ktime_t overflow checking into a new
timespec_valid_strict function, and converts the timekeeping codes
internal checking to use this more strict function.
Reported-and-tested-by: Andreas Bombe <aeb@debian.org>
Cc: Zhouping Liu <zliu@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Prarit Bhargava <prarit@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 4e8b14526c upstream
Unexpected behavior could occur if the time is set to a value large
enough to overflow a 64bit ktime_t (which is something larger then the
year 2262).
Also unexpected behavior could occur if large negative offsets are
injected via adjtimex.
So this patch improves the sanity check timekeeping inputs by
improving the timespec_valid() check, and then makes better use of
timespec_valid() to make sure we don't set the time to an invalid
negative value or one that overflows ktime_t.
Note: This does not protect from setting the time close to overflowing
ktime_t and then letting natural accumulation cause the overflow.
Reported-by: CAI Qian <caiqian@redhat.com>
Reported-by: Sasha Levin <levinsasha928@gmail.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Prarit Bhargava <prarit@redhat.com>
Cc: Zhouping Liu <zliu@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Link: http://lkml.kernel.org/r/1344454580-17031-1-git-send-email-john.stultz@linaro.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 8323f26ce3 upstream.
Stefan reported a crash on a kernel before a3e5d1091c ("sched:
Don't call task_group() too many times in set_task_rq()"), he
found the reason to be that the multiple task_group()
invocations in set_task_rq() returned different values.
Looking at all that I found a lack of serialization and plain
wrong comments.
The below tries to fix it using an extra pointer which is
updated under the appropriate scheduler locks. Its not pretty,
but I can't really see another way given how all the cgroup
stuff works.
Reported-and-tested-by: Stefan Bader <stefan.bader@canonical.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1340364965.18025.71.camel@twins
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[Fixed upstream by commits 2955b47d2c and
a4683487f9 from Dan Williams, but they are much
more intrusive than this tiny fix, according to Andrew - gregkh]
This patch tries to fix a dead loop in async_synchronize_full(), which
could be seen when preemption is disabled on a single cpu machine.
void async_synchronize_full(void)
{
do {
async_synchronize_cookie(next_cookie);
} while (!list_empty(&async_running) || !
list_empty(&async_pending));
}
async_synchronize_cookie() calls async_synchronize_cookie_domain() with
&async_running as the default domain to synchronize.
However, there might be some works in the async_pending list from other
domains. On a single cpu system, without preemption, there is no chance
for the other works to finish, so async_synchronize_full() enters a dead
loop.
It seems async_synchronize_full() wants to synchronize all entries in
all running lists(domains), so maybe we could just check the entry_count
to know whether all works are finished.
Currently, async_synchronize_cookie_domain() expects a non-NULL running
list ( if NULL, there would be NULL pointer dereference ), so maybe a
NULL pointer could be used as an indication for the functions to
synchronize all works in all domains.
Reported-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Li Zhong <zhong@linux.vnet.ibm.com>
Tested-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Christian Kujau <lists@nerdbynature.de>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Dan Williams <dan.j.williams@gmail.com>
Cc: Christian Kujau <lists@nerdbynature.de>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 96e65306b8 upstream.
The compiler may compile the following code into TWO write/modify
instructions.
worker->flags &= ~WORKER_UNBOUND;
worker->flags |= WORKER_REBIND;
so the other CPU may temporarily see worker->flags which doesn't have
either WORKER_UNBOUND or WORKER_REBIND set and perform local wakeup
prematurely.
Fix it by using single explicit assignment via ACCESS_ONCE().
Because idle workers have another WORKER_NOT_RUNNING flag, this bug
doesn't exist for them; however, update it to use the same pattern for
consistency.
tj: Applied the change to idle workers too and updated comments and
patch description a bit.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit a6fa941d94 upstream.
Don't mess with file refcounts (or keep a reference to file, for
that matter) in perf_event. Use explicit refcount of its own
instead. Deal with the race between the final reference to event
going away and new children getting created for it by use of
atomic_long_inc_not_zero() in inherit_event(); just have the
latter free what it had allocated and return NULL, that works
out just fine (children of siblings of something doomed are
created as singletons, same as if the child of leader had been
created and immediately killed).
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20120820135925.GG23464@ZenIV.linux.org.uk
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit ed48ece27c upstream.
The existing work_on_cpu() implementation is hugely inefficient. It
creates a new kthread, execute that single function and then let the
kthread die on each invocation.
Now that system_wq can handle concurrent executions, there's no
advantage of doing this. Reimplement work_on_cpu() using system_wq
which makes it simpler and way more efficient.
stable: While this isn't a fix in itself, it's needed to fix a
workqueue related bug in cpufreq/powernow-k8. AFAICS, this
shouldn't break other existing users.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Jiri Kosina <jkosina@suse.cz>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Cc: Len Brown <lenb@kernel.org>
Cc: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit a2140fc0cb upstream.
Refcounting of fsnotify_mark in audit tree is broken. E.g:
refcount
create_chunk
alloc_chunk 1
fsnotify_add_mark 2
untag_chunk
fsnotify_get_mark 3
fsnotify_destroy_mark
audit_tree_freeing_mark 2
fsnotify_put_mark 1
fsnotify_put_mark 0
via destroy_list
fsnotify_mark_destroy -1
This was reported by various people as triggering Oops when stopping auditd.
We could just remove the put_mark from audit_tree_freeing_mark() but that would
break freeing via inode destruction. So this patch simply omits a put_mark
after calling destroy_mark or adds a get_mark before.
The additional get_mark is necessary where there's no other put_mark after
fsnotify_destroy_mark() since it assumes that the caller is holding a reference
(or the inode is keeping the mark pinned, not the case here AFAICS).
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Reported-by: Valentin Avram <aval13@gmail.com>
Reported-by: Peter Moody <pmoody@google.com>
Acked-by: Eric Paris <eparis@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 0fe33aae0e upstream.
Don't do free_chunk() after fsnotify_add_mark(). That one does a delayed unref
via the destroy list and this results in use-after-free.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Acked-by: Eric Paris <eparis@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit c5857ccf29 upstream.
With the new interrupt sampling system, we are no longer using the
timer_rand_state structure in the irq descriptor, so we can stop
initializing it now.
[ Merged in fixes from Sedat to find some last missing references to
rand_initialize_irq() ]
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Signed-off-by: Sedat Dilek <sedat.dilek@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 775f4b297b upstream.
We've been moving away from add_interrupt_randomness() for various
reasons: it's too expensive to do on every interrupt, and flooding the
CPU with interrupts could theoretically cause bogus floods of entropy
from a somewhat externally controllable source.
This solves both problems by limiting the actual randomness addition
to just once a second or after 64 interrupts, whicever comes first.
During that time, the interrupt cycle data is buffered up in a per-cpu
pool. Also, we make sure the the nonblocking pool used by urandom is
initialized before we start feeding the normal input pool. This
assures that /dev/urandom is returning unpredictable data as soon as
possible.
(Based on an original patch by Linus, but significantly modified by
tytso.)
Tested-by: Eric Wustrow <ewust@umich.edu>
Reported-by: Eric Wustrow <ewust@umich.edu>
Reported-by: Nadia Heninger <nadiah@cs.ucsd.edu>
Reported-by: Zakir Durumeric <zakir@umich.edu>
Reported-by: J. Alex Halderman <jhalderm@umich.edu>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 6575820221 upstream.
Currently, all workqueue cpu hotplug operations run off
CPU_PRI_WORKQUEUE which is higher than normal notifiers. This is to
ensure that workqueue is up and running while bringing up a CPU before
other notifiers try to use workqueue on the CPU.
Per-cpu workqueues are supposed to remain working and bound to the CPU
for normal CPU_DOWN_PREPARE notifiers. This holds mostly true even
with workqueue offlining running with higher priority because
workqueue CPU_DOWN_PREPARE only creates a bound trustee thread which
runs the per-cpu workqueue without concurrency management without
explicitly detaching the existing workers.
However, if the trustee needs to create new workers, it creates
unbound workers which may wander off to other CPUs while
CPU_DOWN_PREPARE notifiers are in progress. Furthermore, if the CPU
down is cancelled, the per-CPU workqueue may end up with workers which
aren't bound to the CPU.
While reliably reproducible with a convoluted artificial test-case
involving scheduling and flushing CPU burning work items from CPU down
notifiers, this isn't very likely to happen in the wild, and, even
when it happens, the effects are likely to be hidden by the following
successful CPU down.
Fix it by using different priorities for up and down notifiers - high
priority for up operations and low priority for down operations.
Workqueue cpu hotplug operations will soon go through further cleanup.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: "Rafael J. Wysocki" <rjw@sisk.pl>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 443772d408 upstream.
If function tracing is enabled for some of the low-level suspend/resume
functions, it leads to triple fault during resume from suspend, ultimately
ending up in a reboot instead of a resume (or a total refusal to come out
of suspended state, on some machines).
This issue was explained in more detail in commit f42ac38c59 (ftrace:
disable tracing for suspend to ram). However, the changes made by that commit
got reverted by commit cbe2f5a6e8 (tracing: allow tracing of
suspend/resume & hibernation code again). So, unfortunately since things are
not yet robust enough to allow tracing of low-level suspend/resume functions,
suspend/resume is still broken when ftrace is enabled.
So fix this by disabling function tracing during suspend/resume & hibernation.
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit cc9a6c8776 upstream.
Stable note: Not tracked in Bugzilla. [get|put]_mems_allowed() is extremely
expensive and severely impacted page allocator performance. This
is part of a series of patches that reduce page allocator overhead.
Commit c0ff7453bb ("cpuset,mm: fix no node to alloc memory when
changing cpuset's mems") wins a super prize for the largest number of
memory barriers entered into fast paths for one commit.
[get|put]_mems_allowed is incredibly heavy with pairs of full memory
barriers inserted into a number of hot paths. This was detected while
investigating at large page allocator slowdown introduced some time
after 2.6.32. The largest portion of this overhead was shown by
oprofile to be at an mfence introduced by this commit into the page
allocator hot path.
For extra style points, the commit introduced the use of yield() in an
implementation of what looks like a spinning mutex.
This patch replaces the full memory barriers on both read and write
sides with a sequence counter with just read barriers on the fast path
side. This is much cheaper on some architectures, including x86. The
main bulk of the patch is the retry logic if the nodemask changes in a
manner that can cause a false failure.
While updating the nodemask, a check is made to see if a false failure
is a risk. If it is, the sequence number gets bumped and parallel
allocators will briefly stall while the nodemask update takes place.
In a page fault test microbenchmark, oprofile samples from
__alloc_pages_nodemask went from 4.53% of all samples to 1.15%. The
actual results were
3.3.0-rc3 3.3.0-rc3
rc3-vanilla nobarrier-v2r1
Clients 1 UserTime 0.07 ( 0.00%) 0.08 (-14.19%)
Clients 2 UserTime 0.07 ( 0.00%) 0.07 ( 2.72%)
Clients 4 UserTime 0.08 ( 0.00%) 0.07 ( 3.29%)
Clients 1 SysTime 0.70 ( 0.00%) 0.65 ( 6.65%)
Clients 2 SysTime 0.85 ( 0.00%) 0.82 ( 3.65%)
Clients 4 SysTime 1.41 ( 0.00%) 1.41 ( 0.32%)
Clients 1 WallTime 0.77 ( 0.00%) 0.74 ( 4.19%)
Clients 2 WallTime 0.47 ( 0.00%) 0.45 ( 3.73%)
Clients 4 WallTime 0.38 ( 0.00%) 0.37 ( 1.58%)
Clients 1 Flt/sec/cpu 497620.28 ( 0.00%) 520294.53 ( 4.56%)
Clients 2 Flt/sec/cpu 414639.05 ( 0.00%) 429882.01 ( 3.68%)
Clients 4 Flt/sec/cpu 257959.16 ( 0.00%) 258761.48 ( 0.31%)
Clients 1 Flt/sec 495161.39 ( 0.00%) 517292.87 ( 4.47%)
Clients 2 Flt/sec 820325.95 ( 0.00%) 850289.77 ( 3.65%)
Clients 4 Flt/sec 1020068.93 ( 0.00%) 1022674.06 ( 0.26%)
MMTests Statistics: duration
Sys Time Running Test (seconds) 135.68 132.17
User+Sys Time Running Test (seconds) 164.2 160.13
Total Elapsed Time (seconds) 123.46 120.87
The overall improvement is small but the System CPU time is much
improved and roughly in correlation to what oprofile reported (these
performance figures are without profiling so skew is expected). The
actual number of page faults is noticeably improved.
For benchmarks like kernel builds, the overall benefit is marginal but
the system CPU time is slightly reduced.
To test the actual bug the commit fixed I opened two terminals. The
first ran within a cpuset and continually ran a small program that
faulted 100M of anonymous data. In a second window, the nodemask of the
cpuset was continually randomised in a loop.
Without the commit, the program would fail every so often (usually
within 10 seconds) and obviously with the commit everything worked fine.
With this patch applied, it also worked fine so the fix should be
functionally equivalent.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit b246272ecc upstream.
Stable note: Not tracked in Bugzilla. [get|put]_mems_allowed() is extremely
expensive and severely impacted page allocator performance. This is
part of a series of patches that reduce page allocator overhead.
Kernels where MAX_NUMNODES > BITS_PER_LONG may temporarily see an empty
nodemask in a tsk's mempolicy if its previous nodemask is remapped onto a
new set of allowed cpuset nodes where the two nodemasks, as a result of
the remap, are now disjoint.
c0ff7453bb ("cpuset,mm: fix no node to alloc memory when changing
cpuset's mems") adds get_mems_allowed() to prevent the set of allowed
nodes from changing for a thread. This causes any update to a set of
allowed nodes to stall until put_mems_allowed() is called.
This stall is unncessary, however, if at least one node remains unchanged
in the update to the set of allowed nodes. This was addressed by
89e8a244b9 ("cpusets: avoid looping when storing to mems_allowed if one
node remains set"), but it's still possible that an empty nodemask may be
read from a mempolicy because the old nodemask may be remapped to the new
nodemask during rebind. To prevent this, only avoid the stall if there is
no mempolicy for the thread being changed.
This is a temporary solution until all reads from mempolicy nodemasks can
be guaranteed to not be empty without the get_mems_allowed()
synchronization.
Also moves the check for nodemask intersection inside task_lock() so that
tsk->mems_allowed cannot change. This ensures that nothing can set this
tsk's mems_allowed out from under us and also protects tsk->mempolicy.
Reported-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Paul Menage <paul@paulmenage.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>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 89e8a244b9 upstream.
Stable note: Not tracked in Bugzilla. [get|put]_mems_allowed() is
extremely expensive and severely impacted page allocator performance.
This is part of a series of patches that reduce page allocator
overhead.
{get,put}_mems_allowed() exist so that general kernel code may locklessly
access a task's set of allowable nodes without having the chance that a
concurrent write will cause the nodemask to be empty on configurations
where MAX_NUMNODES > BITS_PER_LONG.
This could incur a significant delay, however, especially in low memory
conditions because the page allocator is blocking and reclaim requires
get_mems_allowed() itself. It is not atypical to see writes to
cpuset.mems take over 2 seconds to complete, for example. In low memory
conditions, this is problematic because it's one of the most imporant
times to change cpuset.mems in the first place!
The only way a task's set of allowable nodes may change is through cpusets
by writing to cpuset.mems and when attaching a task to a generic code is
not reading the nodemask with get_mems_allowed() at the same time, and
then clearing all the old nodes. This prevents the possibility that a
reader will see an empty nodemask at the same time the writer is storing a
new nodemask.
If at least one node remains unchanged, though, it's possible to simply
set all new nodes and then clear all the old nodes. Changing a task's
nodemask is protected by cgroup_mutex so it's guaranteed that two threads
are not changing the same task's nodemask at the same time, so the
nodemask is guaranteed to be stored before another thread changes it and
determines whether a node remains set or not.
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Paul Menage <paul@paulmenage.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
