While parallel applications tend to align their data on the cache
boundary, they tend not to align on the page or THP boundary.
Consequently tasks that partition their data can still "false-share"
pages presenting a problem for optimal NUMA placement.
This patch uses NUMA hinting faults to chain tasks together into
numa_groups. As well as storing the NID a task was running on when
accessing a page a truncated representation of the faulting PID is
stored. If subsequent faults are from different PIDs it is reasonable
to assume that those two tasks share a page and are candidates for
being grouped together. Note that this patch makes no scheduling
decisions based on the grouping information.
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/1381141781-10992-44-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Use the new stop_two_cpus() to implement migrate_swap(), a function that
flips two tasks between their respective cpus.
I'm fairly sure there's a less crude way than employing the stop_two_cpus()
method, but everything I tried either got horribly fragile and/or complex. So
keep it simple for now.
The notable detail is how we 'migrate' tasks that aren't runnable
anymore. We'll make it appear like we migrated them before they went to
sleep. The sole difference is the previous cpu in the wakeup path, so we
override this.
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Link: http://lkml.kernel.org/r/1381141781-10992-39-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When a preferred node is selected for a tasks there is an attempt to migrate
the task to a CPU there. This may fail in which case the task will only
migrate if the active load balancer takes action. This may never happen if
the conditions are not right. This patch will check at NUMA hinting fault
time if another attempt should be made to migrate the task. It will only
make an attempt once every five seconds.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1381141781-10992-34-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch favours moving tasks towards NUMA node that recorded a higher
number of NUMA faults during active load balancing. Ideally this is
self-reinforcing as the longer the task runs on that node, the more faults
it should incur causing task_numa_placement to keep the task running on that
node. In reality a big weakness is that the nodes CPUs can be overloaded
and it would be more efficient to queue tasks on an idle node and migrate
to the new node. This would require additional smarts in the balancer so
for now the balancer will simply prefer to place the task on the preferred
node for a PTE scans which is controlled by the numa_balancing_settle_count
sysctl. Once the settle_count number of scans has complete the schedule
is free to place the task on an alternative node if the load is imbalanced.
[srikar@linux.vnet.ibm.com: Fixed statistics]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
[ Tunable and use higher faults instead of preferred. ]
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1381141781-10992-23-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
NUMA hinting fault counts and placement decisions are both recorded in the
same array which distorts the samples in an unpredictable fashion. The values
linearly accumulate during the scan and then decay creating a sawtooth-like
pattern in the per-node counts. It also means that placement decisions are
time sensitive. At best it means that it is very difficult to state that
the buffer holds a decaying average of past faulting behaviour. At worst,
it can confuse the load balancer if it sees one node with an artifically high
count due to very recent faulting activity and may create a bouncing effect.
This patch adds a second array. numa_faults stores the historical data
which is used for placement decisions. numa_faults_buffer holds the
fault activity during the current scan window. When the scan completes,
numa_faults decays and the values from numa_faults_buffer are copied
across.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1381141781-10992-22-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The NUMA PTE scan rate is controlled with a combination of the
numa_balancing_scan_period_min, numa_balancing_scan_period_max and
numa_balancing_scan_size. This scan rate is independent of the size
of the task and as an aside it is further complicated by the fact that
numa_balancing_scan_size controls how many pages are marked pte_numa and
not how much virtual memory is scanned.
In combination, it is almost impossible to meaningfully tune the min and
max scan periods and reasoning about performance is complex when the time
to complete a full scan is is partially a function of the tasks memory
size. This patch alters the semantic of the min and max tunables to be
about tuning the length time it takes to complete a scan of a tasks occupied
virtual address space. Conceptually this is a lot easier to understand. There
is a "sanity" check to ensure the scan rate is never extremely fast based on
the amount of virtual memory that should be scanned in a second. The default
of 2.5G seems arbitrary but it is to have the maximum scan rate after the
patch roughly match the maximum scan rate before the patch was applied.
On a similar note, numa_scan_period is in milliseconds and not
jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies
to numa_scan_period means that the rate scanning slows depends on HZ which
is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When using per-cpu preempt_count variables we need to save/restore the
preempt_count on context switch (into per task storage; for instance
the old thread_info::preempt_count variable) because of
PREEMPT_ACTIVE.
However, this means that on fork() the preempt_count value of the last
context switch gets copied and if we had a PREEMPT_ACTIVE switch right
before cloning a child task the child task will now too have
PREEMPT_ACTIVE set and start its life with an extra PREEMPT_ACTIVE
count.
Therefore we need to make init_task_preempt_count() unconditional;
this resets whatever preempt_count we inherited from our parent
process.
Doing so for !per-cpu implementations is harmless.
For !PREEMPT_COUNT kernels we need to be careful not to start life
with an increased preempt_count.
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/n/tip-4k0b7oy1rcdyzochwiixuwi9@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Rewrite the preempt_count macros in order to extract the 3 basic
preempt_count value modifiers:
__preempt_count_add()
__preempt_count_sub()
and the new:
__preempt_count_dec_and_test()
And since we're at it anyway, replace the unconventional
$op_preempt_count names with the more conventional preempt_count_$op.
Since these basic operators are equivalent to the previous _notrace()
variants, do away with the _notrace() versions.
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/n/tip-ewbpdbupy9xpsjhg960zwbv8@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In order to combine the preemption and need_resched test we need to
fold the need_resched information into the preempt_count value.
Since the NEED_RESCHED flag is set across CPUs this needs to be an
atomic operation, however we very much want to avoid making
preempt_count atomic, therefore we keep the existing TIF_NEED_RESCHED
infrastructure in place but at 3 sites test it and fold its value into
preempt_count; namely:
- resched_task() when setting TIF_NEED_RESCHED on the current task
- scheduler_ipi() when resched_task() sets TIF_NEED_RESCHED on a
remote task it follows it up with a reschedule IPI
and we can modify the cpu local preempt_count from
there.
- cpu_idle_loop() for when resched_task() found tsk_is_polling().
We use an inverted bitmask to indicate need_resched so that a 0 means
both need_resched and !atomic.
Also remove the barrier() in preempt_enable() between
preempt_enable_no_resched() and preempt_check_resched() to avoid
having to reload the preemption value and allow the compiler to use
the flags of the previuos decrement. I couldn't come up with any sane
reason for this barrier() to be there as preempt_enable_no_resched()
already has a barrier() before doing the decrement.
Suggested-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/n/tip-7a7m5qqbn5pmwnd4wko9u6da@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Mike reported that commit 7d1a9417 ("x86: Use generic idle loop")
regressed several workloads and caused excessive reschedule
interrupts.
The patch in question failed to notice that the x86 code had an
inverted sense of the polling state versus the new generic code (x86:
default polling, generic: default !polling).
Fix the two prominent x86 mwait based idle drivers and introduce a few
new generic polling helpers (fixing the wrong smp_mb__after_clear_bit
usage).
Also switch the idle routines to using tif_need_resched() which is an
immediate TIF_NEED_RESCHED test as opposed to need_resched which will
end up being slightly different.
Reported-by: Mike Galbraith <bitbucket@online.de>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: lenb@kernel.org
Cc: tglx@linutronix.de
Link: http://lkml.kernel.org/n/tip-nc03imb0etuefmzybzj7sprf@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In this patch, we keep track of the max cost we spend doing idle load balancing
for each sched domain. If the avg time the CPU remains idle is less then the
time we have already spent on idle balancing + the max cost of idle balancing
in the sched domain, then we don't continue to attempt the balance. We also
keep a per rq variable, max_idle_balance_cost, which keeps track of the max
time spent on newidle load balances throughout all its domains so that we can
determine the avg_idle's max value.
By using the max, we avoid overrunning the average. This further reduces the
chance we attempt balancing when the CPU is not idle for longer than the cost
to balance.
Signed-off-by: Jason Low <jason.low2@hp.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1379096813-3032-3-git-send-email-jason.low2@hp.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The memcg OOM handling is incredibly fragile and can deadlock. When a
task fails to charge memory, it invokes the OOM killer and loops right
there in the charge code until it succeeds. Comparably, any other task
that enters the charge path at this point will go to a waitqueue right
then and there and sleep until the OOM situation is resolved. The problem
is that these tasks may hold filesystem locks and the mmap_sem; locks that
the selected OOM victim may need to exit.
For example, in one reported case, the task invoking the OOM killer was
about to charge a page cache page during a write(), which holds the
i_mutex. The OOM killer selected a task that was just entering truncate()
and trying to acquire the i_mutex:
OOM invoking task:
mem_cgroup_handle_oom+0x241/0x3b0
mem_cgroup_cache_charge+0xbe/0xe0
add_to_page_cache_locked+0x4c/0x140
add_to_page_cache_lru+0x22/0x50
grab_cache_page_write_begin+0x8b/0xe0
ext3_write_begin+0x88/0x270
generic_file_buffered_write+0x116/0x290
__generic_file_aio_write+0x27c/0x480
generic_file_aio_write+0x76/0xf0 # takes ->i_mutex
do_sync_write+0xea/0x130
vfs_write+0xf3/0x1f0
sys_write+0x51/0x90
system_call_fastpath+0x18/0x1d
OOM kill victim:
do_truncate+0x58/0xa0 # takes i_mutex
do_last+0x250/0xa30
path_openat+0xd7/0x440
do_filp_open+0x49/0xa0
do_sys_open+0x106/0x240
sys_open+0x20/0x30
system_call_fastpath+0x18/0x1d
The OOM handling task will retry the charge indefinitely while the OOM
killed task is not releasing any resources.
A similar scenario can happen when the kernel OOM killer for a memcg is
disabled and a userspace task is in charge of resolving OOM situations.
In this case, ALL tasks that enter the OOM path will be made to sleep on
the OOM waitqueue and wait for userspace to free resources or increase
the group's limit. But a userspace OOM handler is prone to deadlock
itself on the locks held by the waiting tasks. For example one of the
sleeping tasks may be stuck in a brk() call with the mmap_sem held for
writing but the userspace handler, in order to pick an optimal victim,
may need to read files from /proc/<pid>, which tries to acquire the same
mmap_sem for reading and deadlocks.
This patch changes the way tasks behave after detecting a memcg OOM and
makes sure nobody loops or sleeps with locks held:
1. When OOMing in a user fault, invoke the OOM killer and restart the
fault instead of looping on the charge attempt. This way, the OOM
victim can not get stuck on locks the looping task may hold.
2. When OOMing in a user fault but somebody else is handling it
(either the kernel OOM killer or a userspace handler), don't go to
sleep in the charge context. Instead, remember the OOMing memcg in
the task struct and then fully unwind the page fault stack with
-ENOMEM. pagefault_out_of_memory() will then call back into the
memcg code to check if the -ENOMEM came from the memcg, and then
either put the task to sleep on the memcg's OOM waitqueue or just
restart the fault. The OOM victim can no longer get stuck on any
lock a sleeping task may hold.
Debugged by Michal Hocko.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: azurIt <azurit@pobox.sk>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Ingo Molnar
Rik van Riel
Mel Gorman
Peter Zijlstra
Jason Low
Johannes Weiner
Oleg Nesterov