RESOURCE_MAX is far too general name, change it to RES_COUNTER_MAX.
Signed-off-by: Sha Zhengju <handai.szj@taobao.com>
Signed-off-by: Qiang Huang <h.huangqiang@huawei.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Jeff Liu <jeff.liu@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.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>
The memcg OOM handler open-codes a sleeping lock for OOM serialization
(trylock, wait, repeat) because the required locking is so specific to
memcg hierarchies. However, it would be nice if this construct would be
clearly recognizable and not be as obfuscated as it is right now. Clean
up as follows:
1. Remove the return value of mem_cgroup_oom_unlock()
2. Rename mem_cgroup_oom_lock() to mem_cgroup_oom_trylock().
3. Pull the prepare_to_wait() out of the memcg_oom_lock scope. This
makes it more obvious that the task has to be on the waitqueue
before attempting to OOM-trylock the hierarchy, to not miss any
wakeups before going to sleep. It just didn't matter until now
because it was all lumped together into the global memcg_oom_lock
spinlock section.
4. Pull the mem_cgroup_oom_notify() out of the memcg_oom_lock scope.
It is proctected by the hierarchical OOM-lock.
5. The memcg_oom_lock spinlock is only required to propagate the OOM
lock in any given hierarchy atomically. Restrict its scope to
mem_cgroup_oom_(trylock|unlock).
6. Do not wake up the waitqueue unconditionally at the end of the
function. Only the lockholder has to wake up the next in line
after releasing the lock.
Note that the lockholder kicks off the OOM-killer, which in turn
leads to wakeups from the uncharges of the exiting task. But a
contender is not guaranteed to see them if it enters the OOM path
after the OOM kills but before the lockholder releases the lock.
Thus there has to be an explicit wakeup after releasing the lock.
7. Put the OOM task on the waitqueue before marking the hierarchy as
under OOM as that is the point where we start to receive wakeups.
No point in listening before being on the waitqueue.
8. Likewise, unmark the hierarchy before finishing the sleep, for
symmetry.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: azurIt <azurit@pobox.sk>
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>
System calls and kernel faults (uaccess, gup) can handle an out of memory
situation gracefully and just return -ENOMEM.
Enable the memcg OOM killer only for user faults, where it's really the
only option available.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: azurIt <azurit@pobox.sk>
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>
Clean up some mess made by the "Soft limit rework" series, and a few other
things.
Cc: Michal Hocko <mhocko@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
shrink_zone starts with soft reclaim pass first and then falls back to
regular reclaim if nothing has been scanned. This behavior is natural
but there is a catch. Memcg iterators, when used with the reclaim
cookie, are designed to help to prevent from over reclaim by
interleaving reclaimers (per node-zone-priority) so the tree walk might
miss many (even all) nodes in the hierarchy e.g. when there are direct
reclaimers racing with each other or with kswapd in the global case or
multiple allocators reaching the limit for the target reclaim case. To
make it even more complicated, targeted reclaim doesn't do the whole
tree walk because it stops reclaiming once it reclaims sufficient pages.
As a result groups over the limit might be missed, thus nothing is
scanned, and reclaim would fall back to the reclaim all mode.
This patch checks for the incomplete tree walk in shrink_zone. If no
group has been visited and the hierarchy is soft reclaimable then we
must have missed some groups, in which case the __shrink_zone is called
again. This doesn't guarantee there will be some progress of course
because the current reclaimer might be still racing with others but it
would at least give a chance to start the walk without a big risk of
reclaim latencies.
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Glauber Costa <glommer@openvz.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Children in soft limit excess are currently tracked up the hierarchy in
memcg->children_in_excess. Nevertheless there still might exist tons of
groups that are not in hierarchy relation to the root cgroup (e.g. all
first level groups if root_mem_cgroup->use_hierarchy == false).
As the whole tree walk has to be done when the iteration starts at
root_mem_cgroup the iterator should be able to skip the walk if there is
no child above the limit without iterating them. This can be done
easily if the root tracks all children rather than only hierarchical
children. This is done by this patch which updates root_mem_cgroup
children_in_excess if root_mem_cgroup->use_hierarchy == false so the
root knows about all children in excess.
Please note that this is not an issue for inner memcgs which have
use_hierarchy == false because then only the single group is visited so
no special optimization is necessary.
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Glauber Costa <glommer@openvz.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mem_cgroup_should_soft_reclaim controls whether soft reclaim pass is
done and it always says yes currently. Memcg iterators are clever to
skip nodes that are not soft reclaimable quite efficiently but
mem_cgroup_should_soft_reclaim can be more clever and do not start the
soft reclaim pass at all if it knows that nothing would be scanned
anyway.
In order to do that, simply reuse mem_cgroup_soft_reclaim_eligible for
the target group of the reclaim and allow the pass only if the whole
subtree wouldn't be skipped.
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Glauber Costa <glommer@openvz.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The caller of the iterator might know that some nodes or even subtrees
should be skipped but there is no way to tell iterators about that so the
only choice left is to let iterators to visit each node and do the
selection outside of the iterating code. This, however, doesn't scale
well with hierarchies with many groups where only few groups are
interesting.
This patch adds mem_cgroup_iter_cond variant of the iterator with a
callback which gets called for every visited node. There are three
possible ways how the callback can influence the walk. Either the node is
visited, it is skipped but the tree walk continues down the tree or the
whole subtree of the current group is skipped.
[hughd@google.com: fix memcg-less page reclaim]
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Glauber Costa <glommer@openvz.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Soft reclaim has been done only for the global reclaim (both background
and direct). Since "memcg: integrate soft reclaim tighter with zone
shrinking code" there is no reason for this limitation anymore as the soft
limit reclaim doesn't use any special code paths and it is a part of the
zone shrinking code which is used by both global and targeted reclaims.
From the semantic point of view it is natural to consider soft limit
before touching all groups in the hierarchy tree which is touching the
hard limit because soft limit tells us where to push back when there is a
memory pressure. It is not important whether the pressure comes from the
limit or imbalanced zones.
This patch simply enables soft reclaim unconditionally in
mem_cgroup_should_soft_reclaim so it is enabled for both global and
targeted reclaim paths. mem_cgroup_soft_reclaim_eligible needs to learn
about the root of the reclaim to know where to stop checking soft limit
state of parents up the hierarchy. Say we have
A (over soft limit)
\
B (below s.l., hit the hard limit)
/ \
C D (below s.l.)
B is the source of the outside memory pressure now for D but we shouldn't
soft reclaim it because it is behaving well under B subtree and we can
still reclaim from C (pressumably it is over the limit).
mem_cgroup_soft_reclaim_eligible should therefore stop climbing up the
hierarchy at B (root of the memory pressure).
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Glauber Costa <glommer@openvz.org>
Reviewed-by: Tejun Heo <tj@kernel.org>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Now that the soft limit is integrated to the reclaim directly the whole
soft-limit tree infrastructure is not needed anymore. Rip it out.
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Glauber Costa <glommer@openvz.org>
Reviewed-by: Tejun Heo <tj@kernel.org>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patchset is sitting out of tree for quite some time without any
objections. I would be really happy if it made it into 3.12. I do not
want to push it too hard but I think this work is basically ready and
waiting more doesn't help.
The basic idea is quite simple. Pull soft reclaim into shrink_zone in the
first step and get rid of the previous soft reclaim infrastructure.
shrink_zone is done in two passes now. First it tries to do the soft
limit reclaim and it falls back to reclaim-all mode if no group is over
the limit or no pages have been scanned. The second pass happens at the
same priority so the only time we waste is the memcg tree walk which has
been updated in the third step to have only negligible overhead.
As a bonus we will get rid of a _lot_ of code by this and soft reclaim
will not stand out like before when it wasn't integrated into the zone
shrinking code and it reclaimed at priority 0 (the testing results show
that some workloads suffers from such an aggressive reclaim). The clean
up is in a separate patch because I felt it would be easier to review that
way.
The second step is soft limit reclaim integration into targeted reclaim.
It should be rather straight forward. Soft limit has been used only for
the global reclaim so far but it makes sense for any kind of pressure
coming from up-the-hierarchy, including targeted reclaim.
The third step (patches 4-8) addresses the tree walk overhead by enhancing
memcg iterators to enable skipping whole subtrees and tracking number of
over soft limit children at each level of the hierarchy. This information
is updated same way the old soft limit tree was updated (from
memcg_check_events) so we shouldn't see an additional overhead. In fact
mem_cgroup_update_soft_limit is much simpler than tree manipulation done
previously.
__shrink_zone uses mem_cgroup_soft_reclaim_eligible as a predicate for
mem_cgroup_iter so the decision whether a particular group should be
visited is done at the iterator level which allows us to decide to skip
the whole subtree as well (if there is no child in excess). This reduces
the tree walk overhead considerably.
* TEST 1
========
My primary test case was a parallel kernel build with 2 groups (make is
running with -j8 with a distribution .config in a separate cgroup without
any hard limit) on a 32 CPU machine booted with 1GB memory and both builds
run taskset to Node 0 cpus.
I was mostly interested in 2 setups. Default - no soft limit set and -
and 0 soft limit set to both groups. The first one should tell us whether
the rework regresses the default behavior while the second one should show
us improvements in an extreme case where both workloads are always over
the soft limit.
/usr/bin/time -v has been used to collect the statistics and each
configuration had 3 runs after fresh boot without any other load on the
system.
base is mmotm-2013-07-18-16-40
rework all 8 patches applied on top of base
* No-limit
User
no-limit/base: min: 651.92 max: 672.65 avg: 664.33 std: 8.01 runs: 6
no-limit/rework: min: 657.34 [100.8%] max: 668.39 [99.4%] avg: 663.13 [99.8%] std: 3.61 runs: 6
System
no-limit/base: min: 69.33 max: 71.39 avg: 70.32 std: 0.79 runs: 6
no-limit/rework: min: 69.12 [99.7%] max: 71.05 [99.5%] avg: 70.04 [99.6%] std: 0.59 runs: 6
Elapsed
no-limit/base: min: 398.27 max: 422.36 avg: 408.85 std: 7.74 runs: 6
no-limit/rework: min: 386.36 [97.0%] max: 438.40 [103.8%] avg: 416.34 [101.8%] std: 18.85 runs: 6
The results are within noise. Elapsed time has a bigger variance but the
average looks good.
* 0-limit
User
0-limit/base: min: 573.76 max: 605.63 avg: 585.73 std: 12.21 runs: 6
0-limit/rework: min: 645.77 [112.6%] max: 666.25 [110.0%] avg: 656.97 [112.2%] std: 7.77 runs: 6
System
0-limit/base: min: 69.57 max: 71.13 avg: 70.29 std: 0.54 runs: 6
0-limit/rework: min: 68.68 [98.7%] max: 71.40 [100.4%] avg: 69.91 [99.5%] std: 0.87 runs: 6
Elapsed
0-limit/base: min: 1306.14 max: 1550.17 avg: 1430.35 std: 90.86 runs: 6
0-limit/rework: min: 404.06 [30.9%] max: 465.94 [30.1%] avg: 434.81 [30.4%] std: 22.68 runs: 6
The improvement is really huge here (even bigger than with my previous
testing and I suspect that this highly depends on the storage). Page
fault statistics tell us at least part of the story:
Minor
0-limit/base: min: 37180461.00 max: 37319986.00 avg: 37247470.00 std: 54772.71 runs: 6
0-limit/rework: min: 36751685.00 [98.8%] max: 36805379.00 [98.6%] avg: 36774506.33 [98.7%] std: 17109.03 runs: 6
Major
0-limit/base: min: 170604.00 max: 221141.00 avg: 196081.83 std: 18217.01 runs: 6
0-limit/rework: min: 2864.00 [1.7%] max: 10029.00 [4.5%] avg: 5627.33 [2.9%] std: 2252.71 runs: 6
Same as with my previous testing Minor faults are more or less within
noise but Major fault count is way bellow the base kernel.
While this looks as a nice win it is fair to say that 0-limit
configuration is quite artificial. So I was playing with 0-no-limit
loads as well.
* TEST 2
========
The following results are from 2 groups configuration on a 16GB machine
(single NUMA node).
- A running stream IO (dd if=/dev/zero of=local.file bs=1024) with
2*TotalMem with 0 soft limit.
- B running a mem_eater which consumes TotalMem-1G without any limit. The
mem_eater consumes the memory in 100 chunks with 1s nap after each
mmap+poppulate so that both loads have chance to fight for the memory.
The expected result is that B shouldn't be reclaimed and A shouldn't see
a big dropdown in elapsed time.
User
base: min: 2.68 max: 2.89 avg: 2.76 std: 0.09 runs: 3
rework: min: 3.27 [122.0%] max: 3.74 [129.4%] avg: 3.44 [124.6%] std: 0.21 runs: 3
System
base: min: 86.26 max: 88.29 avg: 87.28 std: 0.83 runs: 3
rework: min: 81.05 [94.0%] max: 84.96 [96.2%] avg: 83.14 [95.3%] std: 1.61 runs: 3
Elapsed
base: min: 317.28 max: 332.39 avg: 325.84 std: 6.33 runs: 3
rework: min: 281.53 [88.7%] max: 298.16 [89.7%] avg: 290.99 [89.3%] std: 6.98 runs: 3
System time improved slightly as well as Elapsed. My previous testing
has shown worse numbers but this again seem to depend on the storage
speed.
My theory is that the writeback doesn't catch up and prio-0 soft reclaim
falls into wait on writeback page too often in the base kernel. The
patched kernel doesn't do that because the soft reclaim is done from the
kswapd/direct reclaim context. This can be seen on the following graph
nicely. The A's group usage_in_bytes regurarly drops really low very often.
All 3 runs
http://labs.suse.cz/mhocko/soft_limit_rework/stream_io-vs-mem_eater/stream.png
resp. a detail of the single run
http://labs.suse.cz/mhocko/soft_limit_rework/stream_io-vs-mem_eater/stream-one-run.png
mem_eater seems to be doing better as well. It gets to the full
allocation size faster as can be seen on the following graph:
http://labs.suse.cz/mhocko/soft_limit_rework/stream_io-vs-mem_eater/mem_eater-one-run.png
/proc/meminfo collected during the test also shows that rework kernel
hasn't swapped that much (well almost not at all):
base: max: 123900 K avg: 56388.29 K
rework: max: 300 K avg: 128.68 K
kswapd and direct reclaim statistics are of no use unfortunatelly because
soft reclaim is not accounted properly as the counters are hidden by
global_reclaim() checks in the base kernel.
* TEST 3
========
Another test was the same configuration as TEST2 except the stream IO was
replaced by a single kbuild (16 parallel jobs bound to Node0 cpus same as
in TEST1) and mem_eater allocated TotalMem-200M so kbuild had only 200MB
left.
Kbuild did better with the rework kernel here as well:
User
base: min: 860.28 max: 872.86 avg: 868.03 std: 5.54 runs: 3
rework: min: 880.81 [102.4%] max: 887.45 [101.7%] avg: 883.56 [101.8%] std: 2.83 runs: 3
System
base: min: 84.35 max: 85.06 avg: 84.79 std: 0.31 runs: 3
rework: min: 85.62 [101.5%] max: 86.09 [101.2%] avg: 85.79 [101.2%] std: 0.21 runs: 3
Elapsed
base: min: 135.36 max: 243.30 avg: 182.47 std: 45.12 runs: 3
rework: min: 110.46 [81.6%] max: 116.20 [47.8%] avg: 114.15 [62.6%] std: 2.61 runs: 3
Minor
base: min: 36635476.00 max: 36673365.00 avg: 36654812.00 std: 15478.03 runs: 3
rework: min: 36639301.00 [100.0%] max: 36695541.00 [100.1%] avg: 36665511.00 [100.0%] std: 23118.23 runs: 3
Major
base: min: 14708.00 max: 53328.00 avg: 31379.00 std: 16202.24 runs: 3
rework: min: 302.00 [2.1%] max: 414.00 [0.8%] avg: 366.33 [1.2%] std: 47.22 runs: 3
Again we can see a significant improvement in Elapsed (it also seems to
be more stable), there is a huge dropdown for the Major page faults and
much more swapping:
base: max: 583736 K avg: 112547.43 K
rework: max: 4012 K avg: 124.36 K
Graphs from all three runs show the variability of the kbuild quite
nicely. It even seems that it took longer after every run with the base
kernel which would be quite surprising as the source tree for the build is
removed and caches are dropped after each run so the build operates on a
freshly extracted sources everytime.
http://labs.suse.cz/mhocko/soft_limit_rework/stream_io-vs-mem_eater/kbuild-mem_eater.png
My other testing shows that this is just a matter of timing and other runs
behave differently the std for Elapsed time is similar ~50. Example of
other three runs:
http://labs.suse.cz/mhocko/soft_limit_rework/stream_io-vs-mem_eater/kbuild-mem_eater2.png
So to wrap this up. The series is still doing good and improves the soft
limit.
The testing results for bunch of cgroups with both stream IO and kbuild
loads can be found in "memcg: track children in soft limit excess to
improve soft limit".
This patch:
Memcg soft reclaim has been traditionally triggered from the global
reclaim paths before calling shrink_zone. mem_cgroup_soft_limit_reclaim
then picked up a group which exceeds the soft limit the most and reclaimed
it with 0 priority to reclaim at least SWAP_CLUSTER_MAX pages.
The infrastructure requires per-node-zone trees which hold over-limit
groups and keep them up-to-date (via memcg_check_events) which is not cost
free. Although this overhead hasn't turned out to be a bottle neck the
implementation is suboptimal because mem_cgroup_update_tree has no idea
which zones consumed memory over the limit so we could easily end up
having a group on a node-zone tree having only few pages from that
node-zone.
This patch doesn't try to fix node-zone trees management because it seems
that integrating soft reclaim into zone shrinking sounds much easier and
more appropriate for several reasons. First of all 0 priority reclaim was
a crude hack which might lead to big stalls if the group's LRUs are big
and hard to reclaim (e.g. a lot of dirty/writeback pages). Soft reclaim
should be applicable also to the targeted reclaim which is awkward right
now without additional hacks. Last but not least the whole infrastructure
eats quite some code.
After this patch shrink_zone is done in 2 passes. First it tries to do
the soft reclaim if appropriate (only for global reclaim for now to keep
compatible with the original state) and fall back to ignoring soft limit
if no group is eligible to soft reclaim or nothing has been scanned during
the first pass. Only groups which are over their soft limit or any of
their parents up the hierarchy is over the limit are considered eligible
during the first pass.
Soft limit tree which is not necessary anymore will be removed in the
follow up patch to make this patch smaller and easier to review.
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Glauber Costa <glommer@openvz.org>
Reviewed-by: Tejun Heo <tj@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Ying Han <yinghan@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Glauber Costa <glommer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
vfs guarantees the cgroup won't be destroyed, so it's redundant to get a
css reference.
Signed-off-by: Li Zefan <lizefan@huawei.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Conditionally call the appropriate fs_init function and fill_super
functions. Add a use once guard to shmem_init() to simply succeed on a
second call.
(Note that IS_ENABLED() is a compile time constant so dead code
elimination removes unused function calls when CONFIG_TMPFS is disabled.)
Signed-off-by: Rob Landley <rob@landley.net>
Cc: Jeff Layton <jlayton@redhat.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Stephen Warren <swarren@nvidia.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Jim Cromie <jim.cromie@gmail.com>
Cc: Sam Ravnborg <sam@ravnborg.org>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
With users of radix_tree_preload() run from interrupt (block/blk-ioc.c is
one such possible user), the following race can happen:
radix_tree_preload()
...
radix_tree_insert()
radix_tree_node_alloc()
if (rtp->nr) {
ret = rtp->nodes[rtp->nr - 1];
<interrupt>
...
radix_tree_preload()
...
radix_tree_insert()
radix_tree_node_alloc()
if (rtp->nr) {
ret = rtp->nodes[rtp->nr - 1];
And we give out one radix tree node twice. That clearly results in radix
tree corruption with different results (usually OOPS) depending on which
two users of radix tree race.
We fix the problem by making radix_tree_node_alloc() always allocate fresh
radix tree nodes when in interrupt. Using preloading when in interrupt
doesn't make sense since all the allocations have to be atomic anyway and
we cannot steal nodes from process-context users because some users rely
on radix_tree_insert() succeeding after radix_tree_preload().
in_interrupt() check is somewhat ugly but we cannot simply key off passed
gfp_mask as that is acquired from root_gfp_mask() and thus the same for
all preload users.
Another part of the fix is to avoid node preallocation in
radix_tree_preload() when passed gfp_mask doesn't allow waiting. Again,
preallocation in such case doesn't make sense and when preallocation would
happen in interrupt we could possibly leak some allocated nodes. However,
some users of radix_tree_preload() require following radix_tree_insert()
to succeed. To avoid unexpected effects for these users,
radix_tree_preload() only warns if passed gfp mask doesn't allow waiting
and we provide a new function radix_tree_maybe_preload() for those users
which get different gfp mask from different call sites and which are
prepared to handle radix_tree_insert() failure.
Signed-off-by: Jan Kara <jack@suse.cz>
Cc: Jens Axboe <jaxboe@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
A memory cgroup with (1) multiple threshold notifications and (2) at least
one threshold >=2G was not reliable. Specifically the notifications would
either not fire or would not fire in the proper order.
The __mem_cgroup_threshold() signaling logic depends on keeping 64 bit
thresholds in sorted order. mem_cgroup_usage_register_event() sorts them
with compare_thresholds(), which returns the difference of two 64 bit
thresholds as an int. If the difference is positive but has bit[31] set,
then sort() treats the difference as negative and breaks sort order.
This fix compares the two arbitrary 64 bit thresholds returning the
classic -1, 0, 1 result.
The test below sets two notifications (at 0x1000 and 0x81001000):
cd /sys/fs/cgroup/memory
mkdir x
for x in 4096 2164264960; do
cgroup_event_listener x/memory.usage_in_bytes $x | sed "s/^/$x listener:/" &
done
echo $$ > x/cgroup.procs
anon_leaker 500M
v3.11-rc7 fails to signal the 4096 event listener:
Leaking...
Done leaking pages.
Patched v3.11-rc7 properly notifies:
Leaking...
4096 listener:2013:8:31:14:13:36
Done leaking pages.
The fixed bug is old. It appears to date back to the introduction of
memcg threshold notifications in v2.6.34-rc1-116-g2e72b6347c94 "memcg:
implement memory thresholds"
Signed-off-by: Greg Thelen <gthelen@google.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Use the helper function instead of __GFP_ZERO.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
pgoff is not used after the statement "pgoff = vma->vm_pgoff;", so the
assignment is redundant.
Signed-off-by: Yanchuan Nian <ycnian@gmail.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
madvise_hwpoison() has two locals called "ret". Fix it all up.
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The return value outside for loop is always zero which means
madvise_hwpoison return success, however, this is not truth for
soft_offline_page w/ failure return value.
Signed-off-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
madvise hwpoison inject will poison the read-only empty zero page if there
is no write access before poison. Empty zero page reference count will be
increased for hwpoison, subsequent poison zero page will return directly
since page has already been set PG_hwpoison, however, page reference count
is still increased by get_user_pages_fast. The unpoison process will
unpoison the empty zero page and decrease the reference count successfully
for the fist time, however, subsequent unpoison empty zero page will
return directly since page has already been unpoisoned and without
decrease the page reference count of empty zero page.
This patch fixes it by make madvise_hwpoison() put a page and return
immediately (without calling memory_failure() or soft_offline_page()) when
the page is already hwpoisoned.
Testcase:
#define _GNU_SOURCE
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
#include <errno.h>
#define PAGES_TO_TEST 3
#define PAGE_SIZE 4096
int main(void)
{
char *mem;
int i;
mem = mmap(NULL, PAGES_TO_TEST * PAGE_SIZE,
PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
if (madvise(mem, PAGES_TO_TEST * PAGE_SIZE, MADV_HWPOISON) == -1)
return -1;
munmap(mem, PAGES_TO_TEST * PAGE_SIZE);
return 0;
}
Add printk to dump page reference count:
[ 93.075959] Injecting memory failure for page 0x19d0 at 0xb77d8000
[ 93.076207] MCE 0x19d0: non LRU page recovery: Ignored
[ 93.076209] pfn 0x19d0, page count = 1 after memory failure
[ 93.076220] Injecting memory failure for page 0x19d0 at 0xb77d9000
[ 93.076221] MCE 0x19d0: already hardware poisoned
[ 93.076222] pfn 0x19d0, page count = 2 after memory failure
[ 93.076224] Injecting memory failure for page 0x19d0 at 0xb77da000
[ 93.076224] MCE 0x19d0: already hardware poisoned
[ 93.076225] pfn 0x19d0, page count = 3 after memory failure
Signed-off-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Suggested-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Set pageblock migration type will hold zone->lock which is heavy contended
in system to avoid race. However, soft offline page will set pageblock
migration type twice during get page if the page is in used, not hugetlbfs
page and not on lru list. There is unnecessary to set the pageblock
migration type and hold heavy contended zone->lock again if the first
round get page have already set the pageblock to right migration type.
The trick here is migration type is MIGRATE_ISOLATE. There are other two
parts can change MIGRATE_ISOLATE except hwpoison. One is memory hoplug,
however, we hold lock_memory_hotplug() which avoid race. The second is
CMA which umovable page allocation requst can't fallback to. So it's safe
here.
Signed-off-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Tony Luck <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Replace atomic_long_sub() with atomic_long_dec() since the page is normal
page instead of hugetlbfs page or thp.
Signed-off-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Tony Luck <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There is a race between hwpoison page and unpoison page, memory_failure
set the page hwpoison and increase num_poisoned_pages without hold page
lock, and one page count will be accounted against thp for
num_poisoned_pages. However, unpoison can occur before memory_failure
hold page lock and split transparent hugepage, unpoison will decrease
num_poisoned_pages by 1 << compound_order since memory_failure has not yet
split transparent hugepage with page lock held. That means we account one
page for hwpoison and 1 << compound_order for unpoison. This patch fix it
by inserting a PageTransHuge check before doing TestClearPageHWPoison,
unpoison failed without clearing PageHWPoison and decreasing
num_poisoned_pages.
A B
memory_failue
TestSetPageHWPoison(p);
if (PageHuge(p))
nr_pages = 1 << compound_order(hpage);
else
nr_pages = 1;
atomic_long_add(nr_pages, &num_poisoned_pages);
unpoison_memory
nr_pages = 1<< compound_trans_order(page);
if(TestClearPageHWPoison(p))
atomic_long_sub(nr_pages, &num_poisoned_pages);
lock page
if (!PageHWPoison(p))
unlock page and return
hwpoison_user_mappings
if (PageTransHuge(hpage))
split_huge_page(hpage);
Signed-off-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Suggested-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Tony Luck <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
compound lock is introduced by commit e9da73d67("thp: compound_lock."), it
is used to serialize put_page against __split_huge_page_refcount(). In
addition, transparent hugepages will be splitted in hwpoison handler and
just one subpage will be poisoned. There is unnecessary to hold compound
lock for hugetlbfs page. This patch replace compound_trans_order by
compond_order in the place where the page is hugetlbfs page.
Signed-off-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Tony Luck <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
memory_failure() store the page flag of the error page before doing unmap,
and (only) if the first check with page flags at the time decided the
error page is unknown, it do the second check with the stored page flag
since memory_failure() does unmapping of the error pages before doing
page_action(). This unmapping changes the page state, especially
page_remove_rmap() (called from try_to_unmap_one()) clears PG_mlocked, so
page_action() can't catch mlocked pages after that.
However, memory_failure() can't handle memory errors on dirty mlocked
pages correctly. try_to_unmap_one will move the dirty bit from pte to the
physical page, the second check lose it since it check the stored page
flag. This patch fix it by restore PG_dirty flag to stored page flag if
the page is dirty.
Testcase:
#define _GNU_SOURCE
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#include <sys/types.h>
#include <errno.h>
#define PAGES_TO_TEST 2
#define PAGE_SIZE 4096
int main(void)
{
char *mem;
int i;
mem = mmap(NULL, PAGES_TO_TEST * PAGE_SIZE,
PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_LOCKED, 0, 0);
for (i = 0; i < PAGES_TO_TEST; i++)
mem[i * PAGE_SIZE] = 'a';
if (madvise(mem, PAGES_TO_TEST * PAGE_SIZE, MADV_HWPOISON) == -1)
return -1;
return 0;
}
Before patch:
[ 912.839247] Injecting memory failure for page 7dfb8 at 7f6b4e37b000
[ 912.839257] MCE 0x7dfb8: clean mlocked LRU page recovery: Recovered
[ 912.845550] MCE 0x7dfb8: clean mlocked LRU page still referenced by 1 users
[ 912.852586] Injecting memory failure for page 7e6aa at 7f6b4e37c000
[ 912.852594] MCE 0x7e6aa: clean mlocked LRU page recovery: Recovered
[ 912.858936] MCE 0x7e6aa: clean mlocked LRU page still referenced by 1 users
After patch:
[ 163.590225] Injecting memory failure for page 91bc2f at 7f9f5b0e5000
[ 163.590264] MCE 0x91bc2f: dirty mlocked LRU page recovery: Recovered
[ 163.596680] MCE 0x91bc2f: dirty mlocked LRU page still referenced by 1 users
[ 163.603831] Injecting memory failure for page 91cdd3 at 7f9f5b0e6000
[ 163.603852] MCE 0x91cdd3: dirty mlocked LRU page recovery: Recovered
[ 163.610305] MCE 0x91cdd3: dirty mlocked LRU page still referenced by 1 users
Signed-off-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Tony Luck <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Soft offline code expects that MIGRATE_ISOLATE is set on the target page
only during soft offlining work. But currenly it doesn't work as expected
when get_any_page() fails and returns negative value. In the result, end
users can have unexpectedly isolated pages. This patch just fixes it.
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Fengguang Wu <fengguang.wu@intel.com>
Cc: Tony Luck <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Set _mapcount PAGE_BUDDY_MAPCOUNT_VALUE to make the page buddy. Not the
magic number -2.
Signed-off-by: Wang Sheng-Hui <shhuiw@gmail.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
'*lenp' may be less than "sizeof(kbuf)" so we must check this before the
next copy_to_user().
pdflush_proc_obsolete() is called by sysctl which 'procname' is
"nr_pdflush_threads", if the user passes buffer length less than
"sizeof(kbuf)", it will cause issue.
Signed-off-by: Chen Gang <gang.chen@asianux.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In alloc_new_pmd(), if pud_alloc() was called successfully, but
pmd_alloc() fails, avoid leaking `pud'.
Signed-off-by: Chen Gang <gang.chen@asianux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
After commit 9bdac91424 ("sparsemem: Put mem map for one node
together."), vmemmap for one node will be allocated together, its logic
is similar as memory allocation for pageblock flags. This patch
introduces alloc_usemap_and_memmap to extract the same logic of memory
alloction for pageblock flags and vmemmap.
Signed-off-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Fengguang Wu <fengguang.wu@intel.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiri Kosina <jkosina@suse.cz>
Cc: Yinghai Lu <yinghai@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patch is based on KOSAKI's work and I add a little more description,
please refer https://lkml.org/lkml/2012/6/14/74.
Currently, I found system can enter a state that there are lots of free
pages in a zone but only order-0 and order-1 pages which means the zone is
heavily fragmented, then high order allocation could make direct reclaim
path's long stall(ex, 60 seconds) especially in no swap and no compaciton
enviroment. This problem happened on v3.4, but it seems issue still lives
in current tree, the reason is do_try_to_free_pages enter live lock:
kswapd will go to sleep if the zones have been fully scanned and are still
not balanced. As kswapd thinks there's little point trying all over again
to avoid infinite loop. Instead it changes order from high-order to
0-order because kswapd think order-0 is the most important. Look at
73ce02e9 in detail. If watermarks are ok, kswapd will go back to sleep
and may leave zone->all_unreclaimable =3D 0. It assume high-order users
can still perform direct reclaim if they wish.
Direct reclaim continue to reclaim for a high order which is not a
COSTLY_ORDER without oom-killer until kswapd turn on
zone->all_unreclaimble= . This is because to avoid too early oom-kill.
So it means direct_reclaim depends on kswapd to break this loop.
In worst case, direct-reclaim may continue to page reclaim forever when
kswapd sleeps forever until someone like watchdog detect and finally kill
the process. As described in:
http://thread.gmane.org/gmane.linux.kernel.mm/103737
We can't turn on zone->all_unreclaimable from direct reclaim path because
direct reclaim path don't take any lock and this way is racy. Thus this
patch removes zone->all_unreclaimable field completely and recalculates
zone reclaimable state every time.
Note: we can't take the idea that direct-reclaim see zone->pages_scanned
directly and kswapd continue to use zone->all_unreclaimable. Because, it
is racy. commit 929bea7c71 (vmscan: all_unreclaimable() use
zone->all_unreclaimable as a name) describes the detail.
[akpm@linux-foundation.org: uninline zone_reclaimable_pages() and zone_reclaimable()]
Cc: Aaditya Kumar <aaditya.kumar.30@gmail.com>
Cc: Ying Han <yinghan@google.com>
Cc: Nick Piggin <npiggin@gmail.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Bob Liu <lliubbo@gmail.com>
Cc: Neil Zhang <zhangwm@marvell.com>
Cc: Russell King - ARM Linux <linux@arm.linux.org.uk>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Minchan Kim <minchan@kernel.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Lisa Du <cldu@marvell.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently munlock_vma_pages_range() calls follow_page_mask() to obtain
each individual struct page. This entails repeated full page table
translations and page table lock taken for each page separately.
This patch avoids the costly follow_page_mask() where possible, by
iterating over ptes within single pmd under single page table lock. The
first pte is obtained by get_locked_pte() for non-THP page acquired by the
initial follow_page_mask(). The rest of the on-stack pagevec for munlock
is filled up using pte_walk as long as pte_present() and vm_normal_page()
are sufficient to obtain the struct page.
After this patch, a 14% speedup was measured for munlocking a 56GB large
memory area with THP disabled.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Jörn Engel <joern@logfs.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michel Lespinasse <walken@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The performance of the fast path in munlock_vma_range() can be further
improved by avoiding atomic ops of a redundant get_page()/put_page() pair.
When calling get_page() during page isolation, we already have the pin
from follow_page_mask(). This pin will be then returned by
__pagevec_lru_add(), after which we do not reference the pages anymore.
After this patch, an 8% speedup was measured for munlocking a 56GB large
memory area with THP disabled.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Jörn Engel <joern@logfs.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Michel Lespinasse <walken@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
After introducing batching by pagevecs into munlock_vma_range(), we can
further improve performance by bypassing the copying into per-cpu pagevec
and the get_page/put_page pair associated with that. Instead we perform
LRU putback directly from our pagevec. However, this is possible only for
single-mapped pages that are evictable after munlock. Unevictable pages
require rechecking after putting on the unevictable list, so for those we
fallback to putback_lru_page(), hich handles that.
After this patch, a 13% speedup was measured for munlocking a 56GB large
memory area with THP disabled.
[akpm@linux-foundation.org:clarify comment]
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Jörn Engel <joern@logfs.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Michel Lespinasse <walken@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Depending on previous batch which introduced batched isolation in
munlock_vma_range(), we can batch also the updates of NR_MLOCK page stats.
After the whole pagevec is processed for page isolation, the stats are
updated only once with the number of successful isolations. There were
however no measurable perfomance gains.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Jörn Engel <joern@logfs.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Michel Lespinasse <walken@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently, munlock_vma_range() calls munlock_vma_page on each page in a
loop, which results in repeated taking and releasing of the lru_lock
spinlock for isolating pages one by one. This patch batches the munlock
operations using an on-stack pagevec, so that isolation is done under
single lru_lock. For THP pages, the old behavior is preserved as they
might be split while putting them into the pagevec. After this patch, a
9% speedup was measured for munlocking a 56GB large memory area with THP
disabled.
A new function __munlock_pagevec() is introduced that takes a pagevec and:
1) It clears PageMlocked and isolates all pages under lru_lock. Zone page
stats can be also updated using the variant which assumes disabled
interrupts. 2) It finishes the munlock and lru putback on all pages under
their lock_page. Note that previously, lock_page covered also the
PageMlocked clearing and page isolation, but it is not needed for those
operations.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Jörn Engel <joern@logfs.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Michel Lespinasse <walken@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In munlock_vma_range(), lru_add_drain() is currently called in a loop
before each munlock_vma_page() call.
This is suboptimal for performance when munlocking many pages. The
benefits of per-cpu pagevec for batching the LRU putback are removed since
the pagevec only holds at most one page from the previous loop's
iteration.
The lru_add_drain() call also does not serve any purposes for correctness
- it does not even drain pagavecs of all cpu's. The munlock code already
expects and handles situations where a page cannot be isolated from the
LRU (e.g. because it is on some per-cpu pagevec).
The history of the (not commented) call also suggest that it appears there
as an oversight rather than intentionally. Before commit ff6a6da6 ("mm:
accelerate munlock() treatment of THP pages") the call happened only once
upon entering the function. The commit has moved the call into the while
loope. So while the other changes in the commit improved munlock
performance for THP pages, it introduced the abovementioned suboptimal
per-cpu pagevec usage.
Further in history, before commit 408e82b7 ("mm: munlock use
follow_page"), munlock_vma_pages_range() was just a wrapper around
__mlock_vma_pages_range which performed both mlock and munlock depending
on a flag. However, before ba470de4 ("mmap: handle mlocked pages during
map, remap, unmap") the function handled only mlock, not munlock. The
lru_add_drain call thus comes from the implementation in commit b291f000
("mlock: mlocked pages are unevictable" and was intended only for
mlocking, not munlocking. The original intention of draining the LRU
pagevec at mlock time was to ensure the pages were on the LRU before the
lock operation so that they could be placed on the unevictable list
immediately. There is very little motivation to do the same in the
munlock path this, particularly for every single page.
This patch therefore removes the call completely. After removing the
call, a 10% speedup was measured for munlock() of a 56GB large memory area
with THP disabled.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Jörn Engel <joern@logfs.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Michel Lespinasse <walken@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The goal of this patch series is to improve performance of munlock() of
large mlocked memory areas on systems without THP. This is motivated by
reported very long times of crash recovery of processes with such areas,
where munlock() can take several seconds. See
http://lwn.net/Articles/548108/
The work was driven by a simple benchmark (to be included in mmtests) that
mmaps() e.g. 56GB with MAP_LOCKED | MAP_POPULATE and measures the time of
munlock(). Profiling was performed by attaching operf --pid to the
process and sending a signal to trigger the munlock() part and then notify
bach the monitoring wrapper to stop operf, so that only munlock() appears
in the profile.
The profiles have shown that CPU time is spent mostly by atomic operations
and repeated locking per single pages. This series aims to reduce both, starting
from simpler to more complex changes.
Patch 1 performs a simple cleanup in putback_lru_page() so that page lru base
type is not determined without being actually needed.
Patch 2 removes an unnecessary call to lru_add_drain() which drains the per-cpu
pagevec after each munlocked page is put there.
Patch 3 changes munlock_vma_range() to use an on-stack pagevec for isolating
multiple non-THP pages under a single lru_lock instead of locking and
processing each page separately.
Patch 4 changes the NR_MLOCK accounting to be called only once per the pvec
introduced by previous patch.
Patch 5 uses the introduced pagevec to batch also the work of putback_lru_page
when possible, bypassing the per-cpu pvec and associated overhead.
Patch 6 removes a redundant get_page/put_page pair which saves costly atomic
operations.
Patch 7 avoids calling follow_page_mask() on each individual page, and obtains
multiple page references under a single page table lock where possible.
Measurements were made using 3.11-rc3 as a baseline. The first set of
measurements shows the possibly ideal conditions where batching should
help the most. All memory is allocated from a single NUMA node and THP is
disabled.
timedmunlock
3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3
0 1 2 3 4 5 6 7
Elapsed min 3.38 ( 0.00%) 3.39 ( -0.13%) 3.00 ( 11.33%) 2.70 ( 20.20%) 2.67 ( 21.11%) 2.37 ( 29.88%) 2.20 ( 34.91%) 1.91 ( 43.59%)
Elapsed mean 3.39 ( 0.00%) 3.40 ( -0.23%) 3.01 ( 11.33%) 2.70 ( 20.26%) 2.67 ( 21.21%) 2.38 ( 29.88%) 2.21 ( 34.93%) 1.92 ( 43.46%)
Elapsed stddev 0.01 ( 0.00%) 0.01 (-43.09%) 0.01 ( 15.42%) 0.01 ( 23.42%) 0.00 ( 89.78%) 0.01 ( -7.15%) 0.00 ( 76.69%) 0.02 (-91.77%)
Elapsed max 3.41 ( 0.00%) 3.43 ( -0.52%) 3.03 ( 11.29%) 2.72 ( 20.16%) 2.67 ( 21.63%) 2.40 ( 29.50%) 2.21 ( 35.21%) 1.96 ( 42.39%)
Elapsed range 0.03 ( 0.00%) 0.04 (-51.16%) 0.02 ( 6.27%) 0.02 ( 14.67%) 0.00 ( 88.90%) 0.03 (-19.18%) 0.01 ( 73.70%) 0.06 (-113.35%
The second set of measurements simulates the worst possible conditions for
batching by using numactl --interleave, so that there is in fact only one
page per pagevec. Even in this case the series seems to improve
performance thanks to reduced atomic operations and removal of
lru_add_drain().
timedmunlock
3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3
0 1 2 3 4 5 6 7
Elapsed min 4.00 ( 0.00%) 4.04 ( -0.93%) 3.87 ( 3.37%) 3.72 ( 6.94%) 3.81 ( 4.72%) 3.69 ( 7.82%) 3.64 ( 8.92%) 3.41 ( 14.81%)
Elapsed mean 4.17 ( 0.00%) 4.15 ( 0.51%) 4.03 ( 3.49%) 3.89 ( 6.84%) 3.86 ( 7.48%) 3.89 ( 6.69%) 3.70 ( 11.27%) 3.48 ( 16.59%)
Elapsed stddev 0.16 ( 0.00%) 0.08 ( 50.76%) 0.10 ( 41.58%) 0.16 ( 4.59%) 0.05 ( 72.38%) 0.19 (-12.91%) 0.05 ( 68.09%) 0.06 ( 66.03%)
Elapsed max 4.34 ( 0.00%) 4.32 ( 0.56%) 4.19 ( 3.62%) 4.12 ( 5.15%) 3.91 ( 9.88%) 4.12 ( 5.25%) 3.80 ( 12.58%) 3.56 ( 18.08%)
Elapsed range 0.34 ( 0.00%) 0.28 ( 17.91%) 0.32 ( 6.45%) 0.40 (-15.73%) 0.10 ( 70.06%) 0.43 (-24.84%) 0.15 ( 55.32%) 0.15 ( 56.16%)
For completeness, a third set of measurements shows the situation where
THP is enabled and allocations are again done on a single NUMA node. Here
munlock() is already very fast thanks to huge pages, and this series does
not compromise that performance. It seems that the removal of call to
lru_add_drain() still helps a bit.
timedmunlock
3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3 3.11-rc3
0 1 2 3 4 5 6 7
Elapsed min 0.01 ( 0.00%) 0.01 ( -0.11%) 0.01 ( 6.59%) 0.01 ( 5.41%) 0.01 ( 5.45%) 0.01 ( 5.03%) 0.01 ( 6.08%) 0.01 ( 5.20%)
Elapsed mean 0.01 ( 0.00%) 0.01 ( -0.27%) 0.01 ( 6.39%) 0.01 ( 5.30%) 0.01 ( 5.32%) 0.01 ( 5.03%) 0.01 ( 5.97%) 0.01 ( 5.22%)
Elapsed stddev 0.00 ( 0.00%) 0.00 ( -9.59%) 0.00 ( 10.77%) 0.00 ( 3.24%) 0.00 ( 24.42%) 0.00 ( 31.86%) 0.00 ( -7.46%) 0.00 ( 6.11%)
Elapsed max 0.01 ( 0.00%) 0.01 ( -0.01%) 0.01 ( 6.83%) 0.01 ( 5.42%) 0.01 ( 5.79%) 0.01 ( 5.53%) 0.01 ( 6.08%) 0.01 ( 5.26%)
Elapsed range 0.00 ( 0.00%) 0.00 ( 7.30%) 0.00 ( 24.38%) 0.00 ( 6.10%) 0.00 ( 30.79%) 0.00 ( 42.52%) 0.00 ( 6.11%) 0.00 ( 10.07%)
This patch (of 7):
In putback_lru_page() since commit c53954a092 (""mm: remove lru parameter
from __lru_cache_add and lru_cache_add_lru") it is no longer needed to
determine lru list via page_lru_base_type().
This patch replaces it with simple flag is_unevictable which says that the
page was put on the inevictable list. This is the only information that
matters in subsequent tests.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Jörn Engel <joern@logfs.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Michel Lespinasse <walken@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pavel reported that in case if vma area get unmapped and then mapped (or
expanded) in-place, the soft dirty tracker won't be able to recognize this
situation since it works on pte level and ptes are get zapped on unmap,
loosing soft dirty bit of course.
So to resolve this situation we need to track actions on vma level, there
VM_SOFTDIRTY flag comes in. When new vma area created (or old expanded)
we set this bit, and keep it here until application calls for clearing
soft dirty bit.
Thus when user space application track memory changes now it can detect if
vma area is renewed.
Reported-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Rob Landley <rob@landley.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
cpuset_zone_allowed is changed to cpuset_zone_allowed_softwall and the
comment is moved to __cpuset_node_allowed_softwall. So fix this comment.
Signed-off-by: SeungHun Lee <waydi1@gmail.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
I am working with a tool that simulates oracle database I/O workload.
This tool (orion to be specific -
<http://docs.oracle.com/cd/E11882_01/server.112/e16638/iodesign.htm#autoId24>)
allocates hugetlbfs pages using shmget() with SHM_HUGETLB flag. It then
does aio into these pages from flash disks using various common block
sizes used by database. I am looking at performance with two of the most
common block sizes - 1M and 64K. aio performance with these two block
sizes plunged after Transparent HugePages was introduced in the kernel.
Here are performance numbers:
pre-THP 2.6.39 3.11-rc5
1M read 8384 MB/s 5629 MB/s 6501 MB/s
64K read 7867 MB/s 4576 MB/s 4251 MB/s
I have narrowed the performance impact down to the overheads introduced by
THP in __get_page_tail() and put_compound_page() routines. perf top shows
>40% of cycles being spent in these two routines. Every time direct I/O
to hugetlbfs pages starts, kernel calls get_page() to grab a reference to
the pages and calls put_page() when I/O completes to put the reference
away. THP introduced significant amount of locking overhead to get_page()
and put_page() when dealing with compound pages because hugepages can be
split underneath get_page() and put_page(). It added this overhead
irrespective of whether it is dealing with hugetlbfs pages or transparent
hugepages. This resulted in 20%-45% drop in aio performance when using
hugetlbfs pages.
Since hugetlbfs pages can not be split, there is no reason to go through
all the locking overhead for these pages from what I can see. I added
code to __get_page_tail() and put_compound_page() to bypass all the
locking code when working with hugetlbfs pages. This improved performance
significantly. Performance numbers with this patch:
pre-THP 3.11-rc5 3.11-rc5 + Patch
1M read 8384 MB/s 6501 MB/s 8371 MB/s
64K read 7867 MB/s 4251 MB/s 6510 MB/s
Performance with 64K read is still lower than what it was before THP, but
still a 53% improvement. It does mean there is more work to be done but I
will take a 53% improvement for now.
Please take a look at the following patch and let me know if it looks
reasonable.
[akpm@linux-foundation.org: tweak comments]
Signed-off-by: Khalid Aziz <khalid.aziz@oracle.com>
Cc: Pravin B Shelar <pshelar@nicira.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Sha Zhengju
Johannes Weiner
Andrew Morton
Michal Hocko
Li Zefan
Rob Landley
Jan Kara
Cody P Schafer
Greg Thelen
Joe Perches
Yanchuan Nian
Wanpeng Li
Naoya Horiguchi
Wang Sheng-Hui
Maxim Patlasov
Chen Gang
Lisa Du
Vlastimil Babka
Cyrill Gorcunov
SeungHun Lee
Khalid Aziz
