Introduce a CONFIG_DEBUG_VM_VMACACHE option to enable counting the cache
hit rate -- exported in /proc/vmstat.
Any updates to the caching scheme needs this kind of data, thus it can
save some work re-implementing the counting all the time.
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
Cc: Aswin Chandramouleeswaran <aswin@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
vm counters are allowed to be racy. Use raw_cpu_ops to avoid the
local_irq_disable overhead and to avoid preemption checks which will be
added to the __this_cpu operations.
[akpm@linux-foundation.org: Add comment. Again.]
Signed-off-by: Christoph Lameter <cl@linux.com>
Reported-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Cc: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Summary:
The VM maintains cached filesystem pages on two types of lists. One
list holds the pages recently faulted into the cache, the other list
holds pages that have been referenced repeatedly on that first list.
The idea is to prefer reclaiming young pages over those that have shown
to benefit from caching in the past. We call the recently used list
"inactive list" and the frequently used list "active list".
Currently, the VM aims for a 1:1 ratio between the lists, which is the
"perfect" trade-off between the ability to *protect* frequently used
pages and the ability to *detect* frequently used pages. This means
that working set changes bigger than half of cache memory go undetected
and thrash indefinitely, whereas working sets bigger than half of cache
memory are unprotected against used-once streams that don't even need
caching.
This happens on file servers and media streaming servers, where the
popular files and file sections change over time. Even though the
individual files might be smaller than half of memory, concurrent access
to many of them may still result in their inter-reference distance being
greater than half of memory. It's also been reported as a problem on
database workloads that switch back and forth between tables that are
bigger than half of memory. In these cases the VM never recognizes the
new working set and will for the remainder of the workload thrash disk
data which could easily live in memory.
Historically, every reclaim scan of the inactive list also took a
smaller number of pages from the tail of the active list and moved them
to the head of the inactive list. This model gave established working
sets more gracetime in the face of temporary use-once streams, but
ultimately was not significantly better than a FIFO policy and still
thrashed cache based on eviction speed, rather than actual demand for
cache.
This series solves the problem by maintaining a history of pages evicted
from the inactive list, enabling the VM to detect frequently used pages
regardless of inactive list size and facilitate working set transitions.
Tests:
The reported database workload is easily demonstrated on a 8G machine
with two filesets a 6G. This fio workload operates on one set first,
then switches to the other. The VM should obviously always cache the
set that the workload is currently using.
This test is based on a problem encountered by Citus Data customers:
http://citusdata.com/blog/72-linux-memory-manager-and-your-big-data
unpatched:
db1: READ: io=98304MB, aggrb=885559KB/s, minb=885559KB/s, maxb=885559KB/s, mint= 113672msec, maxt= 113672msec
db2: READ: io=98304MB, aggrb= 66169KB/s, minb= 66169KB/s, maxb= 66169KB/s, mint=1521302msec, maxt=1521302msec
sdb: ios=835750/4, merge=2/1, ticks=4659739/60016, in_queue=4719203, util=98.92%
real 27m15.541s
user 0m19.059s
sys 0m51.459s
patched:
db1: READ: io=98304MB, aggrb=877783KB/s, minb=877783KB/s, maxb=877783KB/s, mint=114679msec, maxt=114679msec
db2: READ: io=98304MB, aggrb=397449KB/s, minb=397449KB/s, maxb=397449KB/s, mint=253273msec, maxt=253273msec
sdb: ios=170587/4, merge=2/1, ticks=954910/61123, in_queue=1015923, util=90.40%
real 6m8.630s
user 0m14.714s
sys 0m31.233s
As can be seen, the unpatched kernel simply never adapts to the
workingset change and db2 is stuck indefinitely with secondary storage
speed. The patched kernel needs 2-3 iterations over db2 before it
replaces db1 and reaches full memory speed. Given the unbounded
negative affect of the existing VM behavior, these patches should be
considered correctness fixes rather than performance optimizations.
Another test resembles a fileserver or streaming server workload, where
data in excess of memory size is accessed at different frequencies.
There is very hot data accessed at a high frequency. Machines should be
fitted so that the hot set of such a workload can be fully cached or all
bets are off. Then there is a very big (compared to available memory)
set of data that is used-once or at a very low frequency; this is what
drives the inactive list and does not really benefit from caching.
Lastly, there is a big set of warm data in between that is accessed at
medium frequencies and benefits from caching the pages between the first
and last streamer of each burst.
unpatched:
hot: READ: io=128000MB, aggrb=160693KB/s, minb=160693KB/s, maxb=160693KB/s, mint=815665msec, maxt=815665msec
warm: READ: io= 81920MB, aggrb=109853KB/s, minb= 27463KB/s, maxb= 29244KB/s, mint=717110msec, maxt=763617msec
cold: READ: io= 30720MB, aggrb= 35245KB/s, minb= 35245KB/s, maxb= 35245KB/s, mint=892530msec, maxt=892530msec
sdb: ios=797960/4, merge=11763/1, ticks=4307910/796, in_queue=4308380, util=100.00%
patched:
hot: READ: io=128000MB, aggrb=160678KB/s, minb=160678KB/s, maxb=160678KB/s, mint=815740msec, maxt=815740msec
warm: READ: io= 81920MB, aggrb=147747KB/s, minb= 36936KB/s, maxb= 40960KB/s, mint=512000msec, maxt=567767msec
cold: READ: io= 30720MB, aggrb= 40960KB/s, minb= 40960KB/s, maxb= 40960KB/s, mint=768000msec, maxt=768000msec
sdb: ios=596514/4, merge=9341/1, ticks=2395362/997, in_queue=2396484, util=79.18%
In both kernels, the hot set is propagated to the active list and then
served from cache.
In both kernels, the beginning of the warm set is propagated to the
active list as well, but in the unpatched case the active list
eventually takes up half of memory and no new pages from the warm set
get activated, despite repeated access, and despite most of the active
list soon being stale. The patched kernel on the other hand detects the
thrashing and manages to keep this cache window rolling through the data
set. This frees up enough IO bandwidth that the cold set is served at
full speed as well and disk utilization even drops by 20%.
For reference, this same test was performed with the traditional
demotion mechanism, where deactivation is coupled to inactive list
reclaim. However, this had the same outcome as the unpatched kernel:
while the warm set does indeed get activated continuously, it is forced
out of the active list by inactive list pressure, which is dictated
primarily by the unrelated cold set. The warm set is evicted before
subsequent streamers can benefit from it, even though there would be
enough space available to cache the pages of interest.
Costs:
Page reclaim used to shrink the radix trees but now the tree nodes are
reused for shadow entries, where the cost depends heavily on the page
cache access patterns. However, with workloads that maintain spatial or
temporal locality, the shadow entries are either refaulted quickly or
reclaimed along with the inode object itself. Workloads that will
experience a memory cost increase are those that don't really benefit
from caching in the first place.
A more predictable alternative would be a fixed-cost separate pool of
shadow entries, but this would incur relatively higher memory cost for
well-behaved workloads at the benefit of cornercases. It would also
make the shadow entry lookup more costly compared to storing them
directly in the cache structure.
Future:
To simplify the merging process, this patch set is implementing thrash
detection on a global per-zone level only for now, but the design is
such that it can be extended to memory cgroups as well. All we need to
do is store the unique cgroup ID along the node and zone identifier
inside the eviction cookie to identify the lruvec.
Right now we have a fixed ratio (50:50) between inactive and active list
but we already have complaints about working sets exceeding half of
memory being pushed out of the cache by simple streaming in the
background. Ultimately, we want to adjust this ratio and allow for a
much smaller inactive list. These patches are an essential step in this
direction because they decouple the VMs ability to detect working set
changes from the inactive list size. This would allow us to base the
inactive list size on the combined readahead window size for example and
potentially protect a much bigger working set.
It's also a big step towards activating pages with a reuse distance
larger than memory, as long as they are the most frequently used pages
in the workload. This will require knowing more about the access
frequency of active pages than what we measure right now, so it's also
deferred in this series.
Another possibility of having thrashing information would be to revisit
the idea of local reclaim in the form of zero-config memory control
groups. Instead of having allocating tasks go straight to global
reclaim, they could try to reclaim the pages in the memcg they are part
of first as long as the group is not thrashing. This would allow a user
to drop e.g. a back-up job in an otherwise unconfigured memcg and it
would only inflate (and possibly do global reclaim) until it has enough
memory to do proper readahead. But once it reaches that point and stops
thrashing it would just recycle its own used-once pages without kicking
out the cache of any other tasks in the system more than necessary.
This patch (of 10):
Fengguang Wu's build testing spotted problems with inc_zone_state() and
dec_zone_state() on UP configurations in out-of-tree patches.
inc_zone_state() is declared but not defined, dec_zone_state() is
missing entirely.
Just like with *_zone_page_state(), they can be defined like their
preemption-unsafe counterparts on UP.
[akpm@linux-foundation.org: make it build]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
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 VM is currently heavily tuned to avoid swapping. Whether that is
good or bad is a separate discussion, but as long as the VM won't swap
to make room for dirty cache, we can not consider anonymous pages when
calculating the amount of dirtyable memory, the baseline to which
dirty_background_ratio and dirty_ratio are applied.
A simple workload that occupies a significant size (40+%, depending on
memory layout, storage speeds etc.) of memory with anon/tmpfs pages and
uses the remainder for a streaming writer demonstrates this problem. In
that case, the actual cache pages are a small fraction of what is
considered dirtyable overall, which results in an relatively large
portion of the cache pages to be dirtied. As kswapd starts rotating
these, random tasks enter direct reclaim and stall on IO.
Only consider free pages and file pages dirtyable.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Tejun Heo <tj@kernel.org>
Tested-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Cc: <stable@vger.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>
The main idea behind this patchset is to reduce the vmstat update overhead
by avoiding interrupt enable/disable and the use of per cpu atomics.
This patch (of 3):
It is better to have a separate folding function because
refresh_cpu_vm_stats() also does other things like expire pages in the
page allocator caches.
If we have a separate function then refresh_cpu_vm_stats() is only called
from the local cpu which allows additional optimizations.
The folding function is only called when a cpu is being downed and
therefore no other processor will be accessing the counters. Also
simplifies synchronization.
[akpm@linux-foundation.org: fix UP build]
Signed-off-by: Christoph Lameter <cl@linux.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
CC: Tejun Heo <tj@kernel.org>
Cc: Joonsoo Kim <js1304@gmail.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The current definitions for count_vm_numa_events() is wrong for
!CONFIG_NUMA_BALANCING as the following would miss the side-effect.
count_vm_numa_events(NUMA_FOO, bar++);
There are no such users of count_vm_numa_events() but this patch fixes
it as it is a potential pitfall. Ideally both would be converted to
static inline but NUMA_PTE_UPDATES is not defined if
!CONFIG_NUMA_BALANCING and creating dummy constants just to have a
static inline would be similarly clumsy.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Simon Jeons <simon.jeons@gmail.com>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It is tricky to quantify the basic cost of automatic NUMA placement in a
meaningful manner. This patch adds some vmstats that can be used as part
of a basic costing model.
u = basic unit = sizeof(void *)
Ca = cost of struct page access = sizeof(struct page) / u
Cpte = Cost PTE access = Ca
Cupdate = Cost PTE update = (2 * Cpte) + (2 * Wlock)
where Cpte is incurred twice for a read and a write and Wlock
is a constant representing the cost of taking or releasing a
lock
Cnumahint = Cost of a minor page fault = some high constant e.g. 1000
Cpagerw = Cost to read or write a full page = Ca + PAGE_SIZE/u
Ci = Cost of page isolation = Ca + Wi
where Wi is a constant that should reflect the approximate cost
of the locking operation
Cpagecopy = Cpagerw + (Cpagerw * Wnuma) + Ci + (Ci * Wnuma)
where Wnuma is the approximate NUMA factor. 1 is local. 1.2
would imply that remote accesses are 20% more expensive
Balancing cost = Cpte * numa_pte_updates +
Cnumahint * numa_hint_faults +
Ci * numa_pages_migrated +
Cpagecopy * numa_pages_migrated
Note that numa_pages_migrated is used as a measure of how many pages
were isolated even though it would miss pages that failed to migrate. A
vmstat counter could have been added for it but the isolation cost is
pretty marginal in comparison to the overall cost so it seemed overkill.
The ideal way to measure automatic placement benefit would be to count
the number of remote accesses versus local accesses and do something like
benefit = (remote_accesses_before - remove_access_after) * Wnuma
but the information is not readily available. As a workload converges, the
expection would be that the number of remote numa hints would reduce to 0.
convergence = numa_hint_faults_local / numa_hint_faults
where this is measured for the last N number of
numa hints recorded. When the workload is fully
converged the value is 1.
This can measure if the placement policy is converging and how fast it is
doing it.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
During memory-hotplug, I found NR_ISOLATED_[ANON|FILE] are increasing,
causing the kernel to hang. When the system doesn't have enough free
pages, it enters reclaim but never reclaim any pages due to
too_many_isolated()==true and loops forever.
The cause is that when we do memory-hotadd after memory-remove,
__zone_pcp_update() clears a zone's ZONE_STAT_ITEMS in setup_pageset()
although the vm_stat_diff of all CPUs still have values.
In addtion, when we offline all pages of the zone, we reset them in
zone_pcp_reset without draining so we loss some zone stat item.
Reviewed-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@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>
Add NR_FREE_CMA_PAGES counter to be later used for checking watermark in
__zone_watermark_ok(). For simplicity and to avoid #ifdef hell make this
counter always available (not only when CONFIG_CMA=y).
[akpm@linux-foundation.org: use conventional migratetype naming]
Signed-off-by: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
I found it difficult to make sense of transparent huge pages without
having any counters for its actions. Add some counters to vmstat for
allocation of transparent hugepages and fallback to smaller pages.
Optional patch, but useful for development and understanding the system.
Contains improvements from Andrea Arcangeli and Johannes Weiner
[akpm@linux-foundation.org: coding-style fixes]
[hannes@cmpxchg.org: fix vmstat_text[] entries]
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Add a new __GFP_OTHER_NODE flag to tell the low level numa statistics in
zone_statistics() that an allocation is on behalf of another thread. This
way the local and remote counters can be still correct, even when
background daemons like khugepaged are changing memory mappings.
This only affects the accounting, but I think it's worth doing that right
to avoid confusing users.
I first tried to just pass down the right node, but this required a lot of
changes to pass down this parameter and at least one addition of a 10th
argument to a 9 argument function. Using the flag is a lot less
intrusive.
Open: should be also used for migration?
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
reduce_pgdat_percpu_threshold() and restore_pgdat_percpu_threshold() exist
to adjust the per-cpu vmstat thresholds while kswapd is awake to avoid
errors due to counter drift. The functions duplicate some code so this
patch replaces them with a single set_pgdat_percpu_threshold() that takes
a callback function to calculate the desired threshold as a parameter.
[akpm@linux-foundation.org: readability tweak]
[kosaki.motohiro@jp.fujitsu.com: set_pgdat_percpu_threshold(): don't use for_each_online_cpu]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: 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>
Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory
is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To
avoid synchronization overhead, these counters are maintained on a per-cpu
basis and drained both periodically and when a threshold is above a
threshold. On large CPU systems, the difference between the estimate and
real value of NR_FREE_PAGES can be very high. The system can get into a
case where pages are allocated far below the min watermark potentially
causing livelock issues. The commit solved the problem by taking a better
reading of NR_FREE_PAGES when memory was low.
Unfortately, as reported by Shaohua Li this accurate reading can consume a
large amount of CPU time on systems with many sockets due to cache line
bouncing. This patch takes a different approach. For large machines
where counter drift might be unsafe and while kswapd is awake, the per-cpu
thresholds for the target pgdat are reduced to limit the level of drift to
what should be a safe level. This incurs a performance penalty in heavy
memory pressure by a factor that depends on the workload and the machine
but the machine should function correctly without accidentally exhausting
all memory on a node. There is an additional cost when kswapd wakes and
sleeps but the event is not expected to be frequent - in Shaohua's test
case, there was one recorded sleep and wake event at least.
To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is
introduced that takes a more accurate reading of NR_FREE_PAGES when called
from wakeup_kswapd, when deciding whether it is really safe to go back to
sleep in sleeping_prematurely() and when deciding if a zone is really
balanced or not in balance_pgdat(). We are still using an expensive
function but limiting how often it is called.
When the test case is reproduced, the time spent in the watermark
functions is reduced. The following report is on the percentage of time
spent cumulatively spent in the functions zone_nr_free_pages(),
zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(),
zone_page_state_snapshot(), zone_page_state().
vanilla 11.6615%
disable-threshold 0.2584%
David said:
: We had to pull aa454840 "mm: page allocator: calculate a better estimate
: of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36
: internally because tests showed that it would cause the machine to stall
: as the result of heavy kswapd activity. I merged it back with this fix as
: it is pending in the -mm tree and it solves the issue we were seeing, so I
: definitely think this should be pushed to -stable (and I would seriously
: consider it for 2.6.37 inclusion even at this late date).
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reported-by: Shaohua Li <shaohua.li@intel.com>
Reviewed-by: Christoph Lameter <cl@linux.com>
Tested-by: Nicolas Bareil <nico@chdir.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Kyle McMartin <kyle@mcmartin.ca>
Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Ordinarily watermark checks are based on the vmstat NR_FREE_PAGES as it is
cheaper than scanning a number of lists. To avoid synchronization
overhead, counter deltas are maintained on a per-cpu basis and drained
both periodically and when the delta is above a threshold. On large CPU
systems, the difference between the estimated and real value of
NR_FREE_PAGES can be very high. If NR_FREE_PAGES is much higher than
number of real free page in buddy, the VM can allocate pages below min
watermark, at worst reducing the real number of pages to zero. Even if
the OOM killer kills some victim for freeing memory, it may not free
memory if the exit path requires a new page resulting in livelock.
This patch introduces a zone_page_state_snapshot() function (courtesy of
Christoph) that takes a slightly more accurate view of an arbitrary vmstat
counter. It is used to read NR_FREE_PAGES while kswapd is awake to avoid
the watermark being accidentally broken. The estimate is not perfect and
may result in cache line bounces but is expected to be lighter than the
IPI calls necessary to continually drain the per-cpu counters while kswapd
is awake.
Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Ordinarily when a high-order allocation fails, direct reclaim is entered
to free pages to satisfy the allocation. With this patch, it is
determined if an allocation failed due to external fragmentation instead
of low memory and if so, the calling process will compact until a suitable
page is freed. Compaction by moving pages in memory is considerably
cheaper than paging out to disk and works where there are locked pages or
no swap. If compaction fails to free a page of a suitable size, then
reclaim will still occur.
Direct compaction returns as soon as possible. As each block is
compacted, it is checked if a suitable page has been freed and if so, it
returns.
[akpm@linux-foundation.org: Fix build errors]
[aarcange@redhat.com: fix count_vm_event preempt in memory compaction direct reclaim]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patch is the core of a mechanism which compacts memory in a zone by
relocating movable pages towards the end of the zone.
A single compaction run involves a migration scanner and a free scanner.
Both scanners operate on pageblock-sized areas in the zone. The migration
scanner starts at the bottom of the zone and searches for all movable
pages within each area, isolating them onto a private list called
migratelist. The free scanner starts at the top of the zone and searches
for suitable areas and consumes the free pages within making them
available for the migration scanner. The pages isolated for migration are
then migrated to the newly isolated free pages.
[aarcange@redhat.com: Fix unsafe optimisation]
[mel@csn.ul.ie: do not schedule work on other CPUs for compaction]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
