| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Kmem accounting of memcg is unusable now, because it lacks slab shrinker
support. That means when we hit the limit we will get ENOMEM w/o any
chance to recover. What we should do then is to call shrink_slab, which
would reclaim old inode/dentry caches from this cgroup. This is what
this patch set is intended to do.
Basically, it does two things. First, it introduces the notion of
per-memcg slab shrinker. A shrinker that wants to reclaim objects per
cgroup should mark itself as SHRINKER_MEMCG_AWARE. Then it will be
passed the memory cgroup to scan from in shrink_control->memcg. For
such shrinkers shrink_slab iterates over the whole cgroup subtree under
the target cgroup and calls the shrinker for each kmem-active memory
cgroup.
Secondly, this patch set makes the list_lru structure per-memcg. It's
done transparently to list_lru users - everything they have to do is to
tell list_lru_init that they want memcg-aware list_lru. Then the
list_lru will automatically distribute objects among per-memcg lists
basing on which cgroup the object is accounted to. This way to make FS
shrinkers (icache, dcache) memcg-aware we only need to make them use
memcg-aware list_lru, and this is what this patch set does.
As before, this patch set only enables per-memcg kmem reclaim when the
pressure goes from memory.limit, not from memory.kmem.limit. Handling
memory.kmem.limit is going to be tricky due to GFP_NOFS allocations, and
it is still unclear whether we will have this knob in the unified
hierarchy.
This patch (of 9):
NUMA aware slab shrinkers use the list_lru structure to distribute
objects coming from different NUMA nodes to different lists. Whenever
such a shrinker needs to count or scan objects from a particular node,
it issues commands like this:
count = list_lru_count_node(lru, sc->nid);
freed = list_lru_walk_node(lru, sc->nid, isolate_func,
isolate_arg, &sc->nr_to_scan);
where sc is an instance of the shrink_control structure passed to it
from vmscan.
To simplify this, let's add special list_lru functions to be used by
shrinkers, list_lru_shrink_count() and list_lru_shrink_walk(), which
consolidate the nid and nr_to_scan arguments in the shrink_control
structure.
This will also allow us to avoid patching shrinkers that use list_lru
when we make shrink_slab() per-memcg - all we will have to do is extend
the shrink_control structure to include the target memcg and make
list_lru_shrink_{count,walk} handle this appropriately.
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Suggested-by: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Glauber Costa <glommer@gmail.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
If a PTE or PMD is already marked NUMA when scanning to mark entries for
NUMA hinting then it is not necessary to update the entry and incur a TLB
flush penalty. Avoid the avoidhead where possible.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kirill Shutemov <kirill.shutemov@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
pte_protnone_numa is only safe to use after VMA checks for PROT_NONE are
complete. Treating a real PROT_NONE PTE as a NUMA hinting fault is going
to result in strangeness so add a check for it. BUG_ON looks like
overkill but if this is hit then it's a serious bug that could result in
corruption so do not even try recovering. It would have been more
comprehensive to check VMA flags in pte_protnone_numa but it would have
made the API ugly just for a debugging check.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kirill Shutemov <kirill.shutemov@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Faults on the huge zero page are pointless and there is a BUG_ON to catch
them during fault time. This patch reintroduces a check that avoids
marking the zero page PAGE_NONE.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kirill Shutemov <kirill.shutemov@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
With PROT_NONE, the traditional page table manipulation functions are
sufficient.
[andre.przywara@arm.com: fix compiler warning in pmdp_invalidate()]
[akpm@linux-foundation.org: fix build with STRICT_MM_TYPECHECKS]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Acked-by: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Tested-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kirill Shutemov <kirill.shutemov@linux.intel.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Convert existing users of pte_numa and friends to the new helper. Note
that the kernel is broken after this patch is applied until the other page
table modifiers are also altered. This patch layout is to make review
easier.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Acked-by: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Tested-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Dave Jones <davej@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kirill Shutemov <kirill.shutemov@linux.intel.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Automatic NUMA balancing depends on being able to protect PTEs to trap a
fault and gather reference locality information. Very broadly speaking
it would mark PTEs as not present and use another bit to distinguish
between NUMA hinting faults and other types of faults. It was
universally loved by everybody and caused no problems whatsoever. That
last sentence might be a lie.
This series is very heavily based on patches from Linus and Aneesh to
replace the existing PTE/PMD NUMA helper functions with normal change
protections. I did alter and add parts of it but I consider them
relatively minor contributions. At their suggestion, acked-bys are in
there but I've no problem converting them to Signed-off-by if requested.
AFAIK, this has received no testing on ppc64 and I'm depending on Aneesh
for that. I tested trinity under kvm-tool and passed and ran a few
other basic tests. At the time of writing, only the short-lived tests
have completed but testing of V2 indicated that long-term testing had no
surprises. In most cases I'm leaving out detail as it's not that
interesting.
specjbb single JVM: There was negligible performance difference in the
benchmark itself for short runs. However, system activity is
higher and interrupts are much higher over time -- possibly TLB
flushes. Migrations are also higher. Overall, this is more overhead
but considering the problems faced with the old approach I think
we just have to suck it up and find another way of reducing the
overhead.
specjbb multi JVM: Negligible performance difference to the actual benchmark
but like the single JVM case, the system overhead is noticeably
higher. Again, interrupts are a major factor.
autonumabench: This was all over the place and about all that can be
reasonably concluded is that it's different but not necessarily
better or worse.
autonumabench
3.18.0-rc5 3.18.0-rc5
mmotm-20141119 protnone-v3r3
User NUMA01 32380.24 ( 0.00%) 21642.92 ( 33.16%)
User NUMA01_THEADLOCAL 22481.02 ( 0.00%) 22283.22 ( 0.88%)
User NUMA02 3137.00 ( 0.00%) 3116.54 ( 0.65%)
User NUMA02_SMT 1614.03 ( 0.00%) 1543.53 ( 4.37%)
System NUMA01 322.97 ( 0.00%) 1465.89 (-353.88%)
System NUMA01_THEADLOCAL 91.87 ( 0.00%) 49.32 ( 46.32%)
System NUMA02 37.83 ( 0.00%) 14.61 ( 61.38%)
System NUMA02_SMT 7.36 ( 0.00%) 7.45 ( -1.22%)
Elapsed NUMA01 716.63 ( 0.00%) 599.29 ( 16.37%)
Elapsed NUMA01_THEADLOCAL 553.98 ( 0.00%) 539.94 ( 2.53%)
Elapsed NUMA02 83.85 ( 0.00%) 83.04 ( 0.97%)
Elapsed NUMA02_SMT 86.57 ( 0.00%) 79.15 ( 8.57%)
CPU NUMA01 4563.00 ( 0.00%) 3855.00 ( 15.52%)
CPU NUMA01_THEADLOCAL 4074.00 ( 0.00%) 4136.00 ( -1.52%)
CPU NUMA02 3785.00 ( 0.00%) 3770.00 ( 0.40%)
CPU NUMA02_SMT 1872.00 ( 0.00%) 1959.00 ( -4.65%)
System CPU usage of NUMA01 is worse but it's an adverse workload on this
machine so I'm reluctant to conclude that it's a problem that matters. On
the other workloads that are sensible on this machine, system CPU usage is
great. Overall time to complete the benchmark is comparable
3.18.0-rc5 3.18.0-rc5
mmotm-20141119protnone-v3r3
User 59612.50 48586.44
System 460.22 1537.45
Elapsed 1442.20 1304.29
NUMA alloc hit 5075182 5743353
NUMA alloc miss 0 0
NUMA interleave hit 0 0
NUMA alloc local 5075174 5743339
NUMA base PTE updates 637061448 443106883
NUMA huge PMD updates 1243434 864747
NUMA page range updates 1273699656 885857347
NUMA hint faults 1658116 1214277
NUMA hint local faults 959487 754113
NUMA hint local percent 57 62
NUMA pages migrated 5467056 61676398
The NUMA pages migrated look terrible but when I looked at a graph of the
activity over time I see that the massive spike in migration activity was
during NUMA01. This correlates with high system CPU usage and could be
simply down to bad luck but any modifications that affect that workload
would be related to scan rates and migrations, not the protection
mechanism. For all other workloads, migration activity was comparable.
Overall, headline performance figures are comparable but the overhead is
higher, mostly in interrupts. To some extent, higher overhead from this
approach was anticipated but not to this degree. It's going to be
necessary to reduce this again with a separate series in the future. It's
still worth going ahead with this series though as it's likely to avoid
constant headaches with Xen and is probably easier to maintain.
This patch (of 10):
A transhuge NUMA hinting fault may find the page is migrating and should
wait until migration completes. The check is race-prone because the pmd
is deferenced outside of the page lock and while the race is tiny, it'll
be larger if the PMD is cleared while marking PMDs for hinting fault.
This patch closes the race.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kirill Shutemov <kirill.shutemov@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
I noticed that "allowed" can easily overflow by falling below 0, because
(total_vm / 32) can be larger than "allowed". The problem occurs in
OVERCOMMIT_NONE mode.
In this case, a huge allocation can success and overcommit the system
(despite OVERCOMMIT_NONE mode). All subsequent allocations will fall
(system-wide), so system become unusable.
The problem was masked out by commit c9b1d0981fcc
("mm: limit growth of 3% hardcoded other user reserve"),
but it's easy to reproduce it on older kernels:
1) set overcommit_memory sysctl to 2
2) mmap() large file multiple times (with VM_SHARED flag)
3) try to malloc() large amount of memory
It also can be reproduced on newer kernels, but miss-configured
sysctl_user_reserve_kbytes is required.
Fix this issue by switching to signed arithmetic here.
Signed-off-by: Roman Gushchin <klamm@yandex-team.ru>
Cc: Andrew Shewmaker <agshew@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
I noticed, that "allowed" can easily overflow by falling below 0,
because (total_vm / 32) can be larger than "allowed". The problem
occurs in OVERCOMMIT_NONE mode.
In this case, a huge allocation can success and overcommit the system
(despite OVERCOMMIT_NONE mode). All subsequent allocations will fall
(system-wide), so system become unusable.
The problem was masked out by commit c9b1d0981fcc
("mm: limit growth of 3% hardcoded other user reserve"),
but it's easy to reproduce it on older kernels:
1) set overcommit_memory sysctl to 2
2) mmap() large file multiple times (with VM_SHARED flag)
3) try to malloc() large amount of memory
It also can be reproduced on newer kernels, but miss-configured
sysctl_user_reserve_kbytes is required.
Fix this issue by switching to signed arithmetic here.
[akpm@linux-foundation.org: use min_t]
Signed-off-by: Roman Gushchin <klamm@yandex-team.ru>
Cc: Andrew Shewmaker <agshew@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
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>
|
|
It was noted that the vm stat shepherd runs every 2 seconds and that the
vmstat update is then scheduled 2 seconds in the future.
This yields an interval of double the time interval which is not desired.
Change the shepherd so that it does not delay the vmstat update on the
other cpu. We stil have to use schedule_delayed_work since we are using a
delayed_work_struct but we can set the delay to 0.
Signed-off-by: Christoph Lameter <cl@linux.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Page owner uses the page_ext structure to keep meta-information for every
page in the system. The structure also contains a field of type 'struct
stack_trace', page owner uses this field during invocation of the function
save_stack_trace. It is easy to notice that keeping a copy of this
structure for every page in the system is very inefficiently in terms of
memory.
The patch removes this unnecessary field of page_ext and forces page owner
to use a stack_trace structure allocated on the stack.
[akpm@linux-foundation.org: use struct initializers]
Signed-off-by: Sergei Rogachev <rogachevsergei@gmail.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
This patch aims to improve THP collapse rates, by allowing THP collapse in
the presence of read-only ptes, like those left in place by do_swap_page
after a read fault.
Currently THP can collapse 4kB pages into a THP when there are up to
khugepaged_max_ptes_none pte_none ptes in a 2MB range. This patch applies
the same limit for read-only ptes.
The patch was tested with a test program that allocates 800MB of memory,
writes to it, and then sleeps. I force the system to swap out all but
190MB of the program by touching other memory. Afterwards, the test
program does a mix of reads and writes to its memory, and the memory gets
swapped back in.
Without the patch, only the memory that did not get swapped out remained
in THPs, which corresponds to 24% of the memory of the program. The
percentage did not increase over time.
With this patch, after 5 minutes of waiting khugepaged had collapsed 50%
of the program's memory back into THPs.
Test results:
With the patch:
After swapped out:
cat /proc/pid/smaps:
Anonymous: 100464 kB
AnonHugePages: 100352 kB
Swap: 699540 kB
Fraction: 99,88
cat /proc/meminfo:
AnonPages: 1754448 kB
AnonHugePages: 1716224 kB
Fraction: 97,82
After swapped in:
In a few seconds:
cat /proc/pid/smaps:
Anonymous: 800004 kB
AnonHugePages: 145408 kB
Swap: 0 kB
Fraction: 18,17
cat /proc/meminfo:
AnonPages: 2455016 kB
AnonHugePages: 1761280 kB
Fraction: 71,74
In 5 minutes:
cat /proc/pid/smaps
Anonymous: 800004 kB
AnonHugePages: 407552 kB
Swap: 0 kB
Fraction: 50,94
cat /proc/meminfo:
AnonPages: 2456872 kB
AnonHugePages: 2023424 kB
Fraction: 82,35
Without the patch:
After swapped out:
cat /proc/pid/smaps:
Anonymous: 190660 kB
AnonHugePages: 190464 kB
Swap: 609344 kB
Fraction: 99,89
cat /proc/meminfo:
AnonPages: 1740456 kB
AnonHugePages: 1667072 kB
Fraction: 95,78
After swapped in:
cat /proc/pid/smaps:
Anonymous: 800004 kB
AnonHugePages: 190464 kB
Swap: 0 kB
Fraction: 23,80
cat /proc/meminfo:
AnonPages: 2350032 kB
AnonHugePages: 1667072 kB
Fraction: 70,93
I waited 10 minutes the fractions did not change without the patch.
Signed-off-by: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Acked-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Sasha Levin <sasha.levin@oracle.com>
Cc: Hugh Dickins <hughd@google.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>
|
|
Vinayak Menon has reported that an excessive number of tasks was throttled
in the direct reclaim inside too_many_isolated() because NR_ISOLATED_FILE
was relatively high compared to NR_INACTIVE_FILE. However it turned out
that the real number of NR_ISOLATED_FILE was 0 and the per-cpu
vm_stat_diff wasn't transferred into the global counter.
vmstat_work which is responsible for the sync is defined as deferrable
delayed work which means that the defined timeout doesn't wake up an idle
CPU. A CPU might stay in an idle state for a long time and general effort
is to keep such a CPU in this state as long as possible which might lead
to all sorts of troubles for vmstat consumers as can be seen with the
excessive direct reclaim throttling.
This patch basically reverts 39bf6270f524 ("VM statistics: Make timer
deferrable") but it shouldn't cause any problems for idle CPUs because
only CPUs with an active per-cpu drift are woken up since 7cc36bbddde5
("vmstat: on-demand vmstat workers v8") and CPUs which are idle for a
longer time shouldn't have per-cpu drift.
Fixes: 39bf6270f524 (VM statistics: Make timer deferrable)
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Reported-by: Vinayak Menon <vinmenon@codeaurora.org>
Acked-by: Christoph Lameter <cl@linux.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
When allocation falls back to stealing free pages of another migratetype,
it can decide to steal extra pages, or even the whole pageblock in order
to reduce fragmentation, which could happen if further allocation
fallbacks pick a different pageblock. In try_to_steal_freepages(), one of
the situations where extra pages are stolen happens when we are trying to
allocate a MIGRATE_RECLAIMABLE page.
However, MIGRATE_UNMOVABLE allocations are not treated the same way,
although spreading such allocation over multiple fallback pageblocks is
arguably even worse than it is for RECLAIMABLE allocations. To minimize
fragmentation, we should minimize the number of such fallbacks, and thus
steal as much as is possible from each fallback pageblock.
Note that in theory this might put more pressure on movable pageblocks and
cause movable allocations to steal back from unmovable pageblocks.
However, movable allocations are not as aggressive with stealing, and do
not cause permanent fragmentation, so the tradeoff is reasonable, and
evaluation seems to support the change.
This patch thus adds a check for MIGRATE_UNMOVABLE to the decision to
steal extra free pages. When evaluating with stress-highalloc from
mmtests, this has reduced the number of MIGRATE_UNMOVABLE fallbacks to
roughly 1/6. The number of these fallbacks stealing from MIGRATE_MOVABLE
block is reduced to 1/3. There was no observation of growing number of
unmovable pageblocks over time, and also not of increased movable
allocation fallbacks.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Hocko <mhocko@suse.cz>
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>
|
|
When allocation falls back to another migratetype, it will steal a page
with highest available order, and (depending on this order and desired
migratetype), it might also steal the rest of free pages from the same
pageblock.
Given the preference of highest available order, it is likely that it will
be higher than the desired order, and result in the stolen buddy page
being split. The remaining pages after split are currently stolen only
when the rest of the free pages are stolen. This can however lead to
situations where for MOVABLE allocations we split e.g. order-4 fallback
UNMOVABLE page, but steal only order-0 page. Then on the next MOVABLE
allocation (which may be batched to fill the pcplists) we split another
order-3 or higher page, etc. By stealing all pages that we have split, we
can avoid further stealing.
This patch therefore adjusts the page stealing so that buddy pages created
by split are always stolen. This has effect only on MOVABLE allocations,
as RECLAIMABLE and UNMOVABLE allocations already always do that in
addition to stealing the rest of free pages from the pageblock. The
change also allows to simplify try_to_steal_freepages() and factor out CMA
handling.
According to Mel, it has been intended since the beginning that buddy
pages after split would be stolen always, but it doesn't seem like it was
ever the case until commit 47118af076f6 ("mm: mmzone: MIGRATE_CMA
migration type added"). The commit has unintentionally introduced this
behavior, but was reverted by commit 0cbef29a7821 ("mm:
__rmqueue_fallback() should respect pageblock type"). Neither included
evaluation.
My evaluation with stress-highalloc from mmtests shows about 2.5x
reduction of page stealing events for MOVABLE allocations, without
affecting the page stealing events for other allocation migratetypes.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Hocko <mhocko@suse.cz>
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>
|
|
When studying page stealing, I noticed some weird looking decisions in
try_to_steal_freepages(). The first I assume is a bug (Patch 1), the
following two patches were driven by evaluation.
Testing was done with stress-highalloc of mmtests, using the
mm_page_alloc_extfrag tracepoint and postprocessing to get counts of how
often page stealing occurs for individual migratetypes, and what
migratetypes are used for fallbacks. Arguably, the worst case of page
stealing is when UNMOVABLE allocation steals from MOVABLE pageblock.
RECLAIMABLE allocation stealing from MOVABLE allocation is also not ideal,
so the goal is to minimize these two cases.
The evaluation of v2 wasn't always clear win and Joonsoo questioned the
results. Here I used different baseline which includes RFC compaction
improvements from [1]. I found that the compaction improvements reduce
variability of stress-highalloc, so there's less noise in the data.
First, let's look at stress-highalloc configured to do sync compaction,
and how these patches reduce page stealing events during the test. First
column is after fresh reboot, other two are reiterations of test without
reboot. That was all accumulater over 5 re-iterations (so the benchmark
was run 5x3 times with 5 fresh restarts).
Baseline:
3.19-rc4 3.19-rc4 3.19-rc4
5-nothp-1 5-nothp-2 5-nothp-3
Page alloc extfrag event 10264225 8702233 10244125
Extfrag fragmenting 10263271 8701552 10243473
Extfrag fragmenting for unmovable 13595 17616 15960
Extfrag fragmenting unmovable placed with movable 7989 12193 8447
Extfrag fragmenting for reclaimable 658 1840 1817
Extfrag fragmenting reclaimable placed with movable 558 1677 1679
Extfrag fragmenting for movable 10249018 8682096 10225696
With Patch 1:
3.19-rc4 3.19-rc4 3.19-rc4
6-nothp-1 6-nothp-2 6-nothp-3
Page alloc extfrag event 11834954 9877523 9774860
Extfrag fragmenting 11833993 9876880 9774245
Extfrag fragmenting for unmovable 7342 16129 11712
Extfrag fragmenting unmovable placed with movable 4191 10547 6270
Extfrag fragmenting for reclaimable 373 1130 923
Extfrag fragmenting reclaimable placed with movable 302 906 738
Extfrag fragmenting for movable 11826278 9859621 9761610
With Patch 2:
3.19-rc4 3.19-rc4 3.19-rc4
7-nothp-1 7-nothp-2 7-nothp-3
Page alloc extfrag event 4725990 3668793 3807436
Extfrag fragmenting 4725104 3668252 3806898
Extfrag fragmenting for unmovable 6678 7974 7281
Extfrag fragmenting unmovable placed with movable 2051 3829 4017
Extfrag fragmenting for reclaimable 429 1208 1278
Extfrag fragmenting reclaimable placed with movable 369 976 1034
Extfrag fragmenting for movable 4717997 3659070 3798339
With Patch 3:
3.19-rc4 3.19-rc4 3.19-rc4
8-nothp-1 8-nothp-2 8-nothp-3
Page alloc extfrag event 5016183 4700142 3850633
Extfrag fragmenting 5015325 4699613 3850072
Extfrag fragmenting for unmovable 1312 3154 3088
Extfrag fragmenting unmovable placed with movable 1115 2777 2714
Extfrag fragmenting for reclaimable 437 1193 1097
Extfrag fragmenting reclaimable placed with movable 330 969 879
Extfrag fragmenting for movable 5013576 4695266 3845887
In v2 we've seen apparent regression with Patch 1 for unmovable events,
this is now gone, suggesting it was indeed noise. Here, each patch
improves the situation for unmovable events. Reclaimable is improved by
patch 1 and then either the same modulo noise, or perhaps sligtly worse -
a small price for unmovable improvements, IMHO. The number of movable
allocations falling back to other migratetypes is most noisy, but it's
reduced to half at Patch 2 nevertheless. These are least critical as
compaction can move them around.
If we look at success rates, the patches don't affect them, that didn't change.
Baseline:
3.19-rc4 3.19-rc4 3.19-rc4
5-nothp-1 5-nothp-2 5-nothp-3
Success 1 Min 49.00 ( 0.00%) 42.00 ( 14.29%) 41.00 ( 16.33%)
Success 1 Mean 51.00 ( 0.00%) 45.00 ( 11.76%) 42.60 ( 16.47%)
Success 1 Max 55.00 ( 0.00%) 51.00 ( 7.27%) 46.00 ( 16.36%)
Success 2 Min 53.00 ( 0.00%) 47.00 ( 11.32%) 44.00 ( 16.98%)
Success 2 Mean 59.60 ( 0.00%) 50.80 ( 14.77%) 48.20 ( 19.13%)
Success 2 Max 64.00 ( 0.00%) 56.00 ( 12.50%) 52.00 ( 18.75%)
Success 3 Min 84.00 ( 0.00%) 82.00 ( 2.38%) 78.00 ( 7.14%)
Success 3 Mean 85.60 ( 0.00%) 82.80 ( 3.27%) 79.40 ( 7.24%)
Success 3 Max 86.00 ( 0.00%) 83.00 ( 3.49%) 80.00 ( 6.98%)
Patch 1:
3.19-rc4 3.19-rc4 3.19-rc4
6-nothp-1 6-nothp-2 6-nothp-3
Success 1 Min 49.00 ( 0.00%) 44.00 ( 10.20%) 44.00 ( 10.20%)
Success 1 Mean 51.80 ( 0.00%) 46.00 ( 11.20%) 45.80 ( 11.58%)
Success 1 Max 54.00 ( 0.00%) 49.00 ( 9.26%) 49.00 ( 9.26%)
Success 2 Min 58.00 ( 0.00%) 49.00 ( 15.52%) 48.00 ( 17.24%)
Success 2 Mean 60.40 ( 0.00%) 51.80 ( 14.24%) 50.80 ( 15.89%)
Success 2 Max 63.00 ( 0.00%) 54.00 ( 14.29%) 55.00 ( 12.70%)
Success 3 Min 84.00 ( 0.00%) 81.00 ( 3.57%) 79.00 ( 5.95%)
Success 3 Mean 85.00 ( 0.00%) 81.60 ( 4.00%) 79.80 ( 6.12%)
Success 3 Max 86.00 ( 0.00%) 82.00 ( 4.65%) 82.00 ( 4.65%)
Patch 2:
3.19-rc4 3.19-rc4 3.19-rc4
7-nothp-1 7-nothp-2 7-nothp-3
Success 1 Min 50.00 ( 0.00%) 44.00 ( 12.00%) 39.00 ( 22.00%)
Success 1 Mean 52.80 ( 0.00%) 45.60 ( 13.64%) 42.40 ( 19.70%)
Success 1 Max 55.00 ( 0.00%) 46.00 ( 16.36%) 47.00 ( 14.55%)
Success 2 Min 52.00 ( 0.00%) 48.00 ( 7.69%) 45.00 ( 13.46%)
Success 2 Mean 53.40 ( 0.00%) 49.80 ( 6.74%) 48.80 ( 8.61%)
Success 2 Max 57.00 ( 0.00%) 52.00 ( 8.77%) 52.00 ( 8.77%)
Success 3 Min 84.00 ( 0.00%) 81.00 ( 3.57%) 79.00 ( 5.95%)
Success 3 Mean 85.00 ( 0.00%) 82.40 ( 3.06%) 79.60 ( 6.35%)
Success 3 Max 86.00 ( 0.00%) 83.00 ( 3.49%) 80.00 ( 6.98%)
Patch 3:
3.19-rc4 3.19-rc4 3.19-rc4
8-nothp-1 8-nothp-2 8-nothp-3
Success 1 Min 46.00 ( 0.00%) 44.00 ( 4.35%) 42.00 ( 8.70%)
Success 1 Mean 50.20 ( 0.00%) 45.60 ( 9.16%) 44.00 ( 12.35%)
Success 1 Max 52.00 ( 0.00%) 47.00 ( 9.62%) 47.00 ( 9.62%)
Success 2 Min 53.00 ( 0.00%) 49.00 ( 7.55%) 48.00 ( 9.43%)
Success 2 Mean 55.80 ( 0.00%) 50.60 ( 9.32%) 49.00 ( 12.19%)
Success 2 Max 59.00 ( 0.00%) 52.00 ( 11.86%) 51.00 ( 13.56%)
Success 3 Min 84.00 ( 0.00%) 80.00 ( 4.76%) 79.00 ( 5.95%)
Success 3 Mean 85.40 ( 0.00%) 81.60 ( 4.45%) 80.40 ( 5.85%)
Success 3 Max 87.00 ( 0.00%) 83.00 ( 4.60%) 82.00 ( 5.75%)
While there's no improvement here, I consider reduced fragmentation events
to be worth on its own. Patch 2 also seems to reduce scanning for free
pages, and migrations in compaction, suggesting it has somewhat less work
to do:
Patch 1:
Compaction stalls 4153 3959 3978
Compaction success 1523 1441 1446
Compaction failures 2630 2517 2531
Page migrate success 4600827 4943120 5104348
Page migrate failure 19763 16656 17806
Compaction pages isolated 9597640 10305617 10653541
Compaction migrate scanned 77828948 86533283 87137064
Compaction free scanned 517758295 521312840 521462251
Compaction cost 5503 5932 6110
Patch 2:
Compaction stalls 3800 3450 3518
Compaction success 1421 1316 1317
Compaction failures 2379 2134 2201
Page migrate success 4160421 4502708 4752148
Page migrate failure 19705 14340 14911
Compaction pages isolated 8731983 9382374 9910043
Compaction migrate scanned 98362797 96349194 98609686
Compaction free scanned 496512560 469502017 480442545
Compaction cost 5173 5526 5811
As with v2, /proc/pagetypeinfo appears unaffected with respect to numbers
of unmovable and reclaimable pageblocks.
Configuring the benchmark to allocate like THP page fault (i.e. no sync
compaction) gives much noisier results for iterations 2 and 3 after
reboot. This is not so surprising given how [1] offers lower improvements
in this scenario due to less restarts after deferred compaction which
would change compaction pivot.
Baseline:
3.19-rc4 3.19-rc4 3.19-rc4
5-thp-1 5-thp-2 5-thp-3
Page alloc extfrag event 8148965 6227815 6646741
Extfrag fragmenting 8147872 6227130 6646117
Extfrag fragmenting for unmovable 10324 12942 15975
Extfrag fragmenting unmovable placed with movable 5972 8495 10907
Extfrag fragmenting for reclaimable 601 1707 2210
Extfrag fragmenting reclaimable placed with movable 520 1570 2000
Extfrag fragmenting for movable 8136947 6212481 6627932
Patch 1:
3.19-rc4 3.19-rc4 3.19-rc4
6-thp-1 6-thp-2 6-thp-3
Page alloc extfrag event 8345457 7574471 7020419
Extfrag fragmenting 8343546 7573777 7019718
Extfrag fragmenting for unmovable 10256 18535 30716
Extfrag fragmenting unmovable placed with movable 6893 11726 22181
Extfrag fragmenting for reclaimable 465 1208 1023
Extfrag fragmenting reclaimable placed with movable 353 996 843
Extfrag fragmenting for movable 8332825 7554034 6987979
Patch 2:
3.19-rc4 3.19-rc4 3.19-rc4
7-thp-1 7-thp-2 7-thp-3
Page alloc extfrag event 3512847 3020756 2891625
Extfrag fragmenting 3511940 3020185 2891059
Extfrag fragmenting for unmovable 9017 6892 6191
Extfrag fragmenting unmovable placed with movable 1524 3053 2435
Extfrag fragmenting for reclaimable 445 1081 1160
Extfrag fragmenting reclaimable placed with movable 375 918 986
Extfrag fragmenting for movable 3502478 3012212 2883708
Patch 3:
3.19-rc4 3.19-rc4 3.19-rc4
8-thp-1 8-thp-2 8-thp-3
Page alloc extfrag event 3181699 3082881 2674164
Extfrag fragmenting 3180812 3082303 2673611
Extfrag fragmenting for unmovable 1201 4031 4040
Extfrag fragmenting unmovable placed with movable 974 3611 3645
Extfrag fragmenting for reclaimable 478 1165 1294
Extfrag fragmenting reclaimable placed with movable 387 985 1030
Extfrag fragmenting for movable 3179133 3077107 2668277
The improvements for first iteration are clear, the rest is much noisier
and can appear like regression for Patch 1. Anyway, patch 2 rectifies it.
Allocation success rates are again unaffected so there's no point in
making this e-mail any longer.
[1] http://marc.info/?l=linux-mm&m=142166196321125&w=2
This patch (of 3):
When __rmqueue_fallback() is called to allocate a page of order X, it will
find a page of order Y >= X of a fallback migratetype, which is different
from the desired migratetype. With the help of try_to_steal_freepages(),
it may change the migratetype (to the desired one) also of:
1) all currently free pages in the pageblock containing the fallback page
2) the fallback pageblock itself
3) buddy pages created by splitting the fallback page (when Y > X)
These decisions take the order Y into account, as well as the desired
migratetype, with the goal of preventing multiple fallback allocations
that could e.g. distribute UNMOVABLE allocations among multiple
pageblocks.
Originally, decision for 1) has implied the decision for 3). Commit
47118af076f6 ("mm: mmzone: MIGRATE_CMA migration type added") changed that
(probably unintentionally) so that the buddy pages in case 3) are always
changed to the desired migratetype, except for CMA pageblocks.
Commit fef903efcf0c ("mm/page_allo.c: restructure free-page stealing code
and fix a bug") did some refactoring and added a comment that the case of
3) is intended. Commit 0cbef29a7821 ("mm: __rmqueue_fallback() should
respect pageblock type") removed the comment and tried to restore the
original behavior where 1) implies 3), but due to the previous
refactoring, the result is instead that only 2) implies 3) - and the
conditions for 2) are less frequently met than conditions for 1). This
may increase fragmentation in situations where the code decides to steal
all free pages from the pageblock (case 1)), but then gives back the buddy
pages produced by splitting.
This patch restores the original intended logic where 1) implies 3).
During testing with stress-highalloc from mmtests, this has shown to
decrease the number of events where UNMOVABLE and RECLAIMABLE allocations
steal from MOVABLE pageblocks, which can lead to permanent fragmentation.
In some cases it has increased the number of events when MOVABLE
allocations steal from UNMOVABLE or RECLAIMABLE pageblocks, but these are
fixable by sync compaction and thus less harmful.
Note that evaluation has shown that the behavior introduced by
47118af076f6 for buddy pages in case 3) is actually even better than the
original logic, so the following patch will introduce it properly once
again. For stable backports of this patch it makes thus sense to only fix
versions containing 0cbef29a7821.
[iamjoonsoo.kim@lge.com: tracepoint fix]
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: <stable@vger.kernel.org> [3.13+ containing 0cbef29a7821]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
This patch makes do_mincore() use walk_page_vma(), which reduces many
lines of code by using common page table walk code.
[daeseok.youn@gmail.com: remove unneeded variable 'err']
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Daeseok Youn <daeseok.youn@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
walk_page_range() silently skips vma having VM_PFNMAP set, which leads to
undesirable behaviour at client end (who called walk_page_range). For
example for pagemap_read(), when no callbacks are called against VM_PFNMAP
vma, pagemap_read() may prepare pagemap data for next virtual address
range at wrong index. That could confuse and/or break userspace
applications.
This patch avoid this misbehavior caused by vma(VM_PFNMAP) like follows:
- for pagemap_read() which has its own ->pte_hole(), call the ->pte_hole()
over vma(VM_PFNMAP),
- for clear_refs and queue_pages which have their own ->tests_walk,
just return 1 and skip vma(VM_PFNMAP). This is no problem because
these are not interested in hole regions,
- for other callers, just skip the vma(VM_PFNMAP) as a default behavior.
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Shiraz Hashim <shashim@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
queue_pages_range() does page table walking in its own way now, but there
is some code duplicate. This patch applies page table walker to reduce
lines of code.
queue_pages_range() has to do some precheck to determine whether we really
walk over the vma or just skip it. Now we have test_walk() callback in
mm_walk for this purpose, so we can do this replacement cleanly.
queue_pages_test_walk() depends on not only the current vma but also the
previous one, so queue_pages->prev is introduced to remember it.
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
pagewalk.c can handle vma in itself, so we don't have to pass vma via
walk->private. And both of mem_cgroup_count_precharge() and
mem_cgroup_move_charge() do for each vma loop themselves, but now it's
done in pagewalk.c, so let's clean up them.
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Introduce walk_page_vma(), which is useful for the callers which want to
walk over a given vma. It's used by later patches.
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Current implementation of page table walker has a fundamental problem in
vma handling, which started when we tried to handle vma(VM_HUGETLB).
Because it's done in pgd loop, considering vma boundary makes code
complicated and bug-prone.
From the users viewpoint, some user checks some vma-related condition to
determine whether the user really does page walk over the vma.
In order to solve these, this patch moves vma check outside pgd loop and
introduce a new callback ->test_walk().
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Currently no user of page table walker sets ->pgd_entry() or
->pud_entry(), so checking their existence in each loop is just wasting
CPU cycle. So let's remove it to reduce overhead.
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
This allows those get_user_pages calls to pass FAULT_FLAG_ALLOW_RETRY to
the page fault in order to release the mmap_sem during the I/O.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
This allows the get_user_pages_fast slow path to release the mmap_sem
before blocking.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Some callers (like KVM) may want to set the gup_flags like FOLL_HWPOSION
to get a proper -EHWPOSION retval instead of -EFAULT to take a more
appropriate action if get_user_pages runs into a memory failure.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Peter Feiner <pfeiner@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
FAULT_FOLL_ALLOW_RETRY allows the page fault to drop the mmap_sem for
reading to reduce the mmap_sem contention (for writing), like while
waiting for I/O completion. The problem is that right now practically no
get_user_pages call uses FAULT_FOLL_ALLOW_RETRY, so we're not leveraging
that nifty feature.
Andres fixed it for the KVM page fault. However get_user_pages_fast
remains uncovered, and 99% of other get_user_pages aren't using it either
(the only exception being FOLL_NOWAIT in KVM which is really nonblocking
and in fact it doesn't even release the mmap_sem).
So this patchsets extends the optimization Andres did in the KVM page
fault to the whole kernel. It makes most important places (including
gup_fast) to use FAULT_FOLL_ALLOW_RETRY to reduce the mmap_sem hold times
during I/O.
The only few places that remains uncovered are drivers like v4l and other
exceptions that tends to work on their own memory and they're not working
on random user memory (for example like O_DIRECT that uses gup_fast and is
fully covered by this patch).
A follow up patch should probably also add a printk_once warning to
get_user_pages that should go obsolete and be phased out eventually. The
"vmas" parameter of get_user_pages makes it fundamentally incompatible
with FAULT_FOLL_ALLOW_RETRY (vmas array becomes meaningless the moment the
mmap_sem is released).
While this is just an optimization, this becomes an absolute requirement
for the userfaultfd feature http://lwn.net/Articles/615086/ .
The userfaultfd allows to block the page fault, and in order to do so I
need to drop the mmap_sem first. So this patch also ensures that all
memory where userfaultfd could be registered by KVM, the very first fault
(no matter if it is a regular page fault, or a get_user_pages) always has
FAULT_FOLL_ALLOW_RETRY set. Then the userfaultfd blocks and it is waken
only when the pagetable is already mapped. The second fault attempt after
the wakeup doesn't need FAULT_FOLL_ALLOW_RETRY, so it's ok to retry
without it.
This patch (of 5):
We can leverage the VM_FAULT_RETRY functionality in the page fault paths
better by using either get_user_pages_locked or get_user_pages_unlocked.
The former allows conversion of get_user_pages invocations that will have
to pass a "&locked" parameter to know if the mmap_sem was dropped during
the call. Example from:
down_read(&mm->mmap_sem);
do_something()
get_user_pages(tsk, mm, ..., pages, NULL);
up_read(&mm->mmap_sem);
to:
int locked = 1;
down_read(&mm->mmap_sem);
do_something()
get_user_pages_locked(tsk, mm, ..., pages, &locked);
if (locked)
up_read(&mm->mmap_sem);
The latter is suitable only as a drop in replacement of the form:
down_read(&mm->mmap_sem);
get_user_pages(tsk, mm, ..., pages, NULL);
up_read(&mm->mmap_sem);
into:
get_user_pages_unlocked(tsk, mm, ..., pages);
Where tsk, mm, the intermediate "..." paramters and "pages" can be any
value as before. Just the last parameter of get_user_pages (vmas) must be
NULL for get_user_pages_locked|unlocked to be usable (the latter original
form wouldn't have been safe anyway if vmas wasn't null, for the former we
just make it explicit by dropping the parameter).
If vmas is not NULL these two methods cannot be used.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: Andres Lagar-Cavilla <andreslc@google.com>
Reviewed-by: Peter Feiner <pfeiner@google.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
The previous commit ("mm/thp: Allocate transparent hugepages on local
node") introduced alloc_hugepage_vma() to mm/mempolicy.c to perform a
special policy for THP allocations. The function has the same interface
as alloc_pages_vma(), shares a lot of boilerplate code and a long
comment.
This patch merges the hugepage special case into alloc_pages_vma. The
extra if condition should be cheap enough price to pay. We also prevent
a (however unlikely) race with parallel mems_allowed update, which could
make hugepage allocation restart only within the fallback call to
alloc_hugepage_vma() and not reconsider the special rule in
alloc_hugepage_vma().
Also by making sure mpol_cond_put(pol) is always called before actual
allocation attempt, we can use a single exit path within the function.
Also update the comment for missing node parameter and obsolete reference
to mm_sem.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
This make sure that we try to allocate hugepages from local node if
allowed by mempolicy. If we can't, we fallback to small page allocation
based on mempolicy. This is based on the observation that allocating
pages on local node is more beneficial than allocating hugepages on remote
node.
With this patch applied we may find transparent huge page allocation
failures if the current node doesn't have enough freee hugepages. Before
this patch such failures result in us retrying the allocation on other
nodes in the numa node mask.
[akpm@linux-foundation.org: fix comment, add CONFIG_TRANSPARENT_HUGEPAGE dependency]
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Compaction deferring logic is heavy hammer that block the way to the
compaction. It doesn't consider overall system state, so it could prevent
user from doing compaction falsely. In other words, even if system has
enough range of memory to compact, compaction would be skipped due to
compaction deferring logic. This patch add new tracepoint to understand
work of deferring logic. This will also help to check compaction success
and fail.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
It is not well analyzed that when/why compaction start/finish or not.
With these new tracepoints, we can know much more about start/finish
reason of compaction. I can find following bug with these tracepoint.
http://www.spinics.net/lists/linux-mm/msg81582.html
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
It'd be useful to know current range where compaction work for detailed
analysis. With it, we can know pageblock where we actually scan and
isolate, and, how much pages we try in that pageblock and can guess why it
doesn't become freepage with pageblock order roughly.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
We now have tracepoint for begin event of compaction and it prints start
position of both scanners, but, tracepoint for end event of compaction
doesn't print finish position of both scanners. It'd be also useful to
know finish position of both scanners so this patch add it. It will help
to find odd behavior or problem on compaction internal logic.
And mode is added to both begin/end tracepoint output, since according to
mode, compaction behavior is quite different.
And lastly, status format is changed to string rather than status number
for readability.
[akpm@linux-foundation.org: fix sparse warning]
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Helper account_page_redirty() fixes dirty pages counter for redirtied
pages. This patch puts it after dirtying and prevents temporary
underflows of dirtied pages counters on zone/bdi and current->nr_dirtied.
Signed-off-by: Konstantin Khebnikov <khlebnikov@yandex-team.ru>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
The problem is that we check nr_ptes/nr_pmds in exit_mmap() which happens
*before* pgd_free(). And if an arch does pte/pmd allocation in
pgd_alloc() and frees them in pgd_free() we see offset in counters by the
time of the checks.
We tried to workaround this by offsetting expected counter value according
to FIRST_USER_ADDRESS for both nr_pte and nr_pmd in exit_mmap(). But it
doesn't work in some cases:
1. ARM with LPAE enabled also has non-zero USER_PGTABLES_CEILING, but
upper addresses occupied with huge pmd entries, so the trick with
offsetting expected counter value will get really ugly: we will have
to apply it nr_pmds, but not nr_ptes.
2. Metag has non-zero FIRST_USER_ADDRESS, but doesn't do allocation
pte/pmd page tables allocation in pgd_alloc(), just setup a pgd entry
which is allocated at boot and shared accross all processes.
The proposal is to move the check to check_mm() which happens *after*
pgd_free() and do proper accounting during pgd_alloc() and pgd_free()
which would bring counters to zero if nothing leaked.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: Tyler Baker <tyler.baker@linaro.org>
Tested-by: Tyler Baker <tyler.baker@linaro.org>
Tested-by: Nishanth Menon <nm@ti.com>
Cc: Russell King <linux@arm.linux.org.uk>
Cc: James Hogan <james.hogan@imgtec.com>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Dave noticed that unprivileged process can allocate significant amount of
memory -- >500 MiB on x86_64 -- and stay unnoticed by oom-killer and
memory cgroup. The trick is to allocate a lot of PMD page tables. Linux
kernel doesn't account PMD tables to the process, only PTE.
The use-cases below use few tricks to allocate a lot of PMD page tables
while keeping VmRSS and VmPTE low. oom_score for the process will be 0.
#include <errno.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <sys/mman.h>
#include <sys/prctl.h>
#define PUD_SIZE (1UL << 30)
#define PMD_SIZE (1UL << 21)
#define NR_PUD 130000
int main(void)
{
char *addr = NULL;
unsigned long i;
prctl(PR_SET_THP_DISABLE);
for (i = 0; i < NR_PUD ; i++) {
addr = mmap(addr + PUD_SIZE, PUD_SIZE, PROT_WRITE|PROT_READ,
MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
if (addr == MAP_FAILED) {
perror("mmap");
break;
}
*addr = 'x';
munmap(addr, PMD_SIZE);
mmap(addr, PMD_SIZE, PROT_WRITE|PROT_READ,
MAP_ANONYMOUS|MAP_PRIVATE|MAP_FIXED, -1, 0);
if (addr == MAP_FAILED)
perror("re-mmap"), exit(1);
}
printf("PID %d consumed %lu KiB in PMD page tables\n",
getpid(), i * 4096 >> 10);
return pause();
}
The patch addresses the issue by account PMD tables to the process the
same way we account PTE.
The main place where PMD tables is accounted is __pmd_alloc() and
free_pmd_range(). But there're few corner cases:
- HugeTLB can share PMD page tables. The patch handles by accounting
the table to all processes who share it.
- x86 PAE pre-allocates few PMD tables on fork.
- Architectures with FIRST_USER_ADDRESS > 0. We need to adjust sanity
check on exit(2).
Accounting only happens on configuration where PMD page table's level is
present (PMD is not folded). As with nr_ptes we use per-mm counter. The
counter value is used to calculate baseline for badness score by
oom-killer.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Reviewed-by: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Pavel Emelyanov <xemul@openvz.org>
Cc: David Rientjes <rientjes@google.com>
Tested-by: Sedat Dilek <sedat.dilek@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
The swap controller code is scattered all over the file. Gather all
the code that isn't directly needed by the memory controller at the
end of the file in its own CONFIG_MEMCG_SWAP section.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
The initialization code for the per-cpu charge stock and the soft
limit tree is compact enough to inline it into mem_cgroup_init().
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
- No need to test the node for N_MEMORY. node_online() is enough for
node fallback to work in slab, use NUMA_NO_NODE for everything else.
- Remove the BUG_ON() for allocation failure. A NULL pointer crash is
just as descriptive, and the absent return value check is obvious.
- Move local variables to the inner-most blocks.
- Point to the tree structure after its initialized, not before, it's
just more logical that way.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
The totalcma_pages variable is not updated to account for CMA regions
defined via device tree reserved-memory sub-nodes. Fix this omission by
moving the calculation of totalcma_pages into cma_init_reserved_mem()
instead of cma_declare_contiguous() such that it will include reserved
memory used by all CMA regions.
Signed-off-by: George G. Davis <george_davis@mentor.com>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Commit 5695be142e20 ("OOM, PM: OOM killed task shouldn't escape PM
suspend") has left a race window when OOM killer manages to
note_oom_kill after freeze_processes checks the counter. The race
window is quite small and really unlikely and partial solution deemed
sufficient at the time of submission.
Tejun wasn't happy about this partial solution though and insisted on a
full solution. That requires the full OOM and freezer's task freezing
exclusion, though. This is done by this patch which introduces oom_sem
RW lock and turns oom_killer_disable() into a full OOM barrier.
oom_killer_disabled check is moved from the allocation path to the OOM
level and we take oom_sem for reading for both the check and the whole
OOM invocation.
oom_killer_disable() takes oom_sem for writing so it waits for all
currently running OOM killer invocations. Then it disable all the further
OOMs by setting oom_killer_disabled and checks for any oom victims.
Victims are counted via mark_tsk_oom_victim resp. unmark_oom_victim. The
last victim wakes up all waiters enqueued by oom_killer_disable().
Therefore this function acts as the full OOM barrier.
The page fault path is covered now as well although it was assumed to be
safe before. As per Tejun, "We used to have freezing points deep in file
system code which may be reacheable from page fault." so it would be
better and more robust to not rely on freezing points here. Same applies
to the memcg OOM killer.
out_of_memory tells the caller whether the OOM was allowed to trigger and
the callers are supposed to handle the situation. The page allocation
path simply fails the allocation same as before. The page fault path will
retry the fault (more on that later) and Sysrq OOM trigger will simply
complain to the log.
Normally there wouldn't be any unfrozen user tasks after
try_to_freeze_tasks so the function will not block. But if there was an
OOM killer racing with try_to_freeze_tasks and the OOM victim didn't
finish yet then we have to wait for it. This should complete in a finite
time, though, because
- the victim cannot loop in the page fault handler (it would die
on the way out from the exception)
- it cannot loop in the page allocator because all the further
allocation would fail and __GFP_NOFAIL allocations are not
acceptable at this stage
- it shouldn't be blocked on any locks held by frozen tasks
(try_to_freeze expects lockless context) and kernel threads and
work queues are not frozen yet
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Suggested-by: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Cong Wang <xiyou.wangcong@gmail.com>
Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
oom_kill_process only sets TIF_MEMDIE flag and sends a signal to the
victim. This is basically noop when the task is frozen though because the
task sleeps in the uninterruptible sleep. The victim is eventually thawed
later when oom_scan_process_thread meets the task again in a later OOM
invocation so the OOM killer doesn't live lock. But this is less than
optimal.
Let's add __thaw_task into mark_tsk_oom_victim after we set TIF_MEMDIE to
the victim. We are not checking whether the task is frozen because that
would be racy and __thaw_task does that already. oom_scan_process_thread
doesn't need to care about freezer anymore as TIF_MEMDIE and freezer are
excluded completely now.
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Cong Wang <xiyou.wangcong@gmail.com>
Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
This patchset addresses a race which was described in the changelog for
5695be142e20 ("OOM, PM: OOM killed task shouldn't escape PM suspend"):
: PM freezer relies on having all tasks frozen by the time devices are
: getting frozen so that no task will touch them while they are getting
: frozen. But OOM killer is allowed to kill an already frozen task in order
: to handle OOM situtation. In order to protect from late wake ups OOM
: killer is disabled after all tasks are frozen. This, however, still keeps
: a window open when a killed task didn't manage to die by the time
: freeze_processes finishes.
The original patch hasn't closed the race window completely because that
would require a more complex solution as it can be seen by this patchset.
The primary motivation was to close the race condition between OOM killer
and PM freezer _completely_. As Tejun pointed out, even though the race
condition is unlikely the harder it would be to debug weird bugs deep in
the PM freezer when the debugging options are reduced considerably. I can
only speculate what might happen when a task is still runnable
unexpectedly.
On a plus side and as a side effect the oom enable/disable has a better
(full barrier) semantic without polluting hot paths.
I have tested the series in KVM with 100M RAM:
- many small tasks (20M anon mmap) which are triggering OOM continually
- s2ram which resumes automatically is triggered in a loop
echo processors > /sys/power/pm_test
while true
do
echo mem > /sys/power/state
sleep 1s
done
- simple module which allocates and frees 20M in 8K chunks. If it sees
freezing(current) then it tries another round of allocation before calling
try_to_freeze
- debugging messages of PM stages and OOM killer enable/disable/fail added
and unmark_oom_victim is delayed by 1s after it clears TIF_MEMDIE and before
it wakes up waiters.
- rebased on top of the current mmotm which means some necessary updates
in mm/oom_kill.c. mark_tsk_oom_victim is now called under task_lock but
I think this should be OK because __thaw_task shouldn't interfere with any
locking down wake_up_process. Oleg?
As expected there are no OOM killed tasks after oom is disabled and
allocations requested by the kernel thread are failing after all the tasks
are frozen and OOM disabled. I wasn't able to catch a race where
oom_killer_disable would really have to wait but I kinda expected the race
is really unlikely.
[ 242.609330] Killed process 2992 (mem_eater) total-vm:24412kB, anon-rss:2164kB, file-rss:4kB
[ 243.628071] Unmarking 2992 OOM victim. oom_victims: 1
[ 243.636072] (elapsed 2.837 seconds) done.
[ 243.641985] Trying to disable OOM killer
[ 243.643032] Waiting for concurent OOM victims
[ 243.644342] OOM killer disabled
[ 243.645447] Freezing remaining freezable tasks ... (elapsed 0.005 seconds) done.
[ 243.652983] Suspending console(s) (use no_console_suspend to debug)
[ 243.903299] kmem_eater: page allocation failure: order:1, mode:0x204010
[...]
[ 243.992600] PM: suspend of devices complete after 336.667 msecs
[ 243.993264] PM: late suspend of devices complete after 0.660 msecs
[ 243.994713] PM: noirq suspend of devices complete after 1.446 msecs
[ 243.994717] ACPI: Preparing to enter system sleep state S3
[ 243.994795] PM: Saving platform NVS memory
[ 243.994796] Disabling non-boot CPUs ...
The first 2 patches are simple cleanups for OOM. They should go in
regardless the rest IMO.
Patches 3 and 4 are trivial printk -> pr_info conversion and they should
go in ditto.
The main patch is the last one and I would appreciate acks from Tejun and
Rafael. I think the OOM part should be OK (except for __thaw_task vs.
task_lock where a look from Oleg would appreciated) but I am not so sure I
haven't screwed anything in the freezer code. I have found several
surprises there.
This patch (of 5):
This patch is just a preparatory and it doesn't introduce any functional
change.
Note:
I am utterly unhappy about lowmemory killer abusing TIF_MEMDIE just to
wait for the oom victim and to prevent from new killing. This is
just a side effect of the flag. The primary meaning is to give the oom
victim access to the memory reserves and that shouldn't be necessary
here.
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Cong Wang <xiyou.wangcong@gmail.com>
Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Turn the move type enum into flags and give the flags field a shorter
name. Once that is done, move_anon() and move_file() are simple enough to
just fold them into the callsites.
[akpm@linux-foundation.org: tweak MOVE_MASK definition, per Michal]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
The unified hierarchy interface for memory cgroups will no longer use "-1"
to mean maximum possible resource value. In preparation for this, make
the string an argument and let the caller supply it.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Especially on 32 bit kernels memory node ranges are printed with 32 bit
wide addresses only. Use u64 types and %llx specifiers to print full
width of addresses.
Signed-off-by: Juergen Gross <jgross@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Use BUILD_BUG_ON() to compile assert that memcg string tables are in sync
with corresponding enums. There aren't currently any issues with these
tables. This is just defensive.
Signed-off-by: Greg Thelen <gthelen@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Since commit b2052564e66d ("mm: memcontrol: continue cache reclaim from
offlined groups") pages charged to a memory cgroup are not reparented when
the cgroup is removed. Instead, they are supposed to be reclaimed in a
regular way, along with pages accounted to online memory cgroups.
However, an lruvec of an offline memory cgroup will sooner or later get so
small that it will be scanned only at low scan priorities (see
get_scan_count()). Therefore, if there are enough reclaimable pages in
big lruvecs, pages accounted to offline memory cgroups will never be
scanned at all, wasting memory.
Fix this by unconditionally forcing scanning dead lruvecs from kswapd.
[akpm@linux-foundation.org: fix build]
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: 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>
|
|
Although it was not called, destroy_compound_page() did some potentially
useful checks. Let's re-introduce them in free_pages_prepare(), where
they can be actually triggered when CONFIG_DEBUG_VM=y.
compound_order() assert is already in free_pages_prepare(). We have few
checks for tail pages left.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
