Memory pressure can put dirty pages at the end of the LRU without
anybody running into dirty limits. Don't start writing individual pages
from kswapd while the flushers might be asleep.
Unlike the old direct reclaim flusher wakeup (removed in the next patch)
that flushes the number of pages just scanned, this patch wakes the
flushers for all outstanding dirty pages. That seemed to perform better
in a synthetic test that pushes dirty pages to the end of the LRU and
into reclaim, because we know LRU aging outstrips writeback already, and
this way we give younger dirty pages a headstart rather than wait until
reclaim runs into them as well. It also means less plugging and risk of
exhausting the struct request pool from reclaim.
There is a concern that this will cause temporary files that used to get
dirtied and truncated before writeback to now get written to disk under
memory pressure. If this turns out to be a real problem, we'll have to
revisit this and tame the reclaim flusher wakeups.
[hannes@cmpxchg.org: mention dirty expiration as a condition]
Link: http://lkml.kernel.org/r/20170126174739.GA30636@cmpxchg.org
Link: http://lkml.kernel.org/r/20170123181641.23938-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
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>
Patch series "mm: vmscan: fix kswapd writeback regression".
We noticed a regression on multiple hadoop workloads when moving from
3.10 to 4.0 and 4.6, which involves kswapd getting tangled up in page
writeout, causing direct reclaim herds that also don't make progress.
I tracked it down to the thrash avoidance efforts after 3.10 that make
the kernel better at keeping use-once cache and use-many cache sorted on
the inactive and active list, with more aggressive protection of the
active list as long as there is inactive cache. Unfortunately, our
workload's use-once cache is mostly from streaming writes. Waiting for
writes to avoid potential reloads in the future is not a good tradeoff.
These patches do the following:
1. Wake the flushers when kswapd sees a lump of dirty pages. It's
possible to be below the dirty background limit and still have cache
velocity push them through the LRU. So start a-flushin'.
2. Let kswapd only write pages that have been rotated twice. This makes
sure we really tried to get all the clean pages on the inactive list
before resorting to horrible LRU-order writeback.
3. Move rotating dirty pages off the inactive list. Instead of churning
or waiting on page writeback, we'll go after clean active cache. This
might lead to thrashing, but in this state memory demand outstrips IO
speed anyway, and reads are faster than writes.
Mel backported the series to 4.10-rc5 with one minor conflict and ran a
couple of tests on it. Mix of read/write random workload didn't show
anything interesting. Write-only database didn't show much difference
in performance but there were slight reductions in IO -- probably in the
noise.
simoop did show big differences although not as big as Mel expected.
This is Chris Mason's workload that similate the VM activity of hadoop.
Mel won't go through the full details but over the samples measured
during an hour it reported
4.10.0-rc5 4.10.0-rc5
vanilla johannes-v1r1
Amean p50-Read 21346531.56 ( 0.00%) 21697513.24 ( -1.64%)
Amean p95-Read 24700518.40 ( 0.00%) 25743268.98 ( -4.22%)
Amean p99-Read 27959842.13 ( 0.00%) 28963271.11 ( -3.59%)
Amean p50-Write 1138.04 ( 0.00%) 989.82 ( 13.02%)
Amean p95-Write 1106643.48 ( 0.00%) 12104.00 ( 98.91%)
Amean p99-Write 1569213.22 ( 0.00%) 36343.38 ( 97.68%)
Amean p50-Allocation 85159.82 ( 0.00%) 79120.70 ( 7.09%)
Amean p95-Allocation 204222.58 ( 0.00%) 129018.43 ( 36.82%)
Amean p99-Allocation 278070.04 ( 0.00%) 183354.43 ( 34.06%)
Amean final-p50-Read 21266432.00 ( 0.00%) 21921792.00 ( -3.08%)
Amean final-p95-Read 24870912.00 ( 0.00%) 26116096.00 ( -5.01%)
Amean final-p99-Read 28147712.00 ( 0.00%) 29523968.00 ( -4.89%)
Amean final-p50-Write 1130.00 ( 0.00%) 977.00 ( 13.54%)
Amean final-p95-Write 1033216.00 ( 0.00%) 2980.00 ( 99.71%)
Amean final-p99-Write 1517568.00 ( 0.00%) 32672.00 ( 97.85%)
Amean final-p50-Allocation 86656.00 ( 0.00%) 78464.00 ( 9.45%)
Amean final-p95-Allocation 211712.00 ( 0.00%) 116608.00 ( 44.92%)
Amean final-p99-Allocation 287232.00 ( 0.00%) 168704.00 ( 41.27%)
The latencies are actually completely horrific in comparison to 4.4 (and
4.10-rc5 is worse than 4.9 according to historical data for reasons Mel
hasn't analysed yet).
Still, 95% of write latency (p95-write) is halved by the series and
allocation latency is way down. Direct reclaim activity is one fifth of
what it was according to vmstats. Kswapd activity is higher but this is
not necessarily surprising. Kswapd efficiency is unchanged at 99% (99%
of pages scanned were reclaimed) but direct reclaim efficiency went from
77% to 99%
In the vanilla kernel, 627MB of data was written back from reclaim
context. With the series, no data was written back. With or without
the patch, pages are being immediately reclaimed after writeback
completes. However, with the patch, only 1/8th of the pages are
reclaimed like this.
This patch (of 5):
We have an elaborate dirty/writeback throttling mechanism inside the
reclaim scanner, but for that to work the pages have to go through
shrink_page_list() and get counted for what they are. Otherwise, we
mess up the LRU order and don't match reclaim speed to writeback.
Especially during deactivation, there is never a reason to skip dirty
pages; nothing is even trying to write them out from there. Don't mess
up the LRU order for nothing, shuffle these pages along.
Link: http://lkml.kernel.org/r/20170123181641.23938-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
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>
Patch series "userfaultfd: non-cooperative: add madvise() event for
MADV_REMOVE request".
These patches add notification of madvise(MADV_REMOVE) event to
non-cooperative userfaultfd monitor.
The first pacth renames EVENT_MADVDONTNEED to EVENT_REMOVE along with
relevant functions and structures. Using _REMOVE instead of
_MADVDONTNEED describes the event semantics more clearly and I hope it's
not too late for such change in the ABI.
This patch (of 3):
The UFFD_EVENT_MADVDONTNEED purpose is to notify uffd monitor about
removal of certain range from address space tracked by userfaultfd.
Hence, UFFD_EVENT_REMOVE seems to better reflect the operation
semantics. Respectively, 'madv_dn' field of uffd_msg is renamed to
'remove' and the madvise_userfault_dontneed callback is renamed to
userfaultfd_remove.
Link: http://lkml.kernel.org/r/1484814154-1557-2-git-send-email-rppt@linux.vnet.ibm.com
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Some architectures have a set of zero pages (coloured zero pages)
instead of only one zero page, in order to improve the cache
performance. In those cases, the kernel samepage merger (KSM) would
merge all the allocated pages that happen to be filled with zeroes to
the same deduplicated page, thus losing all the advantages of coloured
zero pages.
This behaviour is noticeable when a process accesses large arrays of
allocated pages containing zeroes. A test I conducted on s390 shows
that there is a speed penalty when KSM merges such pages, compared to
not merging them or using actual zero pages from the start without
breaking the COW.
This patch fixes this behaviour. When coloured zero pages are present,
the checksum of a zero page is calculated during initialisation, and
compared with the checksum of the current canditate during merging. In
case of a match, the normal merging routine is used to merge the page
with the correct coloured zero page, which ensures the candidate page is
checked to be equal to the target zero page.
A sysfs entry is also added to toggle this behaviour, since it can
potentially introduce performance regressions, especially on
architectures without coloured zero pages. The default value is
disabled, for backwards compatibility.
With this patch, the performance with KSM is the same as with non
COW-broken actual zero pages, which is also the same as without KSM.
[akpm@linux-foundation.org: make zero_checksum and ksm_use_zero_pages __read_mostly, per Andrea]
[imbrenda@linux.vnet.ibm.com: documentation for coloured zero pages deduplication]
Link: http://lkml.kernel.org/r/1484927522-1964-1-git-send-email-imbrenda@linux.vnet.ibm.com
Link: http://lkml.kernel.org/r/1484850953-23941-1-git-send-email-imbrenda@linux.vnet.ibm.com
Signed-off-by: Claudio Imbrenda <imbrenda@linux.vnet.ibm.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
Cc: Hugh Dickins <hughd@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>
The patch "mm, page_alloc: warn_alloc print nodemask" implicitly sets
the allocation nodemask to cpuset_current_mems_allowed when there is no
effective mempolicy. cpuset_current_mems_allowed is only effective when
cpusets are enabled, which is also printed by warn_alloc(), so setting
the nodemask to cpuset_current_mems_allowed is redundant and prevents
debugging issues where ac->nodemask is not set properly in the page
allocator.
This provides better debugging output since
cpuset_print_current_mems_allowed() is already provided.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701181347320.142399@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Now that __GFP_NOFAIL doesn't override decisions to skip the oom killer
we are left with requests which require to loop inside the allocator
without invoking the oom killer (e.g. GFP_NOFS|__GFP_NOFAIL used by fs
code) and so they might, in very unlikely situations, loop for ever -
e.g. other parallel request could starve them.
This patch tries to limit the likelihood of such a lockup by giving
these __GFP_NOFAIL requests a chance to move on by consuming a small
part of memory reserves. We are using ALLOC_HARDER which should be
enough to prevent from the starvation by regular allocation requests,
yet it shouldn't consume enough from the reserves to disrupt high
priority requests (ALLOC_HIGH).
While we are at it, let's introduce a helper __alloc_pages_cpuset_fallback
which enforces the cpusets but allows to fallback to ignore them if the
first attempt fails. __GFP_NOFAIL requests can be considered important
enough to allow cpuset runaway in order for the system to move on. It
is highly unlikely that any of these will be GFP_USER anyway.
Link: http://lkml.kernel.org/r/20161220134904.21023-4-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
__alloc_pages_may_oom makes sure to skip the OOM killer depending on the
allocation request. This includes lowmem requests, costly high order
requests and others. For a long time __GFP_NOFAIL acted as an override
for all those rules. This is not documented and it can be quite
surprising as well. E.g. GFP_NOFS requests are not invoking the OOM
killer but GFP_NOFS|__GFP_NOFAIL does so if we try to convert some of
the existing open coded loops around allocator to nofail request (and we
have done that in the past) then such a change would have a non trivial
side effect which is far from obvious. Note that the primary motivation
for skipping the OOM killer is to prevent from pre-mature invocation.
The exception has been added by commit 82553a937f ("oom: invoke oom
killer for __GFP_NOFAIL"). The changelog points out that the oom killer
has to be invoked otherwise the request would be looping for ever. But
this argument is rather weak because the OOM killer doesn't really
guarantee a forward progress for those exceptional cases:
- it will hardly help to form costly order which in turn can result in
the system panic because of no oom killable task in the end - I believe
we certainly do not want to put the system down just because there is a
nasty driver asking for order-9 page with GFP_NOFAIL not realizing all
the consequences. It is much better this request would loop for ever
than the massive system disruption
- lowmem is also highly unlikely to be freed during OOM killer
- GFP_NOFS request could trigger while there is still a lot of memory
pinned by filesystems.
This patch simply removes the __GFP_NOFAIL special case in order to have a
more clear semantic without surprising side effects.
Signed-off-by: Michal Hocko <mhocko@suse.com>
Reported-by: Nils Holland <nholland@tisys.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Tetsuo Handa has pointed out that commit 0a0337e0d1 ("mm, oom: rework
oom detection") has subtly changed semantic for costly high order
requests with __GFP_NOFAIL and withtout __GFP_REPEAT and those can fail
right now. My code inspection didn't reveal any such users in the tree
but it is true that this might lead to unexpected allocation failures
and subsequent OOPs.
__alloc_pages_slowpath wrt. GFP_NOFAIL is hard to follow currently.
There are few special cases but we are lacking a catch all place to be
sure we will not miss any case where the non failing allocation might
fail. This patch reorganizes the code a bit and puts all those special
cases under nopage label which is the generic go-to-fail path. Non
failing allocations are retried or those that cannot retry like
non-sleeping allocation go to the failure point directly. This should
make the code flow much easier to follow and make it less error prone
for future changes.
While we are there we have to move the stall check up to catch
potentially looping non-failing allocations.
[akpm@linux-foundation.org: fix alloc_flags may-be-used-uninitalized]
Link: http://lkml.kernel.org/r/20161220134904.21023-2-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
show_mem() allows to filter out node specific data which is irrelevant
to the allocation request via SHOW_MEM_FILTER_NODES. The filtering is
done in skip_free_areas_node which skips all nodes which are not in the
mems_allowed of the current process. This works most of the time as
expected because the nodemask shouldn't be outside of the allocating
task but there are some exceptions. E.g. memory hotplug might want to
request allocations from outside of the allowed nodes (see
new_node_page).
Get rid of this hardcoded behavior and push the allocation mask down the
show_mem path and use it instead of cpuset_current_mems_allowed. NULL
nodemask is interpreted as cpuset_current_mems_allowed.
[akpm@linux-foundation.org: coding-style fixes]
Link: http://lkml.kernel.org/r/20170117091543.25850-5-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patch series "show_mem updates", v2.
This is a mixture of one bug fix (patch 1), an enhancement (patch 2) and
cleanups (the rest of the series). First two patches should be really
straightforward. Patch 3 removes some arch specific show_mem
implementations because I think they are quite outdated and do not
really serve any useful purpose anymore. I think we should really
strive to have a consistent show_mem output regardless of the
architecture. If some architecture is really special and wants to dump
something additional we should do that via an arch specific hook.
The last patch adds nodemask parameter so that we do not rely on the
hardcoded mems_allowed of the current task when doing the node
filtering. I consider this more a cleanup than a fix because basically
all users use a nodemask which is a subset of mems_allowed. There is
only one call path in the memory hotplug which doesn't comply with this
but that is hardly something to worry about.
This patch (of 4):
Commit 599d0c954f ("mm, vmscan: move LRU lists to node") has added per
numa node statistics to show_mem but it forgot to add
skip_free_areas_node to filter out nodes which are outside of the
allocating task numa policy. Add this check to not pollute the output
with the pointless information.
Link: http://lkml.kernel.org/r/20170117091543.25850-2-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Johannes Weiner
Mike Rapoport
Heiko Carstens
Andrew Morton
Dan Williams
David Rientjes
Claudio Imbrenda
zhong jiang
Miles Chen
Kirill A. Shutemov
Michal Hocko