10 years agodrbd: add major number to major.h
Philipp Reisner [Tue, 16 Jun 2009 22:33:38 +0000]
drbd: add major number to major.h

Since we have had a LANANA major number for years, and it is documented in
devices.txt, I think that this first patch can go upstream.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoheaders: move module_bug_finalize()/module_bug_cleanup() definitions into module.h
Andrew Morton [Tue, 16 Jun 2009 22:33:37 +0000]
headers: move module_bug_finalize()/module_bug_cleanup() definitions into module.h

They're in linux/bug.h at present, which causes include order tangles.  In
particular, linux/bug.h cannot be used by linux/atomic.h because,
according to Nikanth:

linux/bug.h pulls in linux/module.h => linux/spinlock.h => asm/spinlock.h
(which uses atomic_inc) => asm/atomic.h.

bug.h is a pretty low-level thing and module.h is a higher-level thing,
IMO.

Cc: Nikanth Karthikesan <knikanth@novell.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopoll: avoid extra wakeups in select/poll
Eric Dumazet [Tue, 16 Jun 2009 22:33:36 +0000]
poll: avoid extra wakeups in select/poll

After introduction of keyed wakeups Davide Libenzi did on epoll, we are
able to avoid spurious wakeups in poll()/select() code too.

For example, typical use of poll()/select() is to wait for incoming
network frames on many sockets.  But TX completion for UDP/TCP frames call
sock_wfree() which in turn schedules thread.

When scheduled, thread does a full scan of all polled fds and can sleep
again, because nothing is really available.  If number of fds is large,
this cause significant load.

This patch makes select()/poll() aware of keyed wakeups and useless
wakeups are avoided.  This reduces number of context switches by about 50%
on some setups, and work performed by sofirq handlers.

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Acked-by: David S. Miller <davem@davemloft.net>
Acked-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Davide Libenzi <davidel@xmailserver.org>
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>

10 years agontfs: use is_power_of_2() function for clarity.
Robert P. J. Day [Tue, 16 Jun 2009 22:33:35 +0000]
ntfs: use is_power_of_2() function for clarity.

Signed-off-by: Robert P. J. Day <rpjday@crashcourse.ca>
Cc: Anton Altaparmakov <aia21@cantab.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agokernel/kfifo.c: replace conditional test with is_power_of_2()
Robert P. J. Day [Tue, 16 Jun 2009 22:33:34 +0000]
kernel/kfifo.c: replace conditional test with is_power_of_2()

Signed-off-by: Robert P. J. Day <rpjday@crashcourse.ca>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoatomic: only take lock when the counter drops to zero on UP as well
Jan Blunck [Tue, 16 Jun 2009 22:33:33 +0000]
atomic: only take lock when the counter drops to zero on UP as well

_atomic_dec_and_lock() should not unconditionally take the lock before
calling atomic_dec_and_test() in the UP case.  For consistency reasons it
should behave exactly like in the SMP case.

Besides that this works around the problem that with CONFIG_DEBUG_SPINLOCK
this spins in __spin_lock_debug() if the lock is already taken even if the
counter doesn't drop to 0.

Signed-off-by: Jan Blunck <jblunck@suse.de>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Nick Piggin <npiggin@suse.de>
Cc: Valerie Aurora <vaurora@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoutsname.h: make new_utsname fields use the proper length constant
Dan Smith [Tue, 16 Jun 2009 22:33:33 +0000]
utsname.h: make new_utsname fields use the proper length constant

The members of the new_utsname structure are defined with magic numbers
that *should* correspond to the constant __NEW_UTS_LEN+1.  Everywhere
else, code assumes this and uses the constant, so this patch makes the
structure match.

Originally suggested by Serge here:

https://lists.linux-foundation.org/pipermail/containers/2009-March/016258.html

Signed-off-by: Dan Smith <danms@us.ibm.com>
Acked-by: Serge Hallyn <serue@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agouml: bad macro expansion, parameter is member
Roel Kluin [Tue, 16 Jun 2009 22:33:32 +0000]
uml: bad macro expansion, parameter is member

`ELF_CORE_COPY_REGS(x, y)' will make expansions like:
`(y)[0] = (x)->x.gp[0]' but correct is `(y)[0] = (x)->regs.gp[0]'

Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
Cc: WANG Cong <amwang@redhat.com>
Cc: Jeff Dike <jdike@addtoit.com>

Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agouml: fix a section warning
Amerigo Wang [Tue, 16 Jun 2009 22:33:30 +0000]
uml: fix a section warning

When compiling uml on x86_64:

  MODPOST vmlinux.o
WARNING: vmlinux.o (.__syscall_stub.2): unexpected non-allocatable section.
Did you forget to use "ax"/"aw" in a .S file?
Note that for example <linux/init.h> contains
section definitions for use in .S files.

Because modpost checks for missing SHF_ALLOC section flag.  So just add
it.

Signed-off-by: WANG Cong <amwang@redhat.com>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Sam Ravnborg <sam@ravnborg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoum: remove obsolete hw_interrupt_type
Thomas Gleixner [Tue, 16 Jun 2009 22:33:29 +0000]
um: remove obsolete hw_interrupt_type

The defines and typedefs (hw_interrupt_type, no_irq_type, irq_desc_t) have
been kept around for migration reasons.  After more than two years it's
time to remove them finally.

This patch cleans up one of the remaining users.  When all such patches
hit mainline we can remove the defines and typedefs finally.

Impact: cleanup

Convert the last remaining users to struct irq_chip and remove the
define.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Jeff Dike <jdike@addtoit.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agouml: UML net driver does not allow for vlans
Alan Cox [Tue, 16 Jun 2009 22:33:28 +0000]
uml: UML net driver does not allow for vlans

See ancient discussion at
http://marc.info/?l=user-mode-linux-devel&m=101990155831279&w=2

Addresses http://bugzilla.kernel.org/show_bug.cgi?id=7854

Signed-off-by: Alan Cox <alan@linux.intel.com>
Reported-by: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Roland Kletzing <devzero@web.de>
Cc: "David S. Miller" <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agom32r: remove obsolete hw_interrupt_type
Thomas Gleixner [Tue, 16 Jun 2009 22:33:26 +0000]
m32r: remove obsolete hw_interrupt_type

The defines and typedefs (hw_interrupt_type, no_irq_type, irq_desc_t) have
been kept around for migration reasons.  After more than two years it's
time to remove them finally.

This patch cleans up one of the remaining users.  When all such patches
hit mainline we can remove the defines and typedefs finally.

Impact: cleanup

Convert the last remaining users to struct irq_chip and remove the
define.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Hirokazu Takata <takata@linux-m32r.org>

Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoalpha: bad macro expansion, parameter is member
Roel Kluin [Tue, 16 Jun 2009 22:33:25 +0000]
alpha: bad macro expansion, parameter is member

`for_each_mem_cluster(x, y, z)' will expand to
`for ((x) = (y)->x ...' but correct is `for ((x) = (y)->cluster ...'

Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoalpha: remove obsolete hw_interrupt_type
Thomas Gleixner [Tue, 16 Jun 2009 22:33:25 +0000]
alpha: remove obsolete hw_interrupt_type

The defines and typedefs (hw_interrupt_type, no_irq_type, irq_desc_t) have
been kept around for migration reasons.  After more than two years it's
time to remove them finally.

This patch cleans up one of the remaining users.  When all such patches
hit mainline we can remove the defines and typedefs finally.

Impact: cleanup

Convert the last remaining users to struct irq_chip and remove the
define.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Richard Henderson <rth@twiddle.net>

Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: fix lumpy reclaim lru handling at isolate_lru_pages
KAMEZAWA Hiroyuki [Tue, 16 Jun 2009 22:33:24 +0000]
mm: fix lumpy reclaim lru handling at isolate_lru_pages

At lumpy reclaim, a page failed to be taken by __isolate_lru_page() can be
pushed back to "src" list by list_move().  But the page may not be from
"src" list.  This pushes the page back to wrong LRU.  And list_move()
itself is unnecessary because the page is not on top of LRU.  Then, leave
it as it is if __isolate_lru_page() fails.

Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agovmscan: count the number of times zone_reclaim() scans and fails
Mel Gorman [Tue, 16 Jun 2009 22:33:23 +0000]
vmscan: count the number of times zone_reclaim() scans and fails

On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim.  On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.

There is a heuristic that determines if the scan is worthwhile but it is
possible that the heuristic will fail and the CPU gets tied up scanning
uselessly.  Detecting the situation requires some guesswork and
experimentation so this patch adds a counter "zreclaim_failed" to
/proc/vmstat.  If during high CPU utilisation this counter is increasing
rapidly, then the resolution to the problem may be to set
/proc/sys/vm/zone_reclaim_mode to 0.

[akpm@linux-foundation.org: name things consistently]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agovmscan: do not unconditionally treat zones that fail zone_reclaim() as full
Mel Gorman [Tue, 16 Jun 2009 22:33:22 +0000]
vmscan: do not unconditionally treat zones that fail zone_reclaim() as full

On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim.  On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.  The problem is that zone_reclaim() failing at all means the
zone gets marked full.

This can cause situations where a zone is usable, but is being skipped
because it has been considered full.  Take a situation where a large tmpfs
mount is occuping a large percentage of memory overall.  The pages do not
get cleaned or reclaimed by zone_reclaim(), but the zone gets marked full
and the zonelist cache considers them not worth trying in the future.

This patch makes zone_reclaim() return more fine-grained information about
what occured when zone_reclaim() failued.  The zone only gets marked full
if it really is unreclaimable.  If it's a case that the scan did not occur
or if enough pages were not reclaimed with the limited reclaim_mode, then
the zone is simply skipped.

There is a side-effect to this patch.  Currently, if zone_reclaim()
successfully reclaimed SWAP_CLUSTER_MAX, an allocation attempt would go
ahead.  With this patch applied, zone watermarks are rechecked after
zone_reclaim() does some work.

This bug was introduced by commit 9276b1bc96a132f4068fdee00983c532f43d3a26
("memory page_alloc zonelist caching speedup") way back in 2.6.19 when the
zonelist_cache was introduced.  It was not intended that zone_reclaim()
aggressively consider the zone to be full when it failed as full direct
reclaim can still be an option.  Due to the age of the bug, it should be
considered a -stable candidate.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Wu Fengguang <fengguang.wu@intel.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agovmscan: properly account for the number of page cache pages zone_reclaim() can reclaim
Mel Gorman [Tue, 16 Jun 2009 22:33:20 +0000]
vmscan: properly account for the number of page cache pages zone_reclaim() can reclaim

A bug was brought to my attention against a distro kernel but it affects
mainline and I believe problems like this have been reported in various
guises on the mailing lists although I don't have specific examples at the
moment.

The reported problem was that malloc() stalled for a long time (minutes in
some cases) if a large tmpfs mount was occupying a large percentage of
memory overall.  The pages did not get cleaned or reclaimed by
zone_reclaim() because the zone_reclaim_mode was unsuitable, but the lists
are uselessly scanned frequencly making the CPU spin at near 100%.

This patchset intends to address that bug and bring the behaviour of
zone_reclaim() more in line with expectations which were noticed during
investigation.  It is based on top of mmotm and takes advantage of
Kosaki's work with respect to zone_reclaim().

Patch 1 fixes the heuristics that zone_reclaim() uses to determine if the
scan should go ahead. The broken heuristic is what was causing the
malloc() stall as it uselessly scanned the LRU constantly. Currently,
zone_reclaim is assuming zone_reclaim_mode is 1 and historically it
could not deal with tmpfs pages at all. This fixes up the heuristic so
that an unnecessary scan is more likely to be correctly avoided.

Patch 2 notes that zone_reclaim() returning a failure automatically means
the zone is marked full. This is not always true. It could have
failed because the GFP mask or zone_reclaim_mode were unsuitable.

Patch 3 introduces a counter zreclaim_failed that will increment each
time the zone_reclaim scan-avoidance heuristics fail. If that
counter is rapidly increasing, then zone_reclaim_mode should be
set to 0 as a temporarily resolution and a bug reported because
the scan-avoidance heuristic is still broken.

This patch:

On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim.  On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.

There is a heuristic that determines if the scan is worthwhile but the
problem is that the heuristic is not being properly applied and is
basically assuming zone_reclaim_mode is 1 if it is enabled.  The lack of
proper detection can manfiest as high CPU usage as the LRU list is scanned
uselessly.

Historically, once enabled it was depending on NR_FILE_PAGES which may
include swapcache pages that the reclaim_mode cannot deal with.  Patch
vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch by
Kosaki Motohiro noted that zone_page_state(zone, NR_FILE_PAGES) included
pages that were not file-backed such as swapcache and made a calculation
based on the inactive, active and mapped files.  This is far superior when
zone_reclaim==1 but if RECLAIM_SWAP is set, then NR_FILE_PAGES is a
reasonable starting figure.

This patch alters how zone_reclaim() works out how many pages it might be
able to reclaim given the current reclaim_mode.  If RECLAIM_SWAP is set in
the reclaim_mode it will either consider NR_FILE_PAGES as potential
candidates or else use NR_{IN}ACTIVE}_PAGES-NR_FILE_MAPPED to discount
swapcache and other non-file-backed pages.  If RECLAIM_WRITE is not set,
then NR_FILE_DIRTY number of pages are not candidates.  If RECLAIM_SWAP is
not set, then NR_FILE_MAPPED are not.

[kosaki.motohiro@jp.fujitsu.com: Estimate unmapped pages minus tmpfs pages]
[fengguang.wu@intel.com: Fix underflow problem in Kosaki's estimate]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agowriteback: skip new or to-be-freed inodes
Wu Fengguang [Tue, 16 Jun 2009 22:33:17 +0000]
writeback: skip new or to-be-freed inodes

1) I_FREEING tests should be coupled with I_CLEAR

The two I_FREEING tests are racy because clear_inode() can set i_state to
I_CLEAR between the clear of I_SYNC and the test of I_FREEING.

2) skip I_WILL_FREE inodes in generic_sync_sb_inodes() to avoid possible
   races with generic_forget_inode()

generic_forget_inode() sets I_WILL_FREE call writeback on its own, so
generic_sync_sb_inodes() shall not try to step in and create possible races:

  generic_forget_inode
    inode->i_state |= I_WILL_FREE;
    spin_unlock(&inode_lock);
                                       generic_sync_sb_inodes()
                                         spin_lock(&inode_lock);
                                         __iget(inode);
                                         __writeback_single_inode
                                           // see non zero i_count
 may WARN here ==>                         WARN_ON(inode->i_state & I_WILL_FREE);
                                         spin_unlock(&inode_lock);
 may call generic_forget_inode again ==> iput(inode);

The above race and warning didn't turn up because writeback_inodes() holds
the s_umount lock, so generic_forget_inode() finds MS_ACTIVE and returns
early.  But we are not sure the UBIFS calls and future callers will
guarantee that.  So skip I_WILL_FREE inodes for the sake of safety.

Cc: Eric Sandeen <sandeen@sandeen.net>
Acked-by: Jeff Layton <jlayton@redhat.com>
Cc: Masayoshi MIZUMA <m.mizuma@jp.fujitsu.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: Artem Bityutskiy <dedekind1@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Acked-by: Jan Kara <jack@suse.cz>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agooom: only oom kill exiting tasks with attached memory
David Rientjes [Tue, 16 Jun 2009 22:33:16 +0000]
oom: only oom kill exiting tasks with attached memory

When a task is chosen for oom kill and is found to be PF_EXITING,
__oom_kill_task() is called to elevate the task's timeslice and give it
access to memory reserves so that it may quickly exit.

This privilege is unnecessary, however, if the task has already detached
its mm.  Although its possible for the mm to become detached later since
task_lock() is not held, __oom_kill_task() will simply be a no-op in such
circumstances.

Subsequently, it is no longer necessary to warn about killing mm-less
tasks since it is a no-op.

Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agovmscan: handle may_swap more strictly
Daisuke Nishimura [Tue, 16 Jun 2009 22:33:15 +0000]
vmscan: handle may_swap more strictly

Commit 2e2e425989080cc534fc0fca154cae515f971cf5 ("vmscan,memcg:
reintroduce sc->may_swap) add may_swap flag and handle it at
get_scan_ratio().

But the result of get_scan_ratio() is ignored when priority == 0, so anon
lru is scanned even if may_swap == 0 or nr_swap_pages == 0.  IMHO, this is
not an expected behavior.

As for memcg especially, because of this behavior many and many pages are
swapped-out just in vain when oom is invoked by mem+swap limit.

This patch is for handling may_swap flag more strictly.

Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Balbir Singh <balbir@linux.vnet.ibm.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agovmscan: merge duplicate code in shrink_active_list()
Wu Fengguang [Tue, 16 Jun 2009 22:33:13 +0000]
vmscan: merge duplicate code in shrink_active_list()

The "move pages to active list" and "move pages to inactive list" code
blocks are mostly identical and can be served by a function.

Thanks to Andrew Morton for pointing this out.

Note that buffer_heads_over_limit check will also be carried out for
re-activated pages, which is slightly different from pre-2.6.28 kernels.
Also, Rik's "vmscan: evict use-once pages first" patch could totally stop
scans of active file list when memory pressure is low.  So the net effect
could be, the number of buffer heads is now more likely to grow large.

However that's fine according to Johannes' comments:

  I don't think that this could be harmful.  We just preserve the buffer
  mappings of what we consider the working set and with low memory
  pressure, as you say, this set is not big.

  As to stripping of reactivated pages: the only pages we re-activate
  for now are those VM_EXEC mapped ones.  Since we don't expect IO from
  or to these pages, removing the buffer mappings in case they grow too
  large should be okay, I guess.

Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agovmscan: make mapped executable pages the first class citizen
Wu Fengguang [Tue, 16 Jun 2009 22:33:12 +0000]
vmscan: make mapped executable pages the first class citizen

Protect referenced PROT_EXEC mapped pages from being deactivated.

PROT_EXEC(or its internal presentation VM_EXEC) pages normally belong to some
currently running executables and their linked libraries, they shall really be
cached aggressively to provide good user experiences.

Thanks to Johannes Weiner for the advice to reuse the VMA walk in
page_referenced() to get the PROT_EXEC bit.

[more details]

( The consequences of this patch will have to be discussed together with
  Rik van Riel's recent patch "vmscan: evict use-once pages first". )

( Some of the good points and insights are taken into this changelog.
  Thanks to all the involved people for the great LKML discussions. )

the problem
===========

For a typical desktop, the most precious working set is composed of
*actively accessed*
(1) memory mapped executables
(2) and their anonymous pages
(3) and other files
(4) and the dcache/icache/.. slabs
while the least important data are
(5) infrequently used or use-once files

For a typical desktop, one major problem is busty and large amount of (5)
use-once files flushing out the working set.

Inside the working set, (4) dcache/icache have already been too sticky ;-)
So we only have to care (2) anonymous and (1)(3) file pages.

anonymous pages
===============

Anonymous pages are effectively immune to the streaming IO attack, because we
now have separate file/anon LRU lists. When the use-once files crowd into the
file LRU, the list's "quality" is significantly lowered. Therefore the scan
balance policy in get_scan_ratio() will choose to scan the (low quality) file
LRU much more frequently than the anon LRU.

file pages
==========

Rik proposed to *not* scan the active file LRU when the inactive list grows
larger than active list. This guarantees that when there are use-once streaming
IO, and the working set is not too large(so that active_size < inactive_size),
the active file LRU will *not* be scanned at all. So the not-too-large working
set can be well protected.

But there are also situations where the file working set is a bit large so that
(active_size >= inactive_size), or the streaming IOs are not purely use-once.
In these cases, the active list will be scanned slowly. Because the current
shrink_active_list() policy is to deactivate active pages regardless of their
referenced bits. The deactivated pages become susceptible to the streaming IO
attack: the inactive list could be scanned fast (500MB / 50MBps = 10s) so that
the deactivated pages don't have enough time to get re-referenced. Because a
user tend to switch between windows in intervals from seconds to minutes.

This patch holds mapped executable pages in the active list as long as they
are referenced during each full scan of the active list.  Because the active
list is normally scanned much slower, they get longer grace time (eg. 100s)
for further references, which better matches the pace of user operations.

Therefore this patch greatly prolongs the in-cache time of executable code,
when there are moderate memory pressures.

before patch: guaranteed to be cached if reference intervals < I
after  patch: guaranteed to be cached if reference intervals < I+A
      (except when randomly reclaimed by the lumpy reclaim)
where
A = time to fully scan the   active file LRU
I = time to fully scan the inactive file LRU

Note that normally A >> I.

side effects
============

This patch is safe in general, it restores the pre-2.6.28 mmap() behavior
but in a much smaller and well targeted scope.

One may worry about some one to abuse the PROT_EXEC heuristic.  But as
Andrew Morton stated, there are other tricks to getting that sort of boost.

Another concern is the PROT_EXEC mapped pages growing large in rare cases,
and therefore hurting reclaim efficiency. But a sane application targeted for
large audience will never use PROT_EXEC for data mappings. If some home made
application tries to abuse that bit, it shall be aware of the consequences.
If it is abused to scale of 2/3 total memory, it gains nothing but overheads.

benchmarks
==========

1) memory tight desktop

1.1) brief summary

- clock time and major faults are reduced by 50%;
- pswpin numbers are reduced to ~1/3.

That means X desktop responsiveness is doubled under high memory/swap pressure.

1.2) test scenario

- nfsroot gnome desktop with 512M physical memory
- run some programs, and switch between the existing windows
  after starting each new program.

1.3) progress timing (seconds)

  before       after    programs
    0.02        0.02    N xeyes
    0.75        0.76    N firefox
    2.02        1.88    N nautilus
    3.36        3.17    N nautilus --browser
    5.26        4.89    N gthumb
    7.12        6.47    N gedit
    9.22        8.16    N xpdf /usr/share/doc/shared-mime-info/shared-mime-info-spec.pdf
   13.58       12.55    N xterm
   15.87       14.57    N mlterm
   18.63       17.06    N gnome-terminal
   21.16       18.90    N urxvt
   26.24       23.48    N gnome-system-monitor
   28.72       26.52    N gnome-help
   32.15       29.65    N gnome-dictionary
   39.66       36.12    N /usr/games/sol
   43.16       39.27    N /usr/games/gnometris
   48.65       42.56    N /usr/games/gnect
   53.31       47.03    N /usr/games/gtali
   58.60       52.05    N /usr/games/iagno
   65.77       55.42    N /usr/games/gnotravex
   70.76       61.47    N /usr/games/mahjongg
   76.15       67.11    N /usr/games/gnome-sudoku
   86.32       75.15    N /usr/games/glines
   92.21       79.70    N /usr/games/glchess
  103.79       88.48    N /usr/games/gnomine
  113.84       96.51    N /usr/games/gnotski
  124.40      102.19    N /usr/games/gnibbles
  137.41      114.93    N /usr/games/gnobots2
  155.53      125.02    N /usr/games/blackjack
  179.85      135.11    N /usr/games/same-gnome
  224.49      154.50    N /usr/bin/gnome-window-properties
  248.44      162.09    N /usr/bin/gnome-default-applications-properties
  282.62      173.29    N /usr/bin/gnome-at-properties
  323.72      188.21    N /usr/bin/gnome-typing-monitor
  363.99      199.93    N /usr/bin/gnome-at-visual
  394.21      206.95    N /usr/bin/gnome-sound-properties
  435.14      224.49    N /usr/bin/gnome-at-mobility
  463.05      234.11    N /usr/bin/gnome-keybinding-properties
  503.75      248.59    N /usr/bin/gnome-about-me
  554.00      276.27    N /usr/bin/gnome-display-properties
  615.48      304.39    N /usr/bin/gnome-network-preferences
  693.03      342.01    N /usr/bin/gnome-mouse-properties
  759.90      388.58    N /usr/bin/gnome-appearance-properties
  937.90      508.47    N /usr/bin/gnome-control-center
 1109.75      587.57    N /usr/bin/gnome-keyboard-properties
 1399.05      758.16    N : oocalc
 1524.64      830.03    N : oodraw
 1684.31      900.03    N : ooimpress
 1874.04      993.91    N : oomath
 2115.12     1081.89    N : ooweb
 2369.02     1161.99    N : oowriter

Note that the last ": oo*" commands are actually commented out.

1.4) vmstat numbers (some relevant ones are marked with *)

                            before    after
 nr_free_pages              1293      3898
 nr_inactive_anon           59956     53460
 nr_active_anon             26815     30026
 nr_inactive_file           2657      3218
 nr_active_file             2019      2806
 nr_unevictable             4         4
 nr_mlock                   4         4
 nr_anon_pages              26706     27859
*nr_mapped                  3542      4469
 nr_file_pages              72232     67681
 nr_dirty                   1         0
 nr_writeback               123       19
 nr_slab_reclaimable        3375      3534
 nr_slab_unreclaimable      11405     10665
 nr_page_table_pages        8106      7864
 nr_unstable                0         0
 nr_bounce                  0         0
*nr_vmscan_write            394776    230839
 nr_writeback_temp          0         0
 numa_hit                   6843353   3318676
 numa_miss                  0         0
 numa_foreign               0         0
 numa_interleave            1719      1719
 numa_local                 6843353   3318676
 numa_other                 0         0
*pgpgin                     5954683   2057175
*pgpgout                    1578276   922744
*pswpin                     1486615   512238
*pswpout                    394568    230685
 pgalloc_dma                277432    56602
 pgalloc_dma32              6769477   3310348
 pgalloc_normal             0         0
 pgalloc_movable            0         0
 pgfree                     7048396   3371118
 pgactivate                 2036343   1471492
 pgdeactivate               2189691   1612829
 pgfault                    3702176   3100702
*pgmajfault                 452116    201343
 pgrefill_dma               12185     7127
 pgrefill_dma32             334384    653703
 pgrefill_normal            0         0
 pgrefill_movable           0         0
 pgsteal_dma                74214     22179
 pgsteal_dma32              3334164   1638029
 pgsteal_normal             0         0
 pgsteal_movable            0         0
 pgscan_kswapd_dma          1081421   1216199
 pgscan_kswapd_dma32        58979118  46002810
 pgscan_kswapd_normal       0         0
 pgscan_kswapd_movable      0         0
 pgscan_direct_dma          2015438   1086109
 pgscan_direct_dma32        55787823  36101597
 pgscan_direct_normal       0         0
 pgscan_direct_movable      0         0
 pginodesteal               3461      7281
 slabs_scanned              564864    527616
 kswapd_steal               2889797   1448082
 kswapd_inodesteal          14827     14835
 pageoutrun                 43459     21562
 allocstall                 9653      4032
 pgrotated                  384216    228631

1.5) free numbers at the end of the tests

before patch:
                             total       used       free     shared    buffers     cached
                Mem:           474        467          7          0          0        236
                -/+ buffers/cache:        230        243
                Swap:         1023        418        605

after patch:
                             total       used       free     shared    buffers     cached
                Mem:           474        457         16          0          0        236
                -/+ buffers/cache:        221        253
                Swap:         1023        404        619

2) memory flushing in a file server

2.1) brief summary

The number of major faults from 50 to 3 during 10% cache hot reads.

That means this patch successfully stops major faults when the active file
list is slowly scanned when there are partially cache hot streaming IO.

2.2) test scenario

Do 100000 pread(size=110 pages, offset=(i*100) pages), where 10% of the
pages will be activated:

        for i in `seq 0 100 10000000`; do echo $i 110;  done > pattern-hot-10
        iotrace.rb --load pattern-hot-10 --play /b/sparse
vmmon  nr_mapped nr_active_file nr_inactive_file   pgmajfault pgdeactivate pgfree

and monitor /proc/vmstat during the time. The test box has 2G memory.

I carried out tests on fresh booted console as well as X desktop, and
fetched the vmstat numbers on

(1) begin:     shortly after the big read IO starts;
(2) end:       just before the big read IO stops;
(3) restore:   the big read IO stops and the zsh working set restored
(4) restore X: after IO, switch back and forth between the urxvt and firefox
               windows to restore their working set.

2.3) console mode results

        nr_mapped   nr_active_file nr_inactive_file       pgmajfault     pgdeactivate           pgfree

2.6.29 VM_EXEC protection ON:
begin:       2481             2237             8694              630                0           574299
end:          275           231976           233914              633           776271         20933042
restore:      370           232154           234524              691           777183         20958453

2.6.29 VM_EXEC protection ON (second run):
begin:       2434             2237             8493              629                0           574195
end:          284           231970           233536              632           771918         20896129
restore:      399           232218           234789              690           774526         20957909

2.6.30-rc4-mm VM_EXEC protection OFF:
begin:       2479             2344             9659              210                0           579643
end:          284           232010           234142              260           772776         20917184
restore:      379           232159           234371              301           774888         20967849

The above console numbers show that

- The startup pgmajfault of 2.6.30-rc4-mm is merely 1/3 that of 2.6.29.
  I'd attribute that improvement to the mmap readahead improvements :-)

- The pgmajfault increment during the file copy is 633-630=3 vs 260-210=50.
  That's a huge improvement - which means with the VM_EXEC protection logic,
  active mmap pages is pretty safe even under partially cache hot streaming IO.

- when active:inactive file lru size reaches 1:1, their scan rates is 1:20.8
  under 10% cache hot IO. (computed with formula Dpgdeactivate:Dpgfree)
  That roughly means the active mmap pages get 20.8 more chances to get
  re-referenced to stay in memory.

- The absolute nr_mapped drops considerably to 1/9 during the big IO, and the
  dropped pages are mostly inactive ones. The patch has almost no impact in
  this aspect, that means it won't unnecessarily increase memory pressure.
  (In contrast, your 20% mmap protection ratio will keep them all, and
  therefore eliminate the extra 41 major faults to restore working set
  of zsh etc.)

The iotrace.rb read throughput is
151.194384MB/s 284.198252s 100001x 450560b --load pattern-hot-10 --play /b/sparse
which means the inactive list is rotated at the speed of 250MB/s,
so a full scan of which takes about 3.5 seconds, while a full scan
of active file list takes about 77 seconds.

2.4) X mode results

We can reach roughly the same conclusions for X desktop:

        nr_mapped   nr_active_file nr_inactive_file       pgmajfault     pgdeactivate           pgfree

2.6.30-rc4-mm VM_EXEC protection ON:
begin:       9740             8920            64075              561                0           678360
end:          768           218254           220029              565           798953         21057006
restore:      857           218543           220987              606           799462         21075710
restore X:   2414           218560           225344              797           799462         21080795

2.6.30-rc4-mm VM_EXEC protection OFF:
begin:       9368             5035            26389              554                0           633391
end:          770           218449           221230              661           646472         17832500
restore:     1113           218466           220978              710           649881         17905235
restore X:   2687           218650           225484              947           802700         21083584

- the absolute nr_mapped drops considerably (to 1/13 of the original size)
  during the streaming IO.
- the delta of pgmajfault is 3 vs 107 during IO, or 236 vs 393
  during the whole process.

Cc: Elladan <elladan@eskimo.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agovmscan: report vm_flags in page_referenced()
Wu Fengguang [Tue, 16 Jun 2009 22:33:05 +0000]
vmscan: report vm_flags in page_referenced()

Collect vma->vm_flags of the VMAs that actually referenced the page.

This is preparing for more informed reclaim heuristics, eg.  to protect
executable file pages more aggressively.  For now only the VM_EXEC bit
will be used by the caller.

Thanks to Johannes, Peter and Minchan for all the good tips.

Acked-by: Peter Zijlstra <peterz@infradead.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: add a gfp-translate script to help understand page allocation failure reports
Mel Gorman [Tue, 16 Jun 2009 22:33:04 +0000]
mm: add a gfp-translate script to help understand page allocation failure reports

The page allocation failure messages include a line that looks like

page allocation failure. order:1, mode:0x4020

The mode is easy to translate but irritating for the lazy and a bit error
prone.  This patch adds a very simple helper script gfp-translate for the
mode: portion of the page allocation failure messages.  An example usage
looks like

  mel@machina:~/linux-2.6 $ scripts/gfp-translate 0x4020
  Source: /home/mel/linux-2.6
  Parsing: 0x4020
  #define __GFP_HIGH (0x20) /* Should access emergency pools? */
  #define __GFP_COMP (0x4000) /* Add compound page metadata */

The script is not a work of art but it has come in handy for me a few
times so I thought I would share.

[akpm@linux-foundation.org: clarify an error message]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm cleanup: shmem_file_setup: 'char *' -> 'const char *' for name argument
Sergei Trofimovich [Tue, 16 Jun 2009 22:33:02 +0000]
mm cleanup: shmem_file_setup: 'char *' -> 'const char *' for name argument

As function shmem_file_setup does not modify/allocate/free/pass given
filename - mark it as const.

Signed-off-by: Sergei Trofimovich <slyfox@inbox.ru>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: remove file argument from swap_readpage()
Minchan Kim [Tue, 16 Jun 2009 22:33:02 +0000]
mm: remove file argument from swap_readpage()

The file argument resulted from address_space's readpage long time ago.

We don't use it any more.  Let's remove unnecessary argement.

Signed-off-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: remove annotation of gfp_mask in add_to_swap
Minchan Kim [Tue, 16 Jun 2009 22:33:01 +0000]
mm: remove annotation of gfp_mask in add_to_swap

Hugh removed add_to_swap's gfp_mask argument.  (mm: remove gfp_mask from
add_to_swap) So we have to remove annotation of gfp_mask of the function.

Signed-off-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage-allocator: clear N_HIGH_MEMORY map before we set it again
Yinghai Lu [Tue, 16 Jun 2009 22:33:00 +0000]
page-allocator: clear N_HIGH_MEMORY map before we set it again

SRAT tables may contains nodes of very small size.  The arch code may
decide to not activate such a node.  However, currently the early boot
code sets N_HIGH_MEMORY for such nodes.  These nodes therefore seem to be
active although these nodes have no present pages.

For 64bit N_HIGH_MEMORY == N_NORMAL_MEMORY, so that works for 64 bit too

Signed-off-by: Yinghai Lu <Yinghai@kernel.org>
Tested-by: Jack Steiner <steiner@sgi.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: remove __invalidate_mapping_pages variant
Mike Waychison [Tue, 16 Jun 2009 22:32:59 +0000]
mm: remove __invalidate_mapping_pages variant

Remove __invalidate_mapping_pages atomic variant now that its sole caller
can sleep (fixed in eccb95cee4f0d56faa46ef22fb94dd4a3578d3eb ("vfs: fix
lock inversion in drop_pagecache_sb()")).

This fixes softlockups that can occur while in the drop_caches path.

Signed-off-by: Mike Waychison <mikew@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Acked-by: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agooom: invoke oom killer for __GFP_NOFAIL
David Rientjes [Tue, 16 Jun 2009 22:32:58 +0000]
oom: invoke oom killer for __GFP_NOFAIL

The oom killer must be invoked regardless of the order if the allocation
is __GFP_NOFAIL, otherwise it will loop forever when reclaim fails to free
some memory.

Cc: Nick Piggin <npiggin@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agooom: avoid unnecessary mm locking and scanning for OOM_DISABLE
David Rientjes [Tue, 16 Jun 2009 22:32:57 +0000]
oom: avoid unnecessary mm locking and scanning for OOM_DISABLE

This moves the check for OOM_DISABLE to the badness heuristic so it is
only necessary to hold task_lock() once.  If the mm is OOM_DISABLE, the
score is 0, which is also correctly exported via /proc/pid/oom_score.
This requires that tasks with badness scores of 0 are prohibited from
being oom killed, which makes sense since they would not allow for future
memory freeing anyway.

Since the oom_adj value is a characteristic of an mm and not a task, it is
no longer necessary to check the oom_adj value for threads sharing the
same memory (except when simply issuing SIGKILLs for threads in other
thread groups).

Cc: Nick Piggin <npiggin@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agooom: move oom_adj value from task_struct to mm_struct
David Rientjes [Tue, 16 Jun 2009 22:32:56 +0000]
oom: move oom_adj value from task_struct to mm_struct

The per-task oom_adj value is a characteristic of its mm more than the
task itself since it's not possible to oom kill any thread that shares the
mm.  If a task were to be killed while attached to an mm that could not be
freed because another thread were set to OOM_DISABLE, it would have
needlessly been terminated since there is no potential for future memory
freeing.

This patch moves oomkilladj (now more appropriately named oom_adj) from
struct task_struct to struct mm_struct.  This requires task_lock() on a
task to check its oom_adj value to protect against exec, but it's already
necessary to take the lock when dereferencing the mm to find the total VM
size for the badness heuristic.

This fixes a livelock if the oom killer chooses a task and another thread
sharing the same memory has an oom_adj value of OOM_DISABLE.  This occurs
because oom_kill_task() repeatedly returns 1 and refuses to kill the
chosen task while select_bad_process() will repeatedly choose the same
task during the next retry.

Taking task_lock() in select_bad_process() to check for OOM_DISABLE and in
oom_kill_task() to check for threads sharing the same memory will be
removed in the next patch in this series where it will no longer be
necessary.

Writing to /proc/pid/oom_adj for a kthread will now return -EINVAL since
these threads are immune from oom killing already.  They simply report an
oom_adj value of OOM_DISABLE.

Cc: Nick Piggin <npiggin@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: reuse unused swap entry if necessary
KAMEZAWA Hiroyuki [Tue, 16 Jun 2009 22:32:54 +0000]
mm: reuse unused swap entry if necessary

Presently we can know a swap entry is just used as SwapCache via swap_map,
without looking up swap cache.

Then, we have a chance to reuse swap-cache-only swap entries in
get_swap_pages().

This patch tries to free swap-cache-only swap entries if swap is not
enough.

Note: We hit following path when swap_cluster code cannot find a free
cluster.  Then, vm_swap_full() is not only condition to allow the kernel
to reclaim unused swap.

Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Balbir Singh <balbir@in.ibm.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Dhaval Giani <dhaval@linux.vnet.ibm.com>
Cc: YAMAMOTO Takashi <yamamoto@valinux.co.jp>
Tested-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: modify swap_map and add SWAP_HAS_CACHE flag
KAMEZAWA Hiroyuki [Tue, 16 Jun 2009 22:32:53 +0000]
mm: modify swap_map and add SWAP_HAS_CACHE flag

This is a part of the patches for fixing memcg's swap accountinf leak.
But, IMHO, not a bad patch even if no memcg.

There are 2 kinds of references to swap.
 - reference from swap entry
 - reference from swap cache

Then,

 - If there is swap cache && swap's refcnt is 1, there is only swap cache.
  (*) swapcount(entry) == 1 && find_get_page(swapper_space, entry) != NULL

This counting logic have worked well for a long time.  But considering
that we cannot know there is a _real_ reference or not by swap_map[],
current usage of counter is not very good.

This patch adds a flag SWAP_HAS_CACHE and recored information that a swap
entry has a cache or not.  This will remove -1 magic used in swapfile.c
and be a help to avoid unnecessary find_get_page().

Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Tested-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Dhaval Giani <dhaval@linux.vnet.ibm.com>
Cc: YAMAMOTO Takashi <yamamoto@valinux.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: add swap cache interface for swap reference
KAMEZAWA Hiroyuki [Tue, 16 Jun 2009 22:32:52 +0000]
mm: add swap cache interface for swap reference

In a following patch, the usage of swap cache is recorded into swap_map.
This patch is for necessary interface changes to do that.

2 interfaces:

  - swapcache_prepare()
  - swapcache_free()

are added for allocating/freeing refcnt from swap-cache to existing swap
entries.  But implementation itself is not changed under this patch.  At
adding swapcache_free(), memcg's hook code is moved under
swapcache_free().  This is better than using scattered hooks.

Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Acked-by: Balbir Singh <balbir@in.ibm.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Dhaval Giani <dhaval@linux.vnet.ibm.com>
Cc: YAMAMOTO Takashi <yamamoto@valinux.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: remove CONFIG_UNEVICTABLE_LRU config option
KOSAKI Motohiro [Tue, 16 Jun 2009 22:32:51 +0000]
mm: remove CONFIG_UNEVICTABLE_LRU config option

Currently, nobody wants to turn UNEVICTABLE_LRU off.  Thus this
configurability is unnecessary.

Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Andi Kleen <andi@firstfloor.org>
Acked-by: Minchan Kim <minchan.kim@gmail.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage-allocator: reset wmark_min and inactive ratio of zone when hotplug happens
Minchan Kim [Tue, 16 Jun 2009 22:32:50 +0000]
page-allocator: reset wmark_min and inactive ratio of zone when hotplug happens

Solve two problems.

Whenever memory hotplug sucessfully happens, zone->present_pages
have to be changed.

1) Now memory hotplug calls setup_per_zone_wmark_min only when
   online_pages called, not offline_pages.

   It breaks balance.

2) If zone->present_pages is changed, we also have to change
   zone->inactive_ratio.  That's because inactive_ratio depends on
   zone->present_pages.

Signed-off-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage-allocator: add inactive ratio calculation function of each zone
Minchan Kim [Tue, 16 Jun 2009 22:32:49 +0000]
page-allocator: add inactive ratio calculation function of each zone

Factor the per-zone arithemetic inside setup_per_zone_inactive_ratio()'s
loop into a a separate function, calculate_zone_inactive_ratio().  This
function will be used in a later patch

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Rik van Riel <riel@redhat.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>

10 years agopage-allocator: clean up functions related to pages_min
Minchan Kim [Tue, 16 Jun 2009 22:32:48 +0000]
page-allocator: clean up functions related to pages_min

Change the names of two functions. It doesn't affect behavior.

Presently, setup_per_zone_pages_min() changes low, high of zone as well as
min.  So a better name is setup_per_zone_wmarks().  That's because Mel
changed zone->pages_[hig/low/min] to zone->watermark array in "page
allocator: replace the watermark-related union in struct zone with a
watermark[] array".

 * setup_per_zone_pages_min => setup_per_zone_wmarks

Of course, we have to change init_per_zone_pages_min, too.  There are not
pages_min any more.

 * init_per_zone_pages_min => init_per_zone_wmark_min

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Rik van Riel <riel@redhat.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>

10 years agopage-allocator: use integer fields lookup for gfp_zone and check for errors in flags...
Christoph Lameter [Tue, 16 Jun 2009 22:32:46 +0000]
page-allocator: use integer fields lookup for gfp_zone and check for errors in flags passed to the page allocator

This simplifies the code in gfp_zone() and also keeps the ability of the
compiler to use constant folding to get rid of gfp_zone processing.

The lookup of the zone is done using a bitfield stored in an integer.  So
the code in gfp_zone is a simple extraction of bits from a constant
bitfield.  The compiler is generating a load of a constant into a register
and then performs a shift and mask operation to get the zone from a gfp_t.
 No cachelines are touched and no branches have to be predicted by the
compiler.

We are doing some macro tricks here to convince the compiler to always do
the constant folding if possible.

Signed-off-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Mel Gorman <mel@csn.ul.ie>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: check the argument of kunmap on architectures without highmem
Matthew Wilcox [Tue, 16 Jun 2009 22:32:45 +0000]
mm: check the argument of kunmap on architectures without highmem

If you're using a non-highmem architecture, passing an argument with the
wrong type to kunmap() doesn't give you a warning because the ifdef
doesn't check the type.

Using a static inline function solves the problem nicely.

Reported-by: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Matthew Wilcox <willy@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agovmscan: prevent shrinking of active anon lru list in case of no swap space V3
MinChan Kim [Tue, 16 Jun 2009 22:32:44 +0000]
vmscan: prevent shrinking of active anon lru list in case of no swap space V3

shrink_zone() can deactivate active anon pages even if we don't have a
swap device.  Many embedded products don't have a swap device.  So the
deactivation of anon pages is unnecessary.

This patch prevents unnecessary deactivation of anon lru pages.  But, it
don't prevent aging of anon pages to swap out.

Signed-off-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomigration: only migrate_prep() once per move_pages()
Brice Goglin [Tue, 16 Jun 2009 22:32:43 +0000]
migration: only migrate_prep() once per move_pages()

migrate_prep() is fairly expensive (72us on 16-core barcelona 1.9GHz).
Commit 3140a2273009c01c27d316f35ab76a37e105fdd8 improved move_pages()
throughput by breaking it into chunks, but it also made migrate_prep() be
called once per chunk (every 128pages or so) instead of once per
move_pages().

This patch reverts to calling migrate_prep() only once per chunk as we did
before 2.6.29.  It is also a followup to commit
0aedadf91a70a11c4a3e7c7d99b21e5528af8d5d ("mm: move migrate_prep out from
under mmap_sem").

This improves migration throughput on the above machine from 600MB/s to
750MB/s.

Signed-off-by: Brice Goglin <Brice.Goglin@inria.fr>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Rik van Riel <riel@redhat.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm, PM/Freezer: Disable OOM killer when tasks are frozen
Rafael J. Wysocki [Tue, 16 Jun 2009 22:32:41 +0000]
mm, PM/Freezer: Disable OOM killer when tasks are frozen

Currently, the following scenario appears to be possible in theory:

* Tasks are frozen for hibernation or suspend.
* Free pages are almost exhausted.
* Certain piece of code in the suspend code path attempts to allocate
  some memory using GFP_KERNEL and allocation order less than or
  equal to PAGE_ALLOC_COSTLY_ORDER.
* __alloc_pages_internal() cannot find a free page so it invokes the
  OOM killer.
* The OOM killer attempts to kill a task, but the task is frozen, so
  it doesn't die immediately.
* __alloc_pages_internal() jumps to 'restart', unsuccessfully tries
  to find a free page and invokes the OOM killer.
* No progress can be made.

Although it is now hard to trigger during hibernation due to the memory
shrinking carried out by the hibernation code, it is theoretically
possible to trigger during suspend after the memory shrinking has been
removed from that code path.  Moreover, since memory allocations are
going to be used for the hibernation memory shrinking, it will be even
more likely to happen during hibernation.

To prevent it from happening, introduce the oom_killer_disabled switch
that will cause __alloc_pages_internal() to fail in the situations in
which the OOM killer would have been called and make the freezer set
this switch after tasks have been successfully frozen.

[akpm@linux-foundation.org: be nicer to the namespace]
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Fengguang Wu <fengguang.wu@gmail.com>
Cc: David Rientjes <rientjes@google.com>
Acked-by: Pavel Machek <pavel@ucw.cz>
Cc: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: madvise(): correct return code
Nick Piggin [Tue, 16 Jun 2009 22:32:38 +0000]
mm: madvise(): correct return code

The posix_madvise() function succeeds (and does nothing) when called with
parameters (NULL, 0, -1); according to LSB tests, it should fail with
EINVAL because -1 is not a valid flag.

When called with a valid address and size, it correctly fails.

So perform an initial check for valid flags first.

Reported-by: Jiri Dluhos <jdluhos@novell.com>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Reviewed-and-Tested-by: WANG Cong <xiyou.wangcong@gmail.com>
Cc: Michael Kerrisk <mtk.manpages@googlemail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage-allocator: warn if __GFP_NOFAIL is used for a large allocation
Andrew Morton [Tue, 16 Jun 2009 22:32:37 +0000]
page-allocator: warn if __GFP_NOFAIL is used for a large allocation

__GFP_NOFAIL is a bad fiction.  Allocations _can_ fail, and callers should
detect and suitably handle this (and not by lamely moving the infinite
loop up to the caller level either).

Attempting to use __GFP_NOFAIL for a higher-order allocation is even
worse, so add a once-off runtime check for this to slap people around for
even thinking about trying it.

Cc: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
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>

10 years agovideobuf-dma-contig: zero copy USERPTR support
Magnus Damm [Tue, 16 Jun 2009 22:32:36 +0000]
videobuf-dma-contig: zero copy USERPTR support

Since videobuf-dma-contig is designed to handle physically contiguous
memory, this patch modifies the videobuf-dma-contig code to only accept a
user space pointer to physically contiguous memory.  For now only
VM_PFNMAP vmas are supported, so forget hotplug.

On SuperH Mobile we use this with our sh_mobile_ceu_camera driver together
with various multimedia accelerator blocks that are exported to user space
using UIO.  The UIO kernel code exports physically contiguous memory to
user space and lets the user space application mmap() this memory and pass
a pointer using the USERPTR interface for V4L2 zero copy operation.

With this approach we support zero copy capture, hardware scaling and
various forms of hardware encoding and decoding.

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Magnus Damm <damm@igel.co.jp>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Paul Mundt <lethal@linux-sh.org>
Acked-by: Mauro Carvalho Chehab <mchehab@infradead.org>
Cc: Hans Verkuil <hverkuil@xs4all.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: introduce follow_pfn()
Johannes Weiner [Tue, 16 Jun 2009 22:32:35 +0000]
mm: introduce follow_pfn()

Analoguous to follow_phys(), add a helper that looks up the PFN at a
user virtual address in an IO mapping or a raw PFN mapping.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Christoph Hellwig <hch@infradead.org>
Acked-by: Magnus Damm <magnus.damm@gmail.com>
Cc: Hans Verkuil <hverkuil@xs4all.nl>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: use generic follow_pte() in follow_phys()
Johannes Weiner [Tue, 16 Jun 2009 22:32:34 +0000]
mm: use generic follow_pte() in follow_phys()

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Christoph Hellwig <hch@infradead.org>
Acked-by: Magnus Damm <magnus.damm@gmail.com>
Cc: Hans Verkuil <hverkuil@xs4all.nl>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: introduce follow_pte()
Johannes Weiner [Tue, 16 Jun 2009 22:32:33 +0000]
mm: introduce follow_pte()

A generic readonly page table lookup helper to map an address space and an
address from it to a pte.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Christoph Hellwig <hch@infradead.org>
Acked-by: Magnus Damm <magnus.damm@gmail.com>
Cc: Hans Verkuil <hverkuil@xs4all.nl>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: setup_per_zone_inactive_ratio - fix comment and make it __init
Cyrill Gorcunov [Tue, 16 Jun 2009 22:32:32 +0000]
mm: setup_per_zone_inactive_ratio - fix comment and make it __init

The caller of setup_per_zone_inactive_ratio is an __init function.  There
is no need to keep the callee after it completed as well.  Also fix a
comment.

Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: setup_per_zone_inactive_ratio - do not call for int_sqrt if not needed
Cyrill Gorcunov [Tue, 16 Jun 2009 22:32:32 +0000]
mm: setup_per_zone_inactive_ratio - do not call for int_sqrt if not needed

int_sqrt() returns 0 if its argument is zero so call it if only needed.

Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.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>

10 years agovmscan: ZVC updates in shrink_active_list() can be done once
Wu Fengguang [Tue, 16 Jun 2009 22:32:31 +0000]
vmscan: ZVC updates in shrink_active_list() can be done once

This effectively lifts the unit of updates to nr_inactive_* and
pgdeactivate from PAGEVEC_SIZE=14 to SWAP_CLUSTER_MAX=32, or
MAX_ORDER_NR_PAGES=1024 for reclaim_zone().

Cc: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agovmscan: don't export nr_saved_scan in /proc/zoneinfo
Wu Fengguang [Tue, 16 Jun 2009 22:32:30 +0000]
vmscan: don't export nr_saved_scan in /proc/zoneinfo

The lru->nr_saved_scan's are not meaningful counters for even kernel
developers.  They typically are smaller than 32 and are always 0 for large
lists.  So remove them from /proc/zoneinfo.

Hopefully this interface change won't break too many scripts.
/proc/zoneinfo is too unstructured to be script friendly, and I wonder the
affected scripts - if there are any - are still bleeding since the not
long ago commit "vmscan: split LRU lists into anon & file sets", which
also touched the "scanned" line :)

If we are to re-export accumulated vmscan counts in the future, they can
go to new lines in /proc/zoneinfo instead of the current form, or to
/sys/devices/system/node/node0/meminfo?

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>

Acked-by: Rik van Riel <riel@redhat.com>
Cc: Nick Piggin <npiggin@suse.de>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
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>

10 years agovmscan: cleanup the scan batching code
Wu Fengguang [Tue, 16 Jun 2009 22:32:29 +0000]
vmscan: cleanup the scan batching code

The vmscan batching logic is twisting.  Move it into a standalone function
nr_scan_try_batch() and document it.  No behavior change.

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agovmscan: evict use-once pages first
Rik van Riel [Tue, 16 Jun 2009 22:32:28 +0000]
vmscan: evict use-once pages first

When the file LRU lists are dominated by streaming IO pages, evict those
pages first, before considering evicting other pages.

This should be safe from deadlocks or performance problems
because only three things can happen to an inactive file page:

1) referenced twice and promoted to the active list
2) evicted by the pageout code
3) under IO, after which it will get evicted or promoted

The pages freed in this way can either be reused for streaming IO, or
allocated for something else.  If the pages are used for streaming IO,
this pageout pattern continues.  Otherwise, we will fall back to the
normal pageout pattern.

Signed-off-by: Rik van Riel <riel@redhat.com>
Reported-by: Elladan <elladan@eskimo.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopagemap: add page-types tool
Wu Fengguang [Tue, 16 Jun 2009 22:32:27 +0000]
pagemap: add page-types tool

Add page-types, a handy tool for querying page flags.

It will expand some of the overloaded flags:
PG_slob_free   = PG_private
PG_slub_frozen = PG_active
PG_slub_debug  = PG_error
PG_readahead   = PG_reclaim

and mask out obscure flags except in -raw mode:
PG_reserved
PG_mlocked
PG_mappedtodisk
PG_private
PG_private_2
PG_owner_priv_1
PG_arch_1
PG_uncached
PG_compound* for non hugeTLB pages

[akpm@linux-foundation.org: fix warning]
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopagemap: document 9 more exported page flags
Wu Fengguang [Tue, 16 Jun 2009 22:32:26 +0000]
pagemap: document 9 more exported page flags

Also add short descriptions for all of the 20 exported page flags.

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopagemap: document clarifications
Wu Fengguang [Tue, 16 Jun 2009 22:32:25 +0000]
pagemap: document clarifications

Some bit ranges were inclusive and some not.  Fix them to be consistently
inclusive.

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoproc: export more page flags in /proc/kpageflags
Wu Fengguang [Tue, 16 Jun 2009 22:32:24 +0000]
proc: export more page flags in /proc/kpageflags

Export all page flags faithfully in /proc/kpageflags.

11. KPF_MMAP (pseudo flag) memory mapped page
12. KPF_ANON (pseudo flag) memory mapped page (anonymous)
13. KPF_SWAPCACHE page is in swap cache
14. KPF_SWAPBACKED page is swap/RAM backed
15. KPF_COMPOUND_HEAD (*)
16. KPF_COMPOUND_TAIL (*)
17. KPF_HUGE hugeTLB pages
18. KPF_UNEVICTABLE page is in the unevictable LRU list
19. KPF_HWPOISON(TBD) hardware detected corruption
20. KPF_NOPAGE (pseudo flag) no page frame at the address
32-39. more obscure flags for kernel developers

(*) For compound pages, exporting _both_ head/tail info enables
    users to tell where a compound page starts/ends, and its order.

The accompanying page-types tool will handle the details like decoupling
overloaded flags and hiding obscure flags to normal users.

Thanks to KOSAKI and Andi for their valuable recommendations!

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoproc: kpagecount/kpageflags code cleanup
Wu Fengguang [Tue, 16 Jun 2009 22:32:23 +0000]
proc: kpagecount/kpageflags code cleanup

Move increments of pfn/out to bottom of the loop.

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Andi Kleen <andi@firstfloor.org>
Acked-by: Matt Mackall <mpm@selenic.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: introduce PageHuge() for testing huge/gigantic pages
Wu Fengguang [Tue, 16 Jun 2009 22:32:22 +0000]
mm: introduce PageHuge() for testing huge/gigantic pages

A series of patches to enhance the /proc/pagemap interface and to add a
userspace executable which can be used to present the pagemap data.

Export 10 more flags to end users (and more for kernel developers):

        11. KPF_MMAP            (pseudo flag) memory mapped page
        12. KPF_ANON            (pseudo flag) memory mapped page (anonymous)
        13. KPF_SWAPCACHE       page is in swap cache
        14. KPF_SWAPBACKED      page is swap/RAM backed
        15. KPF_COMPOUND_HEAD   (*)
        16. KPF_COMPOUND_TAIL   (*)
        17. KPF_HUGE hugeTLB pages
        18. KPF_UNEVICTABLE     page is in the unevictable LRU list
        19. KPF_HWPOISON        hardware detected corruption
        20. KPF_NOPAGE          (pseudo flag) no page frame at the address

        (*) For compound pages, exporting _both_ head/tail info enables
            users to tell where a compound page starts/ends, and its order.

a simple demo of the page-types tool

# ./page-types -h
page-types [options]
            -r|--raw                  Raw mode, for kernel developers
            -a|--addr    addr-spec    Walk a range of pages
            -b|--bits    bits-spec    Walk pages with specified bits
            -l|--list                 Show page details in ranges
            -L|--list-each            Show page details one by one
            -N|--no-summary           Don't show summay info
            -h|--help                 Show this usage message
addr-spec:
            N                         one page at offset N (unit: pages)
            N+M                       pages range from N to N+M-1
            N,M                       pages range from N to M-1
            N,                        pages range from N to end
            ,M                        pages range from 0 to M
bits-spec:
            bit1,bit2                 (flags & (bit1|bit2)) != 0
            bit1,bit2=bit1            (flags & (bit1|bit2)) == bit1
            bit1,~bit2                (flags & (bit1|bit2)) == bit1
            =bit1,bit2                flags == (bit1|bit2)
bit-names:
          locked              error         referenced           uptodate
           dirty                lru             active               slab
       writeback            reclaim              buddy               mmap
       anonymous          swapcache         swapbacked      compound_head
   compound_tail               huge        unevictable           hwpoison
          nopage           reserved(r)         mlocked(r)    mappedtodisk(r)
         private(r)       private_2(r)   owner_private(r)            arch(r)
        uncached(r)       readahead(o)       slob_free(o)     slub_frozen(o)
      slub_debug(o)
                                   (r) raw mode bits  (o) overloaded bits

# ./page-types
             flags      page-count       MB  symbolic-flags                     long-symbolic-flags
0x0000000000000000          487369     1903  _________________________________
0x0000000000000014               5        0  __R_D____________________________  referenced,dirty
0x0000000000000020               1        0  _____l___________________________  lru
0x0000000000000024              34        0  __R__l___________________________  referenced,lru
0x0000000000000028            3838       14  ___U_l___________________________  uptodate,lru
0x0001000000000028              48        0  ___U_l_______________________I___  uptodate,lru,readahead
0x000000000000002c            6478       25  __RU_l___________________________  referenced,uptodate,lru
0x000100000000002c              47        0  __RU_l_______________________I___  referenced,uptodate,lru,readahead
0x0000000000000040            8344       32  ______A__________________________  active
0x0000000000000060               1        0  _____lA__________________________  lru,active
0x0000000000000068             348        1  ___U_lA__________________________  uptodate,lru,active
0x0001000000000068              12        0  ___U_lA______________________I___  uptodate,lru,active,readahead
0x000000000000006c             988        3  __RU_lA__________________________  referenced,uptodate,lru,active
0x000100000000006c              48        0  __RU_lA______________________I___  referenced,uptodate,lru,active,readahead
0x0000000000004078               1        0  ___UDlA_______b__________________  uptodate,dirty,lru,active,swapbacked
0x000000000000407c              34        0  __RUDlA_______b__________________  referenced,uptodate,dirty,lru,active,swapbacked
0x0000000000000400             503        1  __________B______________________  buddy
0x0000000000000804               1        0  __R________M_____________________  referenced,mmap
0x0000000000000828            1029        4  ___U_l_____M_____________________  uptodate,lru,mmap
0x0001000000000828              43        0  ___U_l_____M_________________I___  uptodate,lru,mmap,readahead
0x000000000000082c             382        1  __RU_l_____M_____________________  referenced,uptodate,lru,mmap
0x000100000000082c              12        0  __RU_l_____M_________________I___  referenced,uptodate,lru,mmap,readahead
0x0000000000000868             192        0  ___U_lA____M_____________________  uptodate,lru,active,mmap
0x0001000000000868              12        0  ___U_lA____M_________________I___  uptodate,lru,active,mmap,readahead
0x000000000000086c             800        3  __RU_lA____M_____________________  referenced,uptodate,lru,active,mmap
0x000100000000086c              31        0  __RU_lA____M_________________I___  referenced,uptodate,lru,active,mmap,readahead
0x0000000000004878               2        0  ___UDlA____M__b__________________  uptodate,dirty,lru,active,mmap,swapbacked
0x0000000000001000             492        1  ____________a____________________  anonymous
0x0000000000005808               4        0  ___U_______Ma_b__________________  uptodate,mmap,anonymous,swapbacked
0x0000000000005868            2839       11  ___U_lA____Ma_b__________________  uptodate,lru,active,mmap,anonymous,swapbacked
0x000000000000586c              30        0  __RU_lA____Ma_b__________________  referenced,uptodate,lru,active,mmap,anonymous,swapbacked
             total          513968     2007

# ./page-types -r
             flags      page-count       MB  symbolic-flags                     long-symbolic-flags
0x0000000000000000          468002     1828  _________________________________
0x0000000100000000           19102       74  _____________________r___________  reserved
0x0000000000008000              41        0  _______________H_________________  compound_head
0x0000000000010000             188        0  ________________T________________  compound_tail
0x0000000000008014               1        0  __R_D__________H_________________  referenced,dirty,compound_head
0x0000000000010014               4        0  __R_D___________T________________  referenced,dirty,compound_tail
0x0000000000000020               1        0  _____l___________________________  lru
0x0000000800000024              34        0  __R__l__________________P________  referenced,lru,private
0x0000000000000028            3794       14  ___U_l___________________________  uptodate,lru
0x0001000000000028              46        0  ___U_l_______________________I___  uptodate,lru,readahead
0x0000000400000028              44        0  ___U_l_________________d_________  uptodate,lru,mappedtodisk
0x0001000400000028               2        0  ___U_l_________________d_____I___  uptodate,lru,mappedtodisk,readahead
0x000000000000002c            6434       25  __RU_l___________________________  referenced,uptodate,lru
0x000100000000002c              47        0  __RU_l_______________________I___  referenced,uptodate,lru,readahead
0x000000040000002c              14        0  __RU_l_________________d_________  referenced,uptodate,lru,mappedtodisk
0x000000080000002c              30        0  __RU_l__________________P________  referenced,uptodate,lru,private
0x0000000800000040            8124       31  ______A_________________P________  active,private
0x0000000000000040             219        0  ______A__________________________  active
0x0000000800000060               1        0  _____lA_________________P________  lru,active,private
0x0000000000000068             322        1  ___U_lA__________________________  uptodate,lru,active
0x0001000000000068              12        0  ___U_lA______________________I___  uptodate,lru,active,readahead
0x0000000400000068              13        0  ___U_lA________________d_________  uptodate,lru,active,mappedtodisk
0x0000000800000068              12        0  ___U_lA_________________P________  uptodate,lru,active,private
0x000000000000006c             977        3  __RU_lA__________________________  referenced,uptodate,lru,active
0x000100000000006c              48        0  __RU_lA______________________I___  referenced,uptodate,lru,active,readahead
0x000000040000006c               5        0  __RU_lA________________d_________  referenced,uptodate,lru,active,mappedtodisk
0x000000080000006c               3        0  __RU_lA_________________P________  referenced,uptodate,lru,active,private
0x0000000c0000006c               3        0  __RU_lA________________dP________  referenced,uptodate,lru,active,mappedtodisk,private
0x0000000c00000068               1        0  ___U_lA________________dP________  uptodate,lru,active,mappedtodisk,private
0x0000000000004078               1        0  ___UDlA_______b__________________  uptodate,dirty,lru,active,swapbacked
0x000000000000407c              34        0  __RUDlA_______b__________________  referenced,uptodate,dirty,lru,active,swapbacked
0x0000000000000400             538        2  __________B______________________  buddy
0x0000000000000804               1        0  __R________M_____________________  referenced,mmap
0x0000000000000828            1029        4  ___U_l_____M_____________________  uptodate,lru,mmap
0x0001000000000828              43        0  ___U_l_____M_________________I___  uptodate,lru,mmap,readahead
0x000000000000082c             382        1  __RU_l_____M_____________________  referenced,uptodate,lru,mmap
0x000100000000082c              12        0  __RU_l_____M_________________I___  referenced,uptodate,lru,mmap,readahead
0x0000000000000868             192        0  ___U_lA____M_____________________  uptodate,lru,active,mmap
0x0001000000000868              12        0  ___U_lA____M_________________I___  uptodate,lru,active,mmap,readahead
0x000000000000086c             800        3  __RU_lA____M_____________________  referenced,uptodate,lru,active,mmap
0x000100000000086c              31        0  __RU_lA____M_________________I___  referenced,uptodate,lru,active,mmap,readahead
0x0000000000004878               2        0  ___UDlA____M__b__________________  uptodate,dirty,lru,active,mmap,swapbacked
0x0000000000001000             492        1  ____________a____________________  anonymous
0x0000000000005008               2        0  ___U________a_b__________________  uptodate,anonymous,swapbacked
0x0000000000005808               4        0  ___U_______Ma_b__________________  uptodate,mmap,anonymous,swapbacked
0x000000000000580c               1        0  __RU_______Ma_b__________________  referenced,uptodate,mmap,anonymous,swapbacked
0x0000000000005868            2839       11  ___U_lA____Ma_b__________________  uptodate,lru,active,mmap,anonymous,swapbacked
0x000000000000586c              29        0  __RU_lA____Ma_b__________________  referenced,uptodate,lru,active,mmap,anonymous,swapbacked
             total          513968     2007

# ./page-types --raw --list --no-summary --bits reserved
offset  count   flags
0       15      _____________________r___________
31      4       _____________________r___________
159     97      _____________________r___________
4096    2067    _____________________r___________
6752    2390    _____________________r___________
9355    3       _____________________r___________
9728    14526   _____________________r___________

This patch:

Introduce PageHuge(), which identifies huge/gigantic pages by their
dedicated compound destructor functions.

Also move prep_compound_gigantic_page() to hugetlb.c and make
__free_pages_ok() non-static.

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: use alloc_pages_exact() in alloc_large_system_hash() to avoid duplicated logic
Mel Gorman [Tue, 16 Jun 2009 22:32:19 +0000]
mm: use alloc_pages_exact() in alloc_large_system_hash() to avoid duplicated logic

alloc_large_system_hash() has logic for freeing pages at the end of an
excessively large power-of-two buffer that is a duplicate of what is in
alloc_pages_exact().  This patch converts alloc_large_system_hash() to use
alloc_pages_exact().

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: sanity check order in the page allocator slow path
Mel Gorman [Tue, 16 Jun 2009 22:32:18 +0000]
page allocator: sanity check order in the page allocator slow path

Callers may speculatively call different allocators in order of preference
trying to allocate a buffer of a given size.  The order needed to allocate
this may be larger than what the page allocator can normally handle.
While the allocator mostly does the right thing, it should not direct
reclaim or wakeup kswapd with a bogus order.  This patch sanity checks the
order in the slow path and returns NULL if it is too large.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Dave Hansen <dave@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: move free_page_mlock() to page_alloc.c
KOSAKI Motohiro [Tue, 16 Jun 2009 22:32:17 +0000]
page allocator: move free_page_mlock() to page_alloc.c

Currently, free_page_mlock() is only called from page_alloc.c.  Thus, we
can move it to page_alloc.c.

Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: slab: use nr_online_nodes to check for a NUMA platform
Mel Gorman [Tue, 16 Jun 2009 22:32:16 +0000]
page allocator: slab: use nr_online_nodes to check for a NUMA platform

SLAB currently avoids checking a bitmap repeatedly by checking once and
storing a flag.  When the addition of nr_online_nodes as a cheaper version
of num_online_nodes(), this check can be replaced by nr_online_nodes.

(Christoph did a patch that this is lifted almost verbatim from)

Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: use a pre-calculated value instead of num_online_nodes() in fast...
Christoph Lameter [Tue, 16 Jun 2009 22:32:15 +0000]
page allocator: use a pre-calculated value instead of num_online_nodes() in fast paths

num_online_nodes() is called in a number of places but most often by the
page allocator when deciding whether the zonelist needs to be filtered
based on cpusets or the zonelist cache.  This is actually a heavy function
and touches a number of cache lines.

This patch stores the number of online nodes at boot time and updates the
value when nodes get onlined and offlined.  The value is then used in a
number of important paths in place of num_online_nodes().

[rientjes@google.com: do not override definition of node_set_online() with macro]
Signed-off-by: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: get the pageblock migratetype without disabling interrupts
Mel Gorman [Tue, 16 Jun 2009 22:32:14 +0000]
page allocator: get the pageblock migratetype without disabling interrupts

Local interrupts are disabled when freeing pages to the PCP list.  Part of
that free checks what the migratetype of the pageblock the page is in but
it checks this with interrupts disabled and interupts should never be
disabled longer than necessary.  This patch checks the pagetype with
interrupts enabled with the impact that it is possible a page is freed to
the wrong list when a pageblock changes type.  As that block is now
already considered mixed from an anti-fragmentation perspective, it's not
of vital importance.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: update NR_FREE_PAGES only as necessary
Mel Gorman [Tue, 16 Jun 2009 22:32:13 +0000]
page allocator: update NR_FREE_PAGES only as necessary

When pages are being freed to the buddy allocator, the zone NR_FREE_PAGES
counter must be updated.  In the case of bulk per-cpu page freeing, it's
updated once per page.  This retouches cache lines more than necessary.
Update the counters one per per-cpu bulk free.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: use allocation flags as an index to the zone watermark
Mel Gorman [Tue, 16 Jun 2009 22:32:12 +0000]
page allocator: use allocation flags as an index to the zone watermark

ALLOC_WMARK_MIN, ALLOC_WMARK_LOW and ALLOC_WMARK_HIGH determin whether
pages_min, pages_low or pages_high is used as the zone watermark when
allocating the pages.  Two branches in the allocator hotpath determine
which watermark to use.

This patch uses the flags as an array index into a watermark array that is
indexed with WMARK_* defines accessed via helpers.  All call sites that
use zone->pages_* are updated to use the helpers for accessing the values
and the array offsets for setting.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: do not check for compound pages during the page allocator sanity...
Nick Piggin [Tue, 16 Jun 2009 22:32:10 +0000]
page allocator: do not check for compound pages during the page allocator sanity checks

A number of sanity checks are made on each page allocation and free
including that the page count is zero.  page_count() checks for compound
pages and checks the count of the head page if true.  However, in these
paths, we do not care if the page is compound or not as the count of each
tail page should also be zero.

This patch makes two changes to the use of page_count() in the free path.
It converts one check of page_count() to a VM_BUG_ON() as the count should
have been unconditionally checked earlier in the free path.  It also
avoids checking for compound pages.

[mel@csn.ul.ie: Wrote changelog]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Nick Piggin <nickpiggin@yahoo.com.au>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: do not setup zonelist cache when there is only one node
Mel Gorman [Tue, 16 Jun 2009 22:32:09 +0000]
page allocator: do not setup zonelist cache when there is only one node

There is a zonelist cache which is used to track zones that are not in the
allowed cpuset or found to be recently full.  This is to reduce cache
footprint on large machines.  On smaller machines, it just incurs cost for
no gain.  This patch only uses the zonelist cache when there are NUMA
nodes.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: do not disable interrupts in free_page_mlock()
Mel Gorman [Tue, 16 Jun 2009 22:32:08 +0000]
page allocator: do not disable interrupts in free_page_mlock()

free_page_mlock() tests and clears PG_mlocked using locked versions of the
bit operations.  If set, it disables interrupts to update counters and
this happens on every page free even though interrupts are disabled very
shortly afterwards a second time.  This is wasteful.

This patch splits what free_page_mlock() does.  The bit check is still
made.  However, the update of counters is delayed until the interrupts are
disabled and the non-lock version for clearing the bit is used.  One
potential weirdness with this split is that the counters do not get
updated if the bad_page() check is triggered but a system showing bad
pages is getting screwed already.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Acked-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: do not call get_pageblock_migratetype() more than necessary
Mel Gorman [Tue, 16 Jun 2009 22:32:07 +0000]
page allocator: do not call get_pageblock_migratetype() more than necessary

get_pageblock_migratetype() is potentially called twice for every page
free.  Once, when being freed to the pcp lists and once when being freed
back to buddy.  When freeing from the pcp lists, it is known what the
pageblock type was at the time of free so use it rather than rechecking.
In low memory situations under memory pressure, this might skew
anti-fragmentation slightly but the interference is minimal and decisions
that are fragmenting memory are being made anyway.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: inline __rmqueue_fallback()
Mel Gorman [Tue, 16 Jun 2009 22:32:06 +0000]
page allocator: inline __rmqueue_fallback()

__rmqueue_fallback() is in the slow path but has only one call site.
Because there is only one call-site, this function can then be inlined
without causing text bloat.  On an x86-based config, it made no difference
as the savings were padded out by NOP instructions.  Milage varies but
text will either decrease in size or remain static.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: inline buffered_rmqueue()
Mel Gorman [Tue, 16 Jun 2009 22:32:05 +0000]
page allocator: inline buffered_rmqueue()

buffered_rmqueue() is in the fast path so inline it.  Because it only has
one call site, this function can then be inlined without causing text
bloat.  On an x86-based config, it made no difference as the savings were
padded out by NOP instructions.  Milage varies but text will either
decrease in size or remain static.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: inline __rmqueue_smallest()
Mel Gorman [Tue, 16 Jun 2009 22:32:04 +0000]
page allocator: inline __rmqueue_smallest()

Inline __rmqueue_smallest by altering flow very slightly so that there is
only one call site.  Because there is only one call-site, this function
can then be inlined without causing text bloat.  On an x86-based config,
this patch reduces text by 16 bytes.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: remove a branch by assuming __GFP_HIGH == ALLOC_HIGH
Mel Gorman [Tue, 16 Jun 2009 22:32:02 +0000]
page allocator: remove a branch by assuming __GFP_HIGH == ALLOC_HIGH

Allocations that specify __GFP_HIGH get the ALLOC_HIGH flag.  If these
flags are equal to each other, we can eliminate a branch.

[akpm@linux-foundation.org: Suggested the hack]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: calculate the alloc_flags for allocation only once
Peter Zijlstra [Tue, 16 Jun 2009 22:32:02 +0000]
page allocator: calculate the alloc_flags for allocation only once

Factor out the mapping between GFP and alloc_flags only once.  Once
factored out, it only needs to be calculated once but some care must be
taken.

[neilb@suse.de says]
As the test:

-       if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
-                       && !in_interrupt()) {
-               if (!(gfp_mask & __GFP_NOMEMALLOC)) {

has been replaced with a slightly weaker one:

+       if (alloc_flags & ALLOC_NO_WATERMARKS) {

Without care, this would allow recursion into the allocator via direct
reclaim.  This patch ensures we do not recurse when PF_MEMALLOC is set but
TF_MEMDIE callers are now allowed to directly reclaim where they would
have been prevented in the past.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: Neil Brown <neilb@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: calculate the migratetype for allocation only once
Mel Gorman [Tue, 16 Jun 2009 22:32:00 +0000]
page allocator: calculate the migratetype for allocation only once

GFP mask is converted into a migratetype when deciding which pagelist to
take a page from.  However, it is happening multiple times per allocation,
at least once per zone traversed.  Calculate it once.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: calculate the preferred zone for allocation only once
Mel Gorman [Tue, 16 Jun 2009 22:31:59 +0000]
page allocator: calculate the preferred zone for allocation only once

get_page_from_freelist() can be called multiple times for an allocation.
Part of this calculates the preferred_zone which is the first usable zone
in the zonelist but the zone depends on the GFP flags specified at the
beginning of the allocation call.  This patch calculates preferred_zone
once.  It's safe to do this because if preferred_zone is NULL at the start
of the call, no amount of direct reclaim or other actions will change the
fact the allocation will fail.

[akpm@linux-foundation.org: remove (void) casts]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: move check for disabled anti-fragmentation out of fastpath
Mel Gorman [Tue, 16 Jun 2009 22:31:58 +0000]
page allocator: move check for disabled anti-fragmentation out of fastpath

On low-memory systems, anti-fragmentation gets disabled as there is
nothing it can do and it would just incur overhead shuffling pages between
lists constantly.  Currently the check is made in the free page fast path
for every page.  This patch moves it to a slow path.  On machines with low
memory, there will be small amount of additional overhead as pages get
shuffled between lists but it should quickly settle.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: break up the allocator entry point into fast and slow paths
Mel Gorman [Tue, 16 Jun 2009 22:31:57 +0000]
page allocator: break up the allocator entry point into fast and slow paths

The core of the page allocator is one giant function which allocates
memory on the stack and makes calculations that may not be needed for
every allocation.  This patch breaks up the allocator path into fast and
slow paths for clarity.  Note the slow paths are still inlined but the
entry is marked unlikely.  If they were not inlined, it actally increases
text size to generate the as there is only one call site.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: check only once if the zonelist is suitable for the allocation
Mel Gorman [Tue, 16 Jun 2009 22:31:56 +0000]
page allocator: check only once if the zonelist is suitable for the allocation

It is possible with __GFP_THISNODE that no zones are suitable.  This patch
makes sure the check is only made once.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: do not check NUMA node ID when the caller knows the node is valid
Mel Gorman [Tue, 16 Jun 2009 22:31:54 +0000]
page allocator: do not check NUMA node ID when the caller knows the node is valid

Callers of alloc_pages_node() can optionally specify -1 as a node to mean
"allocate from the current node".  However, a number of the callers in
fast paths know for a fact their node is valid.  To avoid a comparison and
branch, this patch adds alloc_pages_exact_node() that only checks the nid
with VM_BUG_ON().  Callers that know their node is valid are then
converted.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Acked-by: Paul Mundt <lethal@linux-sh.org> [for the SLOB NUMA bits]
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: do not sanity check order in the fast path
Mel Gorman [Tue, 16 Jun 2009 22:31:53 +0000]
page allocator: do not sanity check order in the fast path

No user of the allocator API should be passing in an order >= MAX_ORDER
but we check for it on each and every allocation.  Delete this check and
make it a VM_BUG_ON check further down the call path.

[akpm@linux-foundation.org: s/VM_BUG_ON/WARN_ON_ONCE/]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agopage allocator: replace __alloc_pages_internal() with __alloc_pages_nodemask()
Mel Gorman [Tue, 16 Jun 2009 22:31:52 +0000]
page allocator: replace __alloc_pages_internal() with __alloc_pages_nodemask()

The start of a large patch series to clean up and optimise the page
allocator.

The performance improvements are in a wide range depending on the exact
machine but the results I've seen so fair are approximately;

kernbench: 0 to  0.12% (elapsed time)
0.49% to  3.20% (sys time)
aim9: -4% to 30% (for page_test and brk_test)
tbench: -1% to  4%
hackbench: -2.5% to  3.45% (mostly within the noise though)
netperf-udp -1.34%  to  4.06% (varies between machines a bit)
netperf-tcp -0.44%  to  5.22% (varies between machines a bit)

I haven't sysbench figures at hand, but previously they were within the
-0.5% to 2% range.

On netperf, the client and server were bound to opposite number CPUs to
maximise the problems with cache line bouncing of the struct pages so I
expect different people to report different results for netperf depending
on their exact machine and how they ran the test (different machines, same
cpus client/server, shared cache but two threads client/server, different
socket client/server etc).

I also measured the vmlinux sizes for a single x86-based config with
CONFIG_DEBUG_INFO enabled but not CONFIG_DEBUG_VM.  The core of the
.config is based on the Debian Lenny kernel config so I expect it to be
reasonably typical.

This patch:

__alloc_pages_internal is the core page allocator function but essentially
it is an alias of __alloc_pages_nodemask.  Naming a publicly available and
exported function "internal" is also a big ugly.  This patch renames
__alloc_pages_internal() to __alloc_pages_nodemask() and deletes the old
nodemask function.

Warning - This patch renames an exported symbol.  No kernel driver is
affected by external drivers calling __alloc_pages_internal() should
change the call to __alloc_pages_nodemask() without any alteration of
parameters.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: alloc_large_system_hash check order
Hugh Dickins [Tue, 16 Jun 2009 22:31:50 +0000]
mm: alloc_large_system_hash check order

On an x86_64 with 4GB ram, tcp_init()'s call to alloc_large_system_hash(),
to allocate tcp_hashinfo.ehash, is now triggering an mmotm WARN_ON_ONCE on
order >= MAX_ORDER - it's hoping for order 11.  alloc_large_system_hash()
had better make its own check on the order.

Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: David Miller <davem@davemloft.net>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Eric Dumazet <dada1@cosmosbay.com>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agocpuset,mm: update tasks' mems_allowed in time
Miao Xie [Tue, 16 Jun 2009 22:31:49 +0000]
cpuset,mm: update tasks' mems_allowed in time

Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.

In order to update tasks' mems_allowed in time, we must modify the code of
memory policy.  Because the memory policy is applied in the process's
context originally.  After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.

But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression.  But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set.  In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.

[lee.schermerhorn@hp.com:
  The rework of mpol_new() to extract the adjusting of the node mask to
  apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
  with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
  allocation.  Fix this by adding the check for MPOL_PREFERRED and empty
  node mask to mpol_new_mpolicy().

  Remove the now unneeded 'nodes = NULL' from mpol_new().

  Note that mpol_new_mempolicy() is always called with a non-NULL
  'nodes' parameter now that it has been removed from mpol_new().
  Therefore, we don't need to test nodes for NULL before testing it for
  'empty'.  However, just to be extra paranoid, add a VM_BUG_ON() to
  verify this assumption.]
[lee.schermerhorn@hp.com:

  I don't think the function name 'mpol_new_mempolicy' is descriptive
  enough to differentiate it from mpol_new().

  This function applies cpuset set context, usually constraining nodes
  to those allowed by the cpuset.  However, when the 'RELATIVE_NODES flag
  is set, it also translates the nodes.  So I settled on
  'mpol_set_nodemask()', because the comment block for mpol_new() mentions
  that we need to call this function to "set nodes".

  Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agocpusets: update tasks' page/slab spread flags in time
Miao Xie [Tue, 16 Jun 2009 22:31:47 +0000]
cpusets: update tasks' page/slab spread flags in time

Fix the bug that the kernel didn't spread page cache/slab object evenly
over all the allowed nodes when spread flags were set by updating tasks'
page/slab spread flags in time.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agocpusets: restructure the function cpuset_update_task_memory_state()
Miao Xie [Tue, 16 Jun 2009 22:31:46 +0000]
cpusets: restructure the function cpuset_update_task_memory_state()

The kernel still allocates the page caches on old node after modifying its
cpuset's mems when 'memory_spread_page' was set, or it didn't spread the
page cache evenly over all the nodes that faulting task is allowed to usr
after memory_spread_page was set.  it is caused by the old mem_allowed and
flags of the task, the current kernel doesn't updates them unless some
function invokes cpuset_update_task_memory_state(), it is too late
sometimes.We must update the mem_allowed and the flags of the tasks in
time.

Slab has the same problem.

The following patches fix this bug by updating tasks' mem_allowed and
spread flag after its cpuset's mems or spread flag is changed.

This patch:

Extract a function from cpuset_update_task_memory_state().  It will be
used later for update tasks' page/slab spread flags after its cpuset's
flag is set

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm/page-writeback.c: dirty limit type should be unsigned long
H Hartley Sweeten [Tue, 16 Jun 2009 22:31:44 +0000]
mm/page-writeback.c: dirty limit type should be unsigned long

get_dirty_limits() calls clip_bdi_dirty_limit() and task_dirty_limit()
with variable pbdi_dirty as one of the arguments.  This variable is an
unsigned long * but both functions expect it to be a long *.  This causes
the following sparse warnings:

  warning: incorrect type in argument 3 (different signedness)
     expected long *pbdi_dirty
     got unsigned long *pbdi_dirty
  warning: incorrect type in argument 2 (different signedness)
     expected long *pdirty
     got unsigned long *pbdi_dirty

Fix the warnings by changing the long * to unsigned long * in both
functions.

Signed-off-by: H Hartley Sweeten <hsweeten@visionengravers.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agovmscan: low order lumpy reclaim also should use PAGEOUT_IO_SYNC
KOSAKI Motohiro [Tue, 16 Jun 2009 22:31:40 +0000]
vmscan: low order lumpy reclaim also should use PAGEOUT_IO_SYNC

Commit 33c120ed2843090e2bd316de1588b8bf8b96cbde ("more aggressively use
lumpy reclaim") increased how aggressive lumpy reclaim was by isolating
both active and inactive pages for asynchronous lumpy reclaim on
costly-high-order pages and for cheap-high-order when memory pressure is
high.  However, if the system is under heavy pressure and there are dirty
pages, asynchronous IO may not be sufficient to reclaim a suitable page in
time.

This patch causes the caller to enter synchronous lumpy reclaim for
costly-high-order pages and for cheap-high-order pages when under memory
pressure.

Minchan.kim@gmail.com said:

Andy added synchronous lumpy reclaim with
c661b078fd62abe06fd11fab4ac5e4eeafe26b6d.  At that time, lumpy reclaim is
not agressive.  His intension is just for high-order users.(above
PAGE_ALLOC_COSTLY_ORDER).

After some time, Rik added aggressive lumpy reclaim with
33c120ed2843090e2bd316de1588b8bf8b96cbde.  His intention was to do lumpy
reclaim when high-order users and trouble getting a small set of
contiguous pages.

So we also have to add synchronous pageout for small set of contiguous
pages.

Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Andy Whitcroft <apw@shadowen.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Minchan Kim <Minchan.kim@gmail.com>
Reviewed-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agomm: clean up get_user_pages_fast() documentation
Nick Piggin [Tue, 16 Jun 2009 22:31:39 +0000]
mm: clean up get_user_pages_fast() documentation

Move more documentation for get_user_pages_fast into the new kerneldoc comment.
Add some comments for get_user_pages as well.

Also, move get_user_pages_fast declaration up to get_user_pages. It wasn't
there initially because it was once a static inline function.

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Andy Grover <andy.grover@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoreadahead: enforce full sync mmap readahead size
Wu Fengguang [Tue, 16 Jun 2009 22:31:38 +0000]
readahead: enforce full sync mmap readahead size

Now that we do readahead for sequential mmap reads, here is a simple
evaluation of the impacts, and one further optimization.

It's an NFS-root debian desktop system, readahead size = 60 pages.
The numbers are grabbed after a fresh boot into console.

approach        pgmajfault      RA miss ratio   mmap IO count   avg IO size(pages)
   A            383             31.6%           383             11
   B            225             32.4%           390             11
   C            224             32.6%           307             13

case A: mmap sync/async readahead disabled
case B: mmap sync/async readahead enabled, with enforced full async readahead size
case C: mmap sync/async readahead enabled, with enforced full sync/async readahead size
or:
A = vanilla 2.6.30-rc1
B = A plus mmap readahead
C = B plus this patch

The numbers show that
- there are good possibilities for random mmap reads to trigger readahead
- 'pgmajfault' is reduced by 1/3, due to the _async_ nature of readahead
- case C can further reduce IO count by 1/4
- readahead miss ratios are not quite affected

The theory is
- readahead is _good_ for clustered random reads, and can perform
  _better_ than readaround because they could be _async_.
- async readahead size is guaranteed to be larger than readaround
  size, and they are _async_, hence will mostly behave better
However for B
- sync readahead size could be smaller than readaround size, hence may
  make things worse by produce more smaller IOs
which will be fixed by this patch.

Final conclusion:
- mmap readahead reduced major faults by 1/3 and no obvious overheads;
- mmap io can be further reduced by 1/4 with this patch.

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoreadahead: remove redundant test in shrink_readahead_size_eio()
Wu Fengguang [Tue, 16 Jun 2009 22:31:36 +0000]
readahead: remove redundant test in shrink_readahead_size_eio()

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoreadahead: introduce context readahead algorithm
Wu Fengguang [Tue, 16 Jun 2009 22:31:36 +0000]
readahead: introduce context readahead algorithm

Introduce page cache context based readahead algorithm.
This is to better support concurrent read streams in general.

RATIONALE
---------
The current readahead algorithm detects interleaved reads in a _passive_ way.
Given a sequence of interleaved streams 1,1001,2,1002,3,4,1003,5,1004,1005,6,...
By checking for (offset == prev_offset + 1), it will discover the sequentialness
between 3,4 and between 1004,1005, and start doing sequential readahead for the
individual streams since page 4 and page 1005.

The context readahead algorithm guarantees to discover the sequentialness no
matter how the streams are interleaved. For the above example, it will start
sequential readahead since page 2 and 1002.

The trick is to poke for page @offset-1 in the page cache when it has no other
clues on the sequentialness of request @offset: if the current requenst belongs
to a sequential stream, that stream must have accessed page @offset-1 recently,
and the page will still be cached now. So if page @offset-1 is there, we can
take request @offset as a sequential access.

BENEFICIARIES
-------------
- strictly interleaved reads  i.e. 1,1001,2,1002,3,1003,...
  the current readahead will take them as silly random reads;
  the context readahead will take them as two sequential streams.

- cooperative IO processes   i.e. NFS and SCST
  They create a thread pool, farming off (sequential) IO requests to different
  threads which will be performing interleaved IO.

  It was not easy(or possible) to reliably tell from file->f_ra all those
  cooperative processes working on the same sequential stream, since they will
  have different file->f_ra instances. And NFSD's file->f_ra is particularly
  unusable, since their file objects are dynamically created for each request.
  The nfsd does have code trying to restore the f_ra bits, but not satisfactory.

  The new scheme is to detect the sequential pattern via looking up the page
  cache, which provides one single and consistent view of the pages recently
  accessed. That makes sequential detection for cooperative processes possible.

USER REPORT
-----------
Vladislav recommends the addition of context readahead as a result of his SCST
benchmarks. It leads to 6%~40% performance gains in various cases and achieves
equal performance in others.                http://lkml.org/lkml/2009/3/19/239

OVERHEADS
---------
In theory, it introduces one extra page cache lookup per random read.  However
the below benchmark shows context readahead to be slightly faster, wondering..

Randomly reading 200MB amount of data on a sparse file, repeat 20 times for
each block size. The average throughputs are:

                        original ra context ra gain
 4K random reads:  65.561MB/s  65.648MB/s +0.1%
16K random reads: 124.767MB/s 124.951MB/s +0.1%
64K random reads:  162.123MB/s 162.278MB/s +0.1%

Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Tested-by: Vladislav Bolkhovitin <vst@vlnb.net>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoreadahead: move the random read case to bottom
Wu Fengguang [Tue, 16 Jun 2009 22:31:33 +0000]
readahead: move the random read case to bottom

Split all readahead cases, and move the random one to bottom.

No behavior changes.

This is to prepare for the introduction of context readahead, and make it
easy for inserting accounting/tracing points for each case.

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: Vladislav Bolkhovitin <vst@vlnb.net>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

10 years agoradix-tree: add radix_tree_prev_hole()
Wu Fengguang [Tue, 16 Jun 2009 22:31:32 +0000]
radix-tree: add radix_tree_prev_hole()

The counterpart of radix_tree_next_hole(). To be used by context readahead.

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: Vladislav Bolkhovitin <vst@vlnb.net>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>