mm: thp: fix pmd_present for split_huge_page and PROT_NONE with THP

In many places !pmd_present has been converted to pmd_none.  For pmds
that's equivalent and pmd_none is quicker so using pmd_none is better.

However (unless we delete pmd_present) we should provide an accurate
pmd_present too.  This will avoid the risk of code thinking the pmd is non
present because it's under __split_huge_page_map, see the pmd_mknotpresent
there and the comment above it.

If the page has been mprotected as PROT_NONE, it would also lead to a
pmd_present false negative in the same way as the race with
split_huge_page.

Because the PSE bit stays on at all times (both during split_huge_page and
when the _PAGE_PROTNONE bit get set), we could only check for the PSE bit,
but checking the PROTNONE bit too is still good to remember pmd_present
must always keep PROT_NONE into account.

This explains a not reproducible BUG_ON that was seldom reported on the
lists.

The same issue is in pmd_large, it would go wrong with both PROT_NONE and
if it races with split_huge_page.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory.txt: remove stray information

Andi removed some outedated documentation from Documentation/memory.txt
back in 2009 by commit 3b2b9a875ddc ("Documentation/memory.txt: remove
some very outdated recommendations"), but the resulting document is not
in a nice shape either.

It seems to me like we are not losing anything by completely removing the
file now.

Signed-off-by: Jiri Kosina <jkosina@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm, numa: reclaim from all nodes within reclaim distance

RECLAIM_DISTANCE represents the distance between nodes at which it is
deemed too costly to allocate from; it's preferred to try to reclaim from
a local zone before falling back to allocating on a remote node with such
a distance.

To do this, zone_reclaim_mode is set if the distance between any two
nodes on the system is greather than this distance.  This, however, ends
up causing the page allocator to reclaim from every zone regardless of
its affinity.

What we really want is to reclaim only from zones that are closer than
RECLAIM_DISTANCE.  This patch adds a nodemask to each node that
represents the set of nodes that are within this distance.  During the
zone iteration, if the bit for a zone's node is set for the local node,
then reclaim is attempted; otherwise, the zone is skipped.

[akpm@linux-foundation.org: fix CONFIG_NUMA=n build]
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: remove free_page_mlock

We should not be seeing non-0 unevictable_pgs_mlockfreed any longer.  So
remove free_page_mlock() from the page freeing paths: __PG_MLOCKED is
already in PAGE_FLAGS_CHECK_AT_FREE, so free_pages_check() will now be
checking it, reporting "BUG: Bad page state" if it's ever found set.
Comment UNEVICTABLE_MLOCKFREED and unevictable_pgs_mlockfreed always 0.

Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Cc: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: use clear_page_mlock() in page_remove_rmap()

We had thought that pages could no longer get freed while still marked as
mlocked; but Johannes Weiner posted this program to demonstrate that
truncating an mlocked private file mapping containing COWed pages is still
mishandled:

#include <sys/types.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <stdio.h>

int main(void)
{
char *map;
int fd;

system("grep mlockfreed /proc/vmstat");
fd = open("chigurh", O_CREAT|O_EXCL|O_RDWR);
unlink("chigurh");
ftruncate(fd, 4096);
map = mmap(NULL, 4096, PROT_WRITE, MAP_PRIVATE, fd, 0);
map[0] = 11;
mlock(map, sizeof(fd));
ftruncate(fd, 0);
close(fd);
munlock(map, sizeof(fd));
munmap(map, 4096);
system("grep mlockfreed /proc/vmstat");
return 0;
}

The anon COWed pages are not caught by truncation's clear_page_mlock() of
the pagecache pages; but unmap_mapping_range() unmaps them, so we ought to
look out for them there in page_remove_rmap().  Indeed, why should
truncation or invalidation be doing the clear_page_mlock() when removing
from pagecache?  mlock is a property of mapping in userspace, not a
property of pagecache: an mlocked unmapped page is nonsensical.

Reported-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Rik van Riel <riel@redhat.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ying Han <yinghan@google.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>

mm: remove vma arg from page_evictable

page_evictable(page, vma) is an irritant: almost all its callers pass
NULL for vma.  Remove the vma arg and use mlocked_vma_newpage(vma, page)
explicitly in the couple of places it's needed.  But in those places we
don't even need page_evictable() itself!  They're dealing with a freshly
allocated anonymous page, which has no "mapping" and cannot be mlocked yet.

Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Cc: Rik van Riel <riel@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: fix invalidate_complete_page2() lock ordering

In fuzzing with trinity, lockdep protested "possible irq lock inversion
dependency detected" when isolate_lru_page() reenabled interrupts while
still holding the supposedly irq-safe tree_lock:

invalidate_inode_pages2
  invalidate_complete_page2
    spin_lock_irq(&mapping->tree_lock)
    clear_page_mlock
      isolate_lru_page
        spin_unlock_irq(&zone->lru_lock)

isolate_lru_page() is correct to enable interrupts unconditionally:
invalidate_complete_page2() is incorrect to call clear_page_mlock() while
holding tree_lock, which is supposed to nest inside lru_lock.

Both truncate_complete_page() and invalidate_complete_page() call
clear_page_mlock() before taking tree_lock to remove page from radix_tree.
 I guess invalidate_complete_page2() preferred to test PageDirty (again)
under tree_lock before committing to the munlock; but since the page has
already been unmapped, its state is already somewhat inconsistent, and no
worse if clear_page_mlock() moved up.

Reported-by: Sasha Levin <levinsasha928@gmail.com>
Deciphered-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Cc: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ying Han <yinghan@google.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memcg: move mem_cgroup_is_root upwards

kmem code uses this function and it is better to not use forward
declarations for static inline functions as some (older) compilers don't
like it:

gcc version 4.3.4 [gcc-4_3-branch revision 152973] (SUSE Linux)

  mm/memcontrol.c:421: warning: `mem_cgroup_is_root' declared inline after being called
  mm/memcontrol.c:421: warning: previous declaration of `mem_cgroup_is_root' was here

Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: Glauber Costa <glommer@parallels.com>
Cc: Sachin Kamat <sachin.kamat@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memcg: cleanup kmem tcp ifdefs

TCP kmem accounting is currently guarded by CONFIG_MEMCG_KMEM ifdefs but
the code is not used if !CONFIG_INET so we should rather test for both.
The same applies to net/sock.h, net/ip.h and net/tcp_memcontrol.h but
let's keep those outside of any ifdefs because it is considered safer wrt.
 future maintainability.

Tested with
- CONFIG_INET && CONFIG_MEMCG_KMEM
- !CONFIG_INET && CONFIG_MEMCG_KMEM
- CONFIG_INET && !CONFIG_MEMCG_KMEM
- !CONFIG_INET && !CONFIG_MEMCG_KMEM

Signed-off-by: Sachin Kamat <sachin.kamat@linaro.org>
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: Glauber Costa <glommer@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memcg: trivial fixes for Documentation/cgroups/memory.txt

While reading through Documentation/cgroups/memory.txt, I found a number
of minor wordos and typos.  The patch below is a conservative handling of
some of these: it provides just a number of "obviously correct" fixes to
the English that improve the readability of the document somewhat.
Obviously some more significant fixes need to be made to the document, but
some of those may not be in the "obvious correct" category.

Signed-off-by: Michael Kerrisk <mtk.manpages@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: fix-up zone present pages

I think zone->present_pages indicates pages that buddy system can management,
it should be:

zone->present_pages = spanned pages - absent pages - bootmem pages,

but is now:
zone->present_pages = spanned pages - absent pages - memmap pages.

spanned pages: total size, including holes.
absent pages: holes.
bootmem pages: pages used in system boot, managed by bootmem allocator.
memmap pages: pages used by page structs.

This may cause zone->present_pages less than it should be.  For example,
numa node 1 has ZONE_NORMAL and ZONE_MOVABLE, it's memmap and other
bootmem will be allocated from ZONE_MOVABLE, so ZONE_NORMAL's
present_pages should be spanned pages - absent pages, but now it also
minus memmap pages(free_area_init_core), which are actually allocated from
ZONE_MOVABLE.  When offlining all memory of a zone, this will cause
zone->present_pages less than 0, because present_pages is unsigned long
type, it is actually a very large integer, it indirectly caused
zone->watermark[WMARK_MIN] becomes a large
integer(setup_per_zone_wmarks()), than cause totalreserve_pages become a
large integer(calculate_totalreserve_pages()), and finally cause memory
allocating failure when fork process(__vm_enough_memory()).

[root@localhost ~]# dmesg
-bash: fork: Cannot allocate memory

I think the bug described in

  http://marc.info/?l=linux-mm&m=134502182714186&w=2

is also caused by wrong zone present pages.

This patch intends to fix-up zone->present_pages when memory are freed to
buddy system on x86_64 and IA64 platforms.

Signed-off-by: Jianguo Wu <wujianguo@huawei.com>
Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Reported-by: Petr Tesarik <ptesarik@suse.cz>
Tested-by: Petr Tesarik <ptesarik@suse.cz>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Yinghai Lu <yinghai@kernel.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.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>

mm: enable CONFIG_COMPACTION by default

Now that lumpy reclaim has been removed, compaction is the only way left
to free up contiguous memory areas.  It is time to just enable
CONFIG_COMPACTION by default.

Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: Rafael Aquini <aquini@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: thp: fix the update_mmu_cache() last argument passing in mm/huge_memory.c

The update_mmu_cache() takes a pointer (to pte_t by default) as the last
argument but the huge_memory.c passes a pmd_t value.  The patch changes
the argument to the pmd_t * pointer.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Steve Capper <steve.capper@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: thp: fix the pmd_clear() arguments in pmdp_get_and_clear()

The CONFIG_TRANSPARENT_HUGEPAGE implementation of pmdp_get_and_clear()
calls pmd_clear() with 3 arguments instead of 1.

This happens only for !__HAVE_ARCH_PMDP_GET_AND_CLEAR which doesn't seem
to happen because x86 defines this and it uses pmd_update.

[mhocko@suse.cz: changelog addition]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Steve Capper <steve.capper@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: khugepaged_prealloc_page() forgot to reset the page alloc indicator

If NUMA is enabled, the indicator is not reset if the previous page
request failed, ausing us to trigger the BUG_ON() in
khugepaged_alloc_page().

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: don't replace lowmem pages with highmem

The changelog for commit 6a6dccba2fdc ("mm: cma: don't replace lowmem
pages with highmem") mentioned that lowmem pages can be replaced by
highmem pages during CMA migration.  6a6dccba2fdc fixed that issue.

Quote from that changelog:

:   The filesystem layer expects pages in the block device's mapping to not
:   be in highmem (the mapping's gfp mask is set in bdget()), but CMA can
:   currently replace lowmem pages with highmem pages, leading to crashes in
:   filesystem code such as the one below:
:
:     Unable to handle kernel NULL pointer dereference at virtual address 00000400
:     pgd = c0c98000
:     [00000400] *pgd=00c91831, *pte=00000000, *ppte=00000000
:     Internal error: Oops: 817 [#1] PREEMPT SMP ARM
:     CPU: 0    Not tainted  (3.5.0-rc5+ #80)
:     PC is at __memzero+0x24/0x80
:     ...
:     Process fsstress (pid: 323, stack limit = 0xc0cbc2f0)
:     Backtrace:
:     [<c010e3f0>] (ext4_getblk+0x0/0x180) from [<c010e58c>] (ext4_bread+0x1c/0x98)
:     [<c010e570>] (ext4_bread+0x0/0x98) from [<c0117944>] (ext4_mkdir+0x160/0x3bc)
:      r4:c15337f0
:     [<c01177e4>] (ext4_mkdir+0x0/0x3bc) from [<c00c29e0>] (vfs_mkdir+0x8c/0x98)
:     [<c00c2954>] (vfs_mkdir+0x0/0x98) from [<c00c2a60>] (sys_mkdirat+0x74/0xac)
:      r6:00000000 r5:c152eb40 r4:000001ff r3:c14b43f0
:     [<c00c29ec>] (sys_mkdirat+0x0/0xac) from [<c00c2ab8>] (sys_mkdir+0x20/0x24)
:      r6:beccdcf0 r5:00074000 r4:beccdbbc
:     [<c00c2a98>] (sys_mkdir+0x0/0x24) from [<c000e3c0>] (ret_fast_syscall+0x0/0x30)

Memory-hotplug has same problem as CMA has so the same fix can be applied
to memory-hotplug as well.

Fix it by reusing.

Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/page_alloc: refactor out __alloc_contig_migrate_alloc()

__alloc_contig_migrate_alloc() can be used by memory-hotplug so refactor
it out (move + rename as a common name) into page_isolation.c.

[akpm@linux-foundation.org: checkpatch fixes]
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/hugetlb.c: remove duplicate inclusion of header file

Signed-off-by: Sachin Kamat <sachin.kamat@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: compaction: clear PG_migrate_skip based on compaction and reclaim activity

Compaction caches if a pageblock was scanned and no pages were isolated so
that the pageblocks can be skipped in the future to reduce scanning.  This
information is not cleared by the page allocator based on activity due to
the impact it would have to the page allocator fast paths.  Hence there is
a requirement that something clear the cache or pageblocks will be skipped
forever.  Currently the cache is cleared if there were a number of recent
allocation failures and it has not been cleared within the last 5 seconds.
Time-based decisions like this are terrible as they have no relationship
to VM activity and is basically a big hammer.

Unfortunately, accurate heuristics would add cost to some hot paths so
this patch implements a rough heuristic.  There are two cases where the
cache is cleared.

1. If a !kswapd process completes a compaction cycle (migrate and free
   scanner meet), the zone is marked compact_blockskip_flush. When kswapd
   goes to sleep, it will clear the cache. This is expected to be the
   common case where the cache is cleared. It does not really matter if
   kswapd happens to be asleep or going to sleep when the flag is set as
   it will be woken on the next allocation request.

2. If there have been multiple failures recently and compaction just
   finished being deferred then a process will clear the cache and start a
   full scan.  This situation happens if there are multiple high-order
   allocation requests under heavy memory pressure.

The clearing of the PG_migrate_skip bits and other scans is inherently
racy but the race is harmless.  For allocations that can fail such as THP,
they will simply fail.  For requests that cannot fail, they will retry the
allocation.  Tests indicated that scanning rates were roughly similar to
when the time-based heuristic was used and the allocation success rates
were similar.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Cc: Richard Davies <richard@arachsys.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Avi Kivity <avi@redhat.com>
Cc: Rafael Aquini <aquini@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: compaction: Restart compaction from near where it left off

This is almost entirely based on Rik's previous patches and discussions
with him about how this might be implemented.

Order > 0 compaction stops when enough free pages of the correct page
order have been coalesced.  When doing subsequent higher order
allocations, it is possible for compaction to be invoked many times.

However, the compaction code always starts out looking for things to
compact at the start of the zone, and for free pages to compact things to
at the end of the zone.

This can cause quadratic behaviour, with isolate_freepages starting at the
end of the zone each time, even though previous invocations of the
compaction code already filled up all free memory on that end of the zone.
 This can cause isolate_freepages to take enormous amounts of CPU with
certain workloads on larger memory systems.

This patch caches where the migration and free scanner should start from
on subsequent compaction invocations using the pageblock-skip information.
 When compaction starts it begins from the cached restart points and will
update the cached restart points until a page is isolated or a pageblock
is skipped that would have been scanned by synchronous compaction.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Richard Davies <richard@arachsys.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Avi Kivity <avi@redhat.com>
Acked-by: Rafael Aquini <aquini@redhat.com>
Cc: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: compaction: cache if a pageblock was scanned and no pages were isolated

When compaction was implemented it was known that scanning could
potentially be excessive.  The ideal was that a counter be maintained for
each pageblock but maintaining this information would incur a severe
penalty due to a shared writable cache line.  It has reached the point
where the scanning costs are a serious problem, particularly on
long-lived systems where a large process starts and allocates a large
number of THPs at the same time.

Instead of using a shared counter, this patch adds another bit to the
pageblock flags called PG_migrate_skip.  If a pageblock is scanned by
either migrate or free scanner and 0 pages were isolated, the pageblock is
marked to be skipped in the future.  When scanning, this bit is checked
before any scanning takes place and the block skipped if set.

The main difficulty with a patch like this is "when to ignore the cached
information?" If it's ignored too often, the scanning rates will still be
excessive.  If the information is too stale then allocations will fail
that might have otherwise succeeded.  In this patch

o CMA always ignores the information
o If the migrate and free scanner meet then the cached information will
  be discarded if it's at least 5 seconds since the last time the cache
  was discarded
o If there are a large number of allocation failures, discard the cache.

The time-based heuristic is very clumsy but there are few choices for a
better event.  Depending solely on multiple allocation failures still
allows excessive scanning when THP allocations are failing in quick
succession due to memory pressure.  Waiting until memory pressure is
relieved would cause compaction to continually fail instead of using
reclaim/compaction to try allocate the page.  The time-based mechanism is
clumsy but a better option is not obvious.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Richard Davies <richard@arachsys.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Avi Kivity <avi@redhat.com>
Acked-by: Rafael Aquini <aquini@redhat.com>
Cc: Fengguang Wu <fengguang.wu@intel.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Cc: Kyungmin Park <kyungmin.park@samsung.com>
Cc: Mark Brown <broonie@opensource.wolfsonmicro.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

revert "mm: have order > 0 compaction start off where it left"

This reverts commit 7db8889ab05b ("mm: have order > 0 compaction start
off where it left") and commit de74f1cc ("mm: have order > 0 compaction
start near a pageblock with free pages").  These patches were a good
idea and tests confirmed that they massively reduced the amount of
scanning but the implementation is complex and tricky to understand.  A
later patch will cache what pageblocks should be skipped and
reimplements the concept of compact_cached_free_pfn on top for both
migration and free scanners.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Richard Davies <richard@arachsys.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Avi Kivity <avi@redhat.com>
Acked-by: Rafael Aquini <aquini@redhat.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: compaction: acquire the zone->lock as late as possible

Compaction's free scanner acquires the zone->lock when checking for
PageBuddy pages and isolating them.  It does this even if there are no
PageBuddy pages in the range.

This patch defers acquiring the zone lock for as long as possible.  In the
event there are no free pages in the pageblock then the lock will not be
acquired at all which reduces contention on zone->lock.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Richard Davies <richard@arachsys.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Avi Kivity <avi@redhat.com>
Acked-by: Rafael Aquini <aquini@redhat.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Tested-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: compaction: acquire the zone->lru_lock as late as possible

Richard Davies and Shaohua Li have both reported lock contention problems
in compaction on the zone and LRU locks as well as significant amounts of
time being spent in compaction.  This series aims to reduce lock
contention and scanning rates to reduce that CPU usage.  Richard reported
at https://lkml.org/lkml/2012/9/21/91 that this series made a big
different to a problem he reported in August:

   http://marc.info/?l=kvm&m=134511507015614&w=2

Patch 1 defers acquiring the zone->lru_lock as long as possible.

Patch 2 defers acquiring the zone->lock as lock as possible.

Patch 3 reverts Rik's "skip-free" patches as the core concept gets
reimplemented later and the remaining patches are easier to
understand if this is reverted first.

Patch 4 adds a pageblock-skip bit to the pageblock flags to cache what
pageblocks should be skipped by the migrate and free scanners.
This drastically reduces the amount of scanning compaction has
to do.

Patch 5 reimplements something similar to Rik's idea except it uses the
pageblock-skip information to decide where the scanners should
restart from and does not need to wrap around.

I tested this on 3.6-rc6 + linux-next/akpm. Kernels tested were

akpm-20120920 3.6-rc6 + linux-next/akpm as of Septeber 20th, 2012
lesslock Patches 1-6
revert Patches 1-7
cachefail Patches 1-8
skipuseless Patches 1-9

Stress high-order allocation tests looked ok.  Success rates are more or
less the same with the full series applied but there is an expectation
that there is less opportunity to race with other allocation requests if
there is less scanning.  The time to complete the tests did not vary that
much and are uninteresting as were the vmstat statistics so I will not
present them here.

Using ftrace I recorded how much scanning was done by compaction and got this

                            3.6.0-rc6     3.6.0-rc6   3.6.0-rc6  3.6.0-rc6 3.6.0-rc6
                            akpm-20120920 lockless  revert-v2r2  cachefail skipuseless

Total   free    scanned         360753976  515414028  565479007   17103281   18916589
Total   free    isolated          2852429    3597369    4048601     670493     727840
Total   free    efficiency        0.0079%    0.0070%    0.0072%    0.0392%    0.0385%
Total   migrate scanned         247728664  822729112 1004645830   17946827   14118903
Total   migrate isolated          2555324    3245937    3437501     616359     658616
Total   migrate efficiency        0.0103%    0.0039%    0.0034%    0.0343%    0.0466%

The efficiency is worthless because of the nature of the test and the
number of failures.  The really interesting point as far as this patch
series is concerned is the number of pages scanned.  Note that reverting
Rik's patches massively increases the number of pages scanned indicating
that those patches really did make a difference to CPU usage.

However, caching what pageblocks should be skipped has a much higher
impact.  With patches 1-8 applied, free page and migrate page scanning are
both reduced by 95% in comparison to the akpm kernel.  If the basic
concept of Rik's patches are implemened on top then scanning then the free
scanner barely changed but migrate scanning was further reduced.  That
said, tests on 3.6-rc5 indicated that the last patch had greater impact
than what was measured here so it is a bit variable.

One way or the other, this series has a large impact on the amount of
scanning compaction does when there is a storm of THP allocations.

This patch:

Compaction's migrate scanner acquires the zone->lru_lock when scanning a
range of pages looking for LRU pages to acquire.  It does this even if
there are no LRU pages in the range.  If multiple processes are compacting
then this can cause severe locking contention.  To make matters worse
commit b2eef8c0 ("mm: compaction: minimise the time IRQs are disabled
while isolating pages for migration") releases the lru_lock every
SWAP_CLUSTER_MAX pages that are scanned.

This patch makes two changes to how the migrate scanner acquires the LRU
lock.  First, it only releases the LRU lock every SWAP_CLUSTER_MAX pages
if the lock is contended.  This reduces the number of times it
unnecessarily disables and re-enables IRQs.  The second is that it defers
acquiring the LRU lock for as long as possible.  If there are no LRU pages
or the only LRU pages are transhuge then the LRU lock will not be acquired
at all which reduces contention on zone->lru_lock.

[minchan@kernel.org: augment comment]
[akpm@linux-foundation.org: tweak comment text]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Richard Davies <richard@arachsys.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Avi Kivity <avi@redhat.com>
Acked-by: Rafael Aquini <aquini@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: compaction: Update try_to_compact_pages()kerneldoc comment

Parameters were added without documentation, tut tut.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: compaction: move fatal signal check out of compact_checklock_irqsave

Commit c67fe3752abe ("mm: compaction: Abort async compaction if locks
are contended or taking too long") addressed a lock contention problem
in compaction by introducing compact_checklock_irqsave() that effecively
aborting async compaction in the event of compaction.

To preserve existing behaviour it also moved a fatal_signal_pending()
check into compact_checklock_irqsave() but that is very misleading.  It
"hides" the check within a locking function but has nothing to do with
locking as such.  It just happens to work in a desirable fashion.

This patch moves the fatal_signal_pending() check to
isolate_migratepages_range() where it belongs.  Arguably the same check
should also happen when isolating pages for freeing but it's overkill.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: compaction: abort compaction loop if lock is contended or run too long

isolate_migratepages_range() might isolate no pages if for example when
zone->lru_lock is contended and running asynchronous compaction. In this
case, we should abort compaction, otherwise, compact_zone will run a
useless loop and make zone->lru_lock is even contended.

An additional check is added to ensure that cc.migratepages and
cc.freepages get properly drained whan compaction is aborted.

[minchan@kernel.org: Putback pages isolated for migration if aborting]
[akpm@linux-foundation.org: compact_zone_order requires non-NULL arg contended]
[akpm@linux-foundation.org: make compact_zone_order() require non-NULL arg `contended']
[minchan@kernel.org: Putback pages isolated for migration if aborting]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Shaohua Li <shli@fusionio.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/memblock: cleanup early_node_map[] related comments

Commit 0ee332c14518 ("memblock: Kill early_node_map[]") removed
early_node_map[].  Clean up the comments to comply with that change.

Signed-off-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Gavin Shan <shangw@linux.vnet.ibm.com>
Cc: Yinghai Lu <yinghai@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/memblock: use existing interface to set nid

Use the existing interface function to set the NUMA node ID (NID) for the
regions, either memory or reserved region.

Signed-off-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Gavin Shan <shangw@linux.vnet.ibm.com>
Cc: Yinghai Lu <yinghai@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

readahead: fault retry breaks mmap file read random detection

.fault now can retry.  The retry can break state machine of .fault.  In
filemap_fault, if page is miss, ra->mmap_miss is increased.  In the second
try, since the page is in page cache now, ra->mmap_miss is decreased.  And
these are done in one fault, so we can't detect random mmap file access.

Add a new flag to indicate .fault is tried once.  In the second try, skip
ra->mmap_miss decreasing.  The filemap_fault state machine is ok with it.

I only tested x86, didn't test other archs, but looks the change for other
archs is obvious, but who knows :)

Signed-off-by: Shaohua Li <shaohua.li@fusionio.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

atomic: implement generic atomic_dec_if_positive()

The x86 implementation of atomic_dec_if_positive is quite generic, so make
it available to all architectures.

This is needed for "swap: add a simple detector for inappropriate swapin
readahead".

[akpm@linux-foundation.org: do the "#define foo foo" trick in the conventional manner]
Signed-off-by: Shaohua Li <shli@fusionio.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Michal Simek <monstr@monstr.eu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: fix pages missed by race rather than failing

If race between allocation and isolation in memory-hotplug offline
happens, some pages could be in MIGRATE_MOVABLE of free_list although the
pageblock's migratetype of the page is MIGRATE_ISOLATE.

The race could be detected by get_freepage_migratetype in
__test_page_isolated_in_pageblock.  If it is detected, now EBUSY gets
bubbled all the way up and the hotplug operations fails.

But better idea is instead of returning and failing memory-hotremove, move
the free page to the correct list at the time it is detected.  It could
enhance memory-hotremove operation success ratio although the race is
really rare.

Suggested by Mel Gorman.

[akpm@linux-foundation.org: small cleanup]
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Xishi Qiu <qiuxishi@huawei.com>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: bug fix race between isolation and allocation

Like below, memory-hotplug makes race between page-isolation
and page-allocation so it can hit BUG_ON in __offline_isolated_pages.

CPU A CPU B

start_isolate_page_range
set_migratetype_isolate
spin_lock_irqsave(zone->lock)

free_hot_cold_page(Page A)
/* without zone->lock */
migratetype = get_pageblock_migratetype(Page A);
/*
 * Page could be moved into MIGRATE_MOVABLE
 * of per_cpu_pages
 */
list_add_tail(&page->lru, &pcp->lists[migratetype]);

set_pageblock_isolate
move_freepages_block
drain_all_pages

/* Page A could be in MIGRATE_MOVABLE of free_list. */

check_pages_isolated
__test_page_isolated_in_pageblock
/*
 * We can't catch freed page which
 * is free_list[MIGRATE_MOVABLE]
 */
if (PageBuddy(page A))
pfn += 1 << page_order(page A);

/* So, Page A could be allocated */

__offline_isolated_pages
/*
 * BUG_ON hit or offline page
 * which is used by someone
 */
BUG_ON(!PageBuddy(page A));

This patch checks page's migratetype in freelist in
__test_page_isolated_in_pageblock.  So now
__test_page_isolated_in_pageblock can check the page caused by above race
and can fail of memory offlining.

Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Xishi Qiu <qiuxishi@huawei.com>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: remain migratetype in freed page

The page allocator caches the pageblock information in page->private while
it is in the PCP freelists but this is overwritten with the order of the
page when freed to the buddy allocator.  This patch stores the migratetype
of the page in the page->index field so that it is available at all times
when the page remain in free_list.

This patch adds a new call site in __free_pages_ok so it might be overhead
a bit but it's for high order allocation.  So I believe damage isn't hurt.

Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Xishi Qiu <qiuxishi@huawei.com>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: page_alloc: use get_freepage_migratetype() instead of page_private()

The page allocator uses set_page_private and page_private for handling
migratetype when it frees page.  Let's replace them with [set|get]
_freepage_migratetype to make it more clear.

Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Xishi Qiu <qiuxishi@huawei.com>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

cma: fix watermark checking

* Add ALLOC_CMA alloc flag and pass it to [__]zone_watermark_ok()
  (from Minchan Kim).

* During watermark check decrease available free pages number by
  free CMA pages number if necessary (unmovable allocations cannot
  use pages from CMA areas).

Signed-off-by: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

cma: count free CMA pages

Add NR_FREE_CMA_PAGES counter to be later used for checking watermark in
__zone_watermark_ok().  For simplicity and to avoid #ifdef hell make this
counter always available (not only when CONFIG_CMA=y).

[akpm@linux-foundation.org: use conventional migratetype naming]
Signed-off-by: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

cma: fix counting of isolated pages

Isolated free pages shouldn't be accounted to NR_FREE_PAGES counter.  Fix
it by properly decreasing/increasing NR_FREE_PAGES counter in
set_migratetype_isolate()/unset_migratetype_isolate() and removing counter
adjustment for isolated pages from free_one_page() and split_free_page().

Signed-off-by: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: fix tracing in free_pcppages_bulk()

page->private gets re-used in __free_one_page() to store page order
(so trace_mm_page_pcpu_drain() may print order instead of migratetype)
thus migratetype value must be cached locally.

Fixes regression introduced in commit a7016235a61d ("mm: fix migratetype
bug which slowed swapping").  This caused incorrect data to be attached
to the mm_page_pcpu_drain trace event.

[akpm@linux-foundation.org: add comment]
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
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>

mm: cma: discard clean pages during contiguous allocation instead of migration

Drop clean cache pages instead of migration during alloc_contig_range() to
minimise allocation latency by reducing the amount of migration that is
necessary.  It's useful for CMA because latency of migration is more
important than evicting the background process's working set.  In
addition, as pages are reclaimed then fewer free pages for migration
targets are required so it avoids memory reclaiming to get free pages,
which is a contributory factor to increased latency.

I measured elapsed time of __alloc_contig_migrate_range() which migrates
10M in 40M movable zone in QEMU machine.

Before - 146ms, After - 7ms

[akpm@linux-foundation.org: fix nommu build]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Signed-off-by: Minchan Kim <minchan@kernel.org>
Reviewed-by: Mel Gorman <mgorman@suse.de>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Cc: Rik van Riel <riel@redhat.com>
Tested-by: Kyungmin Park <kyungmin.park@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: mmu_notifier: make the mmu_notifier srcu static

The variable must be static especially given the variable name.

s/RCU/SRCU/ over a few comments.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Sagi Grimberg <sagig@mellanox.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: build zonelists when offlining pages

online_pages() does build_all_zonelists() and zone_pcp_update(), I think
offline_pages() should do it too.

When the zone has no memory to allocate, remove it from other nodes'
zonelists.  zone_batchsize() depends on zone's present pages, if zone's
present pages are changed, zone's pcp should be updated.

Signed-off-by: Xishi Qiu <qiuxishi@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: avoid taking rmap locks in move_ptes()

During mremap(), the destination VMA is generally placed after the
original vma in rmap traversal order: in move_vma(), we always have
new_pgoff >= vma->vm_pgoff, and as a result new_vma->vm_pgoff >=
vma->vm_pgoff unless vma_merge() merged the new vma with an adjacent one.

When the destination VMA is placed after the original in rmap traversal
order, we can avoid taking the rmap locks in move_ptes().

Essentially, this reintroduces the optimization that had been disabled in
"mm anon rmap: remove anon_vma_moveto_tail".  The difference is that we
don't try to impose the rmap traversal order; instead we just rely on
things being in the desired order in the common case and fall back to
taking locks in the uncommon case.  Also we skip the i_mmap_mutex in
addition to the anon_vma lock: in both cases, the vmas are traversed in
increasing vm_pgoff order with ties resolved in tree insertion order.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm anon rmap: in mremap, set the new vma's position before anon_vma_clone()

anon_vma_clone() expects new_vma->vm_{start,end,pgoff} to be correctly set
so that the new vma can be indexed on the anon interval tree.

copy_vma() was failing to do that, which broke mremap().

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Tested-by: Sasha Levin <levinsasha928@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: add CONFIG_DEBUG_VM_RB build option

Add a CONFIG_DEBUG_VM_RB build option for the previously existing
DEBUG_MM_RB code.  Now that Andi Kleen modified it to avoid using
recursive algorithms, we can expose it a bit more.

Also extend this code to validate_mm() after stack expansion, and to check
that the vma's start and last pgoffs have not changed since the nodes were
inserted on the anon vma interval tree (as it is important that the nodes
be reindexed after each such update).

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm rmap: remove vma_address check for address inside vma

In file and anon rmap, we use interval trees to find potentially relevant
vmas and then call vma_address() to find the virtual address the given
page might be found at in these vmas.  vma_address() used to include a
check that the returned address falls within the limits of the vma, but
this check isn't necessary now that we always use interval trees in rmap:
the interval tree just doesn't return any vmas which this check would find
to be irrelevant.  As a result, we can replace the use of -EFAULT error
code (which then needed to be checked in every call site) with a
VM_BUG_ON().

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm anon rmap: replace same_anon_vma linked list with an interval tree.

When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list.  This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.

By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.

While the change is large, all of its pieces are fairly simple.

Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead.  The exception here is ksm, where the
page's index is not known.  It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.

When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree.  This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
 The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm anon rmap: remove anon_vma_moveto_tail

mremap() had a clever optimization where move_ptes() did not take the
anon_vma lock to avoid a race with anon rmap users such as page migration.
 Instead, the avc's were ordered in such a way that the origin vma was
always visited by rmap before the destination.  This ordering and the use
of page table locks rmap usage safe.  However, we want to replace the use
of linked lists in anon rmap with an interval tree, and this will make it
harder to impose such ordering as the interval tree will always be sorted
by the avc->vma->vm_pgoff value.  For now, let's replace the
anon_vma_moveto_tail() ordering function with proper anon_vma locking in
move_ptes().  Once we have the anon interval tree in place, we will
re-introduce an optimization to avoid taking these locks in the most
common cases.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: interval tree updates

Update the generic interval tree code that was introduced in "mm: replace
vma prio_tree with an interval tree".

Changes:

- fixed 'endpoing' typo noticed by Andrew Morton

- replaced include/linux/interval_tree_tmpl.h, which was used as a
  template (including it automatically defined the interval tree
  functions) with include/linux/interval_tree_generic.h, which only
  defines a preprocessor macro INTERVAL_TREE_DEFINE(), which itself
  defines the interval tree functions when invoked. Now that is a very
  long macro which is unfortunate, but it does make the usage sites
  (lib/interval_tree.c and mm/interval_tree.c) a bit nicer than previously.

- make use of RB_DECLARE_CALLBACKS() in the INTERVAL_TREE_DEFINE() macro,
  instead of duplicating that code in the interval tree template.

- replaced vma_interval_tree_add(), which was actually handling the
  nonlinear and interval tree cases, with vma_interval_tree_insert_after()
  which handles only the interval tree case and has an API that is more
  consistent with the other interval tree handling functions.
  The nonlinear case is now handled explicitly in kernel/fork.c dup_mmap().

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: move augmented rbtree functionality to rbtree_augmented.h

Provide rb_insert_augmented() and rb_erase_augmented() through a new
rbtree_augmented.h include file.  rb_erase_augmented() is defined there as
an __always_inline function, in order to allow inlining of augmented
rbtree callbacks into it.  Since this generates a relatively large
function, each augmented rbtree user should make sure to have a single
call site.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

prio_tree: remove

After both prio_tree users have been converted to use red-black trees,
there is no need to keep around the prio tree library anymore.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

kmemleak: use rbtree instead of prio tree

kmemleak uses a tree where each node represents an allocated memory object
in order to quickly find out what object a given address is part of.
However, the objects don't overlap, so rbtrees are a better choice than
prio tree for this use.  They are both faster and have lower memory
overhead.

Tested by booting a kernel with kmemleak enabled, loading the
kmemleak_test module, and looking for the expected messages.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: replace vma prio_tree with an interval tree

Implement an interval tree as a replacement for the VMA prio_tree.  The
algorithms are similar to lib/interval_tree.c; however that code can't be
directly reused as the interval endpoints are not explicitly stored in the
VMA.  So instead, the common algorithm is moved into a template and the
details (node type, how to get interval endpoints from the node, etc) are
filled in using the C preprocessor.

Once the interval tree functions are available, using them as a
replacement to the VMA prio tree is a relatively simple, mechanical job.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: add prio tree and interval tree tests

Patch 1 implements support for interval trees, on top of the augmented
rbtree API. It also adds synthetic tests to compare the performance of
interval trees vs prio trees. Short answers is that interval trees are
slightly faster (~25%) on insert/erase, and much faster (~2.4 - 3x)
on search. It is debatable how realistic the synthetic test is, and I have
not made such measurements yet, but my impression is that interval trees
would still come out faster.

Patch 2 uses a preprocessor template to make the interval tree generic,
and uses it as a replacement for the vma prio_tree.

Patch 3 takes the other prio_tree user, kmemleak, and converts it to use
a basic rbtree. We don't actually need the augmented rbtree support here
because the intervals are always non-overlapping.

Patch 4 removes the now-unused prio tree library.

Patch 5 proposes an additional optimization to rb_erase_augmented, now
providing it as an inline function so that the augmented callbacks can be
inlined in. This provides an additional 5-10% performance improvement
for the interval tree insert/erase benchmark. There is a maintainance cost
as it exposes augmented rbtree users to some of the rbtree library internals;
however I think this cost shouldn't be too high as I expect the augmented
rbtree will always have much less users than the base rbtree.

I should probably add a quick summary of why I think it makes sense to
replace prio trees with augmented rbtree based interval trees now.  One of
the drivers is that we need augmented rbtrees for Rik's vma gap finding
code, and once you have them, it just makes sense to use them for interval
trees as well, as this is the simpler and more well known algorithm.  prio
trees, in comparison, seem *too* clever: they impose an additional 'heap'
constraint on the tree, which they use to guarantee a faster worst-case
complexity of O(k+log N) for stabbing queries in a well-balanced prio
tree, vs O(k*log N) for interval trees (where k=number of matches,
N=number of intervals).  Now this sounds great, but in practice prio trees
don't realize this theorical benefit.  First, the additional constraint
makes them harder to update, so that the kernel implementation has to
simplify things by balancing them like a radix tree, which is not always
ideal.  Second, the fact that there are both index and heap properties
makes both tree manipulation and search more complex, which results in a
higher multiplicative time constant.  As it turns out, the simple interval
tree algorithm ends up running faster than the more clever prio tree.

This patch:

Add two test modules:

- prio_tree_test measures the performance of lib/prio_tree.c, both for
  insertion/removal and for stabbing searches

- interval_tree_test measures the performance of a library of equivalent
  functionality, built using the augmented rbtree support.

In order to support the second test module, lib/interval_tree.c is
introduced. It is kept separate from the interval_tree_test main file
for two reasons: first we don't want to provide an unfair advantage
over prio_tree_test by having everything in a single compilation unit,
and second there is the possibility that the interval tree functionality
could get some non-test users in kernel over time.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: add RB_DECLARE_CALLBACKS() macro

As proposed by Peter Zijlstra, this makes it easier to define the augmented
rbtree callbacks.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: remove prior augmented rbtree implementation

convert arch/x86/mm/pat_rbtree.c to the proposed augmented rbtree api
and remove the old augmented rbtree implementation.

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: faster augmented rbtree manipulation

Introduce new augmented rbtree APIs that allow minimal recalculation of
augmented node information.

A new callback is added to the rbtree insertion and erase rebalancing
functions, to be called on each tree rotations. Such rotations preserve
the subtree's root augmented value, but require recalculation of the one
child that was previously located at the subtree root.

In the insertion case, the handcoded search phase must be updated to
maintain the augmented information on insertion, and then the rbtree
coloring/rebalancing algorithms keep it up to date.

In the erase case, things are more complicated since it is library
code that manipulates the rbtree in order to remove internal nodes.
This requires a couple additional callbacks to copy a subtree's
augmented value when a new root is stitched in, and to recompute
augmented values down the ancestry path when a node is removed from
the tree.

In order to preserve maximum speed for the non-augmented case,
we provide two versions of each tree manipulation function.
rb_insert_augmented() is the augmented equivalent of rb_insert_color(),
and rb_erase_augmented() is the augmented equivalent of rb_erase().

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: augmented rbtree test

Small test to measure the performance of augmented rbtrees.

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: low level optimizations in rb_erase()

Various minor optimizations in rb_erase():
- Avoid multiple loading of node->__rb_parent_color when computing parent
  and color information (possibly not in close sequence, as there might
  be further branches in the algorithm)
- In the 1-child subcase of case 1, copy the __rb_parent_color field from
  the erased node to the child instead of recomputing it from the desired
  parent and color
- When searching for the erased node's successor, differentiate between
  cases 2 and 3 based on whether any left links were followed. This avoids
  a condition later down.
- In case 3, keep a pointer to the erased node's right child so we don't
  have to refetch it later to adjust its parent.
- In the no-childs subcase of cases 2 and 3, place the rebalance assigment
  last so that the compiler can remove the following if(rebalance) test.

Also, added some comments to illustrate cases 2 and 3.

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: handle 1-child recoloring in rb_erase() instead of rb_erase_color()

An interesting observation for rb_erase() is that when a node has
exactly one child, the node must be black and the child must be red.
An interesting consequence is that removing such a node can be done by
simply replacing it with its child and making the child black,
which we can do efficiently in rb_erase(). __rb_erase_color() then
only needs to handle the no-childs case and can be modified accordingly.

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: place easiest case first in rb_erase()

In rb_erase, move the easy case (node to erase has no more than
1 child) first. I feel the code reads easier that way.

Signed-off-by: Michel Lespinasse <walken@google.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: add __rb_change_child() helper function

Add __rb_change_child() as an inline helper function to replace code that
would otherwise be duplicated 4 times in the source.

No changes to binary size or speed.

Signed-off-by: Michel Lespinasse <walken@google.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree test: fix sparse warning about 64-bit constant

Just a small fix to make sparse happy.

Signed-off-by: Michel Lespinasse <walken@google.com>
Reported-by: Fengguang Wu <wfg@linux.intel.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: optimize fetching of sibling node

When looking to fetch a node's sibling, we went through a sequence of:
- check if node is the parent's left child
- if it is, then fetch the parent's right child

This can be replaced with:
- fetch the parent's right child as an assumed sibling
- check that node is NOT the fetched child

This avoids fetching the parent's left child when node is actually
that child. Saves a bit on code size, though it doesn't seem to make
a large difference in speed.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <David.Woodhouse@intel.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: coding style adjustments

Set comment and indentation style to be consistent with linux coding style
and the rest of the file, as suggested by Peter Zijlstra

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: low level optimizations in __rb_erase_color()

In __rb_erase_color(), we often already have pointers to the nodes being
rotated and/or know what their colors must be, so we can generate more
efficient code than the generic __rb_rotate_left() and __rb_rotate_right()
functions.

Also when the current node is red or when flipping the sibling's color,
the parent is already known so we can use the more efficient
rb_set_parent_color() function to set the desired color.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: optimize case selection logic in __rb_erase_color()

In __rb_erase_color(), we have to select one of 3 cases depending on the
color on the 'other' node children.  If both children are black, we flip a
few node colors and iterate.  Otherwise, we do either one or two tree
rotations, depending on the color of the 'other' child opposite to 'node',
and then we are done.

The corresponding logic had duplicate checks for the color of the 'other'
child opposite to 'node'.  It was checking it first to determine if both
children are black, and then to determine how many tree rotations are
required.  Rearrange the logic to avoid that extra check.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: adjust node color in __rb_erase_color() only when necessary

In __rb_erase_color(), we were always setting a node to black after
exiting the main loop.  And in one case, after fixing up the tree to
satisfy all rbtree invariants, we were setting the current node to root
just to guarantee a loop exit, at which point the root would be set to
black.  However this is not necessary, as the root of an rbtree is already
known to be black.  The only case where the color flip is required is when
we exit the loop due to the current node being red, and it's easiest to
just do the flip at that point instead of doing it after the loop.

[adrian.hunter@intel.com: perf tools: fix build for another rbtree.c change]
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: low level optimizations in rb_insert_color()

- Use the newly introduced rb_set_parent_color() function to flip the color
  of nodes whose parent is already known.
- Optimize rb_parent() when the node is known to be red - there is no need
  to mask out the color in that case.
- Flipping gparent's color to red requires us to fetch its rb_parent_color
  field, so we can reuse it as the parent value for the next loop iteration.
- Do not use __rb_rotate_left() and __rb_rotate_right() to handle tree
  rotations: we already have pointers to all relevant nodes, and know their
  colors (either because we want to adjust it, or because we've tested it,
  or we can deduce it as black due to the node proximity to a known red node).
  So we can generate more efficient code by making use of the node pointers
  we already have, and setting both the parent and color attributes for
  nodes all at once. Also in Case 2, some node attributes don't have to
  be set because we know another tree rotation (Case 3) will always follow
  and override them.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: adjust root color in rb_insert_color() only when necessary

The root node of an rbtree must always be black.  However,
rb_insert_color() only needs to maintain this invariant when it has been
broken - that is, when it exits the loop due to the current (red) node
being the root.  In all other cases (exiting after tree rotations, or
exiting due to an existing black parent) the invariant is already
satisfied, so there is no need to adjust the root node color.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: break out of rb_insert_color loop after tree rotation

It is a well known property of rbtrees that insertion never requires more
than two tree rotations.  In our implementation, after one loop iteration
identified one or two necessary tree rotations, we would iterate and look
for more.  However at that point the node's parent would always be black,
which would cause us to exit the loop.

We can make the code flow more obvious by just adding a break statement
after the tree rotations, where we know we are done.  Additionally, in the
cases where two tree rotations are necessary, we don't have to update the
'node' pointer as it wouldn't be used until the next loop iteration, which
we now avoid due to this break statement.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: performance and correctness test

This small module helps measure the performance of rbtree insert and
erase.

Additionally, we run a few correctness tests to check that the rbtrees
have all desired properties:

- contains the right number of nodes in the order desired,
- never two consecutive red nodes on any path,
- all paths to leaf nodes have the same number of black nodes,
- root node is black

[akpm@linux-foundation.org: fix printk warning: sparc64 cycles_t is unsigned long]
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: move some implementation details from rbtree.h to rbtree.c

rbtree users must use the documented APIs to manipulate the tree
structure.  Low-level helpers to manipulate node colors and parenthood are
not part of that API, so move them to lib/rbtree.c

[dwmw2@infradead.org: fix jffs2 build issue due to renamed __rb_parent_color field]
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: fix incorrect rbtree node insertion in fs/proc/proc_sysctl.c

The recently added code to use rbtrees in sysctl did not follow the proper
rbtree interface on insertion - it was calling rb_link_node() which
inserts a new node into the binary tree, but missed the call to
rb_insert_color() which properly balances the rbtree and establishes all
expected rbtree invariants.

I found out about this only because faulty commit also used
rb_init_node(), which I am removing within this patchset.  But I think
it's an easy mistake to make, and it makes me wonder if we should change
the rbtree API so that insertions would be done with a single rb_insert()
call (even if its implementation could still inline the rb_link_node()
part and call a private __rb_insert_color function to do the rebalancing).

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: empty nodes have no color

Empty nodes have no color.  We can make use of this property to simplify
the code emitted by the RB_EMPTY_NODE and RB_CLEAR_NODE macros.  Also,
we can get rid of the rb_init_node function which had been introduced by
commit 88d19cf37952 ("timers: Add rb_init_node() to allow for stack
allocated rb nodes") to avoid some issue with the empty node's color not
being initialized.

I'm not sure what the RB_EMPTY_NODE checks in rb_prev() / rb_next() are
doing there, though.  axboe introduced them in commit 10fd48f2376d
("rbtree: fixed reversed RB_EMPTY_NODE and rb_next/prev").  The way I
see it, the 'empty node' abstraction is only used by rbtree users to
flag nodes that they haven't inserted in any rbtree, so asking the
predecessor or successor of such nodes doesn't make any sense.

One final rb_init_node() caller was recently added in sysctl code to
implement faster sysctl name lookups.  This code doesn't make use of
RB_EMPTY_NODE at all, and from what I could see it only called
rb_init_node() under the mistaken assumption that such initialization was
required before node insertion.

[sfr@canb.auug.org.au: fix net/ceph/osd_client.c build]
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: John Stultz <john.stultz@linaro.org>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

rbtree: reference Documentation/rbtree.txt for usage instructions

I recently started looking at the rbtree code (with an eye towards
improving the augmented rbtree support, but I haven't gotten there yet).
I noticed a lot of possible speed improvements, which I am now proposing
in this patch set.

Patches 1-4 are preparatory: remove internal functions from rbtree.h so
that users won't be tempted to use them instead of the documented APIs,
clean up some incorrect usages I've noticed (in particular, with the
recently added fs/proc/proc_sysctl.c rbtree usage), reference the
documentation so that people have one less excuse to miss it, etc.

Patch 5 is a small module I wrote to check the rbtree performance.  It
creates 100 nodes with random keys and repeatedly inserts and erases them
from an rbtree.  Additionally, it has code to check for rbtree invariants
after each insert or erase operation.

Patches 6-12 is where the rbtree optimizations are done, and they touch
only that one file, lib/rbtree.c .  I am getting good results out of these
- in my small benchmark doing rbtree insertion (including search) and
erase, I'm seeing a 30% runtime reduction on Sandybridge E5, which is more
than I initially thought would be possible.  (the results aren't as
impressive on my two other test hosts though, AMD barcelona and Intel
Westmere, where I am seeing 14% runtime reduction only).  The code size -
both source (ommiting comments) and compiled - is also shorter after these
changes.  However, I do admit that the updated code is more arduous to
read - one big reason for that is the removal of the tree rotation
helpers, which added some overhead but also made it easier to reason about
things locally.  Overall, I believe this is an acceptable compromise,
given that this code doesn't get modified very often, and that I have good
tests for it.

Upon Peter's suggestion, I added comments showing the rtree configuration
before every rotation.  I think they help; however it's still best to have
a copy of the cormen/leiserson/rivest book when digging into this code.

This patch: reference Documentation/rbtree.txt for usage instructions

include/linux/rbtree.h included some basic usage instructions, while
Documentation/rbtree.txt had some more complete and easier to follow
instructions.  Replacing the former with a reference to the latter.

Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: David Woodhouse <David.Woodhouse@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ipc/mqueue: remove unnecessary rb_init_node() calls

Commit d6629859b36d ("ipc/mqueue: improve performance of send/recv") and
ce2d52cc ("ipc/mqueue: add rbtree node caching support") introduced an
rbtree of message priorities, and usage of rb_init_node() to initialize
the corresponding nodes.  As it turns out, rb_init_node() is unnecessary
here, as the nodes are fully initialized on insertion by rb_link_node()
and the code doesn't access nodes that aren't inserted on the rbtree.

Removing the rb_init_node() calls as I removed that function during
rbtree API cleanups (the only other use of it was in a place that
similarly didn't require it).

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp, s390: architecture backend for thp on s390

This implements the architecture backend for transparent hugepages
on s390.

Signed-off-by: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp, s390: disable thp for kvm host on s390

This patch is part of the architecture backend for thp on s390.  It
disables thp for kvm hosts, because there is no kvm host hugepage support
so far.  Existing thp mappings are split by follow_page() with FOLL_SPLIT,
and future thp mappings are prevented by setting VM_NOHUGEPAGE in
mm->def_flags.

Signed-off-by: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp, s390: thp pagetable pre-allocation for s390

This patch is part of the architecture backend for thp on s390.  It
provides the pagetable pre-allocation functions
pgtable_trans_huge_deposit() and pgtable_trans_huge_withdraw().  Unlike
other archs, s390 has no struct page * as pgtable_t, but rather a pointer
to the page table.  So instead of saving the pagetable pre- allocation
list info inside the struct page, it is being saved within the pagetable
itself.

Signed-off-by: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp, s390: thp splitting backend for s390

This patch is part of the architecture backend for thp on s390.  It
provides the functions related to thp splitting, including serialization
against gup.  Unlike other archs, pmdp_splitting_flush() cannot use a tlb
flushing operation to serialize against gup on s390, because that wouldn't
be stopped by the disabled IRQs.  So instead, smp_call_function() is
called with an empty function, which will have the expected effect.

Signed-off-by: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: make MADV_HUGEPAGE check for mm->def_flags

This adds a check to hugepage_madvise(), to refuse MADV_HUGEPAGE if
VM_NOHUGEPAGE is set in mm->def_flags.  On s390, the VM_NOHUGEPAGE flag
will be set in mm->def_flags for kvm processes, to prevent any future thp
mappings.  In order to also prevent MADV_HUGEPAGE on such an mm,
hugepage_madvise() should check mm->def_flags.

Signed-off-by: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: introduce pmdp_invalidate()

On s390, a valid page table entry must not be changed while it is attached
to any CPU.  So instead of pmd_mknotpresent() and set_pmd_at(), an IDTE
operation would be necessary there.  This patch introduces the
pmdp_invalidate() function, to allow architecture-specific
implementations.

Signed-off-by: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: remove assumptions on pgtable_t type

The thp page table pre-allocation code currently assumes that pgtable_t is
of type "struct page *".  This may not be true for all architectures, so
this patch removes that assumption by replacing the functions
prepare_pmd_huge_pte() and get_pmd_huge_pte() with two new functions that
can be defined architecture-specific.

It also removes two VM_BUG_ON checks for page_count() and page_mapcount()
operating on a pgtable_t.  Apart from the VM_BUG_ON removal, there will be
no functional change introduced by this patch.

Signed-off-by: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp, x86: introduce HAVE_ARCH_TRANSPARENT_HUGEPAGE

Cleanup patch in preparation for transparent hugepage support on s390.
Adding new architectures to the TRANSPARENT_HUGEPAGE config option can
make the "depends" line rather ugly, like "depends on (X86 || (S390 &&
64BIT)) && MMU".

This patch adds a HAVE_ARCH_TRANSPARENT_HUGEPAGE instead.  x86 already has
MMU "def_bool y", so the MMU check is superfluous there and
HAVE_ARCH_TRANSPARENT_HUGEPAGE can be selected in arch/x86/Kconfig.

Signed-off-by: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: fix potential anon_vma locking issue in mprotect()

Fix an anon_vma locking issue in the following situation:

- vma has no anon_vma
- next has an anon_vma
- vma is being shrunk / next is being expanded, due to an mprotect call

We need to take next's anon_vma lock to avoid races with rmap users (such
as page migration) while next is being expanded.

Signed-off-by: Michel Lespinasse <walken@google.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: remove unnecessary set_recommended_min_free_kbytes

Since it is called in start_khugepaged

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: use khugepaged_enabled to remove duplicate code

Use khugepaged_enabled to see whether thp is enabled

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: remove khugepaged_loop

Merge khugepaged_loop into khugepaged

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: introduce khugepaged_prealloc_page and khugepaged_alloc_page

They are used to abstract the difference between NUMA enabled and NUMA
disabled to make the code more readable

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: release page in page pre-alloc path

If NUMA is enabled, we can release the page in the page pre-alloc
operation, then the CONFIG_NUMA dependent code can be reduced

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: merge page pre-alloc in khugepaged_loop into khugepaged_do_scan

There are two pre-alloc operations in these two function, the different is:
- it allows to sleep if page alloc fail in khugepaged_loop
- it exits immediately if page alloc fail in khugepaged_do_scan

Actually, in khugepaged_do_scan, we can allow the pre-alloc to sleep on
the first failure, then the operation in khugepaged_loop can be removed

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: remove some code depend on CONFIG_NUMA

If NUMA is disabled, hpage is used as page pre-alloc, so there are two
cases for hpage:

- it is !NULL, means the page is not consumed otherwise,
- the page has been consumed

If NUMA is enabled, hpage is just used as alloc-fail indicator which is
not a real page, NULL means not fail triggered.

So, we can release the page only if !IS_ERR_OR_NULL

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: remove wake_up_interruptible in the exit path

Add the check of kthread_should_stop() to the conditions which are used to
wakeup on khugepaged_wait, then kthread_stop is enough to let the thread
exit

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: remove unnecessary khugepaged_thread check

Now, khugepaged creation and cancel are completely serial under the
protection of khugepaged_mutex, it is impossible that many khugepaged
entities are running

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: move khugepaged_mutex out of khugepaged

Currently, hugepaged_mutex is used really complexly and hard to
understand, actually, it is just used to serialize start_khugepaged and
khugepaged for these reasons:

- khugepaged_thread is shared between them
- the thp disable path (echo never > transparent_hugepage/enabled) is
  nonblocking, so we need to protect khugepaged_thread to get a stable
  running state

These can be avoided by:

- use the lock to serialize the thread creation and cancel
- thp disable path can not finised until the thread exits

Then khugepaged_thread is fully controlled by start_khugepaged, khugepaged
will be happy without the lock

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: remove unnecessary check in start_khugepaged

The check is unnecessary since if mm_slot_cache or mm_slots_hash
initialize failed, no sysfs interface will be created

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: fix the count of THP_COLLAPSE_ALLOC

THP_COLLAPSE_ALLOC is double counted if NUMA is disabled since it has
already been calculated in khugepaged_alloc_hugepage

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: adjust final #endif position in mm/internal.h

Make sure the #endif that terminates the standard #ifndef / #define /
#endif construct gets labeled, and gets positioned at the end of the file
as is normally the case.

Signed-off-by: Michel Lespinasse <walken@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: hugetlb: add arch hook for clearing page flags before entering pool

The core page allocator ensures that page flags are zeroed when freeing
pages via free_pages_check.  A number of architectures (ARM, PPC, MIPS)
rely on this property to treat new pages as dirty with respect to the data
cache and perform the appropriate flushing before mapping the pages into
userspace.

This can lead to cache synchronisation problems when using hugepages,
since the allocator keeps its own pool of pages above the usual page
allocator and does not reset the page flags when freeing a page into the
pool.

This patch adds a new architecture hook, arch_clear_hugepage_flags, so
that architectures which rely on the page flags being in a particular
state for fresh allocations can adjust the flags accordingly when a page
is freed into the pool.

Signed-off-by: Will Deacon <will.deacon@arm.com>
Cc: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>