memcg: document cgroup dirty memory interfaces

Document cgroup dirty memory interfaces and statistics.

[akpm@linux-foundation.org: fix use_hierarchy description]
Signed-off-by: Andrea Righi <arighi@develer.com>
Signed-off-by: Greg Thelen <gthelen@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: Minchan Kim <minchan.kim@gmail.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>

memcg: add page_cgroup flags for dirty page tracking

This patchset provides the ability for each cgroup to have independent
dirty page limits.

Limiting dirty memory is like fixing the max amount of dirty (hard to
reclaim) page cache used by a cgroup.  So, in case of multiple cgroup
writers, they will not be able to consume more than their designated share
of dirty pages and will be forced to perform write-out if they cross that
limit.

The patches are based on a series proposed by Andrea Righi in Mar 2010.

Overview:

- Add page_cgroup flags to record when pages are dirty, in writeback, or nfs
  unstable.

- Extend mem_cgroup to record the total number of pages in each of the
  interesting dirty states (dirty, writeback, unstable_nfs).

- Add dirty parameters similar to the system-wide  /proc/sys/vm/dirty_*
  limits to mem_cgroup.  The mem_cgroup dirty parameters are accessible
  via cgroupfs control files.

- Consider both system and per-memcg dirty limits in page writeback when
  deciding to queue background writeback or block for foreground writeback.

Known shortcomings:

- When a cgroup dirty limit is exceeded, then bdi writeback is employed to
  writeback dirty inodes.  Bdi writeback considers inodes from any cgroup, not
  just inodes contributing dirty pages to the cgroup exceeding its limit.

- When memory.use_hierarchy is set, then dirty limits are disabled.  This is a
  implementation detail.  An enhanced implementation is needed to check the
  chain of parents to ensure that no dirty limit is exceeded.

Performance data:
- A page fault microbenchmark workload was used to measure performance, which
  can be called in read or write mode:
        f = open(foo. $cpu)
        truncate(f, 4096)
        alarm(60)
        while (1) {
                p = mmap(f, 4096)
                if (write)
*p = 1
else
x = *p
                munmap(p)
        }

- The workload was called for several points in the patch series in different
  modes:
  - s_read is a single threaded reader
  - s_write is a single threaded writer
  - p_read is a 16 thread reader, each operating on a different file
  - p_write is a 16 thread writer, each operating on a different file

- Measurements were collected on a 16 core non-numa system using "perf stat
  --repeat 3".  The -a option was used for parallel (p_*) runs.

- All numbers are page fault rate (M/sec).  Higher is better.

- To compare the performance of a kernel without non-memcg compare the first and
  last rows, neither has memcg configured.  The first row does not include any
  of these memcg patches.

- To compare the performance of using memcg dirty limits, compare the baseline
  (2nd row titled "w/ memcg") with the the code and memcg enabled (2nd to last
  row titled "all patches").

                           root_cgroup                    child_cgroup
                 s_read s_write p_read p_write   s_read s_write p_read p_write
mmotm w/o memcg   0.428  0.390   0.429  0.388
mmotm w/ memcg    0.411  0.378   0.391  0.362     0.412  0.377   0.385  0.363
all patches       0.384  0.360   0.370  0.348     0.381  0.363   0.368  0.347
all patches       0.431  0.402   0.427  0.395
  w/o memcg

This patch:

Add additional flags to page_cgroup to track dirty pages within a
mem_cgroup.

Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrea Righi <arighi@develer.com>
Signed-off-by: Greg Thelen <gthelen@google.com>
Acked-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: Minchan Kim <minchan.kim@gmail.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>

mm: batch activate_page() to reduce lock contention

The zone->lru_lock is heavily contented in workload where activate_page()
is frequently used.  We could do batch activate_page() to reduce the lock
contention.  The batched pages will be added into zone list when the pool
is full or page reclaim is trying to drain them.

For example, in a 4 socket 64 CPU system, create a sparse file and 64
processes, processes shared map to the file.  Each process read access the
whole file and then exit.  The process exit will do unmap_vmas() and cause
a lot of activate_page() call.  In such workload, we saw about 58% total
time reduction with below patch.  Other workloads with a lot of
activate_page also benefits a lot too.

I tested some microbenchmarks:
case-anon-cow-rand-mt 0.58%
case-anon-cow-rand -3.30%
case-anon-cow-seq-mt -0.51%
case-anon-cow-seq -5.68%
case-anon-r-rand-mt 0.23%
case-anon-r-rand 0.81%
case-anon-r-seq-mt -0.71%
case-anon-r-seq -1.99%
case-anon-rx-rand-mt 2.11%
case-anon-rx-seq-mt 3.46%
case-anon-w-rand-mt -0.03%
case-anon-w-rand -0.50%
case-anon-w-seq-mt -1.08%
case-anon-w-seq -0.12%
case-anon-wx-rand-mt -5.02%
case-anon-wx-seq-mt -1.43%
case-fork 1.65%
case-fork-sleep -0.07%
case-fork-withmem 1.39%
case-hugetlb -0.59%
case-lru-file-mmap-read-mt -0.54%
case-lru-file-mmap-read 0.61%
case-lru-file-mmap-read-rand -2.24%
case-lru-file-readonce -0.64%
case-lru-file-readtwice -11.69%
case-lru-memcg -1.35%
case-mmap-pread-rand-mt 1.88%
case-mmap-pread-rand -15.26%
case-mmap-pread-seq-mt 0.89%
case-mmap-pread-seq -69.72%
case-mmap-xread-rand-mt 0.71%
case-mmap-xread-seq-mt 0.38%

The most significent are:
case-lru-file-readtwice -11.69%
case-mmap-pread-rand -15.26%
case-mmap-pread-seq -69.72%

which use activate_page a lot.  others are basically variations because
each run has slightly difference.

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: simplify code of swap.c

Clean up code and remove duplicate code.  Next patch will use
pagevec_lru_move_fn introduced here too.

Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/page_alloc.c: don't cache `current' in a local

It's old-fashioned and unneeded.

akpm:/usr/src/25> size mm/page_alloc.o
   text    data     bss     dec     hex filename
  39884 1241317   18808 1300009  13d629 mm/page_alloc.o (before)
  39838 1241317   18808 1299963  13d5fb mm/page_alloc.o (after)

Acked-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
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 hugepage migration

2.6.37 added an unmap_and_move_huge_page() for memory failure recovery,
but its anon_vma handling was still based around the 2.6.35 conventions.
Update it to use page_lock_anon_vma, get_anon_vma, page_unlock_anon_vma,
drop_anon_vma in the same way as we're now changing unmap_and_move().

I don't particularly like to propose this for stable when I've not seen
its problems in practice nor tested the solution: but it's clearly out of
synch at present.

Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Rik van Riel <riel@redhat.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: "Jun'ichi Nomura" <j-nomura@ce.jp.nec.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: <stable@kernel.org> [2.6.37, 2.6.36]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: fix migration hangs on anon_vma lock

Increased usage of page migration in mmotm reveals that the anon_vma
locking in unmap_and_move() has been deficient since 2.6.36 (or even
earlier).  Review at the time of f18194275c39835cb84563500995e0d503a32d9a
("mm: fix hang on anon_vma->root->lock") missed the issue here: the
anon_vma to which we get a reference may already have been freed back to
its slab (it is in use when we check page_mapped, but that can change),
and so its anon_vma->root may be switched at any moment by reuse in
anon_vma_prepare.

Perhaps we could fix that with a get_anon_vma_unless_zero(), but let's
not: just rely on page_lock_anon_vma() to do all the hard thinking for us,
then we don't need any rcu read locking over here.

In removing the rcu_unlock label: since PageAnon is a bit in
page->mapping, it's impossible for a !page->mapping page to be anon; but
insert VM_BUG_ON in case the implementation ever changes.

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: "Jun'ichi Nomura" <j-nomura@ce.jp.nec.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: <stable@kernel.org> [2.6.37, 2.6.36]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ksm: drain pagevecs to lru

It was hard to explain the page counts which were causing new LTP tests
of KSM to fail: we need to drain the per-cpu pagevecs to LRU occasionally.

Signed-off-by: Hugh Dickins <hughd@google.com>
Reported-by: CAI Qian <caiqian@redhat.com>
Cc:Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

hugetlb: fix handling of parse errors in sysfs

When parsing changes to the huge page pool sizes made from userspace via
the sysfs interface, bogus input values are being covered up by
nr_hugepages_store_common and nr_overcommit_hugepages_store returning 0
when strict_strtoul returns an error.  This can cause an infinite loop in
the nr_hugepages_store code.  This patch changes the return value for
these functions to -EINVAL when strict_strtoul returns an error.

Signed-off-by: Eric B Munson <emunson@mgebm.net>
Reported-by: CAI Qian <caiqian@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Eric B Munson <emunson@mgebm.net>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

hugetlb: do not allow pagesize >= MAX_ORDER pool adjustment

Huge pages with order >= MAX_ORDER must be allocated at boot via the
kernel command line, they cannot be allocated or freed once the kernel is
up and running.  Currently we allow values to be written to the sysfs and
sysctl files controling pool size for these huge page sizes.  This patch
makes the store functions for nr_hugepages and nr_overcommit_hugepages
return -EINVAL when the pool for a page size >= MAX_ORDER is changed.

[akpm@linux-foundation.org: avoid multiple return paths in nr_hugepages_store_common()]
[caiqian@redhat.com: add checking in hugetlb_overcommit_handler()]
Signed-off-by: Eric B Munson <emunson@mgebm.net>
Reported-by: CAI Qian <caiqian@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

hugetlb: check the return value of string conversion in sysctl handler

proc_doulongvec_minmax may fail if the given buffer doesn't represent a
valid number.  If we provide something invalid we will initialize the
resulting value (nr_overcommit_huge_pages in this case) to a random value
from the stack.

The issue was introduced by a3d0c6aa when the default handler has been
replaced by the helper function where we do not check the return value.

Reproducer:
echo "" > /proc/sys/vm/nr_overcommit_hugepages

[akpm@linux-foundation.org: correctly propagate proc_doulongvec_minmax return code]
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: CAI Qian <caiqian@redhat.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

fs/fs-writeback.c: fix sync_inodes_sb() return value kernel-doc

The sync_inodes_sb() function does not have a return value.  Remove the
outdated documentation comment.

Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/dmapool.c: use TASK_UNINTERRUPTIBLE in dma_pool_alloc()

As it stands this code will degenerate into a busy-wait if the calling task
has signal_pending().

Cc: Rolf Eike Beer <eike-kernel@sf-tec.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/dmapool.c: take lock only once in dma_pool_free()

dma_pool_free() scans for the page to free in the pool list holding the
pool lock.  Then it releases the lock basically to acquire it immediately
again.  Modify the code to only take the lock once.

This will do some additional loops and computations with the lock held in
if memory debugging is activated.  If it is not activated the only new
operations with this lock is one if and one substraction.

Signed-off-by: Rolf Eike Beer <eike-kernel@sf-tec.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/page_alloc.c: simplify calculation of combined index of adjacent buddy lists

The previous approach of calucation of combined index was

page_idx & ~(1 << order))

but we have same result with

page_idx & buddy_idx

This reduces instructions slightly as well as enhances readability.

[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix used-unintialised warning]
Signed-off-by: KyongHo Cho <pullip.cho@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

brk: fix min_brk lower bound computation for COMPAT_BRK

Even if CONFIG_COMPAT_BRK is set in the kernel configuration, it can still
be overriden by randomize_va_space sysctl.

If this is the case, the min_brk computation in sys_brk() implementation
is wrong, as it solely takes into account COMPAT_BRK setting, assuming
that brk start is not randomized.  But that might not be the case if
randomize_va_space sysctl has been set to '2' at the time the binary has
been loaded from disk.

In such case, the check has to be done in a same way as in
!CONFIG_COMPAT_BRK case.

In addition to that, the check for the COMPAT_BRK case introduced back in
a5b4592c ("brk: make sys_brk() honor COMPAT_BRK when computing lower
bound") is slightly wrong -- the lower bound shouldn't be mm->end_code,
but mm->end_data instead, as that's where the legacy applications expect
brk section to start (i.e.  immediately after last global variable).

[akpm@linux-foundation.org: fix comment]
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/hugetlb.c: fix error-path memory leak in nr_hugepages_store_common()

The NODEMASK_ALLOC macro may dynamically allocate memory for its second
argument ('nodes_allowed' in this context).

In nr_hugepages_store_common() we may abort early if strict_strtoul()
fails, but in that case we do not free the memory already allocated to
'nodes_allowed', causing a memory leak.

This patch closes the leak by freeing the memory in the error path.

[akpm@linux-foundation.org: use NODEMASK_FREE, per Minchan Kim]
Signed-off-by: Jesper Juhl <jj@chaosbits.net>
Cc: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: migration: use rcu_dereference_protected when dereferencing the radix tree slot during file page migration

migrate_pages() -> unmap_and_move() only calls rcu_read_lock() for
anonymous pages, as introduced by git commit
989f89c57e6361e7d16fbd9572b5da7d313b073d ("fix rcu_read_lock() in page
migraton").  The point of the RCU protection there is part of getting a
stable reference to anon_vma and is only held for anon pages as file pages
are locked which is sufficient protection against freeing.

However, while a file page's mapping is being migrated, the radix tree is
double checked to ensure it is the expected page.  This uses
radix_tree_deref_slot() -> rcu_dereference() without the RCU lock held
triggering the following warning.

[  173.674290] ===================================================
[  173.676016] [ INFO: suspicious rcu_dereference_check() usage. ]
[  173.676016] ---------------------------------------------------
[  173.676016] include/linux/radix-tree.h:145 invoked rcu_dereference_check() without protection!
[  173.676016]
[  173.676016] other info that might help us debug this:
[  173.676016]
[  173.676016]
[  173.676016] rcu_scheduler_active = 1, debug_locks = 0
[  173.676016] 1 lock held by hugeadm/2899:
[  173.676016]  #0:  (&(&inode->i_data.tree_lock)->rlock){..-.-.}, at: [<c10e3d2b>] migrate_page_move_mapping+0x40/0x1ab
[  173.676016]
[  173.676016] stack backtrace:
[  173.676016] Pid: 2899, comm: hugeadm Not tainted 2.6.37-rc5-autobuild
[  173.676016] Call Trace:
[  173.676016]  [<c128cc01>] ? printk+0x14/0x1b
[  173.676016]  [<c1063502>] lockdep_rcu_dereference+0x7d/0x86
[  173.676016]  [<c10e3db5>] migrate_page_move_mapping+0xca/0x1ab
[  173.676016]  [<c10e41ad>] migrate_page+0x23/0x39
[  173.676016]  [<c10e491b>] buffer_migrate_page+0x22/0x107
[  173.676016]  [<c10e48f9>] ? buffer_migrate_page+0x0/0x107
[  173.676016]  [<c10e425d>] move_to_new_page+0x9a/0x1ae
[  173.676016]  [<c10e47e6>] migrate_pages+0x1e7/0x2fa

This patch introduces radix_tree_deref_slot_protected() which calls
rcu_dereference_protected().  Users of it must pass in the
mapping->tree_lock that is protecting this dereference.  Holding the tree
lock protects against parallel updaters of the radix tree meaning that
rcu_dereference_protected is allowable.

[akpm@linux-foundation.org: remove unneeded casts]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Milton Miller <miltonm@bga.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org> [2.6.37.early]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: add compound_trans_head() helper

Cleanup some code with common compound_trans_head helper.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Avi Kivity <avi@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: KSM on THP

This makes KSM full operational with THP pages.  Subpages are scanned
while the hugepage is still in place and delivering max cpu performance,
and only if there's a match and we're going to deduplicate memory, the
single hugepages with the subpage match is split.

There will be no false sharing between ksmd and khugepaged.  khugepaged
won't collapse 2m virtual regions with KSM pages inside.  ksmd also should
only split pages when the checksum matches and we're likely to split an
hugepage for some long living ksm page (usual ksm heuristic to avoid
sharing pages that get de-cowed).

Signed-off-by: Andrea Arcangeli <aarcange@redhat.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>

thp: khugepaged: make khugepaged aware about madvise

MADV_HUGEPAGE and MADV_NOHUGEPAGE were fully effective only if run after
mmap and before touching the memory.  While this is enough for most
usages, it's little effort to make madvise more dynamic at runtime on an
existing mapping by making khugepaged aware about madvise.

MADV_HUGEPAGE: register in khugepaged immediately without waiting a page
fault (that may not ever happen if all pages are already mapped and the
"enabled" knob was set to madvise during the initial page faults).

MADV_NOHUGEPAGE: skip vmas marked VM_NOHUGEPAGE in khugepaged to stop
collapsing pages where not needed.

[akpm@linux-foundation.org: tweak comment]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: madvise(MADV_NOHUGEPAGE)

Add madvise MADV_NOHUGEPAGE to mark regions that are not important to be
hugepage backed.  Return -EINVAL if the vma is not of an anonymous type,
or the feature isn't built into the kernel.  Never silently return
success.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: mm: define MADV_NOHUGEPAGE

Define MADV_NOHUGEPAGE.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: compound_trans_order

Read compound_trans_order safe. Noop for CONFIG_TRANSPARENT_HUGEPAGE=n.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
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>

thp: fix memory-failure hugetlbfs vs THP collision

hugetlbfs was changed to allow memory failure to migrate the hugetlbfs
pages and that broke THP as split_huge_page was then called on hugetlbfs
pages too.

compound_head/order was also run unsafe on THP pages that can be splitted
at any time.

All compound_head() invocations in memory-failure.c that are run on pages
that aren't pinned and that can be freed and reused from under us (while
compound_head is running) are buggy because compound_head can return a
dangling pointer, but I'm not fixing this as this is a generic
memory-failure bug not specific to THP but it applies to hugetlbfs too, so
I can fix it later after THP is merged upstream.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: add debug checks for mapcount related invariants

Add debug checks for invariants that if broken could lead to mapcount vs
page_mapcount debug checks to trigger later in split_huge_page.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: transparent hugepage sysfs meminfo

Add hugepage statistics to per-node sysfs meminfo

Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: scale nr_rotated to balance memory pressure

Make sure we scale up nr_rotated when we encounter a referenced
transparent huge page.  This ensures pageout scanning balance is not
distorted when there are huge pages on the LRU.

Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: fix anon memory statistics with transparent hugepages

Count each transparent hugepage as HPAGE_PMD_NR pages in the LRU
statistics, so the Active(anon) and Inactive(anon) statistics in
/proc/meminfo are correct.

Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: disable transparent hugepages by default on small systems

On small systems, the extra memory used by the anti-fragmentation memory
reserve and simply because huge pages are smaller than large pages can
easily outweigh the benefits of less TLB misses.

A less obvious concern is if run on a NUMA machine with asymmetric node
sizes and one of them is very small.  The reserve could make the node
unusable.

In case of the crashdump kernel, OOMs have been observed due to the
anti-fragmentation memory reserve taking up a large fraction of the
crashdump image.

This patch disables transparent hugepages on systems with less than 1GB of
RAM, but the hugepage subsystem is fully initialized so administrators can
enable THP through /sys if desired.

Signed-off-by: Rik van Riel <riel@redhat.com>
Acked-by: Avi Kiviti <avi@redhat.com>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: use compaction for all allocation orders

It makes no sense not to enable compaction for small order pages as we
don't want to end up with bad order 2 allocations and good and graceful
order 9 allocations.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: use compaction in kswapd for GFP_ATOMIC order > 0

This takes advantage of memory compaction to properly generate pages of
order > 0 if regular page reclaim fails and priority level becomes more
severe and we don't reach the proper watermarks.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: freeze khugepaged and ksmd

It's unclear why schedule friendly kernel threads can't be taken away by
the CPU through the scheduler itself.  It's safer to stop them as they can
trigger memory allocation, if kswapd also freezes itself to avoid
generating I/O they have too.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: mmu_notifier_test_young

For GRU and EPT, we need gup-fast to set referenced bit too (this is why
it's correct to return 0 when shadow_access_mask is zero, it requires
gup-fast to set the referenced bit).  qemu-kvm access already sets the
young bit in the pte if it isn't zero-copy, if it's zero copy or a shadow
paging EPT minor fault we relay on gup-fast to signal the page is in
use...

We also need to check the young bits on the secondary pagetables for NPT
and not nested shadow mmu as the data may never get accessed again by the
primary pte.

Without this closer accuracy, we'd have to remove the heuristic that
avoids collapsing hugepages in hugepage virtual regions that have not even
a single subpage in use.

->test_young is full backwards compatible with GRU and other usages that
don't have young bits in pagetables set by the hardware and that should
nuke the secondary mmu mappings when ->clear_flush_young runs just like
EPT does.

Removing the heuristic that checks the young bit in
khugepaged/collapse_huge_page completely isn't so bad either probably but
I thought it was worth it and this makes it reliable.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: don't allow transparent hugepage support without PSE

Archs implementing Transparent Hugepage Support must implement a function
called has_transparent_hugepage to be sure the virtual or physical CPU
supports Transparent Hugepages.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: avoid breaking huge pmd invariants in case of vma_adjust failures

An huge pmd can only be mapped if the corresponding 2M virtual range is
fully contained in the vma.  At times the VM calls split_vma twice, if the
first split_vma succeeds and the second fail, the first split_vma remains
in effect and it's not rolled back.  For split_vma or vma_adjust to fail
an allocation failure is needed so it's a very unlikely event (the out of
memory killer would normally fire before any allocation failure is visible
to kernel and userland and if an out of memory condition happens it's
unlikely to happen exactly here).  Nevertheless it's safer to ensure that
no huge pmd can be left around if the vma is adjusted in a way that can't
fit hugepages anymore at the new vm_start/vm_end address.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: transhuge isolate_migratepages()

It's not worth migrating transparent hugepages during compaction.  Those
hugepages don't create fragmentation.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: select CONFIG_COMPACTION if TRANSPARENT_HUGEPAGE enabled

With transparent hugepage support we need compaction for the "defrag"
sysfs controls to be effective.

At the moment THP hangs the system if COMPACTION isn't selected, as
without COMPACTION lumpy reclaim wouldn't be entirely disabled.  So at the
moment it's not orthogonal.  When lumpy will be removed from the VM I can
remove the select COMPACTION in theory, but then 99% of THP users would be
still doing a mistake in disabling compaction, even if the mistake won't
return in fatal runtime but just slightly degraded performance.  So from a
theoretical standpoing forcing the below select is not needed (the
dependency isn't strict nor at compile time nor at runtime) but from a
practical standpoint it is safer.

If anybody really wants THP to run without compaction, it'd be such a
weird setup that editing the Kconfig file to allow it will be surely not a
problem.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: transparent hugepage config choice

Allow to choose between the always|madvise default for page faults and
khugepaged at config time.  madvise guarantees zero risk of higher memory
footprint for applications (applications using madvise(MADV_HUGEPAGE)
won't risk to use any more memory by backing their virtual regions with
hugepages).

Initially set the default to N and don't depend on EMBEDDED.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: allocate memory in khugepaged outside of mmap_sem write mode

This tries to be more friendly to filesystem in userland, with userland
backends that allocate memory in the I/O paths and that could deadlock if
khugepaged holds the mmap_sem write mode of the userland backend while
allocating memory.  Memory allocation may wait for writeback I/O
completion from the daemon that may be blocked in the mmap_sem read mode
if a page fault happens and the daemon wasn't using mlock for the memory
required for the I/O submission and completion.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: add numa awareness to hugepage allocations

It's mostly a matter of replacing alloc_pages with alloc_pages_vma after
introducing alloc_pages_vma.  khugepaged needs special handling as the
allocation has to happen inside collapse_huge_page where the vma is known
and an error has to be returned to the outer loop to sleep
alloc_sleep_millisecs in case of failure.  But it retains the more
efficient logic of handling allocation failures in khugepaged in case of
CONFIG_NUMA=n.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: enable direct defrag

With memory compaction in, and lumpy-reclaim disabled, it seems safe
enough to defrag memory during the (synchronous) transparent hugepage page
faults (TRANSPARENT_HUGEPAGE_DEFRAG_FLAG) and not only during khugepaged
(async) hugepage allocations that was already enabled even before memory
compaction was in (TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG).

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: set recommended min free kbytes

If transparent hugepage is enabled initialize min_free_kbytes to an
optimal value by default.  This moves the hugeadm algorithm in kernel.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: mprotect: transparent huge page support

Natively handle huge pmds when changing page tables on behalf of
mprotect().

I left out update_mmu_cache() because we do not need it on x86 anyway but
more importantly the interface works on ptes, not pmds.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: mprotect: pass vma down to page table walkers

Flushing the tlb for huge pmds requires the vma's anon_vma, so pass along
the vma instead of the mm, we can always get the latter when we need it.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: add pmd_modify

Add pmd_modify() for use with mprotect() on huge pmds.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: mincore transparent hugepage support

Handle transparent huge page pmd entries natively instead of splitting
them into subpages.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: add x86 32bit support

Add support for transparent hugepages to x86 32bit.

Share the same VM_ bitflag for VM_MAPPED_COPY.  mm/nommu.c will never
support transparent hugepages.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: remove PG_buddy

PG_buddy can be converted to _mapcount == -2.  So the PG_compound_lock can
be added to page->flags without overflowing (because of the sparse section
bits increasing) with CONFIG_X86_PAE=y and CONFIG_X86_PAT=y.  This also
has to move the memory hotplug code from _mapcount to lru.next to avoid
any risk of clashes.  We can't use lru.next for PG_buddy removal, but
memory hotplug can use lru.next even more easily than the mapcount
instead.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: skip transhuge pages in ksm for now

Skip transhuge pages in ksm for now.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: khugepaged vma merge

register in khugepaged if the vma grows.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: khugepaged

Add khugepaged to relocate fragmented pages into hugepages if new
hugepages become available.  (this is indipendent of the defrag logic that
will have to make new hugepages available)

The fundamental reason why khugepaged is unavoidable, is that some memory
can be fragmented and not everything can be relocated.  So when a virtual
machine quits and releases gigabytes of hugepages, we want to use those
freely available hugepages to create huge-pmd in the other virtual
machines that may be running on fragmented memory, to maximize the CPU
efficiency at all times.  The scan is slow, it takes nearly zero cpu time,
except when it copies data (in which case it means we definitely want to
pay for that cpu time) so it seems a good tradeoff.

In addition to the hugepages being released by other process releasing
memory, we have the strong suspicion that the performance impact of
potentially defragmenting hugepages during or before each page fault could
lead to more performance inconsistency than allocating small pages at
first and having them collapsed into large pages later...  if they prove
themselfs to be long lived mappings (khugepaged scan is slow so short
lived mappings have low probability to run into khugepaged if compared to
long lived mappings).

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: transparent hugepage vmstat

Add hugepage stat information to /proc/vmstat and /proc/meminfo.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: memcg huge memory

Add memcg charge/uncharge to hugepage faults in huge_memory.c.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: transhuge-memcg: commit tail pages at charge

By this patch, when a transparent hugepage is charged, not only the head
page but also all the tail pages are committed, IOW pc->mem_cgroup and
pc->flags of tail pages are set.

Without this patch:

- Tail pages are not linked to any memcg's LRU at splitting. This causes many
  problems, for example, the charged memcg's directory can never be rmdir'ed
  because it doesn't have enough pages to scan to make the usage decrease to 0.
- "rss" field in memory.stat would be incorrect. Moreover, usage_in_bytes in
  root cgroup is calculated by the stat not by res_counter(since 2.6.32),
  it would be incorrect too.

Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: memcg compound

Teach memcg to charge/uncharge compound pages.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: pmd_trans_huge migrate bugcheck

No pmd_trans_huge should ever materialize in migration ptes areas, because
we split the hugepage before migration ptes are instantiated.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: madvise(MADV_HUGEPAGE)

Add madvise MADV_HUGEPAGE to mark regions that are important to be
hugepage backed.  Return -EINVAL if the vma is not of an anonymous type,
or the feature isn't built into the kernel.  Never silently return
success.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: verify pmd_trans_huge isn't leaking

pte_trans_huge must not leak in certain vmas like the mmio special pfn or
filebacked mappings.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: split_huge_page anon_vma ordering dependency

This documents how split_huge_page is safe vs new vma inserctions into the
anon_vma that may have already released the anon_vma->lock but not
established pmds yet when split_huge_page starts.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: transparent hugepage core fixlet

If you configure THP in addition to HUGETLB_PAGE on x86_32 without PAE,
the p?d-folding works out that munlock_vma_pages_range() can crash to
follow_page()'s pud_huge() BUG_ON(flags & FOLL_GET): it needs the same
VM_HUGETLB check already there on the pmd_huge() line.  Conveniently,
openSUSE provides a "blogd" which tests this out at startup!

Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.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>

thp: transparent hugepage core

Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs.  Some of the restrictions I'd like to
see removed:

1) hugepages have to be swappable or the guest physical memory remains
   locked in RAM and can't be paged out to swap

2) if a hugepage allocation fails, regular pages should be allocated
   instead and mixed in the same vma without any failure and without
   userland noticing

3) if some task quits and more hugepages become available in the
   buddy, guest physical memory backed by regular pages should be
   relocated on hugepages automatically in regions under
   madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
   kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
   not null)

4) avoidance of reservation and maximization of use of hugepages whenever
   possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
   1 machine with 1 database with 1 database cache with 1 database cache size
   known at boot time. It's definitely not feasible with a virtualization
   hypervisor usage like RHEV-H that runs an unknown number of virtual machines
   with an unknown size of each virtual machine with an unknown amount of
   pagecache that could be potentially useful in the host for guest not using
   O_DIRECT (aka cache=off).

hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...).  Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario.  So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).

The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas.  This is what this patch tries to achieve in the
least intrusive possible way.  We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).

The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails!  This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...

Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail.  This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM.  Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*.  The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle).  In short the
very value of split_huge_page is that it can't fail.

The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon.  It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode.  collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later.  collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).

The default I like is that transparent hugepages are used at page fault
time.  This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled.  The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used.  /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".

The pmd_trans_splitting/pmd_trans_huge locking is very solid.  The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head.  I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view.  In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...).  And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.

If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet).  But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.

Swap and oom works fine (well just like with regular pages ;).  MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.

NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores.  I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks.  One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault).  Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only.  If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot.  If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time.  It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).

This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone.  Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation.  hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits.  hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.

Some performance result:

vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988

============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>

#define SIZE (3UL*1024*1024*1024)

int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;

gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
       (after.tv_sec-before.tv_sec)*1000000UL +
       after.tv_usec-before.tv_usec);

gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
       (after.tv_sec-before.tv_sec)*1000000UL +
       after.tv_usec-before.tv_usec);

gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
       (after.tv_sec-before.tv_sec)*1000000UL +
       after.tv_usec-before.tv_usec);

gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
       (after.tv_sec-before.tv_sec)*1000000UL +
       after.tv_usec-before.tv_usec);

gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
       (after.tv_sec-before.tv_sec)*1000000UL +
       after.tv_usec-before.tv_usec);

return 0;
}
============

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: don't alloc harder for gfp nomemalloc even if nowait

Not worth throwing away the precious reserved free memory pool for
allocations that can fail gracefully (either through mempool or because
they're transhuge allocations later falling back to 4k allocations).

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: _GFP_NO_KSWAPD

Transparent hugepage allocations must be allowed not to invoke kswapd or
any other kind of indirect reclaim (especially when the defrag sysfs is
control disabled).  It's unacceptable to swap out anonymous pages
(potentially anonymous transparent hugepages) in order to create new
transparent hugepages.  This is true for the MADV_HUGEPAGE areas too
(swapping out a kvm virtual machine and so having it suffer an unbearable
slowdown, so another one with guest physical memory marked MADV_HUGEPAGE
can run 30% faster if it is running memory intensive workloads, makes no
sense).  If a transparent hugepage allocation fails the slowdown is minor
and there is total fallback, so kswapd should never be asked to swapout
memory to allow the high order allocation to succeed.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: kvm mmu transparent hugepage support

This should work for both hugetlbfs and transparent hugepages.

[akpm@linux-foundation.org: bring forward PageTransCompound() addition for bisectability]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Avi Kivity <avi@redhat.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: clear_copy_huge_page

Move the copy/clear_huge_page functions to common code to share between
hugetlb.c and huge_memory.c.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: split_huge_page paging

Paging logic that splits the page before it is unmapped and added to swap
to ensure backwards compatibility with the legacy swap code.  Eventually
swap should natively pageout the hugepages to increase performance and
decrease seeking and fragmentation of swap space.  swapoff can just skip
over huge pmd as they cannot be part of swap yet.  In add_to_swap be
careful to split the page only if we got a valid swap entry so we don't
split hugepages with a full swap.

In theory we could split pages before isolating them during the lru scan,
but for khugepaged to be safe, I'm relying on either mmap_sem write mode,
or PG_lock taken, so split_huge_page has to run either with mmap_sem
read/write mode or PG_lock taken.  Calling it from isolate_lru_page would
make locking more complicated, in addition to that split_huge_page would
deadlock if called by __isolate_lru_page because it has to take the lru
lock to add the tail pages.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: split_huge_page_mm/vma

split_huge_page_pmd compat code.  Each one of those would need to be
expanded to hundred of lines of complex code without a fully reliable
split_huge_page_pmd design.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: add pmd_huge_pte to mm_struct

This increase the size of the mm struct a bit but it is needed to
preallocate one pte for each hugepage so that split_huge_page will not
require a fail path.  Guarantee of success is a fundamental property of
split_huge_page to avoid decrasing swapping reliability and to avoid
adding -ENOMEM fail paths that would otherwise force the hugepage-unaware
VM code to learn rolling back in the middle of its pte mangling operations
(if something we need it to learn handling pmd_trans_huge natively rather
being capable of rollback).  When split_huge_page runs a pte is needed to
succeed the split, to map the newly splitted regular pages with a regular
pte.  This way all existing VM code remains backwards compatible by just
adding a split_huge_page* one liner.  The memory waste of those
preallocated ptes is negligible and so it is worth it.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: clear page compound

split_huge_page must transform a compound page to a regular page and needs
ClearPageCompound.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: add pmd mmu_notifier helpers

Add mmu notifier helpers to handle pmd huge operations.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: pte alloc trans splitting

pte alloc routines must wait for split_huge_page if the pmd is not present
and not null (i.e.  pmd_trans_splitting).  The additional branches are
optimized away at compile time by pmd_trans_splitting if the config option
is off.  However we must pass the vma down in order to know the anon_vma
lock to wait for.

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: bail out gup_fast on splitting pmd

Force gup_fast to take the slow path and block if the pmd is splitting,
not only if it's none.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: add pmd mangling functions to x86

Add needed pmd mangling functions with symmetry with their pte
counterparts.  pmdp_splitting_flush() is the only new addition on the pmd_
methods and it's needed to serialize the VM against split_huge_page.  It
simply atomically sets the splitting bit in a similar way
pmdp_clear_flush_young atomically clears the accessed bit.
pmdp_splitting_flush() also has to flush the tlb to make it effective
against gup_fast, but it wouldn't really require to flush the tlb too.
Just the tlb flush is the simplest operation we can invoke to serialize
pmdp_splitting_flush() against gup_fast.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: add pmd mangling generic functions

Some are needed to build but not actually used on archs not supporting
transparent hugepages.  Others like pmdp_clear_flush are used by x86 too.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: special pmd_trans_* functions

These returns 0 at compile time when the config option is disabled, to
allow gcc to eliminate the transparent hugepage function calls at compile
time without additional #ifdefs (only the export of those functions have
to be visible to gcc but they won't be required at link time and
huge_memory.o can be not built at all).

_PAGE_BIT_UNUSED1 is never used for pmd, only on pte.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: CONFIG_TRANSPARENT_HUGEPAGE

Add config option.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: comment reminder in destroy_compound_page

Warn destroy_compound_page that __split_huge_page_refcount is heavily
dependent on its internal behavior.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: export maybe_mkwrite

huge_memory.c needs it too when it fallbacks in copying hugepages into
regular fragmented pages if hugepage allocation fails during COW.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: no paravirt version of pmd ops

No paravirt version of set_pmd_at/pmd_update/pmd_update_defer.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: add pmd paravirt ops

Paravirt ops pmd_update/pmd_update_defer/pmd_set_at.  Not all might be
necessary (vmware needs pmd_update, Xen needs set_pmd_at, nobody needs
pmd_update_defer), but this is to keep full simmetry with pte paravirt
ops, which looks cleaner and simpler from a common code POV.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: add native_set_pmd_at

Used by paravirt and not paravirt set_pmd_at.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: clear compound mapping

Clear compound mapping for anonymous compound pages like it already
happens for regular anonymous pages.  But crash if mapping is set for any
tail page, also the PageAnon check is meaningless for tail pages.  This
check only makes sense for the head page, for tail page it can only hide
bugs and we definitely don't want to hide bugs.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: update futex compound knowledge

Futex code is smarter than most other gup_fast O_DIRECT code and knows
about the compound internals.  However now doing a put_page(head_page)
will not release the pin on the tail page taken by gup-fast, leading to
all sort of refcounting bugchecks.  Getting a stable head_page is a little
tricky.

page_head = page is there because if this is not a tail page it's also the
page_head.  Only in case this is a tail page, compound_head is called,
otherwise it's guaranteed unnecessary.  And if it's a tail page
compound_head has to run atomically inside irq disabled section
__get_user_pages_fast before returning.  Otherwise ->first_page won't be a
stable pointer.

Disableing irq before __get_user_page_fast and releasing irq after running
compound_head is needed because if __get_user_page_fast returns == 1, it
means the huge pmd is established and cannot go away from under us.
pmdp_splitting_flush_notify in __split_huge_page_splitting will have to
wait for local_irq_enable before the IPI delivery can return.  This means
__split_huge_page_refcount can't be running from under us, and in turn
when we run compound_head(page) we're not reading a dangling pointer from
tailpage->first_page.  Then after we get to stable head page, we are
always safe to call compound_lock and after taking the compound lock on
head page we can finally re-check if the page returned by gup-fast is
still a tail page.  in which case we're set and we didn't need to split
the hugepage in order to take a futex on it.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: put_page: recheck PageHead after releasing the compound_lock

After releasing the compound_lock split_huge_page can still run and release the
page before put_page_testzero runs.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: alter compound get_page/put_page

Alter compound get_page/put_page to keep references on subpages too, in
order to allow __split_huge_page_refcount to split an hugepage even while
subpages have been pinned by one of the get_user_pages() variants.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: compound_lock

Add a new compound_lock() needed to serialize put_page against
__split_huge_page_refcount().

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: mm: define MADV_HUGEPAGE

Define MADV_HUGEPAGE.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: transparent hugepage support documentation

Documentation/vm/transhuge.txt

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: fix bad_page to show the real reason the page is bad

page_count shows the count of the head page, but the actual check is done
on the tail page, so show what is really being checked.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

thp: ksm: free swap when swapcache page is replaced

When a swapcache page is replaced by a ksm page, it's best to free that
swap immediately.

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

writeback: avoid unnecessary determine_dirtyable_memory call

I think determine_dirtyable_memory() is a rather costly function since it
need many atomic reads for gathering zone/global page state.  But when we
use vm_dirty_bytes && dirty_background_bytes, we don't need that costly
calculation.

This patch eliminates such unnecessary overhead.

NOTE : newly added if condition might add overhead in normal path.
       But it should be _really_ small because anyway we need the
       access both vm_dirty_bytes and dirty_background_bytes so it is
       likely to hit the cache.

[akpm@linux-foundation.org: fix used-uninitialised warning]
Signed-off-by: Minchan Kim <minchan.kim@gmail.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: set correct numa_zonelist_order string when configured on the kernel command line

When numa_zonelist_order parameter is set to "node" or "zone" on the
command line it's still showing as "default" in sysctl.  That's because
early_param parsing function changes only user_zonelist_order variable.
Fix this by copying user-provided string to numa_zonelist_order if it was
successfully parsed.

Signed-off-by: Volodymyr G Lukiianyk <volodymyrgl@gmail.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: kswapd: use the classzone idx that kswapd was using for sleeping_prematurely()

When kswapd is woken up for a high-order allocation, it takes account of
the highest usable zone by the caller (the classzone idx).  During
allocation, this index is used to select the lowmem_reserve[] that should
be applied to the watermark calculation in zone_watermark_ok().

When balancing a node, kswapd considers the highest unbalanced zone to be
the classzone index.  This will always be at least be the callers
classzone_idx and can be higher.  However, sleeping_prematurely() always
considers the lowest zone (e.g.  ZONE_DMA) to be the classzone index.
This means that sleeping_prematurely() can consider a zone to be balanced
that is unusable by the allocation request that originally woke kswapd.
This patch changes sleeping_prematurely() to use a classzone_idx matching
the value it used in balance_pgdat().

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: Eric B Munson <emunson@mgebm.net>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Simon Kirby <sim@hostway.ca>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: kswapd: treat zone->all_unreclaimable in sleeping_prematurely similar to balance_pgdat()

After DEF_PRIORITY, balance_pgdat() considers all_unreclaimable zones to
be balanced but sleeping_prematurely does not.  This can force kswapd to
stay awake longer than it should.  This patch fixes it.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Eric B Munson <emunson@mgebm.net>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Simon Kirby <sim@hostway.ca>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: kswapd: reset kswapd_max_order and classzone_idx after reading

When kswapd wakes up, it reads its order and classzone from pgdat and
calls balance_pgdat.  While its awake, it potentially reclaimes at a high
order and a low classzone index.  This might have been a once-off that was
not required by subsequent callers.  However, because the pgdat values
were not reset, they remain artifically high while balance_pgdat() is
running and potentially kswapd enters a second unnecessary reclaim cycle.
Reset the pgdat order and classzone index after reading.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Eric B Munson <emunson@mgebm.net>
Cc: Simon Kirby <sim@hostway.ca>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: kswapd: use the order that kswapd was reclaiming at for sleeping_prematurely()

Before kswapd goes to sleep, it uses sleeping_prematurely() to check if
there was a race pushing a zone below its watermark.  If the race
happened, it stays awake.  However, balance_pgdat() can decide to reclaim
at order-0 if it decides that high-order reclaim is not working as
expected.  This information is not passed back to sleeping_prematurely().
The impact is that kswapd remains awake reclaiming pages long after it
should have gone to sleep.  This patch passes the adjusted order to
sleeping_prematurely and uses the same logic as balance_pgdat to decide if
it's ok to go to sleep.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Eric B Munson <emunson@mgebm.net>
Cc: Simon Kirby <sim@hostway.ca>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: kswapd: keep kswapd awake for high-order allocations until a percentage of the node is balanced

When reclaiming for high-orders, kswapd is responsible for balancing a
node but it should not reclaim excessively.  It avoids excessive reclaim
by considering if any zone in a node is balanced then the node is
balanced.  In the cases where there are imbalanced zone sizes (e.g.
ZONE_DMA with both ZONE_DMA32 and ZONE_NORMAL), kswapd can go to sleep
prematurely as just one small zone was balanced.

This alters the sleep logic of kswapd slightly.  It counts the number of
pages that make up the balanced zones.  If the total number of balanced
pages is more than a quarter of the zone, kswapd will go back to sleep.
This should keep a node balanced without reclaiming an excessive number of
pages.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Eric B Munson <emunson@mgebm.net>
Cc: Simon Kirby <sim@hostway.ca>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: kswapd: stop high-order balancing when any suitable zone is balanced

Simon Kirby reported the following problem

   We're seeing cases on a number of servers where cache never fully
   grows to use all available memory.  Sometimes we see servers with 4 GB
   of memory that never seem to have less than 1.5 GB free, even with a
   constantly-active VM.  In some cases, these servers also swap out while
   this happens, even though they are constantly reading the working set
   into memory.  We have been seeing this happening for a long time; I
   don't think it's anything recent, and it still happens on 2.6.36.

After some debugging work by Simon, Dave Hansen and others, the prevaling
theory became that kswapd is reclaiming order-3 pages requested by SLUB
too aggressive about it.

There are two apparent problems here.  On the target machine, there is a
small Normal zone in comparison to DMA32.  As kswapd tries to balance all
zones, it would continually try reclaiming for Normal even though DMA32
was balanced enough for callers.  The second problem is that
sleeping_prematurely() does not use the same logic as balance_pgdat() when
deciding whether to sleep or not.  This keeps kswapd artifically awake.

A number of tests were run and the figures from previous postings will
look very different for a few reasons.  One, the old figures were forcing
my network card to use GFP_ATOMIC in attempt to replicate Simon's problem.
 Second, I previous specified slub_min_order=3 again in an attempt to
reproduce Simon's problem.  In this posting, I'm depending on Simon to say
whether his problem is fixed or not and these figures are to show the
impact to the ordinary cases.  Finally, the "vmscan" figures are taken
from /proc/vmstat instead of the tracepoints.  There is less information
but recording is less disruptive.

The first test of relevance was postmark with a process running in the
background reading a large amount of anonymous memory in blocks.  The
objective was to vaguely simulate what was happening on Simon's machine
and it's memory intensive enough to have kswapd awake.

POSTMARK
                                            traceonly          kanyzone
Transactions per second:              156.00 ( 0.00%)   153.00 (-1.96%)
Data megabytes read per second:        21.51 ( 0.00%)    21.52 ( 0.05%)
Data megabytes written per second:     29.28 ( 0.00%)    29.11 (-0.58%)
Files created alone per second:       250.00 ( 0.00%)   416.00 (39.90%)
Files create/transact per second:      79.00 ( 0.00%)    76.00 (-3.95%)
Files deleted alone per second:       520.00 ( 0.00%)   420.00 (-23.81%)
Files delete/transact per second:      79.00 ( 0.00%)    76.00 (-3.95%)

MMTests Statistics: duration
User/Sys Time Running Test (seconds)         16.58      17.4
Total Elapsed Time (seconds)                218.48    222.47

VMstat Reclaim Statistics: vmscan
Direct reclaims                                  0          4
Direct reclaim pages scanned                     0        203
Direct reclaim pages reclaimed                   0        184
Kswapd pages scanned                        326631     322018
Kswapd pages reclaimed                      312632     309784
Kswapd low wmark quickly                         1          4
Kswapd high wmark quickly                      122        475
Kswapd skip congestion_wait                      1          0
Pages activated                             700040     705317
Pages deactivated                           212113     203922
Pages written                                 9875       6363

Total pages scanned                         326631    322221
Total pages reclaimed                       312632    309968
%age total pages scanned/reclaimed          95.71%    96.20%
%age total pages scanned/written             3.02%     1.97%

proc vmstat: Faults
Major Faults                                   300       254
Minor Faults                                645183    660284
Page ins                                    493588    486704
Page outs                                  4960088   4986704
Swap ins                                      1230       661
Swap outs                                     9869      6355

Performance is mildly affected because kswapd is no longer doing as much
work and the background memory consumer process is getting in the way.
Note that kswapd scanned and reclaimed fewer pages as it's less aggressive
and overall fewer pages were scanned and reclaimed.  Swap in/out is
particularly reduced again reflecting kswapd throwing out fewer pages.

The slight performance impact is unfortunate here but it looks like a
direct result of kswapd being less aggressive.  As the bug report is about
too many pages being freed by kswapd, it may have to be accepted for now.

The second test is a streaming IO benchmark that was previously used by
Johannes to show regressions in page reclaim.

MICRO
 traceonly  kanyzone
User/Sys Time Running Test (seconds)         29.29     28.87
Total Elapsed Time (seconds)                492.18    488.79

VMstat Reclaim Statistics: vmscan
Direct reclaims                               2128       1460
Direct reclaim pages scanned               2284822    1496067
Direct reclaim pages reclaimed              148919     110937
Kswapd pages scanned                      15450014   16202876
Kswapd pages reclaimed                     8503697    8537897
Kswapd low wmark quickly                      3100       3397
Kswapd high wmark quickly                     1860       7243
Kswapd skip congestion_wait                    708        801
Pages activated                               9635       9573
Pages deactivated                             1432       1271
Pages written                                  223       1130

Total pages scanned                       17734836  17698943
Total pages reclaimed                      8652616   8648834
%age total pages scanned/reclaimed          48.79%    48.87%
%age total pages scanned/written             0.00%     0.01%

proc vmstat: Faults
Major Faults                                   165       221
Minor Faults                               9655785   9656506
Page ins                                      3880      7228
Page outs                                 37692940  37480076
Swap ins                                         0        69
Swap outs                                       19        15

Again fewer pages are scanned and reclaimed as expected and this time the
test completed faster.  Note that kswapd is hitting its watermarks faster
(low and high wmark quickly) which I expect is due to kswapd reclaiming
fewer pages.

I also ran fs-mark, iozone and sysbench but there is nothing interesting
to report in the figures.  Performance is not significantly changed and
the reclaim statistics look reasonable.

Tgis patch:

When the allocator enters its slow path, kswapd is woken up to balance the
node.  It continues working until all zones within the node are balanced.
For order-0 allocations, this makes perfect sense but for higher orders it
can have unintended side-effects.  If the zone sizes are imbalanced,
kswapd may reclaim heavily within a smaller zone discarding an excessive
number of pages.  The user-visible behaviour is that kswapd is awake and
reclaiming even though plenty of pages are free from a suitable zone.

This patch alters the "balance" logic for high-order reclaim allowing
kswapd to stop if any suitable zone becomes balanced to reduce the number
of pages it reclaims from other zones.  kswapd still tries to ensure that
order-0 watermarks for all zones are met before sleeping.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Eric B Munson <emunson@mgebm.net>
Cc: Simon Kirby <sim@hostway.ca>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: remove likely() from grab_cache_page_write_begin()

Running the annotated branch profiler on a box doing average work
(firefox, evolution, xchat, distcc farm), the likely() used in
grab_cache_page_write_begin() was incorrect most of the time:

 correct incorrect  %        Function                  File              Line
 ------- ---------  -        --------                  ----              ----
 1924262 71332401  97 grab_cache_page_write_begin    filemap.c           2206

Adding a trace_printk() and running the function tracer limited to
just this function I can see:

        gconfd-2-2696  [000]  4467.268935: grab_cache_page_write_begin: page=          (null) mapping=ffff8800676a9460 index=7
        gconfd-2-2696  [000]  4467.268946: grab_cache_page_write_begin <-ext3_write_begin
        gconfd-2-2696  [000]  4467.268947: grab_cache_page_write_begin: page=          (null) mapping=ffff8800676a9460 index=8
        gconfd-2-2696  [000]  4467.268959: grab_cache_page_write_begin <-ext3_write_begin
        gconfd-2-2696  [000]  4467.268960: grab_cache_page_write_begin: page=          (null) mapping=ffff8800676a9460 index=9
        gconfd-2-2696  [000]  4467.268972: grab_cache_page_write_begin <-ext3_write_begin
        gconfd-2-2696  [000]  4467.268973: grab_cache_page_write_begin: page=          (null) mapping=ffff8800676a9460 index=10
        gconfd-2-2696  [000]  4467.268991: grab_cache_page_write_begin <-ext3_write_begin
        gconfd-2-2696  [000]  4467.268992: grab_cache_page_write_begin: page=          (null) mapping=ffff8800676a9460 index=11
        gconfd-2-2696  [000]  4467.269005: grab_cache_page_write_begin <-ext3_write_begin

Which shows that a lot of calls from ext3_write_begin will result in the
page returned by "find_lock_page" will be NULL.

Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Acked-by: Nick Piggin <npiggin@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>