ksm: make !merge_across_nodes migration safe

The new KSM NUMA merge_across_nodes knob introduces a problem, when it's
set to non-default 0: if a KSM page is migrated to a different NUMA node,
how do we migrate its stable node to the right tree?  And what if that
collides with an existing stable node?

ksm_migrate_page() can do no more than it's already doing, updating
stable_node->kpfn: the stable tree itself cannot be manipulated without
holding ksm_thread_mutex.  So accept that a stable tree may temporarily
indicate a page belonging to the wrong NUMA node, leave updating until the
next pass of ksmd, just be careful not to merge other pages on to a
misplaced page.  Note nid of holding tree in stable_node, and recognize
that it will not always match nid of kpfn.

A misplaced KSM page is discovered, either when ksm_do_scan() next comes
around to one of its rmap_items (we now have to go to cmp_and_merge_page
even on pages in a stable tree), or when stable_tree_search() arrives at a
matching node for another page, and this node page is found misplaced.

In each case, move the misplaced stable_node to a list of migrate_nodes
(and use the address of migrate_nodes as magic by which to identify them):
we don't need them in a tree.  If stable_tree_search() finds no match for
a page, but it's currently exiled to this list, then slot its stable_node
right there into the tree, bringing all of its mappings with it; otherwise
they get migrated one by one to the original page of the colliding node.
stable_tree_search() is now modelled more like stable_tree_insert(), in
order to handle these insertions of migrated nodes.

remove_node_from_stable_tree(), remove_all_stable_nodes() and
ksm_check_stable_tree() have to handle the migrate_nodes list as well as
the stable tree itself.  Less obviously, we do need to prune the list of
stale entries from time to time (scan_get_next_rmap_item() does it once
each full scan): whereas stale nodes in the stable tree get naturally
pruned as searches try to brush past them, these migrate_nodes may get
forgotten and accumulate.

Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Petr Holasek <pholasek@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Izik Eidus <izik.eidus@ravellosystems.com>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ksm: make KSM page migration possible

KSM page migration is already supported in the case of memory hotremove,
which takes the ksm_thread_mutex across all its migrations to keep life
simple.

But the new KSM NUMA merge_across_nodes knob introduces a problem, when
it's set to non-default 0: if a KSM page is migrated to a different NUMA
node, how do we migrate its stable node to the right tree?  And what if
that collides with an existing stable node?

So far there's no provision for that, and this patch does not attempt to
deal with it either.  But how will I test a solution, when I don't know
how to hotremove memory?  The best answer is to enable KSM page migration
in all cases now, and test more common cases.  With THP and compaction
added since KSM came in, page migration is now mainstream, and it's a
shame that a KSM page can frustrate freeing a page block.

Without worrying about merge_across_nodes 0 for now, this patch gets KSM
page migration working reliably for default merge_across_nodes 1 (but
leave the patch enabling it until near the end of the series).

It's much simpler than I'd originally imagined, and does not require an
additional tier of locking: page migration relies on the page lock, KSM
page reclaim relies on the page lock, the page lock is enough for KSM page
migration too.

Almost all the care has to be in get_ksm_page(): that's the function which
worries about when a stable node is stale and should be freed, now it also
has to worry about the KSM page being migrated.

The only new overhead is an additional put/get/lock/unlock_page when
stable_tree_search() arrives at a matching node: to make sure migration
respects the raised page count, and so does not migrate the page while
we're busy with it here.  That's probably avoidable, either by changing
internal interfaces from using kpage to stable_node, or by moving the
ksm_migrate_page() callsite into a page_freeze_refs() section (even if not
swapcache); but this works well, I've no urge to pull it apart now.

(Descents of the stable tree may pass through nodes whose KSM pages are
under migration: being unlocked, the raised page count does not prevent
that, nor need it: it's safe to memcmp against either old or new page.)

You might worry about mremap, and whether page migration's rmap_walk to
remove migration entries will find all the KSM locations where it inserted
earlier: that should already be handled, by the satisfyingly heavy hammer
of move_vma()'s call to ksm_madvise(,,,MADV_UNMERGEABLE,).

Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Petr Holasek <pholasek@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Izik Eidus <izik.eidus@ravellosystems.com>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ksm: remove old stable nodes more thoroughly

Switching merge_across_nodes after running KSM is liable to oops on stale
nodes still left over from the previous stable tree.  It's not something
that people will often want to do, but it would be lame to demand a reboot
when they're trying to determine which merge_across_nodes setting is best.

How can this happen?  We only permit switching merge_across_nodes when
pages_shared is 0, and usually set run 2 to force that beforehand, which
ought to unmerge everything: yet oopses still occur when you then run 1.

Three causes:

1. The old stable tree (built according to the inverse
   merge_across_nodes) has not been fully torn down.  A stable node
   lingers until get_ksm_page() notices that the page it references no
   longer references it: but the page is not necessarily freed as soon as
   expected, particularly when swapcache.

   Fix this with a pass through the old stable tree, applying
   get_ksm_page() to each of the remaining nodes (most found stale and
   removed immediately), with forced removal of any left over.  Unless the
   page is still mapped: I've not seen that case, it shouldn't occur, but
   better to WARN_ON_ONCE and EBUSY than BUG.

2. __ksm_enter() has a nice little optimization, to insert the new mm
   just behind ksmd's cursor, so there's a full pass for it to stabilize
   (or be removed) before ksmd addresses it.  Nice when ksmd is running,
   but not so nice when we're trying to unmerge all mms: we were missing
   those mms forked and inserted behind the unmerge cursor.  Easily fixed
   by inserting at the end when KSM_RUN_UNMERGE.

3.  It is possible for a KSM page to be faulted back from swapcache
   into an mm, just after unmerge_and_remove_all_rmap_items() scanned past
   it.  Fix this by copying on fault when KSM_RUN_UNMERGE: but that is
   private to ksm.c, so dissolve the distinction between
   ksm_might_need_to_copy() and ksm_does_need_to_copy(), doing it all in
   the one call into ksm.c.

A long outstanding, unrelated bugfix sneaks in with that third fix:
ksm_does_need_to_copy() would copy from a !PageUptodate page (implying I/O
error when read in from swap) to a page which it then marks Uptodate.  Fix
this case by not copying, letting do_swap_page() discover the error.

Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Petr Holasek <pholasek@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Izik Eidus <izik.eidus@ravellosystems.com>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ksm: get_ksm_page locked

In some places where get_ksm_page() is used, we need the page to be locked.

When KSM migration is fully enabled, we shall want that to make sure that
the page just acquired cannot be migrated beneath us (raised page count is
only effective when there is serialization to make sure migration
notices).  Whereas when navigating through the stable tree, we certainly
do not want to lock each node (raised page count is enough to guarantee
the memcmps, even if page is migrated to another node).

Since we're about to add another use case, add the locked argument to
get_ksm_page() now.

Hmm, what's that rcu_read_lock() about?  Complete misunderstanding, I
really got the wrong end of the stick on that!  There's a configuration in
which page_cache_get_speculative() can do something cheaper than
get_page_unless_zero(), relying on its caller's rcu_read_lock() to have
disabled preemption for it.  There's no need for rcu_read_lock() around
get_page_unless_zero() (and mapping checks) here.  Cut out that silliness
before making this any harder to understand.

Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Petr Holasek <pholasek@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Izik Eidus <izik.eidus@ravellosystems.com>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ksm: reorganize ksm_check_stable_tree

Memory hotremove's ksm_check_stable_tree() is pitifully inefficient
(restarting whenever it finds a stale node to remove), but rearrange so
that at least it does not needlessly restart from nid 0 each time.  And
add a couple of comments: here is why we keep pfn instead of page.

Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Petr Holasek <pholasek@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Izik Eidus <izik.eidus@ravellosystems.com>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ksm: trivial tidyups

Add NUMA() and DO_NUMA() macros to minimize blight of #ifdef
CONFIG_NUMAs (but indeed we don't want to expand struct rmap_item by nid
when not NUMA).  Add comment, remove "unsigned" from rmap_item->nid, as
"int nid" elsewhere.  Define ksm_merge_across_nodes 1U when #ifndef NUMA
to help optimizing out.  Use ?: in get_kpfn_nid().  Adjust a few
comments noticed in ongoing work.

Leave stable_tree_insert()'s rb_linkage until after the node has been
set up, as unstable_tree_search_insert() does: ksm_thread_mutex and page
lock make either way safe, but we're going to copy and I prefer this
precedent.

Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Petr Holasek <pholasek@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Izik Eidus <izik.eidus@ravellosystems.com>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ksm: add sysfs ABI Documentation

Add sysfs documentation for Kernel Samepage Merging (KSM) including new
merge_across_nodes knob.

Signed-off-by: Petr Holasek <pholasek@redhat.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Izik Eidus <izik.eidus@ravellosystems.com>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ksm: allow trees per NUMA node

Here's a KSM series, based on mmotm 2013-01-23-17-04: starting with
Petr's v7 "KSM: numa awareness sysfs knob"; then fixing the two issues
we had with that, fully enabling KSM page migration on the way.

(A different kind of KSM/NUMA issue which I've certainly not begun to
address here: when KSM pages are unmerged, there's usually no sense in
preferring to allocate the new pages local to the caller's node.)

This patch:

Introduces new sysfs boolean knob /sys/kernel/mm/ksm/merge_across_nodes
which control merging pages across different numa nodes.  When it is set
to zero only pages from the same node are merged, otherwise pages from
all nodes can be merged together (default behavior).

Typical use-case could be a lot of KVM guests on NUMA machine and cpus
from more distant nodes would have significant increase of access
latency to the merged ksm page.  Sysfs knob was choosen for higher
variability when some users still prefers higher amount of saved
physical memory regardless of access latency.

Every numa node has its own stable & unstable trees because of faster
searching and inserting.  Changing of merge_across_nodes value is
possible only when there are not any ksm shared pages in system.

I've tested this patch on numa machines with 2, 4 and 8 nodes and
measured speed of memory access inside of KVM guests with memory pinned
to one of nodes with this benchmark:

  http://pholasek.fedorapeople.org/alloc_pg.c

Population standard deviations of access times in percentage of average
were following:

merge_across_nodes=1
2 nodes 1.4%
4 nodes 1.6%
8 nodes 1.7%

merge_across_nodes=0
2 nodes 1%
4 nodes 0.32%
8 nodes 0.018%

RFC: https://lkml.org/lkml/2011/11/30/91
v1: https://lkml.org/lkml/2012/1/23/46
v2: https://lkml.org/lkml/2012/6/29/105
v3: https://lkml.org/lkml/2012/9/14/550
v4: https://lkml.org/lkml/2012/9/23/137
v5: https://lkml.org/lkml/2012/12/10/540
v6: https://lkml.org/lkml/2012/12/23/154
v7: https://lkml.org/lkml/2012/12/27/225

Hugh notes that this patch brings two problems, whose solution needs
further support in mm/ksm.c, which follows in subsequent patches:

1) switching merge_across_nodes after running KSM is liable to oops
   on stale nodes still left over from the previous stable tree;

2) memory hotremove may migrate KSM pages, but there is no provision
   here for !merge_across_nodes to migrate nodes to the proper tree.

Signed-off-by: Petr Holasek <pholasek@redhat.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Izik Eidus <izik.eidus@ravellosystems.com>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: rename page struct field helpers

The function names page_xchg_last_nid(), page_last_nid() and
reset_page_last_nid() were judged to be inconsistent so rename them to a
struct_field_op style pattern.  As it looked jarring to have
reset_page_mapcount() and page_nid_reset_last() beside each other in
memmap_init_zone(), this patch also renames reset_page_mapcount() to
page_mapcount_reset().  There are others like init_page_count() but as
it is used throughout the arch code a rename would likely cause more
conflicts than it is worth.

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

memcg: avoid dangling reference count in creation failure.

When use_hierarchy is enabled, we acquire an extra reference count in
our parent during cgroup creation.  We don't release it, though, if any
failure exist in the creation process.

Signed-off-by: Glauber Costa <glommer@parallels.com>
Reported-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hiroyuki Kamezawa <kamezawa.hiroyuki@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memcg: increment static branch right after limit set

We were deferring the kmemcg static branch increment to a later time,
due to a nasty dependency between the cpu_hotplug lock, taken by the
jump label update, and the cgroup_lock.

Now we no longer take the cgroup lock, and we can save ourselves the
trouble.

Signed-off-by: Glauber Costa <glommer@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hiroyuki Kamezawa <kamezawa.hiroyuki@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memcg: replace cgroup_lock with memcg specific memcg_lock

After the preparation work done in earlier patches, the cgroup_lock can
be trivially replaced with a memcg-specific lock.  This is an automatic
translation at every site where the values involved were queried.

The sites where values are written, however, used to be naturally called
under cgroup_lock.  This is the case for instance in the css_online
callback.  For those, we now need to explicitly add the memcg lock.

With this, all the calls to cgroup_lock outside cgroup core are gone.

Signed-off-by: Glauber Costa <glommer@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hiroyuki Kamezawa <kamezawa.hiroyuki@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memcg: fast hierarchy-aware child test

Currently, we use cgroups' provided list of children to verify if it is
safe to proceed with any value change that is dependent on the cgroup
being empty.

This is less than ideal, because it enforces a dependency over cgroup
core that we would be better off without.  The solution proposed here is
to iterate over the child cgroups and if any is found that is already
online, we bounce and return: we don't really care how many children we
have, only if we have any.

This is also made to be hierarchy aware.  IOW, cgroups with hierarchy
disabled, while they still exist, will be considered for the purpose of
this interface as having no children.

[akpm@linux-foundation.org: tweak comments]
Signed-off-by: Glauber Costa <glommer@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hiroyuki Kamezawa <kamezawa.hiroyuki@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memcg: split part of memcg creation to css_online

This patch is a preparatory work for later locking rework to get rid of
big cgroup lock from memory controller code.

The memory controller uses some tunables to adjust its operation.  Those
tunables are inherited from parent to children upon children
intialization.  For most of them, the value cannot be changed after the
parent has a new children.

cgroup core splits initialization in two phases: css_alloc and css_online.
After css_alloc, the memory allocation and basic initialization are done.
But the new group is not yet visible anywhere, not even for cgroup core
code.  It is only somewhere between css_alloc and css_online that it is
inserted into the internal children lists.  Copying tunable values in
css_alloc will lead to inconsistent values: the children will copy the old
parent values, that can change between the copy and the moment in which
the groups is linked to any data structure that can indicate the presence
of children.

Signed-off-by: Glauber Costa <glommer@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hiroyuki Kamezawa <kamezawa.hiroyuki@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memcg: prevent changes to move_charge_at_immigrate during task attach

In memcg, we use the cgroup_lock basically to synchronize against
attaching new children to a cgroup.  We do this because we rely on
cgroup core to provide us with this information.

We need to guarantee that upon child creation, our tunables are
consistent.  For those, the calls to cgroup_lock() all live in handlers
like mem_cgroup_hierarchy_write(), where we change a tunable in the
group that is hierarchy-related.  For instance, the use_hierarchy flag
cannot be changed if the cgroup already have children.

Furthermore, those values are propagated from the parent to the child
when a new child is created.  So if we don't lock like this, we can end
up with the following situation:

A                                   B
 memcg_css_alloc()                       mem_cgroup_hierarchy_write()
 copy use hierarchy from parent          change use hierarchy in parent
 finish creation.

This is mainly because during create, we are still not fully connected
to the css tree.  So all iterators and the such that we could use, will
fail to show that the group has children.

My observation is that all of creation can proceed in parallel with
those tasks, except value assignment.  So what this patch series does is
to first move all value assignment that is dependent on parent values
from css_alloc to css_online, where the iterators all work, and then we
lock only the value assignment.  This will guarantee that parent and
children always have consistent values.  Together with an online test,
that can be derived from the observation that the refcount of an online
memcg can be made to be always positive, we should be able to
synchronize our side without the cgroup lock.

This patch:

Currently, we rely on the cgroup_lock() to prevent changes to
move_charge_at_immigrate during task migration.  However, this is only
needed because the current strategy keeps checking this value throughout
the whole process.  Since all we need is serialization, one needs only
to guarantee that whatever decision we made in the beginning of a
specific migration is respected throughout the process.

We can achieve this by just saving it in mc.  By doing this, no kind of
locking is needed.

Signed-off-by: Glauber Costa <glommer@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hiroyuki Kamezawa <kamezawa.hiroyuki@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memcg: reduce the size of struct memcg 244-fold.

In order to maintain all the memcg bookkeeping, we need per-node
descriptors, which will in turn contain a per-zone descriptor.

Because we want to statically allocate those, this array ends up being
very big.  Part of the reason is that we allocate something large enough
to hold MAX_NUMNODES, the compile time constant that holds the maximum
number of nodes we would ever consider.

However, we can do better in some cases if the firmware help us.  This
is true for modern x86 machines; coincidentally one of the architectures
in which MAX_NUMNODES tends to be very big.

By using the firmware-provided maximum number of nodes instead of
MAX_NUMNODES, we can reduce the memory footprint of struct memcg
considerably.  In the extreme case in which we have only one node, this
reduces the size of the structure from ~ 64k to ~2k.  This is
particularly important because it means that we will no longer resort to
the vmalloc area for the struct memcg on defconfigs.  We also have
enough room for an extra node and still be outside vmalloc.

One also has to keep in mind that with the industry's ability to fit
more processors in a die as fast as the FED prints money, a nodes = 2
configuration is already respectably big.

[akpm@linux-foundation.org: add check for invalid nid, remove inline]
Signed-off-by: Glauber Costa <glommer@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ying Han <yinghan@google.com>
Cc: Mel Gorman <mgorman@suse.de>
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: init: report on last-nid information stored in page->flags

Answering the question "how much space remains in the page->flags" is
time-consuming.  mminit_loglevel can help answer the question but it
does not take last_nid information into account.  This patch corrects it
and while there it corrects the messages related to page flag usage,
pgshifts and node/zone id.  When applied the relevant output looks
something like this but will depend on the kernel configuration.

  mminit::pageflags_layout_widths Section 0 Node 9 Zone 2 Lastnid 9 Flags 25
  mminit::pageflags_layout_shifts Section 19 Node 9 Zone 2 Lastnid 9
  mminit::pageflags_layout_pgshifts Section 0 Node 55 Zone 53 Lastnid 44
  mminit::pageflags_layout_nodezoneid Node/Zone ID: 64 -> 53
  mminit::pageflags_layout_usage location: 64 -> 44 layout 44 -> 25 unused 25 -> 0 page-flags

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: uninline page_xchg_last_nid()

Andrew Morton pointed out that page_xchg_last_nid() and
reset_page_last_nid() were "getting nuttily large" and asked that it be
investigated.

reset_page_last_nid() is on the page free path and it would be
unfortunate to make that path more expensive than it needs to be.  Due
to the internal use of page_xchg_last_nid() it is already too expensive
but fortunately, it should also be impossible for the page->flags to be
updated in parallel when we call reset_page_last_nid().  Instead of
unlining the function, it uses a simplier implementation that assumes no
parallel updates and should now be sufficiently short for inlining.

page_xchg_last_nid() is called in paths that are already quite expensive
(splitting huge page, fault handling, migration) and it is reasonable to
uninline.  There was not really a good place to place the function but
mm/mmzone.c was the closest fit IMO.

This patch saved 128 bytes of text in the vmlinux file for the kernel
configuration I used for testing automatic NUMA balancing.

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>

memcg: clean up swap accounting initialization code

Memcg swap accounting is currently enabled by enable_swap_cgroup when
the root cgroup is created.  mem_cgroup_init acts as a memcg subsystem
initializer which sounds like a much better place for enable_swap_cgroup
as well.  We already register memsw files from there so it makes a lot
of sense to merge those two into a single enable_swap_cgroup function.

This patch doesn't introduce any semantic changes.

Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: Zhouping Liu <zliu@redhat.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Li Zefan <lizefan@huawei.com>
Cc: CAI Qian <caiqian@redhat.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memcg: do not create memsw files if swap accounting is disabled

Zhouping Liu has reported that memsw files are exported even though swap
accounting is runtime disabled if MEMCG_SWAP is enabled.  This behavior
has been introduced by commit af36f906c0f4 ("memcg: always create memsw
files if CGROUP_MEM_RES_CTLR_SWAP") and it causes any attempt to open
the file to return EOPNOTSUPP.  Although EOPNOTSUPP should say be clear
that memsw operations are not supported in the given configuration it is
fair to say that this behavior could be quite confusing.

Let's tear memsw files out of default cgroup files and add them only if
the swap accounting is really enabled (either by MEMCG_SWAP_ENABLED or
swapaccount=1 boot parameter).  We can hook into mem_cgroup_init which
is called when the memcg subsystem is initialized and which happens
after boot command line is processed.

Signed-off-by: Michal Hocko <mhocko@suse.cz>
Reported-by: Zhouping Liu <zliu@redhat.com>
Tested-by: Zhouping Liu <zliu@redhat.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Li Zefan <lizefan@huawei.com>
Cc: CAI Qian <caiqian@redhat.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

page-writeback.c: subtract min_free_kbytes from dirtyable memory

When calculating amount of dirtyable memory, min_free_kbytes should be
subtracted because it is not intended for dirty pages.

Addresses http://bugs.debian.org/695182

[akpm@linux-foundation.org: fix up min_free_kbytes extern declarations]
[akpm@linux-foundation.org: fix min() warning]
Signed-off-by: Paul Szabo <psz@maths.usyd.edu.au>
Acked-by: 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>

mm/rmap: rename anon_vma_unlock() => anon_vma_unlock_write()

The comment in commit 4fc3f1d66b1e ("mm/rmap, migration: Make
rmap_walk_anon() and try_to_unmap_anon() more scalable") says:

| Rename anon_vma_[un]lock() => anon_vma_[un]lock_write(),
| to make it clearer that it's an exclusive write-lock in
| that case - suggested by Rik van Riel.

But that commit renames only anon_vma_lock()

Signed-off-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: Ingo Molnar <mingo@kernel.org>
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>

swap: add per-partition lock for swapfile

swap_lock is heavily contended when I test swap to 3 fast SSD (even
slightly slower than swap to 2 such SSD).  The main contention comes
from swap_info_get().  This patch tries to fix the gap with adding a new
per-partition lock.

Global data like nr_swapfiles, total_swap_pages, least_priority and
swap_list are still protected by swap_lock.

nr_swap_pages is an atomic now, it can be changed without swap_lock.  In
theory, it's possible get_swap_page() finds no swap pages but actually
there are free swap pages.  But sounds not a big problem.

Accessing partition specific data (like scan_swap_map and so on) is only
protected by swap_info_struct.lock.

Changing swap_info_struct.flags need hold swap_lock and
swap_info_struct.lock, because scan_scan_map() will check it.  read the
flags is ok with either the locks hold.

If both swap_lock and swap_info_struct.lock must be hold, we always hold
the former first to avoid deadlock.

swap_entry_free() can change swap_list.  To delete that code, we add a
new highest_priority_index.  Whenever get_swap_page() is called, we
check it.  If it's valid, we use it.

It's a pity get_swap_page() still holds swap_lock().  But in practice,
swap_lock() isn't heavily contended in my test with this patch (or I can
say there are other much more heavier bottlenecks like TLB flush).  And
BTW, looks get_swap_page() doesn't really need the lock.  We never free
swap_info[] and we check SWAP_WRITEOK flag.  The only risk without the
lock is we could swapout to some low priority swap, but we can quickly
recover after several rounds of swap, so sounds not a big deal to me.
But I'd prefer to fix this if it's a real problem.

"swap: make each swap partition have one address_space" improved the
swapout speed from 1.7G/s to 2G/s.  This patch further improves the
speed to 2.3G/s, so around 15% improvement.  It's a multi-process test,
so TLB flush isn't the biggest bottleneck before the patches.

[arnd@arndb.de: fix it for nommu]
[hughd@google.com: add missing unlock]
[minchan@kernel.org: get rid of lockdep whinge on sys_swapon]
Signed-off-by: Shaohua Li <shli@fusionio.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Seth Jennings <sjenning@linux.vnet.ibm.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Dan Magenheimer <dan.magenheimer@oracle.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

swap: make each swap partition have one address_space

When I use several fast SSD to do swap, swapper_space.tree_lock is
heavily contended.  This makes each swap partition have one
address_space to reduce the lock contention.  There is an array of
address_space for swap.  The swap entry type is the index to the array.

In my test with 3 SSD, this increases the swapout throughput 20%.

[akpm@linux-foundation.org: revert unneeded change to  __add_to_swap_cache]
Signed-off-by: Shaohua Li <shli@fusionio.com>
Cc: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@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: don't inline page_mapping()

According to akpm, this saves 1/2k text and makes things simple for the
next patch.

Numbers from Minchan:

add/remove: 1/0 grow/shrink: 6/22 up/down: 92/-516 (-424)
function                                     old     new   delta
page_mapping                                   -      48     +48
do_task_stat                                2292    2308     +16
page_remove_rmap                             240     248      +8
load_elf_binary                             4500    4508      +8
update_queue                                 532     536      +4
scsi_probe_and_add_lun                      2892    2896      +4
lookup_fast                                  644     648      +4
vcs_read                                    1040    1036      -4
__ip_route_output_key                       1904    1900      -4
ip_route_input_noref                        2508    2500      -8
shmem_file_aio_read                          784     772     -12
__isolate_lru_page                           272     256     -16
shmem_replace_page                           708     688     -20
mark_buffer_dirty                            228     208     -20
__set_page_dirty_buffers                     240     220     -20
__remove_mapping                             276     256     -20
update_mmu_cache                             500     476     -24
set_page_dirty_balance                        92      68     -24
set_page_dirty                               172     148     -24
page_evictable                                88      64     -24
page_cache_pipe_buf_steal                    248     224     -24
clear_page_dirty_for_io                      340     316     -24
test_set_page_writeback                      400     372     -28
test_clear_page_writeback                    516     488     -28
invalidate_inode_page                        156     128     -28
page_mkclean                                 432     400     -32
flush_dcache_page                            360     328     -32
__set_page_dirty_nobuffers                   324     280     -44
shrink_page_list                            2412    2356     -56

Signed-off-by: Shaohua Li <shli@fusionio.com>
Suggested-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
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: numa: cleanup flow of transhuge page migration

When correcting commit 04fa5d6a6547 ("mm: migrate: check page_count of
THP before migrating") Hugh Dickins noted that the control flow for
transhuge migration was difficult to follow.  Unconditionally calling
put_page() in numamigrate_isolate_page() made the failure paths of both
migrate_misplaced_transhuge_page() and migrate_misplaced_page() more
complex that they should be.  Further, he was extremely wary that an
unlock_page() should ever happen after a put_page() even if the
put_page() should never be the final put_page.

Hugh implemented the following cleanup to simplify the path by calling
putback_lru_page() inside numamigrate_isolate_page() if it failed to
isolate and always calling unlock_page() within
migrate_misplaced_transhuge_page().

There is no functional change after this patch is applied but the code
is easier to follow and unlock_page() always happens before put_page().

[mgorman@suse.de: changelog only]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Simon Jeons <simon.jeons@gmail.com>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: fold page->_last_nid into page->flags where possible

page->_last_nid fits into page->flags on 64-bit.  The unlikely 32-bit
NUMA configuration with NUMA Balancing will still need an extra page
field.  As Peter notes "Completely dropping 32bit support for
CONFIG_NUMA_BALANCING would simplify things, but it would also remove
the warning if we grow enough 64bit only page-flags to push the last-cpu
out."

[mgorman@suse.de: minor modifications]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Simon Jeons <simon.jeons@gmail.com>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.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: move page flags layout to separate header

This is a preparation patch for moving page->_last_nid into page->flags
that moves page flag layout information to a separate header.  This
patch is necessary because otherwise there would be a circular
dependency between mm_types.h and mm.h.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Simon Jeons <simon.jeons@gmail.com>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.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: numa: handle side-effects in count_vm_numa_events() for !CONFIG_NUMA_BALANCING

The current definitions for count_vm_numa_events() is wrong for
!CONFIG_NUMA_BALANCING as the following would miss the side-effect.

count_vm_numa_events(NUMA_FOO, bar++);

There are no such users of count_vm_numa_events() but this patch fixes
it as it is a potential pitfall.  Ideally both would be converted to
static inline but NUMA_PTE_UPDATES is not defined if
!CONFIG_NUMA_BALANCING and creating dummy constants just to have a
static inline would be similarly clumsy.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Simon Jeons <simon.jeons@gmail.com>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.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: numa: take THP into account when migrating pages for NUMA balancing

Wanpeng Li pointed out that numamigrate_isolate_page() assumes that only
one base page is being migrated when in fact it can also be checking
THP.

The consequences are that a migration will be attempted when a target
node is nearly full and fail later.  It's unlikely to be user-visible
but it should be fixed.  While we are there, migrate_balanced_pgdat()
should treat nr_migrate_pages as an unsigned long as it is treated as a
watermark.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Suggested-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Simon Jeons <simon.jeons@gmail.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: numa: fix minor typo in numa_next_scan

s/me/be/ and clarify the comment a bit when we're changing it anyway.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Suggested-by: Simon Jeons <simon.jeons@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.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: remove unused memclear_highpage_flush()

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

usb: forbid memory allocation with I/O during bus reset

If one storage interface or usb network interface(iSCSI case) exists in
current configuration, memory allocation with GFP_KERNEL during
usb_device_reset() might trigger I/O transfer on the storage interface
itself and cause deadlock because the 'us->dev_mutex' is held in
.pre_reset() and the storage interface can't do I/O transfer when the
reset is triggered by other interface, or the error handling can't be
completed if the reset is triggered by the storage itself (error
handling path).

Signed-off-by: Ming Lei <ming.lei@canonical.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Decotigny <david.decotigny@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Oliver Neukum <oneukum@suse.de>
Reviewed-by: Jiri Kosina <jkosina@suse.cz>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Rafael J. Wysocki" <rjw@sisk.pl>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

pm / runtime: force memory allocation with no I/O during Runtime PM callbcack

Apply the introduced memalloc_noio_save() and memalloc_noio_restore() to
force memory allocation with no I/O during runtime_resume/runtime_suspend
callback on device with the flag of 'memalloc_noio' set.

Signed-off-by: Ming Lei <ming.lei@canonical.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Decotigny <david.decotigny@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Oliver Neukum <oneukum@suse.de>
Cc: Jiri Kosina <jiri.kosina@suse.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Rafael J. Wysocki" <rjw@sisk.pl>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

net/core: apply pm_runtime_set_memalloc_noio on network devices

Deadlock might be caused by allocating memory with GFP_KERNEL in
runtime_resume and runtime_suspend callback of network devices in iSCSI
situation, so mark network devices and its ancestor as 'memalloc_noio'
with the introduced pm_runtime_set_memalloc_noio().

Signed-off-by: Ming Lei <ming.lei@canonical.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Decotigny <david.decotigny@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Oliver Neukum <oneukum@suse.de>
Cc: Jiri Kosina <jiri.kosina@suse.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Rafael J. Wysocki" <rjw@sisk.pl>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

block/genhd.c: apply pm_runtime_set_memalloc_noio on block devices

Apply the introduced pm_runtime_set_memalloc_noio on block device so
that PM core will teach mm to not allocate memory with GFP_IOFS when
calling the runtime_resume and runtime_suspend callback for block
devices and its ancestors.

Signed-off-by: Ming Lei <ming.lei@canonical.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Oliver Neukum <oneukum@suse.de>
Cc: Jiri Kosina <jiri.kosina@suse.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Rafael J. Wysocki" <rjw@sisk.pl>
Cc: Greg KH <greg@kroah.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Decotigny <david.decotigny@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

pm / runtime: introduce pm_runtime_set_memalloc_noio()

Introduce the flag memalloc_noio in 'struct dev_pm_info' to help PM core
to teach mm not allocating memory with GFP_KERNEL flag for avoiding
probable deadlock.

As explained in the comment, any GFP_KERNEL allocation inside
runtime_resume() or runtime_suspend() on any one of device in the path
from one block or network device to the root device in the device tree
may cause deadlock, the introduced pm_runtime_set_memalloc_noio() sets
or clears the flag on device in the path recursively.

Signed-off-by: Ming Lei <ming.lei@canonical.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Oliver Neukum <oneukum@suse.de>
Cc: Jiri Kosina <jiri.kosina@suse.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Rafael J. Wysocki" <rjw@sisk.pl>
Cc: Greg KH <greg@kroah.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Decotigny <david.decotigny@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: teach mm by current context info to not do I/O during memory allocation

This patch introduces PF_MEMALLOC_NOIO on process flag('flags' field of
'struct task_struct'), so that the flag can be set by one task to avoid
doing I/O inside memory allocation in the task's context.

The patch trys to solve one deadlock problem caused by block device, and
the problem may happen at least in the below situations:

- during block device runtime resume, if memory allocation with
  GFP_KERNEL is called inside runtime resume callback of any one of its
  ancestors(or the block device itself), the deadlock may be triggered
  inside the memory allocation since it might not complete until the block
  device becomes active and the involed page I/O finishes.  The situation
  is pointed out first by Alan Stern.  It is not a good approach to
  convert all GFP_KERNEL[1] in the path into GFP_NOIO because several
  subsystems may be involved(for example, PCI, USB and SCSI may be
  involved for usb mass stoarage device, network devices involved too in
  the iSCSI case)

- during block device runtime suspend, because runtime resume need to
  wait for completion of concurrent runtime suspend.

- during error handling of usb mass storage deivce, USB bus reset will
  be put on the device, so there shouldn't have any memory allocation with
  GFP_KERNEL during USB bus reset, otherwise the deadlock similar with
  above may be triggered.  Unfortunately, any usb device may include one
  mass storage interface in theory, so it requires all usb interface
  drivers to handle the situation.  In fact, most usb drivers don't know
  how to handle bus reset on the device and don't provide .pre_set() and
  .post_reset() callback at all, so USB core has to unbind and bind driver
  for these devices.  So it is still not practical to resort to GFP_NOIO
  for solving the problem.

Also the introduced solution can be used by block subsystem or block
drivers too, for example, set the PF_MEMALLOC_NOIO flag before doing
actual I/O transfer.

It is not a good idea to convert all these GFP_KERNEL in the affected
path into GFP_NOIO because these functions doing that may be implemented
as library and will be called in many other contexts.

In fact, memalloc_noio_flags() can convert some of current static
GFP_NOIO allocation into GFP_KERNEL back in other non-affected contexts,
at least almost all GFP_NOIO in USB subsystem can be converted into
GFP_KERNEL after applying the approach and make allocation with GFP_NOIO
only happen in runtime resume/bus reset/block I/O transfer contexts
generally.

[1], several GFP_KERNEL allocation examples in runtime resume path

- pci subsystem
acpi_os_allocate
<-acpi_ut_allocate
<-ACPI_ALLOCATE_ZEROED
<-acpi_evaluate_object
<-__acpi_bus_set_power
<-acpi_bus_set_power
<-acpi_pci_set_power_state
<-platform_pci_set_power_state
<-pci_platform_power_transition
<-__pci_complete_power_transition
<-pci_set_power_state
<-pci_restore_standard_config
<-pci_pm_runtime_resume
- usb subsystem
usb_get_status
<-finish_port_resume
<-usb_port_resume
<-generic_resume
<-usb_resume_device
<-usb_resume_both
<-usb_runtime_resume

- some individual usb drivers
usblp, uvc, gspca, most of dvb-usb-v2 media drivers, cpia2, az6007, ....

That is just what I have found.  Unfortunately, this allocation can only
be found by human being now, and there should be many not found since
any function in the resume path(call tree) may allocate memory with
GFP_KERNEL.

Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Oliver Neukum <oneukum@suse.de>
Cc: Jiri Kosina <jiri.kosina@suse.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Rafael J. Wysocki" <rjw@sisk.pl>
Cc: Greg KH <greg@kroah.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Decotigny <david.decotigny@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: don't wait on congested zones in balance_pgdat()

From: Zlatko Calusic <zlatko.calusic@iskon.hr>

Commit 92df3a723f84 ("mm: vmscan: throttle reclaim if encountering too
many dirty pages under writeback") introduced waiting on congested zones
based on a sane algorithm in shrink_inactive_list().

What this means is that there's no more need for throttling and
additional heuristics in balance_pgdat().  So, let's remove it and tidy
up the code.

Signed-off-by: Zlatko Calusic <zlatko.calusic@iskon.hr>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Johannes Weiner <jweiner@redhat.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/memory-failure.c: fix wrong num_poisoned_pages in handling memory error on thp

num_poisoned_pages counts up the number of pages isolated by memory
errors.  But for thp, only one subpage is isolated because memory error
handler splits it, so it's wrong to add (1 << compound_trans_order).

[akpm@linux-foundation.org: tweak comment]
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Tony Luck <tony.luck@intel.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/memory-failure.c: clean up soft_offline_page()

Currently soft_offline_page() is hard to maintain because it has many
return points and goto statements.  All of this mess come from
get_any_page().

This function should only get page refcount as the name implies, but it
does some page isolating actions like SetPageHWPoison() and dequeuing
hugepage.  This patch corrects it and introduces some internal
subroutines to make soft offlining code more readable and maintainable.

Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Reviewed-by: Andi Kleen <andi@firstfloor.org>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Xishi Qiu <qiuxishi@huawei.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-failure: use num_poisoned_pages instead of mce_bad_pages

Since MCE is an x86 concept, and this code is in mm/, it would be better
to use the name num_poisoned_pages instead of mce_bad_pages.

[akpm@linux-foundation.org: fix mm/sparse.c]
Signed-off-by: Xishi Qiu <qiuxishi@huawei.com>
Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Suggested-by: Borislav Petkov <bp@alien8.de>
Reviewed-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-failure: do code refactor of soft_offline_page()

There are too many return points randomly intermingled with some "goto
done" return points.  So adjust the function structure, one for the
success path, the other for the failure path.  Use atomic_long_inc
instead of atomic_long_add.

Signed-off-by: Xishi Qiu <qiuxishi@huawei.com>
Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Suggested-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-failure: fix an error of mce_bad_pages statistics

When doing

    $ echo paddr > /sys/devices/system/memory/soft_offline_page

to offline a *free* page, the value of mce_bad_pages will be added, and
the page is set HWPoison flag, but it is still managed by page buddy
alocator.

   $ cat /proc/meminfo | grep HardwareCorrupted

shows the value.

If we offline the same page, the value of mce_bad_pages will be added
*again*, this means the value is incorrect now.  Assume the page is
still free during this short time.

  soft_offline_page()
    get_any_page()
      "else if (is_free_buddy_page(p))" branch return 0
        "goto done";
           "atomic_long_add(1, &mce_bad_pages);"

This patch:

Move poisoned page check at the beginning of the function in order to
fix the error.

Signed-off-by: Xishi Qiu <qiuxishi@huawei.com>
Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Tested-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: remove MIGRATE_ISOLATE check in hotpath

Several functions test MIGRATE_ISOLATE and some of those are hotpath but
MIGRATE_ISOLATE is used only if we enable CONFIG_MEMORY_ISOLATION(ie,
CMA, memory-hotplug and memory-failure) which are not common config
option.  So let's not add unnecessary overhead and code when we don't
enable CONFIG_MEMORY_ISOLATION.

Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: increase totalram_pages when free pages allocated by bootmem allocator

Function put_page_bootmem() is used to free pages allocated by bootmem
allocator, so it should increase totalram_pages when freeing pages into
the buddy system.

Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Maciej Rutecki <maciej.rutecki@gmail.com>
Cc: Chris Clayton <chris2553@googlemail.com>
Cc: "Rafael J . Wysocki" <rjw@sisk.pl>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Jianguo Wu <wujianguo@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: set zone->present_pages to number of existing pages in the zone

Now all users of "number of pages managed by the buddy system" have been
converted to use zone->managed_pages, so set zone->present_pages to what
it should be:

present_pages = spanned_pages - absent_pages;

Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Maciej Rutecki <maciej.rutecki@gmail.com>
Cc: Chris Clayton <chris2553@googlemail.com>
Cc: "Rafael J . Wysocki" <rjw@sisk.pl>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Jianguo Wu <wujianguo@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: use zone->present_pages instead of zone->managed_pages where appropriate

Now we have zone->managed_pages for "pages managed by the buddy system
in the zone", so replace zone->present_pages with zone->managed_pages if
what the user really wants is number of allocatable pages.

Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Maciej Rutecki <maciej.rutecki@gmail.com>
Cc: Chris Clayton <chris2553@googlemail.com>
Cc: "Rafael J . Wysocki" <rjw@sisk.pl>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Jianguo Wu <wujianguo@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/memblock.c: use CONFIG_HAVE_MEMBLOCK_NODE_MAP to protect movablecore_map in memblock_overlaps_region().

The definition of struct movablecore_map is protected by
CONFIG_HAVE_MEMBLOCK_NODE_MAP but its use in memblock_overlaps_region()
is not.  So add CONFIG_HAVE_MEMBLOCK_NODE_MAP to protect the use of
movablecore_map in memblock_overlaps_region().

Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Reported-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>

acpi, memory-hotplug: support getting hotplug info from SRAT

We now provide an option for users who don't want to specify physical
memory address in kernel commandline.

         /*
          * For movablemem_map=acpi:
          *
          * SRAT:                |_____| |_____| |_________| |_________| ......
          * node id:                0       1         1           2
          * hotpluggable:           n       y         y           n
          * movablemem_map:              |_____| |_________|
          *
          * Using movablemem_map, we can prevent memblock from allocating memory
          * on ZONE_MOVABLE at boot time.
          */

So user just specify movablemem_map=acpi, and the kernel will use
hotpluggable info in SRAT to determine which memory ranges should be set
as ZONE_MOVABLE.

If all the memory ranges in SRAT is hotpluggable, then no memory can be
used by kernel.  But before parsing SRAT, memblock has already reserve
some memory ranges for other purposes, such as for kernel image, and so
on.  We cannot prevent kernel from using these memory.  So we need to
exclude these ranges even if these memory is hotpluggable.

Furthermore, there could be several memory ranges in the single node
which the kernel resides in.  We may skip one range that have memory
reserved by memblock, but if the rest of memory is too small, then the
kernel will fail to boot.  So, make the whole node which the kernel
resides in un-hotpluggable.  Then the kernel has enough memory to use.

NOTE: Using this way will cause NUMA performance down because the
      whole node will be set as ZONE_MOVABLE, and kernel cannot use memory
      on it.  If users don't want to lose NUMA performance, just don't use
      it.

[akpm@linux-foundation.org: fix warning]
[akpm@linux-foundation.org: use strcmp()]
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Len Brown <lenb@kernel.org>
Cc: "Brown, Len" <len.brown@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

acpi, memory-hotplug: extend movablemem_map ranges to the end of node

When implementing movablemem_map boot option, we introduced an array
movablemem_map.map[] to store the memory ranges to be set as
ZONE_MOVABLE.

Since ZONE_MOVABLE is the latst zone of a node, if user didn't specify
the whole node memory range, we need to extend it to the node end so
that we can use it to prevent memblock from allocating memory in the
ranges user didn't specify.

We now implement movablemem_map boot option like this:

        /*
         * For movablemem_map=nn[KMG]@ss[KMG]:
         *
         * SRAT:                |_____| |_____| |_________| |_________| ......
         * node id:                0       1         1           2
         * user specified:                |__|                 |___|
         * movablemem_map:                |___| |_________|    |______| ......
         *
         * Using movablemem_map, we can prevent memblock from allocating memory
         * on ZONE_MOVABLE at boot time.
         *
         * NOTE: In this case, SRAT info will be ingored.
         */

[akpm@linux-foundation.org: clean up code, fix build warning]
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Len Brown <lenb@kernel.org>
Cc: "Brown, Len" <len.brown@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

acpi, memory-hotplug: parse SRAT before memblock is ready

On linux, the pages used by kernel could not be migrated.  As a result,
if a memory range is used by kernel, it cannot be hot-removed.  So if we
want to hot-remove memory, we should prevent kernel from using it.

The way now used to prevent this is specify a memory range by
movablemem_map boot option and set it as ZONE_MOVABLE.

But when the system is booting, memblock will allocate memory, and
reserve the memory for kernel.  And before we parse SRAT, and know the
node memory ranges, memblock is working.  And it may allocate memory in
ranges to be set as ZONE_MOVABLE.  This memory can be used by kernel,
and never be freed.

So, let's parse SRAT before memblock is called first.  And it is early
enough.

The first call of memblock_find_in_range_node() is in:

  setup_arch()
    |-->setup_real_mode()

so, this patch add a function early_parse_srat() to parse SRAT, and call
it before setup_real_mode() is called.

NOTE:

1) early_parse_srat() is called before numa_init(), and has initialized
   numa_meminfo.  So DO NOT clear numa_nodes_parsed in numa_init() and DO
   NOT zero numa_meminfo in numa_init(), otherwise we will lose memory
   numa info.

2) I don't know why using count of memory affinities parsed from SRAT
   as a return value in original acpi_numa_init().  So I add a static
   variable srat_mem_cnt to remember this count and use it as the return
   value of the new acpi_numa_init()

[mhocko@suse.cz: parse SRAT before memblock is ready fix]
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Reviewed-by: Wen Congyang <wency@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Len Brown <lenb@kernel.org>
Cc: "Brown, Len" <len.brown@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

page_alloc: bootmem limit with movablecore_map

Ensure the bootmem will not allocate memory from areas that may be
ZONE_MOVABLE.  The map info is from movablecore_map boot option.

Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Reviewed-by: Wen Congyang <wency@cn.fujitsu.com>
Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Tested-by: Lin Feng <linfeng@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

page_alloc: make movablemem_map have higher priority

If kernelcore or movablecore is specified at the same time with
movablemem_map, movablemem_map will have higher priority to be
satisfied.  This patch will make find_zone_movable_pfns_for_nodes()
calculate zone_movable_pfn[] with the limit from zone_movable_limit[].

Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Reviewed-by: Wen Congyang <wency@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Tested-by: Lin Feng <linfeng@cn.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

page_alloc: introduce zone_movable_limit[] to keep movable limit for nodes

Introduce a new array zone_movable_limit[] to store the ZONE_MOVABLE
limit from movablemem_map boot option for all nodes.  The function
sanitize_zone_movable_limit() will find out to which node the ranges in
movable_map.map[] belongs, and calculates the low boundary of
ZONE_MOVABLE for each node.

Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Signed-off-by: Liu Jiang <jiang.liu@huawei.com>
Reviewed-by: Wen Congyang <wency@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Tested-by: Lin Feng <linfeng@cn.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

page_alloc: add movable_memmap kernel parameter

Add functions to parse movablemem_map boot option.  Since the option
could be specified more then once, all the maps will be stored in the
global variable movablemem_map.map array.

And also, we keep the array in monotonic increasing order by start_pfn.
And merge all overlapped ranges.

[akpm@linux-foundation.org: improve comment]
[akpm@linux-foundation.org: checkpatch fixes]
[akpm@linux-foundation.org: remove unneeded parens]
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Reviewed-by: Wen Congyang <wency@cn.fujitsu.com>
Tested-by: Lin Feng <linfeng@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

x86: get pg_data_t's memory from other node

During the implementation of SRAT support, we met a problem.  In
setup_arch(), we have the following call series:

 1) memblock is ready;
 2) some functions use memblock to allocate memory;
 3) parse ACPI tables, such as SRAT.

Before 3), we don't know which memory is hotpluggable, and as a result,
we cannot prevent memblock from allocating hotpluggable memory.  So, in
2), there could be some hotpluggable memory allocated by memblock.

Now, we are trying to parse SRAT earlier, before memblock is ready.  But
I think we need more investigation on this topic.  So in this v5, I
dropped all the SRAT support, and v5 is just the same as v3, and it is
based on 3.8-rc3.

As we planned, we will support getting info from SRAT without users'
participation at last.  And we will post another patch-set to do so.

And also, I think for now, we can add this boot option as the first step
of supporting movable node.  Since Linux cannot migrate the direct
mapped pages, the only way for now is to limit the whole node containing
only movable memory.

Using SRAT is one way.  But even if we can use SRAT, users still need an
interface to enable/disable this functionality if they don't want to
loose their NUMA performance.  So I think, a user interface is always
needed.

For now, users can disable this functionality by not specifying the boot
option.  Later, we will post SRAT support, and add another option value
"movablecore_map=acpi" to using SRAT.

This patch:

If system can create movable node which all memory of the node is
allocated as ZONE_MOVABLE, setup_node_data() cannot allocate memory for
the node's pg_data_t.  So, use memblock_alloc_try_nid() instead of
memblock_alloc_nid() to retry when the first allocation fails.

Signed-off-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

sched: do not use cpu_to_node() to find an offlined cpu's node.

If a cpu is offline, its nid will be set to -1, and cpu_to_node(cpu)
will return -1.  As a result, cpumask_of_node(nid) will return NULL.  In
this case, find_next_bit() in for_each_cpu will get a NULL pointer and
cause panic.

Here is a call trace:
  Call Trace:
   <IRQ>
    select_fallback_rq+0x71/0x190
    try_to_wake_up+0x2cb/0x2f0
    wake_up_process+0x15/0x20
    hrtimer_wakeup+0x22/0x30
    __run_hrtimer+0x83/0x320
    hrtimer_interrupt+0x106/0x280
    smp_apic_timer_interrupt+0x69/0x99
    apic_timer_interrupt+0x6f/0x80

There is a hrtimer process sleeping, whose cpu has already been
offlined.  When it is waken up, it tries to find another cpu to run, and
get a -1 nid.  As a result, cpumask_of_node(-1) returns NULL, and causes
ernel panic.

This patch fixes this problem by judging if the nid is -1.  If nid is
not -1, a cpu on the same node will be picked.  Else, a online cpu on
another node will be picked.

Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Jiang Liu <liuj97@gmail.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

cpu-hotplug,memory-hotplug: clear cpu_to_node() when offlining the node

When the node is offlined, there is no memory/cpu on the node.  If a
sleep task runs on a cpu of this node, it will be migrated to the cpu on
the other node.  So we can clear cpu-to-node mapping.

[akpm@linux-foundation.org: numa_clear_node() and numa_set_node() can no longer be __cpuinit]
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Jiang Liu <liuj97@gmail.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

cpu-hotplug, memory-hotplug: try offlining the node when hotremoving a cpu

The node will be offlined when all memory/cpu on the node is hotremoved.
So we should try offline the node when hotremoving a cpu on the node.

Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Jiang Liu <liuj97@gmail.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: export the function try_offline_node()

try_offline_node() will be needed in the tristate
drivers/acpi/processor_driver.c.

The node will be offlined when all memory/cpu on the node have been
hotremoved.  So we need the function try_offline_node() in cpu-hotplug
path.

If the memory-hotplug is disabled, and cpu-hotplug is enabled

1. no memory no the node
   we don't online the node, and cpu's node is the nearest node.

2. the node contains some memory
   the node has been onlined, and cpu's node is still needed
   to migrate the sleep task on the cpu to the same node.

So we do nothing in try_offline_node() in this case.

[rientjes@google.com: export the function try_offline_node() fix]
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Jiang Liu <liuj97@gmail.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

cpu_hotplug: clear apicid to node when the cpu is hotremoved

When a cpu is hotpluged, we call acpi_map_cpu2node() in
_acpi_map_lsapic() to store the cpu's node and apicid's node.  But we
don't clear the cpu's node in acpi_unmap_lsapic() when this cpu is
hotremoved.  If the node is also hotremoved, we will get the following
messages:

  kernel BUG at include/linux/gfp.h:329!
  invalid opcode: 0000 [#1] SMP
  Modules linked in: ebtable_nat ebtables ipt_MASQUERADE iptable_nat nf_nat xt_CHECKSUM iptable_mangle bridge stp llc sunrpc ipt_REJECT nf_conntrack_ipv4 nf_defrag_ipv4 iptable_filter ip_tables ip6t_REJECT nf_conntrack_ipv6 nf_defrag_ipv6 xt_state nf_conntrack ip6table_filter ip6_tables binfmt_misc dm_mirror dm_region_hash dm_log dm_mod vhost_net macvtap macvlan tun uinput iTCO_wdt iTCO_vendor_support coretemp kvm_intel kvm crc32c_intel microcode pcspkr i2c_i801 i2c_core lpc_ich mfd_core ioatdma e1000e i7core_edac edac_core sg acpi_memhotplug igb dca sd_mod crc_t10dif megaraid_sas mptsas mptscsih mptbase scsi_transport_sas scsi_mod
  Pid: 3126, comm: init Not tainted 3.6.0-rc3-tangchen-hostbridge+ #13 FUJITSU-SV PRIMEQUEST 1800E/SB
  RIP: 0010:[<ffffffff811bc3fd>]  [<ffffffff811bc3fd>] allocate_slab+0x28d/0x300
  RSP: 0018:ffff88078a049cf8  EFLAGS: 00010246
  RAX: 0000000000000000 RBX: 0000000000000001 RCX: 0000000000000000
  RDX: 0000000000000001 RSI: 0000000000000001 RDI: 0000000000000246
  RBP: ffff88078a049d38 R08: 00000000000040d0 R09: 0000000000000001
  R10: 0000000000000000 R11: 0000000000000b5f R12: 00000000000052d0
  R13: ffff8807c1417300 R14: 0000000000030038 R15: 0000000000000003
  FS:  00007fa9b1b44700(0000) GS:ffff8807c3800000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
  CR2: 00007fa9b09acca0 CR3: 000000078b855000 CR4: 00000000000007e0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
  Process init (pid: 3126, threadinfo ffff88078a048000, task ffff8807bb6f2650)
  Call Trace:
    new_slab+0x30/0x1b0
    __slab_alloc+0x358/0x4c0
    kmem_cache_alloc_node_trace+0xb4/0x1e0
    alloc_fair_sched_group+0xd0/0x1b0
    sched_create_group+0x3e/0x110
    sched_autogroup_create_attach+0x4d/0x180
    sys_setsid+0xd4/0xf0
    system_call_fastpath+0x16/0x1b
  Code: 89 c4 e9 73 fe ff ff 31 c0 89 de 48 c7 c7 45 de 9e 81 44 89 45 c8 e8 22 05 4b 00 85 db 44 8b 45 c8 0f 89 4f ff ff ff 0f 0b eb fe <0f> 0b 90 eb fd 0f 0b eb fe 89 de 48 c7 c7 45 de 9e 81 31 c0 44
  RIP  [<ffffffff811bc3fd>] allocate_slab+0x28d/0x300
   RSP <ffff88078a049cf8>
  ---[ end trace adf84c90f3fea3e5 ]---

The reason is that the cpu's node is not NUMA_NO_NODE, we will call
alloc_pages_exact_node() to alloc memory on the node, but the node is
offlined.

If the node is onlined, we still need cpu's node.  For example: a task
on the cpu is sleeped when the cpu is hotremoved.  We will choose
another cpu to run this task when it is waked up.  If we know the cpu's
node, we will choose the cpu on the same node first.  So we should clear
cpu-to-node mapping when the node is offlined.

This patch only clears apicid-to-node mapping when the cpu is
hotremoved.

[akpm@linux-foundation.org: fix section error]
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Jiang Liu <liuj97@gmail.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mempolicy: fix is_valid_nodemask()

is_valid_nodemask() was introduced by commit 19770b32609b ("mm: filter
based on a nodemask as well as a gfp_mask").  but it does not match its
comments, because it does not check the zone which > policy_zone.

Also in commit b377fd3982ad ("Apply memory policies to top two highest
zones when highest zone is ZONE_MOVABLE"), this commits told us, if
highest zone is ZONE_MOVABLE, we should also apply memory policies to
it.  so ZONE_MOVABLE should be valid zone for policies.
is_valid_nodemask() need to be changed to match it.

Fix: check all zones, even its zoneid > policy_zone.  Use
nodes_intersects() instead open code to check it.

Reported-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.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>

memory-hotplug: consider compound pages when free memmap

usemap could also be allocated as compound pages.  Should also consider
compound pages when freeing memmap.

If we don't fix it, there could be problems when we free vmemmap
pagetables which are stored in compound pages.  The old pagetables will
not be freed properly, and when we add the memory again, no new
pagetable will be created.  And the old pagetable entry is used, than
the kernel will panic.

The call trace is like the following:

  BUG: unable to handle kernel paging request at ffffea0040000000
  IP: [<ffffffff816a483f>] sparse_add_one_section+0xef/0x166
  PGD 7ff7d4067 PUD 78e035067 PMD 78e11d067 PTE 0
  Oops: 0002 [#1] SMP
  Modules linked in: ip6table_filter ip6_tables ebtable_nat ebtables nf_conntrack_ipv4 nf_defrag_ipv4 xt_state nf_conntrack ipt_REJECT xt_CHECKSUM iptable_mangle iptable_filter ip_tables bridge stp llc sunrpc binfmt_misc dm_mirror dm_region_hash dm_log dm_mod vhost_net macvtap macvlan tun uinput iTCO_wdt iTCO_vendor_support coretemp kvm_intel kvm crc32c_intel microcode pcspkr sg lpc_ich mfd_core i2c_i801 i2c_core i7core_edac edac_core ioatdma e1000e igb dca ptp pps_core sd_mod crc_t10dif megaraid_sas mptsas mptscsih mptbase scsi_transport_sas scsi_mod
  CPU 0
  Pid: 4, comm: kworker/0:0 Tainted: G        W 3.8.0-rc3-phy-hot-remove+ #3 FUJITSU-SV PRIMEQUEST 1800E/SB
  RIP: 0010:[<ffffffff816a483f>]  [<ffffffff816a483f>] sparse_add_one_section+0xef/0x166
  RSP: 0018:ffff8807bdcb35d8  EFLAGS: 00010006
  RAX: 0000000000000000 RBX: 0000000000000200 RCX: 0000000000200000
  RDX: ffff88078df01148 RSI: 0000000000000282 RDI: ffffea0040000000
  RBP: ffff8807bdcb3618 R08: 4cf05005b019467a R09: 0cd98fa09631467a
  R10: 0000000000000000 R11: 0000000000030e20 R12: 0000000000008000
  R13: ffffea0040000000 R14: ffff88078df66248 R15: ffff88078ea13b10
  FS:  0000000000000000(0000) GS:ffff8807c1a00000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
  CR2: ffffea0040000000 CR3: 0000000001c0c000 CR4: 00000000000007f0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
  Process kworker/0:0 (pid: 4, threadinfo ffff8807bdcb2000, task ffff8807bde18000)
  Call Trace:
    __add_pages+0x85/0x120
    arch_add_memory+0x71/0xf0
    add_memory+0xd6/0x1f0
    acpi_memory_device_add+0x170/0x20c
    acpi_device_probe+0x50/0x18a
    really_probe+0x6c/0x320
    driver_probe_device+0x47/0xa0
    __device_attach+0x53/0x60
    bus_for_each_drv+0x6c/0xa0
    device_attach+0xa8/0xc0
    bus_probe_device+0xb0/0xe0
    device_add+0x301/0x570
    device_register+0x1e/0x30
    acpi_device_register+0x1d8/0x27c
    acpi_add_single_object+0x1df/0x2b9
    acpi_bus_check_add+0x112/0x18f
    acpi_ns_walk_namespace+0x105/0x255
    acpi_walk_namespace+0xcf/0x118
    acpi_bus_scan+0x5b/0x7c
    acpi_bus_add+0x2a/0x2c
    container_notify_cb+0x112/0x1a9
    acpi_ev_notify_dispatch+0x46/0x61
    acpi_os_execute_deferred+0x27/0x34
    process_one_work+0x20e/0x5c0
    worker_thread+0x12e/0x370
    kthread+0xee/0x100
    ret_from_fork+0x7c/0xb0
  Code: 00 00 48 89 df 48 89 45 c8 e8 3e 71 b1 ff 48 89 c2 48 8b 75 c8 b8 ef ff ff ff f6 02 01 75 4b 49 63 cc 31 c0 4c 89 ef 48 c1 e1 06 <f3> aa 48 8b 02 48 83 c8 01 48 85 d2 48 89 02 74 29 a8 01 74 25
  RIP  [<ffffffff816a483f>] sparse_add_one_section+0xef/0x166
   RSP <ffff8807bdcb35d8>
  CR2: ffffea0040000000
  ---[ end trace e7f94e3a34c442d4 ]---
  Kernel panic - not syncing: Fatal exception

Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.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>

memory-hotplug: do not allocate pgdat if it was not freed when offline.

Since there is no way to guarentee the address of pgdat/zone is not on
stack of any kernel threads or used by other kernel objects without
reference counting or other symchronizing method, we cannot reset
node_data and free pgdat when offlining a node.  Just reset pgdat to 0
and reuse the memory when the node is online again.

The problem is suggested by Kamezawa Hiroyuki.  The idea is from Wen
Congyang.

NOTE: If we don't reset pgdat to 0, the WARN_ON in free_area_init_node()
      will be triggered.

[akpm@linux-foundation.org: fix warning when CONFIG_NEED_MULTIPLE_NODES=n]
[akpm@linux-foundation.org: fix the warning again again]
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Reviewed-by: Wen Congyang <wency@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: free node_data when a node is offlined

We call hotadd_new_pgdat() to allocate memory to store node_data.  So we
should free it when removing a node.

Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Reviewed-by: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: remove sysfs file of node

Introduce a new function try_offline_node() to remove sysfs file of node
when all memory sections of this node are removed.  If some memory
sections of this node are not removed, this function does nothing.

Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory_hotplug: clear zone when removing the memory

When memory is added, we update zone's and pgdat's start_pfn and
spanned_pages in __add_zone().  So we should revert them when the memory
is removed.

The patch adds a new function __remove_zone() to do this.

Signed-off-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: integrated __remove_section() of CONFIG_SPARSEMEM_VMEMMAP.

Currently __remove_section for SPARSEMEM_VMEMMAP does nothing.  But even
if we use SPARSEMEM_VMEMMAP, we can unregister the memory_section.

Signed-off-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: remove memmap of sparse-vmemmap

Introduce a new API vmemmap_free() to free and remove vmemmap
pagetables.  Since pagetable implements are different, each architecture
has to provide its own version of vmemmap_free(), just like
vmemmap_populate().

Note: vmemmap_free() is not implemented for ia64, ppc, s390, and sparc.

[mhocko@suse.cz: fix implicit declaration of remove_pagetable]
Signed-off-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Jianguo Wu <wujianguo@huawei.com>
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: remove page table of x86_64 architecture

Search a page table about the removed memory, and clear page table for
x86_64 architecture.

[akpm@linux-foundation.org: make kernel_physical_mapping_remove() static]
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Jianguo Wu <wujianguo@huawei.com>
Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: common APIs to support page tables hot-remove

When memory is removed, the corresponding pagetables should alse be
removed.  This patch introduces some common APIs to support vmemmap
pagetable and x86_64 architecture direct mapping pagetable removing.

All pages of virtual mapping in removed memory cannot be freed if some
pages used as PGD/PUD include not only removed memory but also other
memory.  So this patch uses the following way to check whether a page
can be freed or not.

1) When removing memory, the page structs of the removed memory are
   filled with 0FD.

2) All page structs are filled with 0xFD on PT/PMD, PT/PMD can be
   cleared.  In this case, the page used as PT/PMD can be freed.

For direct mapping pages, update direct_pages_count[level] when we freed
their pagetables.  And do not free the pages again because they were
freed when offlining.

For vmemmap pages, free the pages and their pagetables.

For larger pages, do not split them into smaller ones because there is
no way to know if the larger page has been split.  As a result, there is
no way to decide when to split.  We deal the larger pages in the
following way:

1) For direct mapped pages, all the pages were freed when they were
   offlined.  And since menmory offline is done section by section, all
   the memory ranges being removed are aligned to PAGE_SIZE.  So only need
   to deal with unaligned pages when freeing vmemmap pages.

2) For vmemmap pages being used to store page_struct, if part of the
   larger page is still in use, just fill the unused part with 0xFD.  And
   when the whole page is fulfilled with 0xFD, then free the larger page.

[akpm@linux-foundation.org: fix typo in comment]
[tangchen@cn.fujitsu.com: do not calculate direct mapping pages when freeing vmemmap pagetables]
[tangchen@cn.fujitsu.com: do not free direct mapping pages twice]
[tangchen@cn.fujitsu.com: do not free page split from hugepage one by one]
[tangchen@cn.fujitsu.com: do not split pages when freeing pagetable pages]
[akpm@linux-foundation.org: use pmd_page_vaddr()]
[akpm@linux-foundation.org: fix used-uninitialised bug]
Signed-off-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Jianguo Wu <wujianguo@huawei.com>
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: move pgdat_resize_lock into sparse_remove_one_section()

In __remove_section(), we locked pgdat_resize_lock when calling
sparse_remove_one_section().  This lock will disable irq.  But we don't
need to lock the whole function.  If we do some work to free pagetables
in free_section_usemap(), we need to call flush_tlb_all(), which need
irq enabled.  Otherwise the WARN_ON_ONCE() in smp_call_function_many()
will be triggered.

If we lock the whole sparse_remove_one_section(), then we come to this call trace:

  ------------[ cut here ]------------
  WARNING: at kernel/smp.c:461 smp_call_function_many+0xbd/0x260()
  Hardware name: PRIMEQUEST 1800E
  ......
  Call Trace:
    smp_call_function_many+0xbd/0x260
    smp_call_function+0x3b/0x50
    on_each_cpu+0x3b/0xc0
    flush_tlb_all+0x1c/0x20
    remove_pagetable+0x14e/0x1d0
    vmemmap_free+0x18/0x20
    sparse_remove_one_section+0xf7/0x100
    __remove_section+0xa2/0xb0
    __remove_pages+0xa0/0xd0
    arch_remove_memory+0x6b/0xc0
    remove_memory+0xb8/0xf0
    acpi_memory_device_remove+0x53/0x96
    acpi_device_remove+0x90/0xb2
    __device_release_driver+0x7c/0xf0
    device_release_driver+0x2f/0x50
    acpi_bus_remove+0x32/0x6d
    acpi_bus_trim+0x91/0x102
    acpi_bus_hot_remove_device+0x88/0x16b
    acpi_os_execute_deferred+0x27/0x34
    process_one_work+0x20e/0x5c0
    worker_thread+0x12e/0x370
    kthread+0xee/0x100
    ret_from_fork+0x7c/0xb0
  ---[ end trace 25e85300f542aa01 ]---

Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: implement register_page_bootmem_info_section of sparse-vmemmap

For removing memmap region of sparse-vmemmap which is allocated bootmem,
memmap region of sparse-vmemmap needs to be registered by
get_page_bootmem().  So the patch searches pages of virtual mapping and
registers the pages by get_page_bootmem().

NOTE: register_page_bootmem_memmap() is not implemented for ia64,
      ppc, s390, and sparc.  So introduce CONFIG_HAVE_BOOTMEM_INFO_NODE
      and revert register_page_bootmem_info_node() when platform doesn't
      support it.

      It's implemented by adding a new Kconfig option named
      CONFIG_HAVE_BOOTMEM_INFO_NODE, which will be automatically selected
      by memory-hotplug feature fully supported archs(currently only on
      x86_64).

      Since we have 2 config options called MEMORY_HOTPLUG and
      MEMORY_HOTREMOVE used for memory hot-add and hot-remove separately,
      and codes in function register_page_bootmem_info_node() are only
      used for collecting infomation for hot-remove, so reside it under
      MEMORY_HOTREMOVE.

      Besides page_isolation.c selected by MEMORY_ISOLATION under
      MEMORY_HOTPLUG is also such case, move it too.

[mhocko@suse.cz: put register_page_bootmem_memmap inside CONFIG_MEMORY_HOTPLUG_SPARSE]
[linfeng@cn.fujitsu.com: introduce CONFIG_HAVE_BOOTMEM_INFO_NODE and revert register_page_bootmem_info_node()]
[mhocko@suse.cz: remove the arch specific functions without any implementation]
[linfeng@cn.fujitsu.com: mm/Kconfig: move auto selects from MEMORY_HOTPLUG to MEMORY_HOTREMOVE as needed]
[rientjes@google.com: fix defined but not used warning]
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Reviewed-by: Wu Jianguo <wujianguo@huawei.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Signed-off-by: Lin Feng <linfeng@cn.fujitsu.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: introduce new arch_remove_memory() for removing page table

For removing memory, we need to remove page tables.  But it depends on
architecture.  So the patch introduce arch_remove_memory() for removing
page table.  Now it only calls __remove_pages().

Note: __remove_pages() for some archtecuture is not implemented
      (I don't know how to implement it for s390).

Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: remove /sys/firmware/memmap/X sysfs

When (hot)adding memory into system, /sys/firmware/memmap/X/{end, start,
type} sysfs files are created.  But there is no code to remove these
files.  This patch implements the function to remove them.

We cannot free firmware_map_entry which is allocated by bootmem because
there is no way to do so when the system is up.  But we can at least
remember the address of that memory and reuse the storage when the
memory is added next time.

This patch also introduces a new list map_entries_bootmem to link the
map entries allocated by bootmem when they are removed, and a lock to
protect it.  And these entries will be reused when the memory is
hot-added again.

The idea is suggestted by Andrew Morton.

NOTE: It is unsafe to return an entry pointer and release the
      map_entries_lock.  So we should not hold the map_entries_lock
      separately in firmware_map_find_entry() and
      firmware_map_remove_entry().  Hold the map_entries_lock across find
      and remove /sys/firmware/memmap/X operation.

       And also, users of these two functions need to be careful to
      hold the lock when using these two functions.

[tangchen@cn.fujitsu.com: Hold spinlock across find|remove /sys operation]
[tangchen@cn.fujitsu.com: fix the wrong comments of map_entries]
[tangchen@cn.fujitsu.com: reuse the storage of /sys/firmware/memmap/X/ allocated by bootmem]
[tangchen@cn.fujitsu.com: fix section mismatch problem]
[tangchen@cn.fujitsu.com: fix the doc format in drivers/firmware/memmap.c]
Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Reviewed-by: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Tang Chen <tangchen@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Julian Calaby <julian.calaby@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: remove redundant codes

offlining memory blocks and checking whether memory blocks are offlined
are very similar.  This patch introduces a new function to remove
redundant codes.

Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Reviewed-by: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: check whether all memory blocks are offlined or not when removing memory

We remove the memory like this:

 1. lock memory hotplug
 2. offline a memory block
 3. unlock memory hotplug
 4. repeat 1-3 to offline all memory blocks
 5. lock memory hotplug
 6. remove memory(TODO)
 7. unlock memory hotplug

All memory blocks must be offlined before removing memory.  But we don't
hold the lock in the whole operation.  So we should check whether all
memory blocks are offlined before step6.  Otherwise, kernel maybe
panicked.

Offlining a memory block and removing a memory device can be two
different operations.  Users can just offline some memory blocks without
removing the memory device.  For this purpose, the kernel has held
lock_memory_hotplug() in __offline_pages().  To reuse the code for
memory hot-remove, we repeat step 1-3 to offline all the memory blocks,
repeatedly lock and unlock memory hotplug, but not hold the memory
hotplug lock in the whole operation.

Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memory-hotplug: try to offline the memory twice to avoid dependence

memory can't be offlined when CONFIG_MEMCG is selected.  For example:
there is a memory device on node 1.  The address range is [1G, 1.5G).
You will find 4 new directories memory8, memory9, memory10, and memory11
under the directory /sys/devices/system/memory/.

If CONFIG_MEMCG is selected, we will allocate memory to store page
cgroup when we online pages.  When we online memory8, the memory stored
page cgroup is not provided by this memory device.  But when we online
memory9, the memory stored page cgroup may be provided by memory8.  So
we can't offline memory8 now.  We should offline the memory in the
reversed order.

When the memory device is hotremoved, we will auto offline memory
provided by this memory device.  But we don't know which memory is
onlined first, so offlining memory may fail.  In such case, iterate
twice to offline the memory.  1st iterate: offline every non primary
memory block.  2nd iterate: offline primary (i.e.  first added) memory
block.

This idea is suggested by KOSAKI Motohiro.

Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Tang Chen <tangchen@cn.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Wu Jianguo <wujianguo@huawei.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: memory_hotplug: no need to check res twice in add_memory

Remove one redundant check of res.

Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: make do_mmap_pgoff return populate as a size in bytes, not as a bool

do_mmap_pgoff() rounds up the desired size to the next PAGE_SIZE
multiple, however there was no equivalent code in mm_populate(), which
caused issues.

This could be fixed by introduced the same rounding in mm_populate(),
however I think it's preferable to make do_mmap_pgoff() return populate
as a size rather than as a boolean, so we don't have to duplicate the
size rounding logic in mm_populate().

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Tested-by: Andy Lutomirski <luto@amacapital.net>
Cc: Greg Ungerer <gregungerer@westnet.com.au>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: introduce VM_POPULATE flag to better deal with racy userspace programs

The vm_populate() code populates user mappings without constantly
holding the mmap_sem.  This makes it susceptible to racy userspace
programs: the user mappings may change while vm_populate() is running,
and in this case vm_populate() may end up populating the new mapping
instead of the old one.

In order to reduce the possibility of userspace getting surprised by
this behavior, this change introduces the VM_POPULATE vma flag which
gets set on vmas we want vm_populate() to work on.  This way
vm_populate() may still end up populating the new mapping after such a
race, but only if the new mapping is also one that the user has
requested (using MAP_SHARED, MAP_LOCKED or mlock) to be populated.

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Tested-by: Andy Lutomirski <luto@amacapital.net>
Cc: Greg Ungerer <gregungerer@westnet.com.au>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: directly use __mlock_vma_pages_range() in find_extend_vma()

In find_extend_vma(), we don't need mlock_vma_pages_range() to verify
the vma type - we know we're working with a stack.  So, we can call
directly into __mlock_vma_pages_range(), and remove the last
make_pages_present() call site.

Note that we don't use mm_populate() here, so we can't release the
mmap_sem while allocating new stack pages.  This is deemed acceptable,
because the stack vmas grow by a bounded number of pages at a time, and
these are anon pages so we don't have to read from disk to populate
them.

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Tested-by: Andy Lutomirski <luto@amacapital.net>
Cc: Greg Ungerer <gregungerer@westnet.com.au>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: remove flags argument to mmap_region

After the MAP_POPULATE handling has been moved to mmap_region() call
sites, the only remaining use of the flags argument is to pass the
MAP_NORESERVE flag.  This can be just as easily handled by
do_mmap_pgoff(), so do that and remove the mmap_region() flags
parameter.

[akpm@linux-foundation.org: remove double parens]
Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Tested-by: Andy Lutomirski <luto@amacapital.net>
Cc: Greg Ungerer <gregungerer@westnet.com.au>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: use mm_populate() for mremap() of VM_LOCKED vmas

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Tested-by: Andy Lutomirski <luto@amacapital.net>
Cc: Greg Ungerer <gregungerer@westnet.com.au>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: use mm_populate() when adjusting brk with MCL_FUTURE in effect

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Tested-by: Andy Lutomirski <luto@amacapital.net>
Cc: Greg Ungerer <gregungerer@westnet.com.au>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: use mm_populate() for blocking remap_file_pages()

Signed-off-by: Michel Lespinasse <walken@google.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Tested-by: Andy Lutomirski <luto@amacapital.net>
Cc: Greg Ungerer <gregungerer@westnet.com.au>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: introduce mm_populate() for populating new vmas

When creating new mappings using the MAP_POPULATE / MAP_LOCKED flags (or
with MCL_FUTURE in effect), we want to populate the pages within the
newly created vmas.  This may take a while as we may have to read pages
from disk, so ideally we want to do this outside of the write-locked
mmap_sem region.

This change introduces mm_populate(), which is used to defer populating
such mappings until after the mmap_sem write lock has been released.
This is implemented as a generalization of the former do_mlock_pages(),
which accomplished the same task but was using during mlock() /
mlockall().

Signed-off-by: Michel Lespinasse <walken@google.com>
Reported-by: Andy Lutomirski <luto@amacapital.net>
Acked-by: Rik van Riel <riel@redhat.com>
Tested-by: Andy Lutomirski <luto@amacapital.net>
Cc: Greg Ungerer <gregungerer@westnet.com.au>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: remap_file_pages() fixes

We have many vma manipulation functions that are fast in the typical
case, but can optionally be instructed to populate an unbounded number
of ptes within the region they work on:

 - mmap with MAP_POPULATE or MAP_LOCKED flags;
 - remap_file_pages() with MAP_NONBLOCK not set or when working on a
   VM_LOCKED vma;
 - mmap_region() and all its wrappers when mlock(MCL_FUTURE) is in
   effect;
 - brk() when mlock(MCL_FUTURE) is in effect.

Current code handles these pte operations locally, while the
sourrounding code has to hold the mmap_sem write side since it's
manipulating vmas.  This means we're doing an unbounded amount of pte
population work with mmap_sem held, and this causes problems as Andy
Lutomirski reported (we've hit this at Google as well, though it's not
entirely clear why people keep trying to use mlock(MCL_FUTURE) in the
first place).

I propose introducing a new mm_populate() function to do this pte
population work after the mmap_sem has been released.  mm_populate()
does need to acquire the mmap_sem read side, but critically, it doesn't
need to hold it continuously for the entire duration of the operation -
it can drop it whenever things take too long (such as when hitting disk
for a file read) and re-acquire it later on.

The following patches are included

- Patches 1 fixes some issues I noticed while working on the existing code.
  If needed, they could potentially go in before the rest of the patches.

- Patch 2 introduces the new mm_populate() function and changes
  mmap_region() call sites to use it after they drop mmap_sem. This is
  inspired from Andy Lutomirski's proposal and is built as an extension
  of the work I had previously done for mlock() and mlockall() around
  v2.6.38-rc1. I had tried doing something similar at the time but had
  given up as there were so many do_mmap() call sites; the recent cleanups
  by Linus and Viro are a tremendous help here.

- Patches 3-5 convert some of the less-obvious places doing unbounded
  pte populates to the new mm_populate() mechanism.

- Patches 6-7 are code cleanups that are made possible by the
  mm_populate() work. In particular, they remove more code than the
  entire patch series added, which should be a good thing :)

- Patch 8 is optional to this entire series. It only helps to deal more
  nicely with racy userspace programs that might modify their mappings
  while we're trying to populate them. It adds a new VM_POPULATE flag
  on the mappings we do want to populate, so that if userspace replaces
  them with mappings it doesn't want populated, mm_populate() won't
  populate those replacement mappings.

This patch:

Assorted small fixes. The first two are quite small:

- Move check for vma->vm_private_data && !(vma->vm_flags & VM_NONLINEAR)
  within existing if (!(vma->vm_flags & VM_NONLINEAR)) block.
  Purely cosmetic.

- In the VM_LOCKED case, when dropping PG_Mlocked for the over-mapped
  range, make sure we own the mmap_sem write lock around the
  munlock_vma_pages_range call as this manipulates the vma's vm_flags.

Last fix requires a longer explanation. remap_file_pages() can do its work
either through VM_NONLINEAR manipulation or by creating extra vmas.
These two cases were inconsistent with each other (and ultimately, both wrong)
as to exactly when did they fault in the newly mapped file pages:

- In the VM_NONLINEAR case, new file pages would be populated if
  the MAP_NONBLOCK flag wasn't passed. If MAP_NONBLOCK was passed,
  new file pages wouldn't be populated even if the vma is already
  marked as VM_LOCKED.

- In the linear (emulated) case, the work is passed to the mmap_region()
  function which would populate the pages if the vma is marked as
  VM_LOCKED, and would not otherwise - regardless of the value of the
  MAP_NONBLOCK flag, because MAP_POPULATE wasn't being passed to
  mmap_region().

The desired behavior is that we want the pages to be populated and locked
if the vma is marked as VM_LOCKED, or to be populated if the MAP_NONBLOCK
flag is not passed to remap_file_pages().

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Tested-by: Andy Lutomirski <luto@amacapital.net>
Cc: Greg Ungerer <gregungerer@westnet.com.au>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: avoid calling pgdat_balanced() needlessly

Now that balance_pgdat() is slightly tidied up, thanks to more capable
pgdat_balanced(), it's become obvious that pgdat_balanced() is called to
check the status, then break the loop if pgdat is balanced, just to be
immediately called again.  The second call is completely unnecessary, of
course.

The patch introduces pgdat_is_balanced boolean, which helps resolve the
above suboptimal behavior, with the added benefit of slightly better
documenting one other place in the function where we jump and skip lots
of code.

Signed-off-by: Zlatko Calusic <zlatko.calusic@iskon.hr>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
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: compaction: make __compact_pgdat() and compact_pgdat() return void

These functions always return 0.  Formalise this.

Cc: Jason Liu <r64343@freescale.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.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: make madvise(MADV_WILLNEED) support swap file prefetch

Make madvise(MADV_WILLNEED) support swap file prefetch.  If memory is
swapout, this syscall can do swapin prefetch.  It has no impact if the
memory isn't swapout.

[akpm@linux-foundation.org: fix CONFIG_SWAP=n build]
[sasha.levin@oracle.com: fix BUG on madvise early failure]
Signed-off-by: Shaohua Li <shli@fusionio.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

memcg,vmscan: do not break out targeted reclaim without reclaimed pages

Targeted (hard resp soft) reclaim has traditionally tried to scan one
group with decreasing priority until nr_to_reclaim (SWAP_CLUSTER_MAX
pages) is reclaimed or all priorities are exhausted.  The reclaim is
then retried until the limit is met.

This approach, however, doesn't work well with deeper hierarchies where
groups higher in the hierarchy do not have any or only very few pages
(this usually happens if those groups do not have any tasks and they
have only re-parented pages after some of their children is removed).
Those groups are reclaimed with decreasing priority pointlessly as there
is nothing to reclaim from them.

An easiest fix is to break out of the memcg iteration loop in
shrink_zone only if the whole hierarchy has been visited or sufficient
pages have been reclaimed.  This is also more natural because the
reclaimer expects that the hierarchy under the given root is reclaimed.
As a result we can simplify the soft limit reclaim which does its own
iteration.

[yinghan@google.com: break out of the hierarchy loop only if nr_reclaimed exceeded nr_to_reclaim]
[akpm@linux-foundation.org: use conventional comparison order]
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Reported-by: Ying Han <yinghan@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <htejun@gmail.com>
Cc: Glauber Costa <glommer@parallels.com>
Cc: Li Zefan <lizefan@huawei.com>
Signed-off-by: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/ksm.c: use new hashtable implementation

Switch ksm to use the new hashtable implementation.  This reduces the
amount of generic unrelated code in the ksm module.

Signed-off-by: Sasha Levin <levinsasha928@gmail.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: Mel Gorman <mgorman@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/huge_memory.c: use new hashtable implementation

Switch hugemem to use the new hashtable implementation.  This reduces
the amount of generic unrelated code in the hugemem.

This also removes the dymanic allocation of the hash table.  The upside
is that we save a pointer dereference when accessing the hashtable, but
we lose 8KB if CONFIG_TRANSPARENT_HUGEPAGE is enabled but the processor
doesn't support hugepages.

Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: 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: do not accidentally skip pageblocks in the migrate scanner

Compaction uses the ALIGN macro incorrectly with the migrate scanner by
adding pageblock_nr_pages to a PFN.  It happened to work when initially
implemented as the starting PFN was also aligned but with caching
restarts and isolating in smaller chunks this is no longer always true.

The impact is that the migrate scanner scans outside its current
pageblock.  As pfn_valid() is still checked properly it does not cause
any failure and the impact of the bug is that in some cases it will scan
more than necessary when it crosses a page boundary but by no more than
COMPACT_CLUSTER_MAX.  It is highly unlikely this is even measurable but
it's still wrong so this patch addresses the problem.

Signed-off-by: Mel Gorman <mgorman@suse.de>
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>

mm/vmscan.c:__zone_reclaim(): replace max_t() with max()

"mm: vmscan: save work scanning (almost) empty LRU lists" made
SWAP_CLUSTER_MAX an unsigned long.

Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: Satoru Moriya <satoru.moriya@hds.com>
Cc: Simon Jeons <simon.jeons@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/page_alloc.c:__setup_per_zone_wmarks: make min_pages unsigned long

`int' is an inappropriate type for a number-of-pages counter.

While we're there, use the clamp() macro.

Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: Satoru Moriya <satoru.moriya@hds.com>
Cc: Simon Jeons <simon.jeons@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: reduce rmap overhead for ex-KSM page copies created on swap faults

When ex-KSM pages are faulted from swap cache, the fault handler is not
capable of re-establishing anon_vma-spanning KSM pages.  In this case, a
copy of the page is created instead, just like during a COW break.

These freshly made copies are known to be exclusive to the faulting VMA
and there is no reason to go look for this page in parent and sibling
processes during rmap operations.

Use page_add_new_anon_rmap() for these copies.  This also puts them on
the proper LRU lists and marks them SwapBacked, so we can get rid of
doing this ad-hoc in the KSM copy code.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Simon Jeons <simon.jeons@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Satoru Moriya <satoru.moriya@hds.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm: vmscan: compaction works against zones, not lruvecs

The restart logic for when reclaim operates back to back with compaction
is currently applied on the lruvec level.  But this does not make sense,
because the container of interest for compaction is a zone as a whole,
not the zone pages that are part of a certain memory cgroup.

Negative impact is bounded.  For one, the code checks that the lruvec
has enough reclaim candidates, so it does not risk getting stuck on a
condition that can not be fulfilled.  And the unfairness of hammering on
one particular memory cgroup to make progress in a zone will be
amortized by the round robin manner in which reclaim goes through the
memory cgroups.  Still, this can lead to unnecessary allocation
latencies when the code elects to restart on a hard to reclaim or small
group when there are other, more reclaimable groups in the zone.

Move this logic to the zone level and restart reclaim for all memory
cgroups in a zone when compaction requires more free pages from it.

[akpm@linux-foundation.org: no need for min_t]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: Satoru Moriya <satoru.moriya@hds.com>
Cc: Simon Jeons <simon.jeons@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>