1.. _ksm:
2
3=======================
4Kernel Samepage Merging
5=======================
6
7KSM is a memory-saving de-duplication feature, enabled by CONFIG_KSM=y,
8added to the Linux kernel in 2.6.32.  See ``mm/ksm.c`` for its implementation,
9and http://lwn.net/Articles/306704/ and http://lwn.net/Articles/330589/
10
11The userspace interface of KSM is described in :ref:`Documentation/admin-guide/mm/ksm.rst <admin_guide_ksm>`
12
13Design
14======
15
16Overview
17--------
18
19.. kernel-doc:: mm/ksm.c
20   :DOC: Overview
21
22Reverse mapping
23---------------
24KSM maintains reverse mapping information for KSM pages in the stable
25tree.
26
27If a KSM page is shared between less than ``max_page_sharing`` VMAs,
28the node of the stable tree that represents such KSM page points to a
29list of :c:type:`struct rmap_item` and the ``page->mapping`` of the
30KSM page points to the stable tree node.
31
32When the sharing passes this threshold, KSM adds a second dimension to
33the stable tree. The tree node becomes a "chain" that links one or
34more "dups". Each "dup" keeps reverse mapping information for a KSM
35page with ``page->mapping`` pointing to that "dup".
36
37Every "chain" and all "dups" linked into a "chain" enforce the
38invariant that they represent the same write protected memory content,
39even if each "dup" will be pointed by a different KSM page copy of
40that content.
41
42This way the stable tree lookup computational complexity is unaffected
43if compared to an unlimited list of reverse mappings. It is still
44enforced that there cannot be KSM page content duplicates in the
45stable tree itself.
46
47The deduplication limit enforced by ``max_page_sharing`` is required
48to avoid the virtual memory rmap lists to grow too large. The rmap
49walk has O(N) complexity where N is the number of rmap_items
50(i.e. virtual mappings) that are sharing the page, which is in turn
51capped by ``max_page_sharing``. So this effectively spreads the linear
52O(N) computational complexity from rmap walk context over different
53KSM pages. The ksmd walk over the stable_node "chains" is also O(N),
54but N is the number of stable_node "dups", not the number of
55rmap_items, so it has not a significant impact on ksmd performance. In
56practice the best stable_node "dup" candidate will be kept and found
57at the head of the "dups" list.
58
59High values of ``max_page_sharing`` result in faster memory merging
60(because there will be fewer stable_node dups queued into the
61stable_node chain->hlist to check for pruning) and higher
62deduplication factor at the expense of slower worst case for rmap
63walks for any KSM page which can happen during swapping, compaction,
64NUMA balancing and page migration.
65
66The ``stable_node_dups/stable_node_chains`` ratio is also affected by the
67``max_page_sharing`` tunable, and an high ratio may indicate fragmentation
68in the stable_node dups, which could be solved by introducing
69fragmentation algorithms in ksmd which would refile rmap_items from
70one stable_node dup to another stable_node dup, in order to free up
71stable_node "dups" with few rmap_items in them, but that may increase
72the ksmd CPU usage and possibly slowdown the readonly computations on
73the KSM pages of the applications.
74
75The whole list of stable_node "dups" linked in the stable_node
76"chains" is scanned periodically in order to prune stale stable_nodes.
77The frequency of such scans is defined by
78``stable_node_chains_prune_millisecs`` sysfs tunable.
79
80Reference
81---------
82.. kernel-doc:: mm/ksm.c
83   :functions: mm_slot ksm_scan stable_node rmap_item
84
85--
86Izik Eidus,
87Hugh Dickins, 17 Nov 2009
88