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