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