ksm: introduce ksm_max_page_sharing per page deduplication limit

Without a max deduplication limit for each KSM page, the list of the
rmap_items associated to each stable_node can grow infinitely large.

During the rmap walk each entry can take up to ~10usec to process
because of IPIs for the TLB flushing (both for the primary MMU and the
secondary MMUs with the MMU notifier).  With only 16GB of address space
shared in the same KSM page, that would amount to dozens of seconds of
kernel runtime.

A ~256 max deduplication factor will reduce the latencies of the rmap
walks on KSM pages to order of a few msec.  Just doing the
cond_resched() during the rmap walks is not enough, the list size must
have a limit too, otherwise the caller could get blocked in (schedule
friendly) kernel computations for seconds, unexpectedly.

There's room for optimization to significantly reduce the IPI delivery
cost during the page_referenced(), but at least for page_migration in
the KSM case (used by hard NUMA bindings, compaction and NUMA balancing)
it may be inevitable to send lots of IPIs if each rmap_item->mm is
active on a different CPU and there are lots of CPUs.  Even if we ignore
the IPI delivery cost, we've still to walk the whole KSM rmap list, so
we can't allow millions or billions (ulimited) number of entries in the
KSM stable_node rmap_item lists.

The limit is enforced efficiently by adding a second dimension to the
stable rbtree.  So there are three types of stable_nodes: the regular
ones (identical as before, living in the first flat dimension of the
stable rbtree), the "chains" and the "dups".

Every "chain" and all "dups" linked into a "chain" enforce the invariant
that they represent the same write protected memory content, even if
each "dup" will be pointed by a different KSM page copy of that content.
This way the stable rbtree lookup computational complexity is unaffected
if compared to an unlimited max_sharing_limit.  It is still enforced
that there cannot be KSM page content duplicates in the stable rbtree
itself.

Adding the second dimension to the stable rbtree only after the
max_page_sharing limit hits, provides for a zero memory footprint
increase on 64bit archs.  The memory overhead of the per-KSM page
stable_tree and per virtual mapping rmap_item is unchanged.  Only after
the max_page_sharing limit hits, we need to allocate a stable_tree
"chain" and rb_replace() the "regular" stable_node with the newly
allocated stable_node "chain".  After that we simply add the "regular"
stable_node to the chain as a stable_node "dup" by linking hlist_dup in
the stable_node_chain->hlist.  This way the "regular" (flat) stable_node
is converted to a stable_node "dup" living in the second dimension of
the stable rbtree.

During stable rbtree lookups the stable_node "chain" is identified as
stable_node->rmap_hlist_len == STABLE_NODE_CHAIN (aka
is_stable_node_chain()).

When dropping stable_nodes, the stable_node "dup" is identified as
stable_node->head == STABLE_NODE_DUP_HEAD (aka is_stable_node_dup()).

The STABLE_NODE_DUP_HEAD must be an unique valid pointer never used
elsewhere in any stable_node->head/node to avoid a clashes with the
stable_node->node.rb_parent_color pointer, and different from
&migrate_nodes.  So the second field of &migrate_nodes is picked and
verified as always safe with a BUILD_BUG_ON in case the list_head
implementation changes in the future.

The STABLE_NODE_DUP is picked as a random negative value in
stable_node->rmap_hlist_len.  rmap_hlist_len cannot become negative when
it's a "regular" stable_node or a stable_node "dup".

The stable_node_chain->nid is irrelevant.  The stable_node_chain->kpfn
is aliased in a union with a time field used to rate limit the
stable_node_chain->hlist prunes.

The garbage collection of the stable_node_chain happens lazily during
stable rbtree lookups (as for all other kind of stable_nodes), or while
disabling KSM with "echo 2 >/sys/kernel/mm/ksm/run" while collecting the
entire stable rbtree.

While the "regular" stable_nodes and the stable_node "dups" must wait
for their underlying tree_page to be freed before they can be freed
themselves, the stable_node "chains" can be freed immediately if the
stable_node->hlist turns empty.  This is because the "chains" are never
pointed by any page->mapping and they're effectively stable rbtree KSM
self contained metadata.

[[email protected]: fix non-NUMA build]
Signed-off-by: Andrea Arcangeli <[email protected]>
Tested-by: Petr Holasek <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Davidlohr Bueso <[email protected]>
Cc: Arjan van de Ven <[email protected]>
Cc: Evgheni Dereveanchin <[email protected]>
Cc: Andrey Ryabinin <[email protected]>
Cc: Gavin Guo <[email protected]>
Cc: Jay Vosburgh <[email protected]>
Cc: Mel Gorman <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>

Author: aagit

Documentation/vm/ksm.txt

@@ -98,6 +98,50 @@ use_zero_pages   - specifies whether empty pages (i.e. allocated pages
                    it is only effective for pages merged after the change.
                    Default: 0 (normal KSM behaviour as in earlier releases)
 
+max_page_sharing - Maximum sharing allowed for each KSM page. This
+                   enforces a deduplication limit to avoid the virtual
+                   memory rmap lists to grow too large. The minimum
+                   value is 2 as a newly created KSM page will have at
+                   least two sharers. The rmap walk has O(N)
+                   complexity where N is the number of rmap_items
+                   (i.e. virtual mappings) that are sharing the page,
+                   which is in turn capped by max_page_sharing. So
+                   this effectively spread the the linear O(N)
+                   computational complexity from rmap walk context
+                   over different KSM pages. The ksmd walk over the
+                   stable_node "chains" is also O(N), but N is the
+                   number of stable_node "dups", not the number of
+                   rmap_items, so it has not a significant impact on
+                   ksmd performance. In practice the best stable_node
+                   "dup" candidate will be kept and found at the head
+                   of the "dups" list. The higher this value the
+                   faster KSM will merge the memory (because there
+                   will be fewer stable_node dups queued into the
+                   stable_node chain->hlist to check for pruning) and
+                   the higher the deduplication factor will be, but
+                   the slowest the worst case rmap walk could be for
+                   any given KSM page. Slowing down the rmap_walk
+                   means there will be higher latency for certain
+                   virtual memory operations happening during
+                   swapping, compaction, NUMA balancing and page
+                   migration, in turn decreasing responsiveness for
+                   the caller of those virtual memory operations. The
+                   scheduler latency of other tasks not involved with
+                   the VM operations doing the rmap walk is not
+                   affected by this parameter as the rmap walks are
+                   always schedule friendly themselves.
+
+stable_node_chains_prune_millisecs - How frequently to walk the whole
+                   list of stable_node "dups" linked in the
+                   stable_node "chains" in order to prune stale
+                   stable_nodes. Smaller milllisecs values will free
+                   up the KSM metadata with lower latency, but they
+                   will make ksmd use more CPU during the scan. This
+                   only applies to the stable_node chains so it's a
+                   noop if not a single KSM page hit the
+                   max_page_sharing yet (there would be no stable_node
+                   chains in such case).
+
 The effectiveness of KSM and MADV_MERGEABLE is shown in /sys/kernel/mm/ksm/:
 
 pages_shared     - how many shared pages are being used
@@ -106,10 +150,29 @@ pages_unshared   - how many pages unique but repeatedly checked for merging
 pages_volatile   - how many pages changing too fast to be placed in a tree
 full_scans       - how many times all mergeable areas have been scanned
 
+stable_node_chains - number of stable node chains allocated, this is
+		     effectively the number of KSM pages that hit the
+		     max_page_sharing limit
+stable_node_dups   - number of stable node dups queued into the
+		     stable_node chains
+
 A high ratio of pages_sharing to pages_shared indicates good sharing, but
 a high ratio of pages_unshared to pages_sharing indicates wasted effort.
 pages_volatile embraces several different kinds of activity, but a high
 proportion there would also indicate poor use of madvise MADV_MERGEABLE.
 
+The maximum possible page_sharing/page_shared ratio is limited by the
+max_page_sharing tunable. To increase the ratio max_page_sharing must
+be increased accordingly.
+
+The stable_node_dups/stable_node_chains ratio is also affected by the
+max_page_sharing tunable, and an high ratio may indicate fragmentation
+in the stable_node dups, which could be solved by introducing
+fragmentation algorithms in ksmd which would refile rmap_items from
+one stable_node dup to another stable_node dup, in order to freeup
+stable_node "dups" with few rmap_items in them, but that may increase
+the ksmd CPU usage and possibly slowdown the readonly computations on
+the KSM pages of the applications.
+
 Izik Eidus,
 Hugh Dickins, 17 Nov 2009