linux/Documentation/vm/ksm.rst

251 lines
11 KiB
ReStructuredText
Raw Normal View History

.. _ksm:
=======================
Kernel Samepage Merging
=======================
Overview
========
KSM is a memory-saving de-duplication feature, enabled by CONFIG_KSM=y,
added to the Linux kernel in 2.6.32. See ``mm/ksm.c`` for its implementation,
and http://lwn.net/Articles/306704/ and http://lwn.net/Articles/330589/
KSM was originally developed for use with KVM (where it was known as
Kernel Shared Memory), to fit more virtual machines into physical memory,
by sharing the data common between them. But it can be useful to any
application which generates many instances of the same data.
The KSM daemon ksmd periodically scans those areas of user memory
which have been registered with it, looking for pages of identical
content which can be replaced by a single write-protected page (which
is automatically copied if a process later wants to update its
content). The amount of pages that KSM daemon scans in a single pass
and the time between the passes are configured using :ref:`sysfs
intraface <ksm_sysfs>`
KSM only merges anonymous (private) pages, never pagecache (file) pages.
KSM's merged pages were originally locked into kernel memory, but can now
be swapped out just like other user pages (but sharing is broken when they
are swapped back in: ksmd must rediscover their identity and merge again).
Controlling KSM with madvise
============================
KSM only operates on those areas of address space which an application
has advised to be likely candidates for merging, by using the madvise(2)
system call::
int madvise(addr, length, MADV_MERGEABLE)
The app may call
::
int madvise(addr, length, MADV_UNMERGEABLE)
to cancel that advice and restore unshared pages: whereupon KSM
unmerges whatever it merged in that range. Note: this unmerging call
may suddenly require more memory than is available - possibly failing
with EAGAIN, but more probably arousing the Out-Of-Memory killer.
If KSM is not configured into the running kernel, madvise MADV_MERGEABLE
and MADV_UNMERGEABLE simply fail with EINVAL. If the running kernel was
built with CONFIG_KSM=y, those calls will normally succeed: even if the
the KSM daemon is not currently running, MADV_MERGEABLE still registers
the range for whenever the KSM daemon is started; even if the range
cannot contain any pages which KSM could actually merge; even if
MADV_UNMERGEABLE is applied to a range which was never MADV_MERGEABLE.
If a region of memory must be split into at least one new MADV_MERGEABLE
or MADV_UNMERGEABLE region, the madvise may return ENOMEM if the process
will exceed ``vm.max_map_count`` (see Documentation/sysctl/vm.txt).
Like other madvise calls, they are intended for use on mapped areas of
the user address space: they will report ENOMEM if the specified range
includes unmapped gaps (though working on the intervening mapped areas),
and might fail with EAGAIN if not enough memory for internal structures.
Applications should be considerate in their use of MADV_MERGEABLE,
restricting its use to areas likely to benefit. KSM's scans may use a lot
of processing power: some installations will disable KSM for that reason.
.. _ksm_sysfs:
KSM daemon sysfs interface
==========================
The KSM daemon is controlled by sysfs files in ``/sys/kernel/mm/ksm/``,
readable by all but writable only by root:
pages_to_scan
how many pages to scan before ksmd goes to sleep
e.g. ``echo 100 > /sys/kernel/mm/ksm/pages_to_scan``.
Default: 100 (chosen for demonstration purposes)
sleep_millisecs
how many milliseconds ksmd should sleep before next scan
e.g. ``echo 20 > /sys/kernel/mm/ksm/sleep_millisecs``
Default: 20 (chosen for demonstration purposes)
merge_across_nodes
specifies if pages from different NUMA nodes can be merged.
When set to 0, ksm merges only pages which physically reside
in the memory area of same NUMA node. That brings lower
latency to access of shared pages. Systems with more nodes, at
significant NUMA distances, are likely to benefit from the
lower latency of setting 0. Smaller systems, which need to
minimize memory usage, are likely to benefit from the greater
sharing of setting 1 (default). You may wish to compare how
your system performs under each setting, before deciding on
which to use. ``merge_across_nodes`` setting can be changed only
when there are no ksm shared pages in the system: set run 2 to
unmerge pages first, then to 1 after changing
``merge_across_nodes``, to remerge according to the new setting.
Default: 1 (merging across nodes as in earlier releases)
run
* set to 0 to stop ksmd from running but keep merged pages,
* set to 1 to run ksmd e.g. ``echo 1 > /sys/kernel/mm/ksm/run``,
* set to 2 to stop ksmd and unmerge all pages currently merged, but
leave mergeable areas registered for next run.
Default: 0 (must be changed to 1 to activate KSM, except if
CONFIG_SYSFS is disabled)
use_zero_pages
specifies whether empty pages (i.e. allocated pages that only
contain zeroes) should be treated specially. When set to 1,
empty pages are merged with the kernel zero page(s) instead of
with each other as it would happen normally. This can improve
the performance on architectures with coloured zero pages,
depending on the workload. Care should be taken when enabling
this setting, as it can potentially degrade the performance of
KSM for some workloads, for example if the checksums of pages
candidate for merging match the checksum of an empty
page. This setting can be changed at any time, 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 spreads 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
pages_sharing
how many more sites are sharing them i.e. how much saved
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
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. [akpm@linux-foundation.org: fix non-NUMA build] Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> Tested-by: Petr Holasek <pholasek@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Evgheni Dereveanchin <ederevea@redhat.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Gavin Guo <gavin.guo@canonical.com> Cc: Jay Vosburgh <jay.vosburgh@canonical.com> Cc: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 06:36:55 +08:00
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 ``pages_sharing/pages_shared`` ratio is limited by the
``max_page_sharing`` tunable. To increase the ratio ``max_page_sharing`` must
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. [akpm@linux-foundation.org: fix non-NUMA build] Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> Tested-by: Petr Holasek <pholasek@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Evgheni Dereveanchin <ederevea@redhat.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Gavin Guo <gavin.guo@canonical.com> Cc: Jay Vosburgh <jay.vosburgh@canonical.com> Cc: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 06:36:55 +08:00
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
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. [akpm@linux-foundation.org: fix non-NUMA build] Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> Tested-by: Petr Holasek <pholasek@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Evgheni Dereveanchin <ederevea@redhat.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Gavin Guo <gavin.guo@canonical.com> Cc: Jay Vosburgh <jay.vosburgh@canonical.com> Cc: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 06:36:55 +08:00
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 free up
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. [akpm@linux-foundation.org: fix non-NUMA build] Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> Tested-by: Petr Holasek <pholasek@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Evgheni Dereveanchin <ederevea@redhat.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Gavin Guo <gavin.guo@canonical.com> Cc: Jay Vosburgh <jay.vosburgh@canonical.com> Cc: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 06:36:55 +08:00
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.
Design
======
Overview
--------
.. kernel-doc:: mm/ksm.c
:DOC: Overview
Reverse mapping
---------------
KSM maintains reverse mapping information for KSM pages in the stable
tree.
If a KSM page is shared between less than ``max_page_sharing`` VMAs,
the node of the stable tree that represents such KSM page points to a
list of :c:type:`struct rmap_item` and the ``page->mapping`` of the
KSM page points to the stable tree node.
When the sharing passes this threshold, KSM adds a second dimension to
the stable tree. The tree node becomes a "chain" that links one or
more "dups". Each "dup" keeps reverse mapping information for a KSM
page with ``page->mapping`` pointing to that "dup".
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 tree lookup computational complexity is unaffected
if compared to an unlimited list of reverse mappings. It is still
enforced that there cannot be KSM page content duplicates in the
stable tree itself.
Reference
---------
.. kernel-doc:: mm/ksm.c
:functions: mm_slot ksm_scan stable_node rmap_item
--
Izik Eidus,
Hugh Dickins, 17 Nov 2009