[173/212] mm/numa: automatically generate node migration order

Commit Message

Andrew Morton Sept. 2, 2021, 9:59 p.m. UTC

From: Dave Hansen <[e-mail]>
Subject: mm/numa: automatically generate node migration order

Patch series "Migrate Pages in lieu of discard", v11.

We're starting to see systems with more and more kinds of memory such as
Intel's implementation of persistent memory.

Let's say you have a system with some DRAM and some persistent memory.
Today, once DRAM fills up, reclaim will start and some of the DRAM
contents will be thrown out.  Allocations will, at some point, start
falling over to the slower persistent memory.

That has two nasty properties.  First, the newer allocations can end up in
the slower persistent memory.  Second, reclaimed data in DRAM are just
discarded even if there are gobs of space in persistent memory that could
be used.

This patchset implements a solution to these problems.  At the end of the
reclaim process in shrink_page_list() just before the last page refcount
is dropped, the page is migrated to persistent memory instead of being
dropped.

While I've talked about a DRAM/PMEM pairing, this approach would function
in any environment where memory tiers exist.

This is not perfect.  It "strands" pages in slower memory and never brings
them back to fast DRAM.  Huang Ying has follow-on work which repurposes
NUMA balancing to promote hot pages back to DRAM.

This is also all based on an upstream mechanism that allows persistent
memory to be onlined and used as if it were volatile:

	http://lkml.kernel.org/r/[e-mail]

With that, the DRAM and PMEM in each socket will be represented as 2
separate NUMA nodes, with the CPUs sit in the DRAM node.  So the
general inter-NUMA demotion mechanism introduced in the patchset can
migrate the cold DRAM pages to the PMEM node.

We have tested the patchset with the postgresql and pgbench.  On a
2-socket server machine with DRAM and PMEM, the kernel with the patchset
can improve the score of pgbench up to 22.1% compared with that of the
DRAM only + disk case.  This comes from the reduced disk read throughput
(which reduces up to 70.8%).

== Open Issues ==

 * Memory policies and cpusets that, for instance, restrict allocations
   to DRAM can be demoted to PMEM whenever they opt in to this
   new mechanism.  A cgroup-level API to opt-in or opt-out of
   these migrations will likely be required as a follow-on.
 * Could be more aggressive about where anon LRU scanning occurs
   since it no longer necessarily involves I/O.  get_scan_count()
   for instance says: "If we have no swap space, do not bother
   scanning anon pages"


This patch (of 9):

Prepare for the kernel to auto-migrate pages to other memory nodes with a
node migration table.  This allows creating single migration target for
each NUMA node to enable the kernel to do NUMA page migrations instead of
simply discarding colder pages.  A node with no target is a "terminal
node", so reclaim acts normally there.  The migration target does not
fundamentally _need_ to be a single node, but this implementation starts
there to limit complexity.

When memory fills up on a node, memory contents can be automatically
migrated to another node.  The biggest problems are knowing when to
migrate and to where the migration should be targeted.

The most straightforward way to generate the "to where" list would be to
follow the page allocator fallback lists.  Those lists already tell us if
memory is full where to look next.  It would also be logical to move
memory in that order.

But, the allocator fallback lists have a fatal flaw: most nodes appear in
all the lists.  This would potentially lead to migration cycles (A->B,
B->A, A->B, ...).

Instead of using the allocator fallback lists directly, keep a separate
node migration ordering.  But, reuse the same data used to generate page
allocator fallback in the first place: find_next_best_node().

This means that the firmware data used to populate node distances
essentially dictates the ordering for now.  It should also be
architecture-neutral since all NUMA architectures have a working
find_next_best_node().

RCU is used to allow lock-less read of node_demotion[] and prevent
demotion cycles been observed.  If multiple reads of node_demotion[] are
performed, a single rcu_read_lock() must be held over all reads to ensure
no cycles are observed.  Details are as follows.

=== What does RCU provide? ===

Imagine a simple loop which walks down the demotion path looking
for the last node:

        terminal_node = start_node;
        while (node_demotion[terminal_node] != NUMA_NO_NODE) {
                terminal_node = node_demotion[terminal_node];
        }

The initial values are:

        node_demotion[0] = 1;
        node_demotion[1] = NUMA_NO_NODE;

and are updated to:

        node_demotion[0] = NUMA_NO_NODE;
        node_demotion[1] = 0;

What guarantees that the cycle is not observed:

        node_demotion[0] = 1;
        node_demotion[1] = 0;

and would loop forever?

With RCU, a rcu_read_lock/unlock() can be placed around the loop.  Since
the write side does a synchronize_rcu(), the loop that observed the old
contents is known to be complete before the synchronize_rcu() has
completed.

RCU, combined with disable_all_migrate_targets(), ensures that the old
migration state is not visible by the time __set_migration_target_nodes()
is called.

=== What does READ_ONCE() provide? ===

READ_ONCE() forbids the compiler from merging or reordering successive
reads of node_demotion[].  This ensures that any updates are *eventually*
observed.

Consider the above loop again.  The compiler could theoretically read the
entirety of node_demotion[] into local storage (registers) and never go
back to memory, and *permanently* observe bad values for node_demotion[].

Note: RCU does not provide any universal compiler-ordering
guarantees:

	https://lore.kernel.org/lkml/[e-mail]/

This code is unused for now.  It will be called later in the
series.

Link: https://lkml.kernel.org/r/[e-mail]
Link: https://lkml.kernel.org/r/[e-mail]
Link: https://lkml.kernel.org/r/[e-mail]
Signed-off-by: Dave Hansen <[e-mail]>
Signed-off-by: "Huang, Ying" <[e-mail]>
Reviewed-by: Yang Shi <[e-mail]>
Reviewed-by: Zi Yan <[e-mail]>
Reviewed-by: Oscar Salvador <[e-mail]>
Cc: Michal Hocko <[e-mail]>
Cc: Wei Xu <[e-mail]>
Cc: David Rientjes <[e-mail]>
Cc: Dan Williams <[e-mail]>
Cc: David Hildenbrand <[e-mail]>
Cc: Greg Thelen <[e-mail]>
Cc: Keith Busch <[e-mail]>
Cc: Yang Shi <[e-mail]>
Signed-off-by: Andrew Morton <[e-mail]>
---

 mm/internal.h   |    5 +
 mm/migrate.c    |  216 ++++++++++++++++++++++++++++++++++++++++++++++
 mm/page_alloc.c |    2 
 3 files changed, 222 insertions(+), 1 deletion(-)

Patch

--- a/mm/internal.h~mm-numa-automatically-generate-node-migration-order
+++ a/mm/internal.h
@@ -543,12 +543,17 @@  static inline void mminit_validate_memmo
 
 #ifdef CONFIG_NUMA
 extern int node_reclaim(struct pglist_data *, gfp_t, unsigned int);
+extern int find_next_best_node(int node, nodemask_t *used_node_mask);
 #else
 static inline int node_reclaim(struct pglist_data *pgdat, gfp_t mask,
 				unsigned int order)
 {
 	return NODE_RECLAIM_NOSCAN;
 }
+static inline int find_next_best_node(int node, nodemask_t *used_node_mask)
+{
+	return NUMA_NO_NODE;
+}
 #endif
 
 extern int hwpoison_filter(struct page *p);
--- a/mm/migrate.c~mm-numa-automatically-generate-node-migration-order
+++ a/mm/migrate.c
@@ -1099,6 +1099,80 @@  out:
 	return rc;
 }
 
+
+/*
+ * node_demotion[] example:
+ *
+ * Consider a system with two sockets.  Each socket has
+ * three classes of memory attached: fast, medium and slow.
+ * Each memory class is placed in its own NUMA node.  The
+ * CPUs are placed in the node with the "fast" memory.  The
+ * 6 NUMA nodes (0-5) might be split among the sockets like
+ * this:
+ *
+ *	Socket A: 0, 1, 2
+ *	Socket B: 3, 4, 5
+ *
+ * When Node 0 fills up, its memory should be migrated to
+ * Node 1.  When Node 1 fills up, it should be migrated to
+ * Node 2.  The migration path start on the nodes with the
+ * processors (since allocations default to this node) and
+ * fast memory, progress through medium and end with the
+ * slow memory:
+ *
+ *	0 -> 1 -> 2 -> stop
+ *	3 -> 4 -> 5 -> stop
+ *
+ * This is represented in the node_demotion[] like this:
+ *
+ *	{  1, // Node 0 migrates to 1
+ *	   2, // Node 1 migrates to 2
+ *	  -1, // Node 2 does not migrate
+ *	   4, // Node 3 migrates to 4
+ *	   5, // Node 4 migrates to 5
+ *	  -1} // Node 5 does not migrate
+ */
+
+/*
+ * Writes to this array occur without locking.  Cycles are
+ * not allowed: Node X demotes to Y which demotes to X...
+ *
+ * If multiple reads are performed, a single rcu_read_lock()
+ * must be held over all reads to ensure that no cycles are
+ * observed.
+ */
+static int node_demotion[MAX_NUMNODES] __read_mostly =
+	{[0 ...  MAX_NUMNODES - 1] = NUMA_NO_NODE};
+
+/**
+ * next_demotion_node() - Get the next node in the demotion path
+ * @node: The starting node to lookup the next node
+ *
+ * @returns: node id for next memory node in the demotion path hierarchy
+ * from @node; NUMA_NO_NODE if @node is terminal.  This does not keep
+ * @node online or guarantee that it *continues* to be the next demotion
+ * target.
+ */
+int next_demotion_node(int node)
+{
+	int target;
+
+	/*
+	 * node_demotion[] is updated without excluding this
+	 * function from running.  RCU doesn't provide any
+	 * compiler barriers, so the READ_ONCE() is required
+	 * to avoid compiler reordering or read merging.
+	 *
+	 * Make sure to use RCU over entire code blocks if
+	 * node_demotion[] reads need to be consistent.
+	 */
+	rcu_read_lock();
+	target = READ_ONCE(node_demotion[node]);
+	rcu_read_unlock();
+
+	return target;
+}
+
 /*
  * Obtain the lock on page, remove all ptes and migrate the page
  * to the newly allocated page in newpage.
@@ -2982,3 +3056,145 @@  void migrate_vma_finalize(struct migrate
 }
 EXPORT_SYMBOL(migrate_vma_finalize);
 #endif /* CONFIG_DEVICE_PRIVATE */
+
+/* Disable reclaim-based migration. */
+static void __disable_all_migrate_targets(void)
+{
+	int node;
+
+	for_each_online_node(node)
+		node_demotion[node] = NUMA_NO_NODE;
+}
+
+static void disable_all_migrate_targets(void)
+{
+	__disable_all_migrate_targets();
+
+	/*
+	 * Ensure that the "disable" is visible across the system.
+	 * Readers will see either a combination of before+disable
+	 * state or disable+after.  They will never see before and
+	 * after state together.
+	 *
+	 * The before+after state together might have cycles and
+	 * could cause readers to do things like loop until this
+	 * function finishes.  This ensures they can only see a
+	 * single "bad" read and would, for instance, only loop
+	 * once.
+	 */
+	synchronize_rcu();
+}
+
+/*
+ * Find an automatic demotion target for 'node'.
+ * Failing here is OK.  It might just indicate
+ * being at the end of a chain.
+ */
+static int establish_migrate_target(int node, nodemask_t *used)
+{
+	int migration_target;
+
+	/*
+	 * Can not set a migration target on a
+	 * node with it already set.
+	 *
+	 * No need for READ_ONCE() here since this
+	 * in the write path for node_demotion[].
+	 * This should be the only thread writing.
+	 */
+	if (node_demotion[node] != NUMA_NO_NODE)
+		return NUMA_NO_NODE;
+
+	migration_target = find_next_best_node(node, used);
+	if (migration_target == NUMA_NO_NODE)
+		return NUMA_NO_NODE;
+
+	node_demotion[node] = migration_target;
+
+	return migration_target;
+}
+
+/*
+ * When memory fills up on a node, memory contents can be
+ * automatically migrated to another node instead of
+ * discarded at reclaim.
+ *
+ * Establish a "migration path" which will start at nodes
+ * with CPUs and will follow the priorities used to build the
+ * page allocator zonelists.
+ *
+ * The difference here is that cycles must be avoided.  If
+ * node0 migrates to node1, then neither node1, nor anything
+ * node1 migrates to can migrate to node0.
+ *
+ * This function can run simultaneously with readers of
+ * node_demotion[].  However, it can not run simultaneously
+ * with itself.  Exclusion is provided by memory hotplug events
+ * being single-threaded.
+ */
+static void __set_migration_target_nodes(void)
+{
+	nodemask_t next_pass	= NODE_MASK_NONE;
+	nodemask_t this_pass	= NODE_MASK_NONE;
+	nodemask_t used_targets = NODE_MASK_NONE;
+	int node;
+
+	/*
+	 * Avoid any oddities like cycles that could occur
+	 * from changes in the topology.  This will leave
+	 * a momentary gap when migration is disabled.
+	 */
+	disable_all_migrate_targets();
+
+	/*
+	 * Allocations go close to CPUs, first.  Assume that
+	 * the migration path starts at the nodes with CPUs.
+	 */
+	next_pass = node_states[N_CPU];
+again:
+	this_pass = next_pass;
+	next_pass = NODE_MASK_NONE;
+	/*
+	 * To avoid cycles in the migration "graph", ensure
+	 * that migration sources are not future targets by
+	 * setting them in 'used_targets'.  Do this only
+	 * once per pass so that multiple source nodes can
+	 * share a target node.
+	 *
+	 * 'used_targets' will become unavailable in future
+	 * passes.  This limits some opportunities for
+	 * multiple source nodes to share a destination.
+	 */
+	nodes_or(used_targets, used_targets, this_pass);
+	for_each_node_mask(node, this_pass) {
+		int target_node = establish_migrate_target(node, &used_targets);
+
+		if (target_node == NUMA_NO_NODE)
+			continue;
+
+		/*
+		 * Visit targets from this pass in the next pass.
+		 * Eventually, every node will have been part of
+		 * a pass, and will become set in 'used_targets'.
+		 */
+		node_set(target_node, next_pass);
+	}
+	/*
+	 * 'next_pass' contains nodes which became migration
+	 * targets in this pass.  Make additional passes until
+	 * no more migrations targets are available.
+	 */
+	if (!nodes_empty(next_pass))
+		goto again;
+}
+
+/*
+ * For callers that do not hold get_online_mems() already.
+ */
+__maybe_unused // <- temporay to prevent warnings during bisects
+static void set_migration_target_nodes(void)
+{
+	get_online_mems();
+	__set_migration_target_nodes();
+	put_online_mems();
+}
--- a/mm/page_alloc.c~mm-numa-automatically-generate-node-migration-order
+++ a/mm/page_alloc.c
@@ -6157,7 +6157,7 @@  static int node_load[MAX_NUMNODES];
  *
  * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
  */
-static int find_next_best_node(int node, nodemask_t *used_node_mask)
+int find_next_best_node(int node, nodemask_t *used_node_mask)
 {
 	int n, val;
 	int min_val = INT_MAX;