1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/slab.h>
3 #include <linux/lockdep.h>
4 #include <linux/sysfs.h>
5 #include <linux/kobject.h>
6 #include <linux/memory.h>
7 #include <linux/memory-tiers.h>
8 #include <linux/notifier.h>
13 /* hierarchy of memory tiers */
14 struct list_head list;
15 /* list of all memory types part of this tier */
16 struct list_head memory_types;
18 * start value of abstract distance. memory tier maps
19 * an abstract distance range,
20 * adistance_start .. adistance_start + MEMTIER_CHUNK_SIZE
24 /* All the nodes that are part of all the lower memory tiers. */
25 nodemask_t lower_tier_mask;
28 struct demotion_nodes {
32 struct node_memory_type_map {
33 struct memory_dev_type *memtype;
37 static DEFINE_MUTEX(memory_tier_lock);
38 static LIST_HEAD(memory_tiers);
39 static struct node_memory_type_map node_memory_types[MAX_NUMNODES];
40 struct memory_dev_type *default_dram_type;
42 static const struct bus_type memory_tier_subsys = {
43 .name = "memory_tiering",
44 .dev_name = "memory_tier",
47 #ifdef CONFIG_MIGRATION
48 static int top_tier_adistance;
50 * node_demotion[] examples:
54 * Node 0 & 1 are CPU + DRAM nodes, node 2 & 3 are PMEM nodes.
66 * node_demotion[0].preferred = 2
67 * node_demotion[1].preferred = 3
68 * node_demotion[2].preferred = <empty>
69 * node_demotion[3].preferred = <empty>
73 * Node 0 & 1 are CPU + DRAM nodes, node 2 is memory-only DRAM node.
83 * node_demotion[0].preferred = <empty>
84 * node_demotion[1].preferred = <empty>
85 * node_demotion[2].preferred = <empty>
89 * Node 0 is CPU + DRAM nodes, Node 1 is HBM node, node 2 is PMEM node.
101 * node_demotion[0].preferred = 2
102 * node_demotion[1].preferred = 0
103 * node_demotion[2].preferred = <empty>
106 static struct demotion_nodes *node_demotion __read_mostly;
107 #endif /* CONFIG_MIGRATION */
109 static BLOCKING_NOTIFIER_HEAD(mt_adistance_algorithms);
111 static bool default_dram_perf_error;
112 static struct access_coordinate default_dram_perf;
113 static int default_dram_perf_ref_nid = NUMA_NO_NODE;
114 static const char *default_dram_perf_ref_source;
116 static inline struct memory_tier *to_memory_tier(struct device *device)
118 return container_of(device, struct memory_tier, dev);
121 static __always_inline nodemask_t get_memtier_nodemask(struct memory_tier *memtier)
123 nodemask_t nodes = NODE_MASK_NONE;
124 struct memory_dev_type *memtype;
126 list_for_each_entry(memtype, &memtier->memory_types, tier_sibling)
127 nodes_or(nodes, nodes, memtype->nodes);
132 static void memory_tier_device_release(struct device *dev)
134 struct memory_tier *tier = to_memory_tier(dev);
136 * synchronize_rcu in clear_node_memory_tier makes sure
137 * we don't have rcu access to this memory tier.
142 static ssize_t nodelist_show(struct device *dev,
143 struct device_attribute *attr, char *buf)
148 mutex_lock(&memory_tier_lock);
149 nmask = get_memtier_nodemask(to_memory_tier(dev));
150 ret = sysfs_emit(buf, "%*pbl\n", nodemask_pr_args(&nmask));
151 mutex_unlock(&memory_tier_lock);
154 static DEVICE_ATTR_RO(nodelist);
156 static struct attribute *memtier_dev_attrs[] = {
157 &dev_attr_nodelist.attr,
161 static const struct attribute_group memtier_dev_group = {
162 .attrs = memtier_dev_attrs,
165 static const struct attribute_group *memtier_dev_groups[] = {
170 static struct memory_tier *find_create_memory_tier(struct memory_dev_type *memtype)
173 bool found_slot = false;
174 struct memory_tier *memtier, *new_memtier;
175 int adistance = memtype->adistance;
176 unsigned int memtier_adistance_chunk_size = MEMTIER_CHUNK_SIZE;
178 lockdep_assert_held_once(&memory_tier_lock);
180 adistance = round_down(adistance, memtier_adistance_chunk_size);
182 * If the memtype is already part of a memory tier,
185 if (!list_empty(&memtype->tier_sibling)) {
186 list_for_each_entry(memtier, &memory_tiers, list) {
187 if (adistance == memtier->adistance_start)
191 return ERR_PTR(-EINVAL);
194 list_for_each_entry(memtier, &memory_tiers, list) {
195 if (adistance == memtier->adistance_start) {
197 } else if (adistance < memtier->adistance_start) {
203 new_memtier = kzalloc(sizeof(struct memory_tier), GFP_KERNEL);
205 return ERR_PTR(-ENOMEM);
207 new_memtier->adistance_start = adistance;
208 INIT_LIST_HEAD(&new_memtier->list);
209 INIT_LIST_HEAD(&new_memtier->memory_types);
211 list_add_tail(&new_memtier->list, &memtier->list);
213 list_add_tail(&new_memtier->list, &memory_tiers);
215 new_memtier->dev.id = adistance >> MEMTIER_CHUNK_BITS;
216 new_memtier->dev.bus = &memory_tier_subsys;
217 new_memtier->dev.release = memory_tier_device_release;
218 new_memtier->dev.groups = memtier_dev_groups;
220 ret = device_register(&new_memtier->dev);
222 list_del(&new_memtier->list);
223 put_device(&new_memtier->dev);
226 memtier = new_memtier;
229 list_add(&memtype->tier_sibling, &memtier->memory_types);
233 static struct memory_tier *__node_get_memory_tier(int node)
237 pgdat = NODE_DATA(node);
241 * Since we hold memory_tier_lock, we can avoid
242 * RCU read locks when accessing the details. No
243 * parallel updates are possible here.
245 return rcu_dereference_check(pgdat->memtier,
246 lockdep_is_held(&memory_tier_lock));
249 #ifdef CONFIG_MIGRATION
250 bool node_is_toptier(int node)
254 struct memory_tier *memtier;
256 pgdat = NODE_DATA(node);
261 memtier = rcu_dereference(pgdat->memtier);
266 if (memtier->adistance_start <= top_tier_adistance)
275 void node_get_allowed_targets(pg_data_t *pgdat, nodemask_t *targets)
277 struct memory_tier *memtier;
280 * pg_data_t.memtier updates includes a synchronize_rcu()
281 * which ensures that we either find NULL or a valid memtier
282 * in NODE_DATA. protect the access via rcu_read_lock();
285 memtier = rcu_dereference(pgdat->memtier);
287 *targets = memtier->lower_tier_mask;
289 *targets = NODE_MASK_NONE;
294 * next_demotion_node() - Get the next node in the demotion path
295 * @node: The starting node to lookup the next node
297 * Return: node id for next memory node in the demotion path hierarchy
298 * from @node; NUMA_NO_NODE if @node is terminal. This does not keep
299 * @node online or guarantee that it *continues* to be the next demotion
302 int next_demotion_node(int node)
304 struct demotion_nodes *nd;
310 nd = &node_demotion[node];
313 * node_demotion[] is updated without excluding this
314 * function from running.
316 * Make sure to use RCU over entire code blocks if
317 * node_demotion[] reads need to be consistent.
321 * If there are multiple target nodes, just select one
322 * target node randomly.
324 * In addition, we can also use round-robin to select
325 * target node, but we should introduce another variable
326 * for node_demotion[] to record last selected target node,
327 * that may cause cache ping-pong due to the changing of
328 * last target node. Or introducing per-cpu data to avoid
329 * caching issue, which seems more complicated. So selecting
330 * target node randomly seems better until now.
332 target = node_random(&nd->preferred);
338 static void disable_all_demotion_targets(void)
340 struct memory_tier *memtier;
343 for_each_node_state(node, N_MEMORY) {
344 node_demotion[node].preferred = NODE_MASK_NONE;
346 * We are holding memory_tier_lock, it is safe
347 * to access pgda->memtier.
349 memtier = __node_get_memory_tier(node);
351 memtier->lower_tier_mask = NODE_MASK_NONE;
354 * Ensure that the "disable" is visible across the system.
355 * Readers will see either a combination of before+disable
356 * state or disable+after. They will never see before and
357 * after state together.
362 static void dump_demotion_targets(void)
366 for_each_node_state(node, N_MEMORY) {
367 struct memory_tier *memtier = __node_get_memory_tier(node);
368 nodemask_t preferred = node_demotion[node].preferred;
373 if (nodes_empty(preferred))
374 pr_info("Demotion targets for Node %d: null\n", node);
376 pr_info("Demotion targets for Node %d: preferred: %*pbl, fallback: %*pbl\n",
377 node, nodemask_pr_args(&preferred),
378 nodemask_pr_args(&memtier->lower_tier_mask));
383 * Find an automatic demotion target for all memory
384 * nodes. Failing here is OK. It might just indicate
385 * being at the end of a chain.
387 static void establish_demotion_targets(void)
389 struct memory_tier *memtier;
390 struct demotion_nodes *nd;
391 int target = NUMA_NO_NODE, node;
392 int distance, best_distance;
393 nodemask_t tier_nodes, lower_tier;
395 lockdep_assert_held_once(&memory_tier_lock);
400 disable_all_demotion_targets();
402 for_each_node_state(node, N_MEMORY) {
404 nd = &node_demotion[node];
406 memtier = __node_get_memory_tier(node);
407 if (!memtier || list_is_last(&memtier->list, &memory_tiers))
410 * Get the lower memtier to find the demotion node list.
412 memtier = list_next_entry(memtier, list);
413 tier_nodes = get_memtier_nodemask(memtier);
415 * find_next_best_node, use 'used' nodemask as a skip list.
416 * Add all memory nodes except the selected memory tier
417 * nodelist to skip list so that we find the best node from the
420 nodes_andnot(tier_nodes, node_states[N_MEMORY], tier_nodes);
423 * Find all the nodes in the memory tier node list of same best distance.
424 * add them to the preferred mask. We randomly select between nodes
425 * in the preferred mask when allocating pages during demotion.
428 target = find_next_best_node(node, &tier_nodes);
429 if (target == NUMA_NO_NODE)
432 distance = node_distance(node, target);
433 if (distance == best_distance || best_distance == -1) {
434 best_distance = distance;
435 node_set(target, nd->preferred);
442 * Promotion is allowed from a memory tier to higher
443 * memory tier only if the memory tier doesn't include
444 * compute. We want to skip promotion from a memory tier,
445 * if any node that is part of the memory tier have CPUs.
446 * Once we detect such a memory tier, we consider that tier
447 * as top tiper from which promotion is not allowed.
449 list_for_each_entry_reverse(memtier, &memory_tiers, list) {
450 tier_nodes = get_memtier_nodemask(memtier);
451 nodes_and(tier_nodes, node_states[N_CPU], tier_nodes);
452 if (!nodes_empty(tier_nodes)) {
454 * abstract distance below the max value of this memtier
455 * is considered toptier.
457 top_tier_adistance = memtier->adistance_start +
458 MEMTIER_CHUNK_SIZE - 1;
463 * Now build the lower_tier mask for each node collecting node mask from
464 * all memory tier below it. This allows us to fallback demotion page
465 * allocation to a set of nodes that is closer the above selected
468 lower_tier = node_states[N_MEMORY];
469 list_for_each_entry(memtier, &memory_tiers, list) {
471 * Keep removing current tier from lower_tier nodes,
472 * This will remove all nodes in current and above
473 * memory tier from the lower_tier mask.
475 tier_nodes = get_memtier_nodemask(memtier);
476 nodes_andnot(lower_tier, lower_tier, tier_nodes);
477 memtier->lower_tier_mask = lower_tier;
480 dump_demotion_targets();
484 static inline void establish_demotion_targets(void) {}
485 #endif /* CONFIG_MIGRATION */
487 static inline void __init_node_memory_type(int node, struct memory_dev_type *memtype)
489 if (!node_memory_types[node].memtype)
490 node_memory_types[node].memtype = memtype;
492 * for each device getting added in the same NUMA node
493 * with this specific memtype, bump the map count. We
494 * Only take memtype device reference once, so that
495 * changing a node memtype can be done by droping the
496 * only reference count taken here.
499 if (node_memory_types[node].memtype == memtype) {
500 if (!node_memory_types[node].map_count++)
501 kref_get(&memtype->kref);
505 static struct memory_tier *set_node_memory_tier(int node)
507 struct memory_tier *memtier;
508 struct memory_dev_type *memtype;
509 pg_data_t *pgdat = NODE_DATA(node);
512 lockdep_assert_held_once(&memory_tier_lock);
514 if (!node_state(node, N_MEMORY))
515 return ERR_PTR(-EINVAL);
517 __init_node_memory_type(node, default_dram_type);
519 memtype = node_memory_types[node].memtype;
520 node_set(node, memtype->nodes);
521 memtier = find_create_memory_tier(memtype);
522 if (!IS_ERR(memtier))
523 rcu_assign_pointer(pgdat->memtier, memtier);
527 static void destroy_memory_tier(struct memory_tier *memtier)
529 list_del(&memtier->list);
530 device_unregister(&memtier->dev);
533 static bool clear_node_memory_tier(int node)
535 bool cleared = false;
537 struct memory_tier *memtier;
539 pgdat = NODE_DATA(node);
544 * Make sure that anybody looking at NODE_DATA who finds
545 * a valid memtier finds memory_dev_types with nodes still
546 * linked to the memtier. We achieve this by waiting for
547 * rcu read section to finish using synchronize_rcu.
548 * This also enables us to free the destroyed memory tier
549 * with kfree instead of kfree_rcu
551 memtier = __node_get_memory_tier(node);
553 struct memory_dev_type *memtype;
555 rcu_assign_pointer(pgdat->memtier, NULL);
557 memtype = node_memory_types[node].memtype;
558 node_clear(node, memtype->nodes);
559 if (nodes_empty(memtype->nodes)) {
560 list_del_init(&memtype->tier_sibling);
561 if (list_empty(&memtier->memory_types))
562 destroy_memory_tier(memtier);
569 static void release_memtype(struct kref *kref)
571 struct memory_dev_type *memtype;
573 memtype = container_of(kref, struct memory_dev_type, kref);
577 struct memory_dev_type *alloc_memory_type(int adistance)
579 struct memory_dev_type *memtype;
581 memtype = kmalloc(sizeof(*memtype), GFP_KERNEL);
583 return ERR_PTR(-ENOMEM);
585 memtype->adistance = adistance;
586 INIT_LIST_HEAD(&memtype->tier_sibling);
587 memtype->nodes = NODE_MASK_NONE;
588 kref_init(&memtype->kref);
591 EXPORT_SYMBOL_GPL(alloc_memory_type);
593 void put_memory_type(struct memory_dev_type *memtype)
595 kref_put(&memtype->kref, release_memtype);
597 EXPORT_SYMBOL_GPL(put_memory_type);
599 void init_node_memory_type(int node, struct memory_dev_type *memtype)
602 mutex_lock(&memory_tier_lock);
603 __init_node_memory_type(node, memtype);
604 mutex_unlock(&memory_tier_lock);
606 EXPORT_SYMBOL_GPL(init_node_memory_type);
608 void clear_node_memory_type(int node, struct memory_dev_type *memtype)
610 mutex_lock(&memory_tier_lock);
611 if (node_memory_types[node].memtype == memtype || !memtype)
612 node_memory_types[node].map_count--;
614 * If we umapped all the attached devices to this node,
615 * clear the node memory type.
617 if (!node_memory_types[node].map_count) {
618 memtype = node_memory_types[node].memtype;
619 node_memory_types[node].memtype = NULL;
620 put_memory_type(memtype);
622 mutex_unlock(&memory_tier_lock);
624 EXPORT_SYMBOL_GPL(clear_node_memory_type);
626 static void dump_hmem_attrs(struct access_coordinate *coord, const char *prefix)
629 "%sread_latency: %u, write_latency: %u, read_bandwidth: %u, write_bandwidth: %u\n",
630 prefix, coord->read_latency, coord->write_latency,
631 coord->read_bandwidth, coord->write_bandwidth);
634 int mt_set_default_dram_perf(int nid, struct access_coordinate *perf,
639 mutex_lock(&memory_tier_lock);
640 if (default_dram_perf_error) {
645 if (perf->read_latency + perf->write_latency == 0 ||
646 perf->read_bandwidth + perf->write_bandwidth == 0) {
651 if (default_dram_perf_ref_nid == NUMA_NO_NODE) {
652 default_dram_perf = *perf;
653 default_dram_perf_ref_nid = nid;
654 default_dram_perf_ref_source = kstrdup(source, GFP_KERNEL);
659 * The performance of all default DRAM nodes is expected to be
660 * same (that is, the variation is less than 10%). And it
661 * will be used as base to calculate the abstract distance of
662 * other memory nodes.
664 if (abs(perf->read_latency - default_dram_perf.read_latency) * 10 >
665 default_dram_perf.read_latency ||
666 abs(perf->write_latency - default_dram_perf.write_latency) * 10 >
667 default_dram_perf.write_latency ||
668 abs(perf->read_bandwidth - default_dram_perf.read_bandwidth) * 10 >
669 default_dram_perf.read_bandwidth ||
670 abs(perf->write_bandwidth - default_dram_perf.write_bandwidth) * 10 >
671 default_dram_perf.write_bandwidth) {
673 "memory-tiers: the performance of DRAM node %d mismatches that of the reference\n"
674 "DRAM node %d.\n", nid, default_dram_perf_ref_nid);
675 pr_info(" performance of reference DRAM node %d:\n",
676 default_dram_perf_ref_nid);
677 dump_hmem_attrs(&default_dram_perf, " ");
678 pr_info(" performance of DRAM node %d:\n", nid);
679 dump_hmem_attrs(perf, " ");
681 " disable default DRAM node performance based abstract distance algorithm.\n");
682 default_dram_perf_error = true;
687 mutex_unlock(&memory_tier_lock);
691 int mt_perf_to_adistance(struct access_coordinate *perf, int *adist)
693 if (default_dram_perf_error)
696 if (default_dram_perf_ref_nid == NUMA_NO_NODE)
699 if (perf->read_latency + perf->write_latency == 0 ||
700 perf->read_bandwidth + perf->write_bandwidth == 0)
703 mutex_lock(&memory_tier_lock);
705 * The abstract distance of a memory node is in direct proportion to
706 * its memory latency (read + write) and inversely proportional to its
707 * memory bandwidth (read + write). The abstract distance, memory
708 * latency, and memory bandwidth of the default DRAM nodes are used as
711 *adist = MEMTIER_ADISTANCE_DRAM *
712 (perf->read_latency + perf->write_latency) /
713 (default_dram_perf.read_latency + default_dram_perf.write_latency) *
714 (default_dram_perf.read_bandwidth + default_dram_perf.write_bandwidth) /
715 (perf->read_bandwidth + perf->write_bandwidth);
716 mutex_unlock(&memory_tier_lock);
720 EXPORT_SYMBOL_GPL(mt_perf_to_adistance);
723 * register_mt_adistance_algorithm() - Register memory tiering abstract distance algorithm
724 * @nb: The notifier block which describe the algorithm
726 * Return: 0 on success, errno on error.
728 * Every memory tiering abstract distance algorithm provider needs to
729 * register the algorithm with register_mt_adistance_algorithm(). To
730 * calculate the abstract distance for a specified memory node, the
731 * notifier function will be called unless some high priority
732 * algorithm has provided result. The prototype of the notifier
733 * function is as follows,
735 * int (*algorithm_notifier)(struct notifier_block *nb,
736 * unsigned long nid, void *data);
738 * Where "nid" specifies the memory node, "data" is the pointer to the
739 * returned abstract distance (that is, "int *adist"). If the
740 * algorithm provides the result, NOTIFY_STOP should be returned.
741 * Otherwise, return_value & %NOTIFY_STOP_MASK == 0 to allow the next
742 * algorithm in the chain to provide the result.
744 int register_mt_adistance_algorithm(struct notifier_block *nb)
746 return blocking_notifier_chain_register(&mt_adistance_algorithms, nb);
748 EXPORT_SYMBOL_GPL(register_mt_adistance_algorithm);
751 * unregister_mt_adistance_algorithm() - Unregister memory tiering abstract distance algorithm
752 * @nb: the notifier block which describe the algorithm
754 * Return: 0 on success, errno on error.
756 int unregister_mt_adistance_algorithm(struct notifier_block *nb)
758 return blocking_notifier_chain_unregister(&mt_adistance_algorithms, nb);
760 EXPORT_SYMBOL_GPL(unregister_mt_adistance_algorithm);
763 * mt_calc_adistance() - Calculate abstract distance with registered algorithms
764 * @node: the node to calculate abstract distance for
765 * @adist: the returned abstract distance
767 * Return: if return_value & %NOTIFY_STOP_MASK != 0, then some
768 * abstract distance algorithm provides the result, and return it via
769 * @adist. Otherwise, no algorithm can provide the result and @adist
770 * will be kept as it is.
772 int mt_calc_adistance(int node, int *adist)
774 return blocking_notifier_call_chain(&mt_adistance_algorithms, node, adist);
776 EXPORT_SYMBOL_GPL(mt_calc_adistance);
778 static int __meminit memtier_hotplug_callback(struct notifier_block *self,
779 unsigned long action, void *_arg)
781 struct memory_tier *memtier;
782 struct memory_notify *arg = _arg;
785 * Only update the node migration order when a node is
786 * changing status, like online->offline.
788 if (arg->status_change_nid < 0)
789 return notifier_from_errno(0);
793 mutex_lock(&memory_tier_lock);
794 if (clear_node_memory_tier(arg->status_change_nid))
795 establish_demotion_targets();
796 mutex_unlock(&memory_tier_lock);
799 mutex_lock(&memory_tier_lock);
800 memtier = set_node_memory_tier(arg->status_change_nid);
801 if (!IS_ERR(memtier))
802 establish_demotion_targets();
803 mutex_unlock(&memory_tier_lock);
807 return notifier_from_errno(0);
810 static int __init memory_tier_init(void)
813 struct memory_tier *memtier;
815 ret = subsys_virtual_register(&memory_tier_subsys, NULL);
817 panic("%s() failed to register memory tier subsystem\n", __func__);
819 #ifdef CONFIG_MIGRATION
820 node_demotion = kcalloc(nr_node_ids, sizeof(struct demotion_nodes),
822 WARN_ON(!node_demotion);
824 mutex_lock(&memory_tier_lock);
826 * For now we can have 4 faster memory tiers with smaller adistance
827 * than default DRAM tier.
829 default_dram_type = alloc_memory_type(MEMTIER_ADISTANCE_DRAM);
830 if (IS_ERR(default_dram_type))
831 panic("%s() failed to allocate default DRAM tier\n", __func__);
834 * Look at all the existing N_MEMORY nodes and add them to
835 * default memory tier or to a tier if we already have memory
838 for_each_node_state(node, N_MEMORY) {
839 memtier = set_node_memory_tier(node);
842 * Continue with memtiers we are able to setup
846 establish_demotion_targets();
847 mutex_unlock(&memory_tier_lock);
849 hotplug_memory_notifier(memtier_hotplug_callback, MEMTIER_HOTPLUG_PRI);
852 subsys_initcall(memory_tier_init);
854 bool numa_demotion_enabled = false;
856 #ifdef CONFIG_MIGRATION
858 static ssize_t demotion_enabled_show(struct kobject *kobj,
859 struct kobj_attribute *attr, char *buf)
861 return sysfs_emit(buf, "%s\n",
862 numa_demotion_enabled ? "true" : "false");
865 static ssize_t demotion_enabled_store(struct kobject *kobj,
866 struct kobj_attribute *attr,
867 const char *buf, size_t count)
871 ret = kstrtobool(buf, &numa_demotion_enabled);
878 static struct kobj_attribute numa_demotion_enabled_attr =
879 __ATTR_RW(demotion_enabled);
881 static struct attribute *numa_attrs[] = {
882 &numa_demotion_enabled_attr.attr,
886 static const struct attribute_group numa_attr_group = {
890 static int __init numa_init_sysfs(void)
893 struct kobject *numa_kobj;
895 numa_kobj = kobject_create_and_add("numa", mm_kobj);
897 pr_err("failed to create numa kobject\n");
900 err = sysfs_create_group(numa_kobj, &numa_attr_group);
902 pr_err("failed to register numa group\n");
908 kobject_put(numa_kobj);
911 subsys_initcall(numa_init_sysfs);
912 #endif /* CONFIG_SYSFS */