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>
9
10#include "internal.h"
11
12struct memory_tier {
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;
17 /*
18 * start value of abstract distance. memory tier maps
19 * an abstract distance range,
20 * adistance_start .. adistance_start + MEMTIER_CHUNK_SIZE
21 */
22 int adistance_start;
23 struct device dev;
24 /* All the nodes that are part of all the lower memory tiers. */
25 nodemask_t lower_tier_mask;
26};
27
28struct demotion_nodes {
29 nodemask_t preferred;
30};
31
32struct node_memory_type_map {
33 struct memory_dev_type *memtype;
34 int map_count;
35};
36
37static DEFINE_MUTEX(memory_tier_lock);
38static LIST_HEAD(memory_tiers);
39static struct node_memory_type_map node_memory_types[MAX_NUMNODES];
40struct memory_dev_type *default_dram_type;
41
42static struct bus_type memory_tier_subsys = {
43 .name = "memory_tiering",
44 .dev_name = "memory_tier",
45};
46
47#ifdef CONFIG_MIGRATION
48static int top_tier_adistance;
49/*
50 * node_demotion[] examples:
51 *
52 * Example 1:
53 *
54 * Node 0 & 1 are CPU + DRAM nodes, node 2 & 3 are PMEM nodes.
55 *
56 * node distances:
57 * node 0 1 2 3
58 * 0 10 20 30 40
59 * 1 20 10 40 30
60 * 2 30 40 10 40
61 * 3 40 30 40 10
62 *
63 * memory_tiers0 = 0-1
64 * memory_tiers1 = 2-3
65 *
66 * node_demotion[0].preferred = 2
67 * node_demotion[1].preferred = 3
68 * node_demotion[2].preferred = <empty>
69 * node_demotion[3].preferred = <empty>
70 *
71 * Example 2:
72 *
73 * Node 0 & 1 are CPU + DRAM nodes, node 2 is memory-only DRAM node.
74 *
75 * node distances:
76 * node 0 1 2
77 * 0 10 20 30
78 * 1 20 10 30
79 * 2 30 30 10
80 *
81 * memory_tiers0 = 0-2
82 *
83 * node_demotion[0].preferred = <empty>
84 * node_demotion[1].preferred = <empty>
85 * node_demotion[2].preferred = <empty>
86 *
87 * Example 3:
88 *
89 * Node 0 is CPU + DRAM nodes, Node 1 is HBM node, node 2 is PMEM node.
90 *
91 * node distances:
92 * node 0 1 2
93 * 0 10 20 30
94 * 1 20 10 40
95 * 2 30 40 10
96 *
97 * memory_tiers0 = 1
98 * memory_tiers1 = 0
99 * memory_tiers2 = 2
100 *
101 * node_demotion[0].preferred = 2
102 * node_demotion[1].preferred = 0
103 * node_demotion[2].preferred = <empty>
104 *
105 */
106static struct demotion_nodes *node_demotion __read_mostly;
107#endif /* CONFIG_MIGRATION */
108
109static BLOCKING_NOTIFIER_HEAD(mt_adistance_algorithms);
110
111static bool default_dram_perf_error;
112static struct node_hmem_attrs default_dram_perf;
113static int default_dram_perf_ref_nid = NUMA_NO_NODE;
114static const char *default_dram_perf_ref_source;
115
116static inline struct memory_tier *to_memory_tier(struct device *device)
117{
118 return container_of(device, struct memory_tier, dev);
119}
120
121static __always_inline nodemask_t get_memtier_nodemask(struct memory_tier *memtier)
122{
123 nodemask_t nodes = NODE_MASK_NONE;
124 struct memory_dev_type *memtype;
125
126 list_for_each_entry(memtype, &memtier->memory_types, tier_sibling)
127 nodes_or(nodes, nodes, memtype->nodes);
128
129 return nodes;
130}
131
132static void memory_tier_device_release(struct device *dev)
133{
134 struct memory_tier *tier = to_memory_tier(device: dev);
135 /*
136 * synchronize_rcu in clear_node_memory_tier makes sure
137 * we don't have rcu access to this memory tier.
138 */
139 kfree(objp: tier);
140}
141
142static ssize_t nodelist_show(struct device *dev,
143 struct device_attribute *attr, char *buf)
144{
145 int ret;
146 nodemask_t nmask;
147
148 mutex_lock(&memory_tier_lock);
149 nmask = get_memtier_nodemask(memtier: to_memory_tier(device: dev));
150 ret = sysfs_emit(buf, fmt: "%*pbl\n", nodemask_pr_args(&nmask));
151 mutex_unlock(lock: &memory_tier_lock);
152 return ret;
153}
154static DEVICE_ATTR_RO(nodelist);
155
156static struct attribute *memtier_dev_attrs[] = {
157 &dev_attr_nodelist.attr,
158 NULL
159};
160
161static const struct attribute_group memtier_dev_group = {
162 .attrs = memtier_dev_attrs,
163};
164
165static const struct attribute_group *memtier_dev_groups[] = {
166 &memtier_dev_group,
167 NULL
168};
169
170static struct memory_tier *find_create_memory_tier(struct memory_dev_type *memtype)
171{
172 int ret;
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;
177
178 lockdep_assert_held_once(&memory_tier_lock);
179
180 adistance = round_down(adistance, memtier_adistance_chunk_size);
181 /*
182 * If the memtype is already part of a memory tier,
183 * just return that.
184 */
185 if (!list_empty(head: &memtype->tier_sibling)) {
186 list_for_each_entry(memtier, &memory_tiers, list) {
187 if (adistance == memtier->adistance_start)
188 return memtier;
189 }
190 WARN_ON(1);
191 return ERR_PTR(error: -EINVAL);
192 }
193
194 list_for_each_entry(memtier, &memory_tiers, list) {
195 if (adistance == memtier->adistance_start) {
196 goto link_memtype;
197 } else if (adistance < memtier->adistance_start) {
198 found_slot = true;
199 break;
200 }
201 }
202
203 new_memtier = kzalloc(size: sizeof(struct memory_tier), GFP_KERNEL);
204 if (!new_memtier)
205 return ERR_PTR(error: -ENOMEM);
206
207 new_memtier->adistance_start = adistance;
208 INIT_LIST_HEAD(list: &new_memtier->list);
209 INIT_LIST_HEAD(list: &new_memtier->memory_types);
210 if (found_slot)
211 list_add_tail(new: &new_memtier->list, head: &memtier->list);
212 else
213 list_add_tail(new: &new_memtier->list, head: &memory_tiers);
214
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;
219
220 ret = device_register(dev: &new_memtier->dev);
221 if (ret) {
222 list_del(entry: &new_memtier->list);
223 put_device(dev: &new_memtier->dev);
224 return ERR_PTR(error: ret);
225 }
226 memtier = new_memtier;
227
228link_memtype:
229 list_add(new: &memtype->tier_sibling, head: &memtier->memory_types);
230 return memtier;
231}
232
233static struct memory_tier *__node_get_memory_tier(int node)
234{
235 pg_data_t *pgdat;
236
237 pgdat = NODE_DATA(node);
238 if (!pgdat)
239 return NULL;
240 /*
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.
244 */
245 return rcu_dereference_check(pgdat->memtier,
246 lockdep_is_held(&memory_tier_lock));
247}
248
249#ifdef CONFIG_MIGRATION
250bool node_is_toptier(int node)
251{
252 bool toptier;
253 pg_data_t *pgdat;
254 struct memory_tier *memtier;
255
256 pgdat = NODE_DATA(node);
257 if (!pgdat)
258 return false;
259
260 rcu_read_lock();
261 memtier = rcu_dereference(pgdat->memtier);
262 if (!memtier) {
263 toptier = true;
264 goto out;
265 }
266 if (memtier->adistance_start <= top_tier_adistance)
267 toptier = true;
268 else
269 toptier = false;
270out:
271 rcu_read_unlock();
272 return toptier;
273}
274
275void node_get_allowed_targets(pg_data_t *pgdat, nodemask_t *targets)
276{
277 struct memory_tier *memtier;
278
279 /*
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();
283 */
284 rcu_read_lock();
285 memtier = rcu_dereference(pgdat->memtier);
286 if (memtier)
287 *targets = memtier->lower_tier_mask;
288 else
289 *targets = NODE_MASK_NONE;
290 rcu_read_unlock();
291}
292
293/**
294 * next_demotion_node() - Get the next node in the demotion path
295 * @node: The starting node to lookup the next node
296 *
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
300 * target.
301 */
302int next_demotion_node(int node)
303{
304 struct demotion_nodes *nd;
305 int target;
306
307 if (!node_demotion)
308 return NUMA_NO_NODE;
309
310 nd = &node_demotion[node];
311
312 /*
313 * node_demotion[] is updated without excluding this
314 * function from running.
315 *
316 * Make sure to use RCU over entire code blocks if
317 * node_demotion[] reads need to be consistent.
318 */
319 rcu_read_lock();
320 /*
321 * If there are multiple target nodes, just select one
322 * target node randomly.
323 *
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.
331 */
332 target = node_random(maskp: &nd->preferred);
333 rcu_read_unlock();
334
335 return target;
336}
337
338static void disable_all_demotion_targets(void)
339{
340 struct memory_tier *memtier;
341 int node;
342
343 for_each_node_state(node, N_MEMORY) {
344 node_demotion[node].preferred = NODE_MASK_NONE;
345 /*
346 * We are holding memory_tier_lock, it is safe
347 * to access pgda->memtier.
348 */
349 memtier = __node_get_memory_tier(node);
350 if (memtier)
351 memtier->lower_tier_mask = NODE_MASK_NONE;
352 }
353 /*
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.
358 */
359 synchronize_rcu();
360}
361
362/*
363 * Find an automatic demotion target for all memory
364 * nodes. Failing here is OK. It might just indicate
365 * being at the end of a chain.
366 */
367static void establish_demotion_targets(void)
368{
369 struct memory_tier *memtier;
370 struct demotion_nodes *nd;
371 int target = NUMA_NO_NODE, node;
372 int distance, best_distance;
373 nodemask_t tier_nodes, lower_tier;
374
375 lockdep_assert_held_once(&memory_tier_lock);
376
377 if (!node_demotion)
378 return;
379
380 disable_all_demotion_targets();
381
382 for_each_node_state(node, N_MEMORY) {
383 best_distance = -1;
384 nd = &node_demotion[node];
385
386 memtier = __node_get_memory_tier(node);
387 if (!memtier || list_is_last(list: &memtier->list, head: &memory_tiers))
388 continue;
389 /*
390 * Get the lower memtier to find the demotion node list.
391 */
392 memtier = list_next_entry(memtier, list);
393 tier_nodes = get_memtier_nodemask(memtier);
394 /*
395 * find_next_best_node, use 'used' nodemask as a skip list.
396 * Add all memory nodes except the selected memory tier
397 * nodelist to skip list so that we find the best node from the
398 * memtier nodelist.
399 */
400 nodes_andnot(tier_nodes, node_states[N_MEMORY], tier_nodes);
401
402 /*
403 * Find all the nodes in the memory tier node list of same best distance.
404 * add them to the preferred mask. We randomly select between nodes
405 * in the preferred mask when allocating pages during demotion.
406 */
407 do {
408 target = find_next_best_node(node, used_node_mask: &tier_nodes);
409 if (target == NUMA_NO_NODE)
410 break;
411
412 distance = node_distance(node, target);
413 if (distance == best_distance || best_distance == -1) {
414 best_distance = distance;
415 node_set(target, nd->preferred);
416 } else {
417 break;
418 }
419 } while (1);
420 }
421 /*
422 * Promotion is allowed from a memory tier to higher
423 * memory tier only if the memory tier doesn't include
424 * compute. We want to skip promotion from a memory tier,
425 * if any node that is part of the memory tier have CPUs.
426 * Once we detect such a memory tier, we consider that tier
427 * as top tiper from which promotion is not allowed.
428 */
429 list_for_each_entry_reverse(memtier, &memory_tiers, list) {
430 tier_nodes = get_memtier_nodemask(memtier);
431 nodes_and(tier_nodes, node_states[N_CPU], tier_nodes);
432 if (!nodes_empty(tier_nodes)) {
433 /*
434 * abstract distance below the max value of this memtier
435 * is considered toptier.
436 */
437 top_tier_adistance = memtier->adistance_start +
438 MEMTIER_CHUNK_SIZE - 1;
439 break;
440 }
441 }
442 /*
443 * Now build the lower_tier mask for each node collecting node mask from
444 * all memory tier below it. This allows us to fallback demotion page
445 * allocation to a set of nodes that is closer the above selected
446 * perferred node.
447 */
448 lower_tier = node_states[N_MEMORY];
449 list_for_each_entry(memtier, &memory_tiers, list) {
450 /*
451 * Keep removing current tier from lower_tier nodes,
452 * This will remove all nodes in current and above
453 * memory tier from the lower_tier mask.
454 */
455 tier_nodes = get_memtier_nodemask(memtier);
456 nodes_andnot(lower_tier, lower_tier, tier_nodes);
457 memtier->lower_tier_mask = lower_tier;
458 }
459}
460
461#else
462static inline void establish_demotion_targets(void) {}
463#endif /* CONFIG_MIGRATION */
464
465static inline void __init_node_memory_type(int node, struct memory_dev_type *memtype)
466{
467 if (!node_memory_types[node].memtype)
468 node_memory_types[node].memtype = memtype;
469 /*
470 * for each device getting added in the same NUMA node
471 * with this specific memtype, bump the map count. We
472 * Only take memtype device reference once, so that
473 * changing a node memtype can be done by droping the
474 * only reference count taken here.
475 */
476
477 if (node_memory_types[node].memtype == memtype) {
478 if (!node_memory_types[node].map_count++)
479 kref_get(kref: &memtype->kref);
480 }
481}
482
483static struct memory_tier *set_node_memory_tier(int node)
484{
485 struct memory_tier *memtier;
486 struct memory_dev_type *memtype;
487 pg_data_t *pgdat = NODE_DATA(node);
488
489
490 lockdep_assert_held_once(&memory_tier_lock);
491
492 if (!node_state(node, state: N_MEMORY))
493 return ERR_PTR(error: -EINVAL);
494
495 __init_node_memory_type(node, memtype: default_dram_type);
496
497 memtype = node_memory_types[node].memtype;
498 node_set(node, memtype->nodes);
499 memtier = find_create_memory_tier(memtype);
500 if (!IS_ERR(ptr: memtier))
501 rcu_assign_pointer(pgdat->memtier, memtier);
502 return memtier;
503}
504
505static void destroy_memory_tier(struct memory_tier *memtier)
506{
507 list_del(entry: &memtier->list);
508 device_unregister(dev: &memtier->dev);
509}
510
511static bool clear_node_memory_tier(int node)
512{
513 bool cleared = false;
514 pg_data_t *pgdat;
515 struct memory_tier *memtier;
516
517 pgdat = NODE_DATA(node);
518 if (!pgdat)
519 return false;
520
521 /*
522 * Make sure that anybody looking at NODE_DATA who finds
523 * a valid memtier finds memory_dev_types with nodes still
524 * linked to the memtier. We achieve this by waiting for
525 * rcu read section to finish using synchronize_rcu.
526 * This also enables us to free the destroyed memory tier
527 * with kfree instead of kfree_rcu
528 */
529 memtier = __node_get_memory_tier(node);
530 if (memtier) {
531 struct memory_dev_type *memtype;
532
533 rcu_assign_pointer(pgdat->memtier, NULL);
534 synchronize_rcu();
535 memtype = node_memory_types[node].memtype;
536 node_clear(node, memtype->nodes);
537 if (nodes_empty(memtype->nodes)) {
538 list_del_init(entry: &memtype->tier_sibling);
539 if (list_empty(head: &memtier->memory_types))
540 destroy_memory_tier(memtier);
541 }
542 cleared = true;
543 }
544 return cleared;
545}
546
547static void release_memtype(struct kref *kref)
548{
549 struct memory_dev_type *memtype;
550
551 memtype = container_of(kref, struct memory_dev_type, kref);
552 kfree(objp: memtype);
553}
554
555struct memory_dev_type *alloc_memory_type(int adistance)
556{
557 struct memory_dev_type *memtype;
558
559 memtype = kmalloc(size: sizeof(*memtype), GFP_KERNEL);
560 if (!memtype)
561 return ERR_PTR(error: -ENOMEM);
562
563 memtype->adistance = adistance;
564 INIT_LIST_HEAD(list: &memtype->tier_sibling);
565 memtype->nodes = NODE_MASK_NONE;
566 kref_init(kref: &memtype->kref);
567 return memtype;
568}
569EXPORT_SYMBOL_GPL(alloc_memory_type);
570
571void put_memory_type(struct memory_dev_type *memtype)
572{
573 kref_put(kref: &memtype->kref, release: release_memtype);
574}
575EXPORT_SYMBOL_GPL(put_memory_type);
576
577void init_node_memory_type(int node, struct memory_dev_type *memtype)
578{
579
580 mutex_lock(&memory_tier_lock);
581 __init_node_memory_type(node, memtype);
582 mutex_unlock(lock: &memory_tier_lock);
583}
584EXPORT_SYMBOL_GPL(init_node_memory_type);
585
586void clear_node_memory_type(int node, struct memory_dev_type *memtype)
587{
588 mutex_lock(&memory_tier_lock);
589 if (node_memory_types[node].memtype == memtype || !memtype)
590 node_memory_types[node].map_count--;
591 /*
592 * If we umapped all the attached devices to this node,
593 * clear the node memory type.
594 */
595 if (!node_memory_types[node].map_count) {
596 memtype = node_memory_types[node].memtype;
597 node_memory_types[node].memtype = NULL;
598 put_memory_type(memtype);
599 }
600 mutex_unlock(lock: &memory_tier_lock);
601}
602EXPORT_SYMBOL_GPL(clear_node_memory_type);
603
604static void dump_hmem_attrs(struct node_hmem_attrs *attrs, const char *prefix)
605{
606 pr_info(
607"%sread_latency: %u, write_latency: %u, read_bandwidth: %u, write_bandwidth: %u\n",
608 prefix, attrs->read_latency, attrs->write_latency,
609 attrs->read_bandwidth, attrs->write_bandwidth);
610}
611
612int mt_set_default_dram_perf(int nid, struct node_hmem_attrs *perf,
613 const char *source)
614{
615 int rc = 0;
616
617 mutex_lock(&memory_tier_lock);
618 if (default_dram_perf_error) {
619 rc = -EIO;
620 goto out;
621 }
622
623 if (perf->read_latency + perf->write_latency == 0 ||
624 perf->read_bandwidth + perf->write_bandwidth == 0) {
625 rc = -EINVAL;
626 goto out;
627 }
628
629 if (default_dram_perf_ref_nid == NUMA_NO_NODE) {
630 default_dram_perf = *perf;
631 default_dram_perf_ref_nid = nid;
632 default_dram_perf_ref_source = kstrdup(s: source, GFP_KERNEL);
633 goto out;
634 }
635
636 /*
637 * The performance of all default DRAM nodes is expected to be
638 * same (that is, the variation is less than 10%). And it
639 * will be used as base to calculate the abstract distance of
640 * other memory nodes.
641 */
642 if (abs(perf->read_latency - default_dram_perf.read_latency) * 10 >
643 default_dram_perf.read_latency ||
644 abs(perf->write_latency - default_dram_perf.write_latency) * 10 >
645 default_dram_perf.write_latency ||
646 abs(perf->read_bandwidth - default_dram_perf.read_bandwidth) * 10 >
647 default_dram_perf.read_bandwidth ||
648 abs(perf->write_bandwidth - default_dram_perf.write_bandwidth) * 10 >
649 default_dram_perf.write_bandwidth) {
650 pr_info(
651"memory-tiers: the performance of DRAM node %d mismatches that of the reference\n"
652"DRAM node %d.\n", nid, default_dram_perf_ref_nid);
653 pr_info(" performance of reference DRAM node %d:\n",
654 default_dram_perf_ref_nid);
655 dump_hmem_attrs(attrs: &default_dram_perf, prefix: " ");
656 pr_info(" performance of DRAM node %d:\n", nid);
657 dump_hmem_attrs(attrs: perf, prefix: " ");
658 pr_info(
659" disable default DRAM node performance based abstract distance algorithm.\n");
660 default_dram_perf_error = true;
661 rc = -EINVAL;
662 }
663
664out:
665 mutex_unlock(lock: &memory_tier_lock);
666 return rc;
667}
668
669int mt_perf_to_adistance(struct node_hmem_attrs *perf, int *adist)
670{
671 if (default_dram_perf_error)
672 return -EIO;
673
674 if (default_dram_perf_ref_nid == NUMA_NO_NODE)
675 return -ENOENT;
676
677 if (perf->read_latency + perf->write_latency == 0 ||
678 perf->read_bandwidth + perf->write_bandwidth == 0)
679 return -EINVAL;
680
681 mutex_lock(&memory_tier_lock);
682 /*
683 * The abstract distance of a memory node is in direct proportion to
684 * its memory latency (read + write) and inversely proportional to its
685 * memory bandwidth (read + write). The abstract distance, memory
686 * latency, and memory bandwidth of the default DRAM nodes are used as
687 * the base.
688 */
689 *adist = MEMTIER_ADISTANCE_DRAM *
690 (perf->read_latency + perf->write_latency) /
691 (default_dram_perf.read_latency + default_dram_perf.write_latency) *
692 (default_dram_perf.read_bandwidth + default_dram_perf.write_bandwidth) /
693 (perf->read_bandwidth + perf->write_bandwidth);
694 mutex_unlock(lock: &memory_tier_lock);
695
696 return 0;
697}
698EXPORT_SYMBOL_GPL(mt_perf_to_adistance);
699
700/**
701 * register_mt_adistance_algorithm() - Register memory tiering abstract distance algorithm
702 * @nb: The notifier block which describe the algorithm
703 *
704 * Return: 0 on success, errno on error.
705 *
706 * Every memory tiering abstract distance algorithm provider needs to
707 * register the algorithm with register_mt_adistance_algorithm(). To
708 * calculate the abstract distance for a specified memory node, the
709 * notifier function will be called unless some high priority
710 * algorithm has provided result. The prototype of the notifier
711 * function is as follows,
712 *
713 * int (*algorithm_notifier)(struct notifier_block *nb,
714 * unsigned long nid, void *data);
715 *
716 * Where "nid" specifies the memory node, "data" is the pointer to the
717 * returned abstract distance (that is, "int *adist"). If the
718 * algorithm provides the result, NOTIFY_STOP should be returned.
719 * Otherwise, return_value & %NOTIFY_STOP_MASK == 0 to allow the next
720 * algorithm in the chain to provide the result.
721 */
722int register_mt_adistance_algorithm(struct notifier_block *nb)
723{
724 return blocking_notifier_chain_register(nh: &mt_adistance_algorithms, nb);
725}
726EXPORT_SYMBOL_GPL(register_mt_adistance_algorithm);
727
728/**
729 * unregister_mt_adistance_algorithm() - Unregister memory tiering abstract distance algorithm
730 * @nb: the notifier block which describe the algorithm
731 *
732 * Return: 0 on success, errno on error.
733 */
734int unregister_mt_adistance_algorithm(struct notifier_block *nb)
735{
736 return blocking_notifier_chain_unregister(nh: &mt_adistance_algorithms, nb);
737}
738EXPORT_SYMBOL_GPL(unregister_mt_adistance_algorithm);
739
740/**
741 * mt_calc_adistance() - Calculate abstract distance with registered algorithms
742 * @node: the node to calculate abstract distance for
743 * @adist: the returned abstract distance
744 *
745 * Return: if return_value & %NOTIFY_STOP_MASK != 0, then some
746 * abstract distance algorithm provides the result, and return it via
747 * @adist. Otherwise, no algorithm can provide the result and @adist
748 * will be kept as it is.
749 */
750int mt_calc_adistance(int node, int *adist)
751{
752 return blocking_notifier_call_chain(nh: &mt_adistance_algorithms, val: node, v: adist);
753}
754EXPORT_SYMBOL_GPL(mt_calc_adistance);
755
756static int __meminit memtier_hotplug_callback(struct notifier_block *self,
757 unsigned long action, void *_arg)
758{
759 struct memory_tier *memtier;
760 struct memory_notify *arg = _arg;
761
762 /*
763 * Only update the node migration order when a node is
764 * changing status, like online->offline.
765 */
766 if (arg->status_change_nid < 0)
767 return notifier_from_errno(err: 0);
768
769 switch (action) {
770 case MEM_OFFLINE:
771 mutex_lock(&memory_tier_lock);
772 if (clear_node_memory_tier(node: arg->status_change_nid))
773 establish_demotion_targets();
774 mutex_unlock(lock: &memory_tier_lock);
775 break;
776 case MEM_ONLINE:
777 mutex_lock(&memory_tier_lock);
778 memtier = set_node_memory_tier(arg->status_change_nid);
779 if (!IS_ERR(ptr: memtier))
780 establish_demotion_targets();
781 mutex_unlock(lock: &memory_tier_lock);
782 break;
783 }
784
785 return notifier_from_errno(err: 0);
786}
787
788static int __init memory_tier_init(void)
789{
790 int ret, node;
791 struct memory_tier *memtier;
792
793 ret = subsys_virtual_register(subsys: &memory_tier_subsys, NULL);
794 if (ret)
795 panic(fmt: "%s() failed to register memory tier subsystem\n", __func__);
796
797#ifdef CONFIG_MIGRATION
798 node_demotion = kcalloc(n: nr_node_ids, size: sizeof(struct demotion_nodes),
799 GFP_KERNEL);
800 WARN_ON(!node_demotion);
801#endif
802 mutex_lock(&memory_tier_lock);
803 /*
804 * For now we can have 4 faster memory tiers with smaller adistance
805 * than default DRAM tier.
806 */
807 default_dram_type = alloc_memory_type(MEMTIER_ADISTANCE_DRAM);
808 if (IS_ERR(ptr: default_dram_type))
809 panic(fmt: "%s() failed to allocate default DRAM tier\n", __func__);
810
811 /*
812 * Look at all the existing N_MEMORY nodes and add them to
813 * default memory tier or to a tier if we already have memory
814 * types assigned.
815 */
816 for_each_node_state(node, N_MEMORY) {
817 memtier = set_node_memory_tier(node);
818 if (IS_ERR(ptr: memtier))
819 /*
820 * Continue with memtiers we are able to setup
821 */
822 break;
823 }
824 establish_demotion_targets();
825 mutex_unlock(lock: &memory_tier_lock);
826
827 hotplug_memory_notifier(memtier_hotplug_callback, MEMTIER_HOTPLUG_PRI);
828 return 0;
829}
830subsys_initcall(memory_tier_init);
831
832bool numa_demotion_enabled = false;
833
834#ifdef CONFIG_MIGRATION
835#ifdef CONFIG_SYSFS
836static ssize_t demotion_enabled_show(struct kobject *kobj,
837 struct kobj_attribute *attr, char *buf)
838{
839 return sysfs_emit(buf, fmt: "%s\n",
840 numa_demotion_enabled ? "true" : "false");
841}
842
843static ssize_t demotion_enabled_store(struct kobject *kobj,
844 struct kobj_attribute *attr,
845 const char *buf, size_t count)
846{
847 ssize_t ret;
848
849 ret = kstrtobool(s: buf, res: &numa_demotion_enabled);
850 if (ret)
851 return ret;
852
853 return count;
854}
855
856static struct kobj_attribute numa_demotion_enabled_attr =
857 __ATTR_RW(demotion_enabled);
858
859static struct attribute *numa_attrs[] = {
860 &numa_demotion_enabled_attr.attr,
861 NULL,
862};
863
864static const struct attribute_group numa_attr_group = {
865 .attrs = numa_attrs,
866};
867
868static int __init numa_init_sysfs(void)
869{
870 int err;
871 struct kobject *numa_kobj;
872
873 numa_kobj = kobject_create_and_add(name: "numa", parent: mm_kobj);
874 if (!numa_kobj) {
875 pr_err("failed to create numa kobject\n");
876 return -ENOMEM;
877 }
878 err = sysfs_create_group(kobj: numa_kobj, grp: &numa_attr_group);
879 if (err) {
880 pr_err("failed to register numa group\n");
881 goto delete_obj;
882 }
883 return 0;
884
885delete_obj:
886 kobject_put(kobj: numa_kobj);
887 return err;
888}
889subsys_initcall(numa_init_sysfs);
890#endif /* CONFIG_SYSFS */
891#endif
892

source code of linux/mm/memory-tiers.c