Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Memory subsystem support
4 *
5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
6 * Dave Hansen <haveblue@us.ibm.com>
7 *
8 * This file provides the necessary infrastructure to represent
9 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
10 * All arch-independent code that assumes MEMORY_HOTPLUG requires
11 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
12 */
13
14#include <linux/module.h>
15#include <linux/init.h>
16#include <linux/topology.h>
17#include <linux/capability.h>
18#include <linux/device.h>
19#include <linux/memory.h>
20#include <linux/memory_hotplug.h>
21#include <linux/mm.h>
22#include <linux/stat.h>
23#include <linux/slab.h>
24#include <linux/xarray.h>
25
26#include <linux/atomic.h>
27#include <linux/uaccess.h>
28
29#define MEMORY_CLASS_NAME "memory"
30
31static const char *const online_type_to_str[] = {
32 [MMOP_OFFLINE] = "offline",
33 [MMOP_ONLINE] = "online",
34 [MMOP_ONLINE_KERNEL] = "online_kernel",
35 [MMOP_ONLINE_MOVABLE] = "online_movable",
36};
37
38int mhp_online_type_from_str(const char *str)
39{
40 int i;
41
42 for (i = 0; i < ARRAY_SIZE(online_type_to_str); i++) {
43 if (sysfs_streq(str, online_type_to_str[i]))
44 return i;
45 }
46 return -EINVAL;
47}
48
49#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
50
51static int sections_per_block;
52
53static inline unsigned long memory_block_id(unsigned long section_nr)
54{
55 return section_nr / sections_per_block;
56}
57
58static inline unsigned long pfn_to_block_id(unsigned long pfn)
59{
60 return memory_block_id(pfn_to_section_nr(pfn));
61}
62
63static inline unsigned long phys_to_block_id(unsigned long phys)
64{
65 return pfn_to_block_id(PFN_DOWN(phys));
66}
67
68static int memory_subsys_online(struct device *dev);
69static int memory_subsys_offline(struct device *dev);
70
71static struct bus_type memory_subsys = {
72 .name = MEMORY_CLASS_NAME,
73 .dev_name = MEMORY_CLASS_NAME,
74 .online = memory_subsys_online,
75 .offline = memory_subsys_offline,
76};
77
78/*
79 * Memory blocks are cached in a local radix tree to avoid
80 * a costly linear search for the corresponding device on
81 * the subsystem bus.
82 */
83static DEFINE_XARRAY(memory_blocks);
84
85/*
86 * Memory groups, indexed by memory group id (mgid).
87 */
88static DEFINE_XARRAY_FLAGS(memory_groups, XA_FLAGS_ALLOC);
89#define MEMORY_GROUP_MARK_DYNAMIC XA_MARK_1
90
91static BLOCKING_NOTIFIER_HEAD(memory_chain);
92
93int register_memory_notifier(struct notifier_block *nb)
94{
95 return blocking_notifier_chain_register(&memory_chain, nb);
96}
97EXPORT_SYMBOL(register_memory_notifier);
98
99void unregister_memory_notifier(struct notifier_block *nb)
100{
101 blocking_notifier_chain_unregister(&memory_chain, nb);
102}
103EXPORT_SYMBOL(unregister_memory_notifier);
104
105static void memory_block_release(struct device *dev)
106{
107 struct memory_block *mem = to_memory_block(dev);
108
109 kfree(mem);
110}
111
112unsigned long __weak memory_block_size_bytes(void)
113{
114 return MIN_MEMORY_BLOCK_SIZE;
115}
116EXPORT_SYMBOL_GPL(memory_block_size_bytes);
117
118/*
119 * Show the first physical section index (number) of this memory block.
120 */
121static ssize_t phys_index_show(struct device *dev,
122 struct device_attribute *attr, char *buf)
123{
124 struct memory_block *mem = to_memory_block(dev);
125 unsigned long phys_index;
126
127 phys_index = mem->start_section_nr / sections_per_block;
128
129 return sysfs_emit(buf, "%08lx\n", phys_index);
130}
131
132/*
133 * Legacy interface that we cannot remove. Always indicate "removable"
134 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
135 */
136static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
137 char *buf)
138{
139 return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
140}
141
142/*
143 * online, offline, going offline, etc.
144 */
145static ssize_t state_show(struct device *dev, struct device_attribute *attr,
146 char *buf)
147{
148 struct memory_block *mem = to_memory_block(dev);
149 const char *output;
150
151 /*
152 * We can probably put these states in a nice little array
153 * so that they're not open-coded
154 */
155 switch (mem->state) {
156 case MEM_ONLINE:
157 output = "online";
158 break;
159 case MEM_OFFLINE:
160 output = "offline";
161 break;
162 case MEM_GOING_OFFLINE:
163 output = "going-offline";
164 break;
165 default:
166 WARN_ON(1);
167 return sysfs_emit(buf, "ERROR-UNKNOWN-%ld\n", mem->state);
168 }
169
170 return sysfs_emit(buf, "%s\n", output);
171}
172
173int memory_notify(unsigned long val, void *v)
174{
175 return blocking_notifier_call_chain(&memory_chain, val, v);
176}
177
178#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
179static unsigned long memblk_nr_poison(struct memory_block *mem);
180#else
181static inline unsigned long memblk_nr_poison(struct memory_block *mem)
182{
183 return 0;
184}
185#endif
186
187static int memory_block_online(struct memory_block *mem)
188{
189 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
190 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
191 unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
192 struct zone *zone;
193 int ret;
194
195 if (memblk_nr_poison(mem))
196 return -EHWPOISON;
197
198 zone = zone_for_pfn_range(mem->online_type, mem->nid, mem->group,
199 start_pfn, nr_pages);
200
201 /*
202 * Although vmemmap pages have a different lifecycle than the pages
203 * they describe (they remain until the memory is unplugged), doing
204 * their initialization and accounting at memory onlining/offlining
205 * stage helps to keep accounting easier to follow - e.g vmemmaps
206 * belong to the same zone as the memory they backed.
207 */
208 if (nr_vmemmap_pages) {
209 ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
210 if (ret)
211 return ret;
212 }
213
214 ret = online_pages(start_pfn + nr_vmemmap_pages,
215 nr_pages - nr_vmemmap_pages, zone, mem->group);
216 if (ret) {
217 if (nr_vmemmap_pages)
218 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
219 return ret;
220 }
221
222 /*
223 * Account once onlining succeeded. If the zone was unpopulated, it is
224 * now already properly populated.
225 */
226 if (nr_vmemmap_pages)
227 adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
228 nr_vmemmap_pages);
229
230 mem->zone = zone;
231 return ret;
232}
233
234static int memory_block_offline(struct memory_block *mem)
235{
236 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
237 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
238 unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
239 int ret;
240
241 if (!mem->zone)
242 return -EINVAL;
243
244 /*
245 * Unaccount before offlining, such that unpopulated zone and kthreads
246 * can properly be torn down in offline_pages().
247 */
248 if (nr_vmemmap_pages)
249 adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
250 -nr_vmemmap_pages);
251
252 ret = offline_pages(start_pfn + nr_vmemmap_pages,
253 nr_pages - nr_vmemmap_pages, mem->zone, mem->group);
254 if (ret) {
255 /* offline_pages() failed. Account back. */
256 if (nr_vmemmap_pages)
257 adjust_present_page_count(pfn_to_page(start_pfn),
258 mem->group, nr_vmemmap_pages);
259 return ret;
260 }
261
262 if (nr_vmemmap_pages)
263 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
264
265 mem->zone = NULL;
266 return ret;
267}
268
269/*
270 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
271 * OK to have direct references to sparsemem variables in here.
272 */
273static int
274memory_block_action(struct memory_block *mem, unsigned long action)
275{
276 int ret;
277
278 switch (action) {
279 case MEM_ONLINE:
280 ret = memory_block_online(mem);
281 break;
282 case MEM_OFFLINE:
283 ret = memory_block_offline(mem);
284 break;
285 default:
286 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
287 "%ld\n", __func__, mem->start_section_nr, action, action);
288 ret = -EINVAL;
289 }
290
291 return ret;
292}
293
294static int memory_block_change_state(struct memory_block *mem,
295 unsigned long to_state, unsigned long from_state_req)
296{
297 int ret = 0;
298
299 if (mem->state != from_state_req)
300 return -EINVAL;
301
302 if (to_state == MEM_OFFLINE)
303 mem->state = MEM_GOING_OFFLINE;
304
305 ret = memory_block_action(mem, to_state);
306 mem->state = ret ? from_state_req : to_state;
307
308 return ret;
309}
310
311/* The device lock serializes operations on memory_subsys_[online|offline] */
312static int memory_subsys_online(struct device *dev)
313{
314 struct memory_block *mem = to_memory_block(dev);
315 int ret;
316
317 if (mem->state == MEM_ONLINE)
318 return 0;
319
320 /*
321 * When called via device_online() without configuring the online_type,
322 * we want to default to MMOP_ONLINE.
323 */
324 if (mem->online_type == MMOP_OFFLINE)
325 mem->online_type = MMOP_ONLINE;
326
327 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
328 mem->online_type = MMOP_OFFLINE;
329
330 return ret;
331}
332
333static int memory_subsys_offline(struct device *dev)
334{
335 struct memory_block *mem = to_memory_block(dev);
336
337 if (mem->state == MEM_OFFLINE)
338 return 0;
339
340 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
341}
342
343static ssize_t state_store(struct device *dev, struct device_attribute *attr,
344 const char *buf, size_t count)
345{
346 const int online_type = mhp_online_type_from_str(buf);
347 struct memory_block *mem = to_memory_block(dev);
348 int ret;
349
350 if (online_type < 0)
351 return -EINVAL;
352
353 ret = lock_device_hotplug_sysfs();
354 if (ret)
355 return ret;
356
357 switch (online_type) {
358 case MMOP_ONLINE_KERNEL:
359 case MMOP_ONLINE_MOVABLE:
360 case MMOP_ONLINE:
361 /* mem->online_type is protected by device_hotplug_lock */
362 mem->online_type = online_type;
363 ret = device_online(&mem->dev);
364 break;
365 case MMOP_OFFLINE:
366 ret = device_offline(&mem->dev);
367 break;
368 default:
369 ret = -EINVAL; /* should never happen */
370 }
371
372 unlock_device_hotplug();
373
374 if (ret < 0)
375 return ret;
376 if (ret)
377 return -EINVAL;
378
379 return count;
380}
381
382/*
383 * Legacy interface that we cannot remove: s390x exposes the storage increment
384 * covered by a memory block, allowing for identifying which memory blocks
385 * comprise a storage increment. Since a memory block spans complete
386 * storage increments nowadays, this interface is basically unused. Other
387 * archs never exposed != 0.
388 */
389static ssize_t phys_device_show(struct device *dev,
390 struct device_attribute *attr, char *buf)
391{
392 struct memory_block *mem = to_memory_block(dev);
393 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
394
395 return sysfs_emit(buf, "%d\n",
396 arch_get_memory_phys_device(start_pfn));
397}
398
399#ifdef CONFIG_MEMORY_HOTREMOVE
400static int print_allowed_zone(char *buf, int len, int nid,
401 struct memory_group *group,
402 unsigned long start_pfn, unsigned long nr_pages,
403 int online_type, struct zone *default_zone)
404{
405 struct zone *zone;
406
407 zone = zone_for_pfn_range(online_type, nid, group, start_pfn, nr_pages);
408 if (zone == default_zone)
409 return 0;
410
411 return sysfs_emit_at(buf, len, " %s", zone->name);
412}
413
414static ssize_t valid_zones_show(struct device *dev,
415 struct device_attribute *attr, char *buf)
416{
417 struct memory_block *mem = to_memory_block(dev);
418 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
419 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
420 struct memory_group *group = mem->group;
421 struct zone *default_zone;
422 int nid = mem->nid;
423 int len = 0;
424
425 /*
426 * Check the existing zone. Make sure that we do that only on the
427 * online nodes otherwise the page_zone is not reliable
428 */
429 if (mem->state == MEM_ONLINE) {
430 /*
431 * If !mem->zone, the memory block spans multiple zones and
432 * cannot get offlined.
433 */
434 default_zone = mem->zone;
435 if (!default_zone)
436 return sysfs_emit(buf, "%s\n", "none");
437 len += sysfs_emit_at(buf, len, "%s", default_zone->name);
438 goto out;
439 }
440
441 default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, group,
442 start_pfn, nr_pages);
443
444 len += sysfs_emit_at(buf, len, "%s", default_zone->name);
445 len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
446 MMOP_ONLINE_KERNEL, default_zone);
447 len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
448 MMOP_ONLINE_MOVABLE, default_zone);
449out:
450 len += sysfs_emit_at(buf, len, "\n");
451 return len;
452}
453static DEVICE_ATTR_RO(valid_zones);
454#endif
455
456static DEVICE_ATTR_RO(phys_index);
457static DEVICE_ATTR_RW(state);
458static DEVICE_ATTR_RO(phys_device);
459static DEVICE_ATTR_RO(removable);
460
461/*
462 * Show the memory block size (shared by all memory blocks).
463 */
464static ssize_t block_size_bytes_show(struct device *dev,
465 struct device_attribute *attr, char *buf)
466{
467 return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
468}
469
470static DEVICE_ATTR_RO(block_size_bytes);
471
472/*
473 * Memory auto online policy.
474 */
475
476static ssize_t auto_online_blocks_show(struct device *dev,
477 struct device_attribute *attr, char *buf)
478{
479 return sysfs_emit(buf, "%s\n",
480 online_type_to_str[mhp_default_online_type]);
481}
482
483static ssize_t auto_online_blocks_store(struct device *dev,
484 struct device_attribute *attr,
485 const char *buf, size_t count)
486{
487 const int online_type = mhp_online_type_from_str(buf);
488
489 if (online_type < 0)
490 return -EINVAL;
491
492 mhp_default_online_type = online_type;
493 return count;
494}
495
496static DEVICE_ATTR_RW(auto_online_blocks);
497
498/*
499 * Some architectures will have custom drivers to do this, and
500 * will not need to do it from userspace. The fake hot-add code
501 * as well as ppc64 will do all of their discovery in userspace
502 * and will require this interface.
503 */
504#ifdef CONFIG_ARCH_MEMORY_PROBE
505static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
506 const char *buf, size_t count)
507{
508 u64 phys_addr;
509 int nid, ret;
510 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
511
512 ret = kstrtoull(buf, 0, &phys_addr);
513 if (ret)
514 return ret;
515
516 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
517 return -EINVAL;
518
519 ret = lock_device_hotplug_sysfs();
520 if (ret)
521 return ret;
522
523 nid = memory_add_physaddr_to_nid(phys_addr);
524 ret = __add_memory(nid, phys_addr,
525 MIN_MEMORY_BLOCK_SIZE * sections_per_block,
526 MHP_NONE);
527
528 if (ret)
529 goto out;
530
531 ret = count;
532out:
533 unlock_device_hotplug();
534 return ret;
535}
536
537static DEVICE_ATTR_WO(probe);
538#endif
539
540#ifdef CONFIG_MEMORY_FAILURE
541/*
542 * Support for offlining pages of memory
543 */
544
545/* Soft offline a page */
546static ssize_t soft_offline_page_store(struct device *dev,
547 struct device_attribute *attr,
548 const char *buf, size_t count)
549{
550 int ret;
551 u64 pfn;
552 if (!capable(CAP_SYS_ADMIN))
553 return -EPERM;
554 if (kstrtoull(buf, 0, &pfn) < 0)
555 return -EINVAL;
556 pfn >>= PAGE_SHIFT;
557 ret = soft_offline_page(pfn, 0);
558 return ret == 0 ? count : ret;
559}
560
561/* Forcibly offline a page, including killing processes. */
562static ssize_t hard_offline_page_store(struct device *dev,
563 struct device_attribute *attr,
564 const char *buf, size_t count)
565{
566 int ret;
567 u64 pfn;
568 if (!capable(CAP_SYS_ADMIN))
569 return -EPERM;
570 if (kstrtoull(buf, 0, &pfn) < 0)
571 return -EINVAL;
572 pfn >>= PAGE_SHIFT;
573 ret = memory_failure(pfn, MF_SW_SIMULATED);
574 if (ret == -EOPNOTSUPP)
575 ret = 0;
576 return ret ? ret : count;
577}
578
579static DEVICE_ATTR_WO(soft_offline_page);
580static DEVICE_ATTR_WO(hard_offline_page);
581#endif
582
583/* See phys_device_show(). */
584int __weak arch_get_memory_phys_device(unsigned long start_pfn)
585{
586 return 0;
587}
588
589/*
590 * A reference for the returned memory block device is acquired.
591 *
592 * Called under device_hotplug_lock.
593 */
594static struct memory_block *find_memory_block_by_id(unsigned long block_id)
595{
596 struct memory_block *mem;
597
598 mem = xa_load(&memory_blocks, block_id);
599 if (mem)
600 get_device(&mem->dev);
601 return mem;
602}
603
604/*
605 * Called under device_hotplug_lock.
606 */
607struct memory_block *find_memory_block(unsigned long section_nr)
608{
609 unsigned long block_id = memory_block_id(section_nr);
610
611 return find_memory_block_by_id(block_id);
612}
613
614static struct attribute *memory_memblk_attrs[] = {
615 &dev_attr_phys_index.attr,
616 &dev_attr_state.attr,
617 &dev_attr_phys_device.attr,
618 &dev_attr_removable.attr,
619#ifdef CONFIG_MEMORY_HOTREMOVE
620 &dev_attr_valid_zones.attr,
621#endif
622 NULL
623};
624
625static const struct attribute_group memory_memblk_attr_group = {
626 .attrs = memory_memblk_attrs,
627};
628
629static const struct attribute_group *memory_memblk_attr_groups[] = {
630 &memory_memblk_attr_group,
631 NULL,
632};
633
634static int __add_memory_block(struct memory_block *memory)
635{
636 int ret;
637
638 memory->dev.bus = &memory_subsys;
639 memory->dev.id = memory->start_section_nr / sections_per_block;
640 memory->dev.release = memory_block_release;
641 memory->dev.groups = memory_memblk_attr_groups;
642 memory->dev.offline = memory->state == MEM_OFFLINE;
643
644 ret = device_register(&memory->dev);
645 if (ret) {
646 put_device(&memory->dev);
647 return ret;
648 }
649 ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
650 GFP_KERNEL));
651 if (ret)
652 device_unregister(&memory->dev);
653
654 return ret;
655}
656
657static struct zone *early_node_zone_for_memory_block(struct memory_block *mem,
658 int nid)
659{
660 const unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
661 const unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
662 struct zone *zone, *matching_zone = NULL;
663 pg_data_t *pgdat = NODE_DATA(nid);
664 int i;
665
666 /*
667 * This logic only works for early memory, when the applicable zones
668 * already span the memory block. We don't expect overlapping zones on
669 * a single node for early memory. So if we're told that some PFNs
670 * of a node fall into this memory block, we can assume that all node
671 * zones that intersect with the memory block are actually applicable.
672 * No need to look at the memmap.
673 */
674 for (i = 0; i < MAX_NR_ZONES; i++) {
675 zone = pgdat->node_zones + i;
676 if (!populated_zone(zone))
677 continue;
678 if (!zone_intersects(zone, start_pfn, nr_pages))
679 continue;
680 if (!matching_zone) {
681 matching_zone = zone;
682 continue;
683 }
684 /* Spans multiple zones ... */
685 matching_zone = NULL;
686 break;
687 }
688 return matching_zone;
689}
690
691#ifdef CONFIG_NUMA
692/**
693 * memory_block_add_nid() - Indicate that system RAM falling into this memory
694 * block device (partially) belongs to the given node.
695 * @mem: The memory block device.
696 * @nid: The node id.
697 * @context: The memory initialization context.
698 *
699 * Indicate that system RAM falling into this memory block (partially) belongs
700 * to the given node. If the context indicates ("early") that we are adding the
701 * node during node device subsystem initialization, this will also properly
702 * set/adjust mem->zone based on the zone ranges of the given node.
703 */
704void memory_block_add_nid(struct memory_block *mem, int nid,
705 enum meminit_context context)
706{
707 if (context == MEMINIT_EARLY && mem->nid != nid) {
708 /*
709 * For early memory we have to determine the zone when setting
710 * the node id and handle multiple nodes spanning a single
711 * memory block by indicate via zone == NULL that we're not
712 * dealing with a single zone. So if we're setting the node id
713 * the first time, determine if there is a single zone. If we're
714 * setting the node id a second time to a different node,
715 * invalidate the single detected zone.
716 */
717 if (mem->nid == NUMA_NO_NODE)
718 mem->zone = early_node_zone_for_memory_block(mem, nid);
719 else
720 mem->zone = NULL;
721 }
722
723 /*
724 * If this memory block spans multiple nodes, we only indicate
725 * the last processed node. If we span multiple nodes (not applicable
726 * to hotplugged memory), zone == NULL will prohibit memory offlining
727 * and consequently unplug.
728 */
729 mem->nid = nid;
730}
731#endif
732
733static int add_memory_block(unsigned long block_id, unsigned long state,
734 unsigned long nr_vmemmap_pages,
735 struct memory_group *group)
736{
737 struct memory_block *mem;
738 int ret = 0;
739
740 mem = find_memory_block_by_id(block_id);
741 if (mem) {
742 put_device(&mem->dev);
743 return -EEXIST;
744 }
745 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
746 if (!mem)
747 return -ENOMEM;
748
749 mem->start_section_nr = block_id * sections_per_block;
750 mem->state = state;
751 mem->nid = NUMA_NO_NODE;
752 mem->nr_vmemmap_pages = nr_vmemmap_pages;
753 INIT_LIST_HEAD(&mem->group_next);
754
755#ifndef CONFIG_NUMA
756 if (state == MEM_ONLINE)
757 /*
758 * MEM_ONLINE at this point implies early memory. With NUMA,
759 * we'll determine the zone when setting the node id via
760 * memory_block_add_nid(). Memory hotplug updated the zone
761 * manually when memory onlining/offlining succeeds.
762 */
763 mem->zone = early_node_zone_for_memory_block(mem, NUMA_NO_NODE);
764#endif /* CONFIG_NUMA */
765
766 ret = __add_memory_block(mem);
767 if (ret)
768 return ret;
769
770 if (group) {
771 mem->group = group;
772 list_add(&mem->group_next, &group->memory_blocks);
773 }
774
775 return 0;
776}
777
778static int __init add_boot_memory_block(unsigned long base_section_nr)
779{
780 int section_count = 0;
781 unsigned long nr;
782
783 for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
784 nr++)
785 if (present_section_nr(nr))
786 section_count++;
787
788 if (section_count == 0)
789 return 0;
790 return add_memory_block(memory_block_id(base_section_nr),
791 MEM_ONLINE, 0, NULL);
792}
793
794static int add_hotplug_memory_block(unsigned long block_id,
795 unsigned long nr_vmemmap_pages,
796 struct memory_group *group)
797{
798 return add_memory_block(block_id, MEM_OFFLINE, nr_vmemmap_pages, group);
799}
800
801static void remove_memory_block(struct memory_block *memory)
802{
803 if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
804 return;
805
806 WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
807
808 if (memory->group) {
809 list_del(&memory->group_next);
810 memory->group = NULL;
811 }
812
813 /* drop the ref. we got via find_memory_block() */
814 put_device(&memory->dev);
815 device_unregister(&memory->dev);
816}
817
818/*
819 * Create memory block devices for the given memory area. Start and size
820 * have to be aligned to memory block granularity. Memory block devices
821 * will be initialized as offline.
822 *
823 * Called under device_hotplug_lock.
824 */
825int create_memory_block_devices(unsigned long start, unsigned long size,
826 unsigned long vmemmap_pages,
827 struct memory_group *group)
828{
829 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
830 unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
831 struct memory_block *mem;
832 unsigned long block_id;
833 int ret = 0;
834
835 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
836 !IS_ALIGNED(size, memory_block_size_bytes())))
837 return -EINVAL;
838
839 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
840 ret = add_hotplug_memory_block(block_id, vmemmap_pages, group);
841 if (ret)
842 break;
843 }
844 if (ret) {
845 end_block_id = block_id;
846 for (block_id = start_block_id; block_id != end_block_id;
847 block_id++) {
848 mem = find_memory_block_by_id(block_id);
849 if (WARN_ON_ONCE(!mem))
850 continue;
851 remove_memory_block(mem);
852 }
853 }
854 return ret;
855}
856
857/*
858 * Remove memory block devices for the given memory area. Start and size
859 * have to be aligned to memory block granularity. Memory block devices
860 * have to be offline.
861 *
862 * Called under device_hotplug_lock.
863 */
864void remove_memory_block_devices(unsigned long start, unsigned long size)
865{
866 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
867 const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
868 struct memory_block *mem;
869 unsigned long block_id;
870
871 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
872 !IS_ALIGNED(size, memory_block_size_bytes())))
873 return;
874
875 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
876 mem = find_memory_block_by_id(block_id);
877 if (WARN_ON_ONCE(!mem))
878 continue;
879 num_poisoned_pages_sub(-1UL, memblk_nr_poison(mem));
880 unregister_memory_block_under_nodes(mem);
881 remove_memory_block(mem);
882 }
883}
884
885static struct attribute *memory_root_attrs[] = {
886#ifdef CONFIG_ARCH_MEMORY_PROBE
887 &dev_attr_probe.attr,
888#endif
889
890#ifdef CONFIG_MEMORY_FAILURE
891 &dev_attr_soft_offline_page.attr,
892 &dev_attr_hard_offline_page.attr,
893#endif
894
895 &dev_attr_block_size_bytes.attr,
896 &dev_attr_auto_online_blocks.attr,
897 NULL
898};
899
900static const struct attribute_group memory_root_attr_group = {
901 .attrs = memory_root_attrs,
902};
903
904static const struct attribute_group *memory_root_attr_groups[] = {
905 &memory_root_attr_group,
906 NULL,
907};
908
909/*
910 * Initialize the sysfs support for memory devices. At the time this function
911 * is called, we cannot have concurrent creation/deletion of memory block
912 * devices, the device_hotplug_lock is not needed.
913 */
914void __init memory_dev_init(void)
915{
916 int ret;
917 unsigned long block_sz, nr;
918
919 /* Validate the configured memory block size */
920 block_sz = memory_block_size_bytes();
921 if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
922 panic("Memory block size not suitable: 0x%lx\n", block_sz);
923 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
924
925 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
926 if (ret)
927 panic("%s() failed to register subsystem: %d\n", __func__, ret);
928
929 /*
930 * Create entries for memory sections that were found
931 * during boot and have been initialized
932 */
933 for (nr = 0; nr <= __highest_present_section_nr;
934 nr += sections_per_block) {
935 ret = add_boot_memory_block(nr);
936 if (ret)
937 panic("%s() failed to add memory block: %d\n", __func__,
938 ret);
939 }
940}
941
942/**
943 * walk_memory_blocks - walk through all present memory blocks overlapped
944 * by the range [start, start + size)
945 *
946 * @start: start address of the memory range
947 * @size: size of the memory range
948 * @arg: argument passed to func
949 * @func: callback for each memory section walked
950 *
951 * This function walks through all present memory blocks overlapped by the
952 * range [start, start + size), calling func on each memory block.
953 *
954 * In case func() returns an error, walking is aborted and the error is
955 * returned.
956 *
957 * Called under device_hotplug_lock.
958 */
959int walk_memory_blocks(unsigned long start, unsigned long size,
960 void *arg, walk_memory_blocks_func_t func)
961{
962 const unsigned long start_block_id = phys_to_block_id(start);
963 const unsigned long end_block_id = phys_to_block_id(start + size - 1);
964 struct memory_block *mem;
965 unsigned long block_id;
966 int ret = 0;
967
968 if (!size)
969 return 0;
970
971 for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
972 mem = find_memory_block_by_id(block_id);
973 if (!mem)
974 continue;
975
976 ret = func(mem, arg);
977 put_device(&mem->dev);
978 if (ret)
979 break;
980 }
981 return ret;
982}
983
984struct for_each_memory_block_cb_data {
985 walk_memory_blocks_func_t func;
986 void *arg;
987};
988
989static int for_each_memory_block_cb(struct device *dev, void *data)
990{
991 struct memory_block *mem = to_memory_block(dev);
992 struct for_each_memory_block_cb_data *cb_data = data;
993
994 return cb_data->func(mem, cb_data->arg);
995}
996
997/**
998 * for_each_memory_block - walk through all present memory blocks
999 *
1000 * @arg: argument passed to func
1001 * @func: callback for each memory block walked
1002 *
1003 * This function walks through all present memory blocks, calling func on
1004 * each memory block.
1005 *
1006 * In case func() returns an error, walking is aborted and the error is
1007 * returned.
1008 */
1009int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
1010{
1011 struct for_each_memory_block_cb_data cb_data = {
1012 .func = func,
1013 .arg = arg,
1014 };
1015
1016 return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
1017 for_each_memory_block_cb);
1018}
1019
1020/*
1021 * This is an internal helper to unify allocation and initialization of
1022 * memory groups. Note that the passed memory group will be copied to a
1023 * dynamically allocated memory group. After this call, the passed
1024 * memory group should no longer be used.
1025 */
1026static int memory_group_register(struct memory_group group)
1027{
1028 struct memory_group *new_group;
1029 uint32_t mgid;
1030 int ret;
1031
1032 if (!node_possible(group.nid))
1033 return -EINVAL;
1034
1035 new_group = kzalloc(sizeof(group), GFP_KERNEL);
1036 if (!new_group)
1037 return -ENOMEM;
1038 *new_group = group;
1039 INIT_LIST_HEAD(&new_group->memory_blocks);
1040
1041 ret = xa_alloc(&memory_groups, &mgid, new_group, xa_limit_31b,
1042 GFP_KERNEL);
1043 if (ret) {
1044 kfree(new_group);
1045 return ret;
1046 } else if (group.is_dynamic) {
1047 xa_set_mark(&memory_groups, mgid, MEMORY_GROUP_MARK_DYNAMIC);
1048 }
1049 return mgid;
1050}
1051
1052/**
1053 * memory_group_register_static() - Register a static memory group.
1054 * @nid: The node id.
1055 * @max_pages: The maximum number of pages we'll have in this static memory
1056 * group.
1057 *
1058 * Register a new static memory group and return the memory group id.
1059 * All memory in the group belongs to a single unit, such as a DIMM. All
1060 * memory belonging to a static memory group is added in one go to be removed
1061 * in one go -- it's static.
1062 *
1063 * Returns an error if out of memory, if the node id is invalid, if no new
1064 * memory groups can be registered, or if max_pages is invalid (0). Otherwise,
1065 * returns the new memory group id.
1066 */
1067int memory_group_register_static(int nid, unsigned long max_pages)
1068{
1069 struct memory_group group = {
1070 .nid = nid,
1071 .s = {
1072 .max_pages = max_pages,
1073 },
1074 };
1075
1076 if (!max_pages)
1077 return -EINVAL;
1078 return memory_group_register(group);
1079}
1080EXPORT_SYMBOL_GPL(memory_group_register_static);
1081
1082/**
1083 * memory_group_register_dynamic() - Register a dynamic memory group.
1084 * @nid: The node id.
1085 * @unit_pages: Unit in pages in which is memory added/removed in this dynamic
1086 * memory group.
1087 *
1088 * Register a new dynamic memory group and return the memory group id.
1089 * Memory within a dynamic memory group is added/removed dynamically
1090 * in unit_pages.
1091 *
1092 * Returns an error if out of memory, if the node id is invalid, if no new
1093 * memory groups can be registered, or if unit_pages is invalid (0, not a
1094 * power of two, smaller than a single memory block). Otherwise, returns the
1095 * new memory group id.
1096 */
1097int memory_group_register_dynamic(int nid, unsigned long unit_pages)
1098{
1099 struct memory_group group = {
1100 .nid = nid,
1101 .is_dynamic = true,
1102 .d = {
1103 .unit_pages = unit_pages,
1104 },
1105 };
1106
1107 if (!unit_pages || !is_power_of_2(unit_pages) ||
1108 unit_pages < PHYS_PFN(memory_block_size_bytes()))
1109 return -EINVAL;
1110 return memory_group_register(group);
1111}
1112EXPORT_SYMBOL_GPL(memory_group_register_dynamic);
1113
1114/**
1115 * memory_group_unregister() - Unregister a memory group.
1116 * @mgid: the memory group id
1117 *
1118 * Unregister a memory group. If any memory block still belongs to this
1119 * memory group, unregistering will fail.
1120 *
1121 * Returns -EINVAL if the memory group id is invalid, returns -EBUSY if some
1122 * memory blocks still belong to this memory group and returns 0 if
1123 * unregistering succeeded.
1124 */
1125int memory_group_unregister(int mgid)
1126{
1127 struct memory_group *group;
1128
1129 if (mgid < 0)
1130 return -EINVAL;
1131
1132 group = xa_load(&memory_groups, mgid);
1133 if (!group)
1134 return -EINVAL;
1135 if (!list_empty(&group->memory_blocks))
1136 return -EBUSY;
1137 xa_erase(&memory_groups, mgid);
1138 kfree(group);
1139 return 0;
1140}
1141EXPORT_SYMBOL_GPL(memory_group_unregister);
1142
1143/*
1144 * This is an internal helper only to be used in core memory hotplug code to
1145 * lookup a memory group. We don't care about locking, as we don't expect a
1146 * memory group to get unregistered while adding memory to it -- because
1147 * the group and the memory is managed by the same driver.
1148 */
1149struct memory_group *memory_group_find_by_id(int mgid)
1150{
1151 return xa_load(&memory_groups, mgid);
1152}
1153
1154/*
1155 * This is an internal helper only to be used in core memory hotplug code to
1156 * walk all dynamic memory groups excluding a given memory group, either
1157 * belonging to a specific node, or belonging to any node.
1158 */
1159int walk_dynamic_memory_groups(int nid, walk_memory_groups_func_t func,
1160 struct memory_group *excluded, void *arg)
1161{
1162 struct memory_group *group;
1163 unsigned long index;
1164 int ret = 0;
1165
1166 xa_for_each_marked(&memory_groups, index, group,
1167 MEMORY_GROUP_MARK_DYNAMIC) {
1168 if (group == excluded)
1169 continue;
1170#ifdef CONFIG_NUMA
1171 if (nid != NUMA_NO_NODE && group->nid != nid)
1172 continue;
1173#endif /* CONFIG_NUMA */
1174 ret = func(group, arg);
1175 if (ret)
1176 break;
1177 }
1178 return ret;
1179}
1180
1181#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
1182void memblk_nr_poison_inc(unsigned long pfn)
1183{
1184 const unsigned long block_id = pfn_to_block_id(pfn);
1185 struct memory_block *mem = find_memory_block_by_id(block_id);
1186
1187 if (mem)
1188 atomic_long_inc(&mem->nr_hwpoison);
1189}
1190
1191void memblk_nr_poison_sub(unsigned long pfn, long i)
1192{
1193 const unsigned long block_id = pfn_to_block_id(pfn);
1194 struct memory_block *mem = find_memory_block_by_id(block_id);
1195
1196 if (mem)
1197 atomic_long_sub(i, &mem->nr_hwpoison);
1198}
1199
1200static unsigned long memblk_nr_poison(struct memory_block *mem)
1201{
1202 return atomic_long_read(&mem->nr_hwpoison);
1203}
1204#endif
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Memory subsystem support
4 *
5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
6 * Dave Hansen <haveblue@us.ibm.com>
7 *
8 * This file provides the necessary infrastructure to represent
9 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
10 * All arch-independent code that assumes MEMORY_HOTPLUG requires
11 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
12 */
13
14#include <linux/module.h>
15#include <linux/init.h>
16#include <linux/topology.h>
17#include <linux/capability.h>
18#include <linux/device.h>
19#include <linux/memory.h>
20#include <linux/memory_hotplug.h>
21#include <linux/mm.h>
22#include <linux/mutex.h>
23#include <linux/stat.h>
24#include <linux/slab.h>
25
26#include <linux/atomic.h>
27#include <linux/uaccess.h>
28
29static DEFINE_MUTEX(mem_sysfs_mutex);
30
31#define MEMORY_CLASS_NAME "memory"
32
33#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
34
35static int sections_per_block;
36
37static inline unsigned long base_memory_block_id(unsigned long section_nr)
38{
39 return section_nr / sections_per_block;
40}
41
42static inline unsigned long pfn_to_block_id(unsigned long pfn)
43{
44 return base_memory_block_id(pfn_to_section_nr(pfn));
45}
46
47static inline unsigned long phys_to_block_id(unsigned long phys)
48{
49 return pfn_to_block_id(PFN_DOWN(phys));
50}
51
52static int memory_subsys_online(struct device *dev);
53static int memory_subsys_offline(struct device *dev);
54
55static struct bus_type memory_subsys = {
56 .name = MEMORY_CLASS_NAME,
57 .dev_name = MEMORY_CLASS_NAME,
58 .online = memory_subsys_online,
59 .offline = memory_subsys_offline,
60};
61
62static BLOCKING_NOTIFIER_HEAD(memory_chain);
63
64int register_memory_notifier(struct notifier_block *nb)
65{
66 return blocking_notifier_chain_register(&memory_chain, nb);
67}
68EXPORT_SYMBOL(register_memory_notifier);
69
70void unregister_memory_notifier(struct notifier_block *nb)
71{
72 blocking_notifier_chain_unregister(&memory_chain, nb);
73}
74EXPORT_SYMBOL(unregister_memory_notifier);
75
76static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
77
78int register_memory_isolate_notifier(struct notifier_block *nb)
79{
80 return atomic_notifier_chain_register(&memory_isolate_chain, nb);
81}
82EXPORT_SYMBOL(register_memory_isolate_notifier);
83
84void unregister_memory_isolate_notifier(struct notifier_block *nb)
85{
86 atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
87}
88EXPORT_SYMBOL(unregister_memory_isolate_notifier);
89
90static void memory_block_release(struct device *dev)
91{
92 struct memory_block *mem = to_memory_block(dev);
93
94 kfree(mem);
95}
96
97unsigned long __weak memory_block_size_bytes(void)
98{
99 return MIN_MEMORY_BLOCK_SIZE;
100}
101EXPORT_SYMBOL_GPL(memory_block_size_bytes);
102
103/*
104 * Show the first physical section index (number) of this memory block.
105 */
106static ssize_t phys_index_show(struct device *dev,
107 struct device_attribute *attr, char *buf)
108{
109 struct memory_block *mem = to_memory_block(dev);
110 unsigned long phys_index;
111
112 phys_index = mem->start_section_nr / sections_per_block;
113 return sprintf(buf, "%08lx\n", phys_index);
114}
115
116/*
117 * Show whether the memory block is likely to be offlineable (or is already
118 * offline). Once offline, the memory block could be removed. The return
119 * value does, however, not indicate that there is a way to remove the
120 * memory block.
121 */
122static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
123 char *buf)
124{
125 struct memory_block *mem = to_memory_block(dev);
126 unsigned long pfn;
127 int ret = 1, i;
128
129 if (mem->state != MEM_ONLINE)
130 goto out;
131
132 for (i = 0; i < sections_per_block; i++) {
133 if (!present_section_nr(mem->start_section_nr + i))
134 continue;
135 pfn = section_nr_to_pfn(mem->start_section_nr + i);
136 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
137 }
138
139out:
140 return sprintf(buf, "%d\n", ret);
141}
142
143/*
144 * online, offline, going offline, etc.
145 */
146static ssize_t state_show(struct device *dev, struct device_attribute *attr,
147 char *buf)
148{
149 struct memory_block *mem = to_memory_block(dev);
150 ssize_t len = 0;
151
152 /*
153 * We can probably put these states in a nice little array
154 * so that they're not open-coded
155 */
156 switch (mem->state) {
157 case MEM_ONLINE:
158 len = sprintf(buf, "online\n");
159 break;
160 case MEM_OFFLINE:
161 len = sprintf(buf, "offline\n");
162 break;
163 case MEM_GOING_OFFLINE:
164 len = sprintf(buf, "going-offline\n");
165 break;
166 default:
167 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
168 mem->state);
169 WARN_ON(1);
170 break;
171 }
172
173 return len;
174}
175
176int memory_notify(unsigned long val, void *v)
177{
178 return blocking_notifier_call_chain(&memory_chain, val, v);
179}
180
181int memory_isolate_notify(unsigned long val, void *v)
182{
183 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
184}
185
186/*
187 * The probe routines leave the pages uninitialized, just as the bootmem code
188 * does. Make sure we do not access them, but instead use only information from
189 * within sections.
190 */
191static bool pages_correctly_probed(unsigned long start_pfn)
192{
193 unsigned long section_nr = pfn_to_section_nr(start_pfn);
194 unsigned long section_nr_end = section_nr + sections_per_block;
195 unsigned long pfn = start_pfn;
196
197 /*
198 * memmap between sections is not contiguous except with
199 * SPARSEMEM_VMEMMAP. We lookup the page once per section
200 * and assume memmap is contiguous within each section
201 */
202 for (; section_nr < section_nr_end; section_nr++) {
203 if (WARN_ON_ONCE(!pfn_valid(pfn)))
204 return false;
205
206 if (!present_section_nr(section_nr)) {
207 pr_warn("section %ld pfn[%lx, %lx) not present\n",
208 section_nr, pfn, pfn + PAGES_PER_SECTION);
209 return false;
210 } else if (!valid_section_nr(section_nr)) {
211 pr_warn("section %ld pfn[%lx, %lx) no valid memmap\n",
212 section_nr, pfn, pfn + PAGES_PER_SECTION);
213 return false;
214 } else if (online_section_nr(section_nr)) {
215 pr_warn("section %ld pfn[%lx, %lx) is already online\n",
216 section_nr, pfn, pfn + PAGES_PER_SECTION);
217 return false;
218 }
219 pfn += PAGES_PER_SECTION;
220 }
221
222 return true;
223}
224
225/*
226 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
227 * OK to have direct references to sparsemem variables in here.
228 */
229static int
230memory_block_action(unsigned long start_section_nr, unsigned long action,
231 int online_type)
232{
233 unsigned long start_pfn;
234 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
235 int ret;
236
237 start_pfn = section_nr_to_pfn(start_section_nr);
238
239 switch (action) {
240 case MEM_ONLINE:
241 if (!pages_correctly_probed(start_pfn))
242 return -EBUSY;
243
244 ret = online_pages(start_pfn, nr_pages, online_type);
245 break;
246 case MEM_OFFLINE:
247 ret = offline_pages(start_pfn, nr_pages);
248 break;
249 default:
250 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
251 "%ld\n", __func__, start_section_nr, action, action);
252 ret = -EINVAL;
253 }
254
255 return ret;
256}
257
258static int memory_block_change_state(struct memory_block *mem,
259 unsigned long to_state, unsigned long from_state_req)
260{
261 int ret = 0;
262
263 if (mem->state != from_state_req)
264 return -EINVAL;
265
266 if (to_state == MEM_OFFLINE)
267 mem->state = MEM_GOING_OFFLINE;
268
269 ret = memory_block_action(mem->start_section_nr, to_state,
270 mem->online_type);
271
272 mem->state = ret ? from_state_req : to_state;
273
274 return ret;
275}
276
277/* The device lock serializes operations on memory_subsys_[online|offline] */
278static int memory_subsys_online(struct device *dev)
279{
280 struct memory_block *mem = to_memory_block(dev);
281 int ret;
282
283 if (mem->state == MEM_ONLINE)
284 return 0;
285
286 /*
287 * If we are called from state_store(), online_type will be
288 * set >= 0 Otherwise we were called from the device online
289 * attribute and need to set the online_type.
290 */
291 if (mem->online_type < 0)
292 mem->online_type = MMOP_ONLINE_KEEP;
293
294 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
295
296 /* clear online_type */
297 mem->online_type = -1;
298
299 return ret;
300}
301
302static int memory_subsys_offline(struct device *dev)
303{
304 struct memory_block *mem = to_memory_block(dev);
305
306 if (mem->state == MEM_OFFLINE)
307 return 0;
308
309 /* Can't offline block with non-present sections */
310 if (mem->section_count != sections_per_block)
311 return -EINVAL;
312
313 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
314}
315
316static ssize_t state_store(struct device *dev, struct device_attribute *attr,
317 const char *buf, size_t count)
318{
319 struct memory_block *mem = to_memory_block(dev);
320 int ret, online_type;
321
322 ret = lock_device_hotplug_sysfs();
323 if (ret)
324 return ret;
325
326 if (sysfs_streq(buf, "online_kernel"))
327 online_type = MMOP_ONLINE_KERNEL;
328 else if (sysfs_streq(buf, "online_movable"))
329 online_type = MMOP_ONLINE_MOVABLE;
330 else if (sysfs_streq(buf, "online"))
331 online_type = MMOP_ONLINE_KEEP;
332 else if (sysfs_streq(buf, "offline"))
333 online_type = MMOP_OFFLINE;
334 else {
335 ret = -EINVAL;
336 goto err;
337 }
338
339 switch (online_type) {
340 case MMOP_ONLINE_KERNEL:
341 case MMOP_ONLINE_MOVABLE:
342 case MMOP_ONLINE_KEEP:
343 /* mem->online_type is protected by device_hotplug_lock */
344 mem->online_type = online_type;
345 ret = device_online(&mem->dev);
346 break;
347 case MMOP_OFFLINE:
348 ret = device_offline(&mem->dev);
349 break;
350 default:
351 ret = -EINVAL; /* should never happen */
352 }
353
354err:
355 unlock_device_hotplug();
356
357 if (ret < 0)
358 return ret;
359 if (ret)
360 return -EINVAL;
361
362 return count;
363}
364
365/*
366 * phys_device is a bad name for this. What I really want
367 * is a way to differentiate between memory ranges that
368 * are part of physical devices that constitute
369 * a complete removable unit or fru.
370 * i.e. do these ranges belong to the same physical device,
371 * s.t. if I offline all of these sections I can then
372 * remove the physical device?
373 */
374static ssize_t phys_device_show(struct device *dev,
375 struct device_attribute *attr, char *buf)
376{
377 struct memory_block *mem = to_memory_block(dev);
378 return sprintf(buf, "%d\n", mem->phys_device);
379}
380
381#ifdef CONFIG_MEMORY_HOTREMOVE
382static void print_allowed_zone(char *buf, int nid, unsigned long start_pfn,
383 unsigned long nr_pages, int online_type,
384 struct zone *default_zone)
385{
386 struct zone *zone;
387
388 zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
389 if (zone != default_zone) {
390 strcat(buf, " ");
391 strcat(buf, zone->name);
392 }
393}
394
395static ssize_t valid_zones_show(struct device *dev,
396 struct device_attribute *attr, char *buf)
397{
398 struct memory_block *mem = to_memory_block(dev);
399 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
400 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
401 unsigned long valid_start_pfn, valid_end_pfn;
402 struct zone *default_zone;
403 int nid;
404
405 /*
406 * Check the existing zone. Make sure that we do that only on the
407 * online nodes otherwise the page_zone is not reliable
408 */
409 if (mem->state == MEM_ONLINE) {
410 /*
411 * The block contains more than one zone can not be offlined.
412 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
413 */
414 if (!test_pages_in_a_zone(start_pfn, start_pfn + nr_pages,
415 &valid_start_pfn, &valid_end_pfn))
416 return sprintf(buf, "none\n");
417 start_pfn = valid_start_pfn;
418 strcat(buf, page_zone(pfn_to_page(start_pfn))->name);
419 goto out;
420 }
421
422 nid = mem->nid;
423 default_zone = zone_for_pfn_range(MMOP_ONLINE_KEEP, nid, start_pfn, nr_pages);
424 strcat(buf, default_zone->name);
425
426 print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_KERNEL,
427 default_zone);
428 print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_MOVABLE,
429 default_zone);
430out:
431 strcat(buf, "\n");
432
433 return strlen(buf);
434}
435static DEVICE_ATTR_RO(valid_zones);
436#endif
437
438static DEVICE_ATTR_RO(phys_index);
439static DEVICE_ATTR_RW(state);
440static DEVICE_ATTR_RO(phys_device);
441static DEVICE_ATTR_RO(removable);
442
443/*
444 * Show the memory block size (shared by all memory blocks).
445 */
446static ssize_t block_size_bytes_show(struct device *dev,
447 struct device_attribute *attr, char *buf)
448{
449 return sprintf(buf, "%lx\n", memory_block_size_bytes());
450}
451
452static DEVICE_ATTR_RO(block_size_bytes);
453
454/*
455 * Memory auto online policy.
456 */
457
458static ssize_t auto_online_blocks_show(struct device *dev,
459 struct device_attribute *attr, char *buf)
460{
461 if (memhp_auto_online)
462 return sprintf(buf, "online\n");
463 else
464 return sprintf(buf, "offline\n");
465}
466
467static ssize_t auto_online_blocks_store(struct device *dev,
468 struct device_attribute *attr,
469 const char *buf, size_t count)
470{
471 if (sysfs_streq(buf, "online"))
472 memhp_auto_online = true;
473 else if (sysfs_streq(buf, "offline"))
474 memhp_auto_online = false;
475 else
476 return -EINVAL;
477
478 return count;
479}
480
481static DEVICE_ATTR_RW(auto_online_blocks);
482
483/*
484 * Some architectures will have custom drivers to do this, and
485 * will not need to do it from userspace. The fake hot-add code
486 * as well as ppc64 will do all of their discovery in userspace
487 * and will require this interface.
488 */
489#ifdef CONFIG_ARCH_MEMORY_PROBE
490static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
491 const char *buf, size_t count)
492{
493 u64 phys_addr;
494 int nid, ret;
495 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
496
497 ret = kstrtoull(buf, 0, &phys_addr);
498 if (ret)
499 return ret;
500
501 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
502 return -EINVAL;
503
504 ret = lock_device_hotplug_sysfs();
505 if (ret)
506 return ret;
507
508 nid = memory_add_physaddr_to_nid(phys_addr);
509 ret = __add_memory(nid, phys_addr,
510 MIN_MEMORY_BLOCK_SIZE * sections_per_block);
511
512 if (ret)
513 goto out;
514
515 ret = count;
516out:
517 unlock_device_hotplug();
518 return ret;
519}
520
521static DEVICE_ATTR_WO(probe);
522#endif
523
524#ifdef CONFIG_MEMORY_FAILURE
525/*
526 * Support for offlining pages of memory
527 */
528
529/* Soft offline a page */
530static ssize_t soft_offline_page_store(struct device *dev,
531 struct device_attribute *attr,
532 const char *buf, size_t count)
533{
534 int ret;
535 u64 pfn;
536 if (!capable(CAP_SYS_ADMIN))
537 return -EPERM;
538 if (kstrtoull(buf, 0, &pfn) < 0)
539 return -EINVAL;
540 pfn >>= PAGE_SHIFT;
541 if (!pfn_valid(pfn))
542 return -ENXIO;
543 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
544 if (!pfn_to_online_page(pfn))
545 return -EIO;
546 ret = soft_offline_page(pfn_to_page(pfn), 0);
547 return ret == 0 ? count : ret;
548}
549
550/* Forcibly offline a page, including killing processes. */
551static ssize_t hard_offline_page_store(struct device *dev,
552 struct device_attribute *attr,
553 const char *buf, size_t count)
554{
555 int ret;
556 u64 pfn;
557 if (!capable(CAP_SYS_ADMIN))
558 return -EPERM;
559 if (kstrtoull(buf, 0, &pfn) < 0)
560 return -EINVAL;
561 pfn >>= PAGE_SHIFT;
562 ret = memory_failure(pfn, 0);
563 return ret ? ret : count;
564}
565
566static DEVICE_ATTR_WO(soft_offline_page);
567static DEVICE_ATTR_WO(hard_offline_page);
568#endif
569
570/*
571 * Note that phys_device is optional. It is here to allow for
572 * differentiation between which *physical* devices each
573 * section belongs to...
574 */
575int __weak arch_get_memory_phys_device(unsigned long start_pfn)
576{
577 return 0;
578}
579
580/* A reference for the returned memory block device is acquired. */
581static struct memory_block *find_memory_block_by_id(unsigned long block_id)
582{
583 struct device *dev;
584
585 dev = subsys_find_device_by_id(&memory_subsys, block_id, NULL);
586 return dev ? to_memory_block(dev) : NULL;
587}
588
589/*
590 * For now, we have a linear search to go find the appropriate
591 * memory_block corresponding to a particular phys_index. If
592 * this gets to be a real problem, we can always use a radix
593 * tree or something here.
594 *
595 * This could be made generic for all device subsystems.
596 */
597struct memory_block *find_memory_block(struct mem_section *section)
598{
599 unsigned long block_id = base_memory_block_id(__section_nr(section));
600
601 return find_memory_block_by_id(block_id);
602}
603
604static struct attribute *memory_memblk_attrs[] = {
605 &dev_attr_phys_index.attr,
606 &dev_attr_state.attr,
607 &dev_attr_phys_device.attr,
608 &dev_attr_removable.attr,
609#ifdef CONFIG_MEMORY_HOTREMOVE
610 &dev_attr_valid_zones.attr,
611#endif
612 NULL
613};
614
615static struct attribute_group memory_memblk_attr_group = {
616 .attrs = memory_memblk_attrs,
617};
618
619static const struct attribute_group *memory_memblk_attr_groups[] = {
620 &memory_memblk_attr_group,
621 NULL,
622};
623
624/*
625 * register_memory - Setup a sysfs device for a memory block
626 */
627static
628int register_memory(struct memory_block *memory)
629{
630 int ret;
631
632 memory->dev.bus = &memory_subsys;
633 memory->dev.id = memory->start_section_nr / sections_per_block;
634 memory->dev.release = memory_block_release;
635 memory->dev.groups = memory_memblk_attr_groups;
636 memory->dev.offline = memory->state == MEM_OFFLINE;
637
638 ret = device_register(&memory->dev);
639 if (ret)
640 put_device(&memory->dev);
641
642 return ret;
643}
644
645static int init_memory_block(struct memory_block **memory,
646 unsigned long block_id, unsigned long state)
647{
648 struct memory_block *mem;
649 unsigned long start_pfn;
650 int ret = 0;
651
652 mem = find_memory_block_by_id(block_id);
653 if (mem) {
654 put_device(&mem->dev);
655 return -EEXIST;
656 }
657 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
658 if (!mem)
659 return -ENOMEM;
660
661 mem->start_section_nr = block_id * sections_per_block;
662 mem->state = state;
663 start_pfn = section_nr_to_pfn(mem->start_section_nr);
664 mem->phys_device = arch_get_memory_phys_device(start_pfn);
665 mem->nid = NUMA_NO_NODE;
666
667 ret = register_memory(mem);
668
669 *memory = mem;
670 return ret;
671}
672
673static int add_memory_block(unsigned long base_section_nr)
674{
675 int ret, section_count = 0;
676 struct memory_block *mem;
677 unsigned long nr;
678
679 for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
680 nr++)
681 if (present_section_nr(nr))
682 section_count++;
683
684 if (section_count == 0)
685 return 0;
686 ret = init_memory_block(&mem, base_memory_block_id(base_section_nr),
687 MEM_ONLINE);
688 if (ret)
689 return ret;
690 mem->section_count = section_count;
691 return 0;
692}
693
694static void unregister_memory(struct memory_block *memory)
695{
696 if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
697 return;
698
699 /* drop the ref. we got via find_memory_block() */
700 put_device(&memory->dev);
701 device_unregister(&memory->dev);
702}
703
704/*
705 * Create memory block devices for the given memory area. Start and size
706 * have to be aligned to memory block granularity. Memory block devices
707 * will be initialized as offline.
708 */
709int create_memory_block_devices(unsigned long start, unsigned long size)
710{
711 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
712 unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
713 struct memory_block *mem;
714 unsigned long block_id;
715 int ret = 0;
716
717 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
718 !IS_ALIGNED(size, memory_block_size_bytes())))
719 return -EINVAL;
720
721 mutex_lock(&mem_sysfs_mutex);
722 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
723 ret = init_memory_block(&mem, block_id, MEM_OFFLINE);
724 if (ret)
725 break;
726 mem->section_count = sections_per_block;
727 }
728 if (ret) {
729 end_block_id = block_id;
730 for (block_id = start_block_id; block_id != end_block_id;
731 block_id++) {
732 mem = find_memory_block_by_id(block_id);
733 mem->section_count = 0;
734 unregister_memory(mem);
735 }
736 }
737 mutex_unlock(&mem_sysfs_mutex);
738 return ret;
739}
740
741/*
742 * Remove memory block devices for the given memory area. Start and size
743 * have to be aligned to memory block granularity. Memory block devices
744 * have to be offline.
745 */
746void remove_memory_block_devices(unsigned long start, unsigned long size)
747{
748 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
749 const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
750 struct memory_block *mem;
751 unsigned long block_id;
752
753 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
754 !IS_ALIGNED(size, memory_block_size_bytes())))
755 return;
756
757 mutex_lock(&mem_sysfs_mutex);
758 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
759 mem = find_memory_block_by_id(block_id);
760 if (WARN_ON_ONCE(!mem))
761 continue;
762 mem->section_count = 0;
763 unregister_memory_block_under_nodes(mem);
764 unregister_memory(mem);
765 }
766 mutex_unlock(&mem_sysfs_mutex);
767}
768
769/* return true if the memory block is offlined, otherwise, return false */
770bool is_memblock_offlined(struct memory_block *mem)
771{
772 return mem->state == MEM_OFFLINE;
773}
774
775static struct attribute *memory_root_attrs[] = {
776#ifdef CONFIG_ARCH_MEMORY_PROBE
777 &dev_attr_probe.attr,
778#endif
779
780#ifdef CONFIG_MEMORY_FAILURE
781 &dev_attr_soft_offline_page.attr,
782 &dev_attr_hard_offline_page.attr,
783#endif
784
785 &dev_attr_block_size_bytes.attr,
786 &dev_attr_auto_online_blocks.attr,
787 NULL
788};
789
790static struct attribute_group memory_root_attr_group = {
791 .attrs = memory_root_attrs,
792};
793
794static const struct attribute_group *memory_root_attr_groups[] = {
795 &memory_root_attr_group,
796 NULL,
797};
798
799/*
800 * Initialize the sysfs support for memory devices...
801 */
802void __init memory_dev_init(void)
803{
804 int ret;
805 int err;
806 unsigned long block_sz, nr;
807
808 /* Validate the configured memory block size */
809 block_sz = memory_block_size_bytes();
810 if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
811 panic("Memory block size not suitable: 0x%lx\n", block_sz);
812 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
813
814 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
815 if (ret)
816 goto out;
817
818 /*
819 * Create entries for memory sections that were found
820 * during boot and have been initialized
821 */
822 mutex_lock(&mem_sysfs_mutex);
823 for (nr = 0; nr <= __highest_present_section_nr;
824 nr += sections_per_block) {
825 err = add_memory_block(nr);
826 if (!ret)
827 ret = err;
828 }
829 mutex_unlock(&mem_sysfs_mutex);
830
831out:
832 if (ret)
833 panic("%s() failed: %d\n", __func__, ret);
834}
835
836/**
837 * walk_memory_blocks - walk through all present memory blocks overlapped
838 * by the range [start, start + size)
839 *
840 * @start: start address of the memory range
841 * @size: size of the memory range
842 * @arg: argument passed to func
843 * @func: callback for each memory section walked
844 *
845 * This function walks through all present memory blocks overlapped by the
846 * range [start, start + size), calling func on each memory block.
847 *
848 * In case func() returns an error, walking is aborted and the error is
849 * returned.
850 */
851int walk_memory_blocks(unsigned long start, unsigned long size,
852 void *arg, walk_memory_blocks_func_t func)
853{
854 const unsigned long start_block_id = phys_to_block_id(start);
855 const unsigned long end_block_id = phys_to_block_id(start + size - 1);
856 struct memory_block *mem;
857 unsigned long block_id;
858 int ret = 0;
859
860 if (!size)
861 return 0;
862
863 for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
864 mem = find_memory_block_by_id(block_id);
865 if (!mem)
866 continue;
867
868 ret = func(mem, arg);
869 put_device(&mem->dev);
870 if (ret)
871 break;
872 }
873 return ret;
874}
875
876struct for_each_memory_block_cb_data {
877 walk_memory_blocks_func_t func;
878 void *arg;
879};
880
881static int for_each_memory_block_cb(struct device *dev, void *data)
882{
883 struct memory_block *mem = to_memory_block(dev);
884 struct for_each_memory_block_cb_data *cb_data = data;
885
886 return cb_data->func(mem, cb_data->arg);
887}
888
889/**
890 * for_each_memory_block - walk through all present memory blocks
891 *
892 * @arg: argument passed to func
893 * @func: callback for each memory block walked
894 *
895 * This function walks through all present memory blocks, calling func on
896 * each memory block.
897 *
898 * In case func() returns an error, walking is aborted and the error is
899 * returned.
900 */
901int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
902{
903 struct for_each_memory_block_cb_data cb_data = {
904 .func = func,
905 .arg = arg,
906 };
907
908 return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
909 for_each_memory_block_cb);
910}