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 memhp_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
85static BLOCKING_NOTIFIER_HEAD(memory_chain);
86
87int register_memory_notifier(struct notifier_block *nb)
88{
89 return blocking_notifier_chain_register(&memory_chain, nb);
90}
91EXPORT_SYMBOL(register_memory_notifier);
92
93void unregister_memory_notifier(struct notifier_block *nb)
94{
95 blocking_notifier_chain_unregister(&memory_chain, nb);
96}
97EXPORT_SYMBOL(unregister_memory_notifier);
98
99static void memory_block_release(struct device *dev)
100{
101 struct memory_block *mem = to_memory_block(dev);
102
103 kfree(mem);
104}
105
106unsigned long __weak memory_block_size_bytes(void)
107{
108 return MIN_MEMORY_BLOCK_SIZE;
109}
110EXPORT_SYMBOL_GPL(memory_block_size_bytes);
111
112/*
113 * Show the first physical section index (number) of this memory block.
114 */
115static ssize_t phys_index_show(struct device *dev,
116 struct device_attribute *attr, char *buf)
117{
118 struct memory_block *mem = to_memory_block(dev);
119 unsigned long phys_index;
120
121 phys_index = mem->start_section_nr / sections_per_block;
122 return sprintf(buf, "%08lx\n", phys_index);
123}
124
125/*
126 * Legacy interface that we cannot remove. Always indicate "removable"
127 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
128 */
129static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
130 char *buf)
131{
132 return sprintf(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
133}
134
135/*
136 * online, offline, going offline, etc.
137 */
138static ssize_t state_show(struct device *dev, struct device_attribute *attr,
139 char *buf)
140{
141 struct memory_block *mem = to_memory_block(dev);
142 ssize_t len = 0;
143
144 /*
145 * We can probably put these states in a nice little array
146 * so that they're not open-coded
147 */
148 switch (mem->state) {
149 case MEM_ONLINE:
150 len = sprintf(buf, "online\n");
151 break;
152 case MEM_OFFLINE:
153 len = sprintf(buf, "offline\n");
154 break;
155 case MEM_GOING_OFFLINE:
156 len = sprintf(buf, "going-offline\n");
157 break;
158 default:
159 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
160 mem->state);
161 WARN_ON(1);
162 break;
163 }
164
165 return len;
166}
167
168int memory_notify(unsigned long val, void *v)
169{
170 return blocking_notifier_call_chain(&memory_chain, val, v);
171}
172
173/*
174 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
175 * OK to have direct references to sparsemem variables in here.
176 */
177static int
178memory_block_action(unsigned long start_section_nr, unsigned long action,
179 int online_type, int nid)
180{
181 unsigned long start_pfn;
182 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
183 int ret;
184
185 start_pfn = section_nr_to_pfn(start_section_nr);
186
187 switch (action) {
188 case MEM_ONLINE:
189 ret = online_pages(start_pfn, nr_pages, online_type, nid);
190 break;
191 case MEM_OFFLINE:
192 ret = offline_pages(start_pfn, nr_pages);
193 break;
194 default:
195 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
196 "%ld\n", __func__, start_section_nr, action, action);
197 ret = -EINVAL;
198 }
199
200 return ret;
201}
202
203static int memory_block_change_state(struct memory_block *mem,
204 unsigned long to_state, unsigned long from_state_req)
205{
206 int ret = 0;
207
208 if (mem->state != from_state_req)
209 return -EINVAL;
210
211 if (to_state == MEM_OFFLINE)
212 mem->state = MEM_GOING_OFFLINE;
213
214 ret = memory_block_action(mem->start_section_nr, to_state,
215 mem->online_type, mem->nid);
216
217 mem->state = ret ? from_state_req : to_state;
218
219 return ret;
220}
221
222/* The device lock serializes operations on memory_subsys_[online|offline] */
223static int memory_subsys_online(struct device *dev)
224{
225 struct memory_block *mem = to_memory_block(dev);
226 int ret;
227
228 if (mem->state == MEM_ONLINE)
229 return 0;
230
231 /*
232 * When called via device_online() without configuring the online_type,
233 * we want to default to MMOP_ONLINE.
234 */
235 if (mem->online_type == MMOP_OFFLINE)
236 mem->online_type = MMOP_ONLINE;
237
238 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
239 mem->online_type = MMOP_OFFLINE;
240
241 return ret;
242}
243
244static int memory_subsys_offline(struct device *dev)
245{
246 struct memory_block *mem = to_memory_block(dev);
247
248 if (mem->state == MEM_OFFLINE)
249 return 0;
250
251 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
252}
253
254static ssize_t state_store(struct device *dev, struct device_attribute *attr,
255 const char *buf, size_t count)
256{
257 const int online_type = memhp_online_type_from_str(buf);
258 struct memory_block *mem = to_memory_block(dev);
259 int ret;
260
261 if (online_type < 0)
262 return -EINVAL;
263
264 ret = lock_device_hotplug_sysfs();
265 if (ret)
266 return ret;
267
268 switch (online_type) {
269 case MMOP_ONLINE_KERNEL:
270 case MMOP_ONLINE_MOVABLE:
271 case MMOP_ONLINE:
272 /* mem->online_type is protected by device_hotplug_lock */
273 mem->online_type = online_type;
274 ret = device_online(&mem->dev);
275 break;
276 case MMOP_OFFLINE:
277 ret = device_offline(&mem->dev);
278 break;
279 default:
280 ret = -EINVAL; /* should never happen */
281 }
282
283 unlock_device_hotplug();
284
285 if (ret < 0)
286 return ret;
287 if (ret)
288 return -EINVAL;
289
290 return count;
291}
292
293/*
294 * phys_device is a bad name for this. What I really want
295 * is a way to differentiate between memory ranges that
296 * are part of physical devices that constitute
297 * a complete removable unit or fru.
298 * i.e. do these ranges belong to the same physical device,
299 * s.t. if I offline all of these sections I can then
300 * remove the physical device?
301 */
302static ssize_t phys_device_show(struct device *dev,
303 struct device_attribute *attr, char *buf)
304{
305 struct memory_block *mem = to_memory_block(dev);
306 return sprintf(buf, "%d\n", mem->phys_device);
307}
308
309#ifdef CONFIG_MEMORY_HOTREMOVE
310static void print_allowed_zone(char *buf, int nid, unsigned long start_pfn,
311 unsigned long nr_pages, int online_type,
312 struct zone *default_zone)
313{
314 struct zone *zone;
315
316 zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
317 if (zone != default_zone) {
318 strcat(buf, " ");
319 strcat(buf, zone->name);
320 }
321}
322
323static ssize_t valid_zones_show(struct device *dev,
324 struct device_attribute *attr, char *buf)
325{
326 struct memory_block *mem = to_memory_block(dev);
327 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
328 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
329 struct zone *default_zone;
330 int nid;
331
332 /*
333 * Check the existing zone. Make sure that we do that only on the
334 * online nodes otherwise the page_zone is not reliable
335 */
336 if (mem->state == MEM_ONLINE) {
337 /*
338 * The block contains more than one zone can not be offlined.
339 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
340 */
341 default_zone = test_pages_in_a_zone(start_pfn,
342 start_pfn + nr_pages);
343 if (!default_zone)
344 return sprintf(buf, "none\n");
345 strcat(buf, default_zone->name);
346 goto out;
347 }
348
349 nid = mem->nid;
350 default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, start_pfn,
351 nr_pages);
352 strcat(buf, default_zone->name);
353
354 print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_KERNEL,
355 default_zone);
356 print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_MOVABLE,
357 default_zone);
358out:
359 strcat(buf, "\n");
360
361 return strlen(buf);
362}
363static DEVICE_ATTR_RO(valid_zones);
364#endif
365
366static DEVICE_ATTR_RO(phys_index);
367static DEVICE_ATTR_RW(state);
368static DEVICE_ATTR_RO(phys_device);
369static DEVICE_ATTR_RO(removable);
370
371/*
372 * Show the memory block size (shared by all memory blocks).
373 */
374static ssize_t block_size_bytes_show(struct device *dev,
375 struct device_attribute *attr, char *buf)
376{
377 return sprintf(buf, "%lx\n", memory_block_size_bytes());
378}
379
380static DEVICE_ATTR_RO(block_size_bytes);
381
382/*
383 * Memory auto online policy.
384 */
385
386static ssize_t auto_online_blocks_show(struct device *dev,
387 struct device_attribute *attr, char *buf)
388{
389 return sprintf(buf, "%s\n",
390 online_type_to_str[memhp_default_online_type]);
391}
392
393static ssize_t auto_online_blocks_store(struct device *dev,
394 struct device_attribute *attr,
395 const char *buf, size_t count)
396{
397 const int online_type = memhp_online_type_from_str(buf);
398
399 if (online_type < 0)
400 return -EINVAL;
401
402 memhp_default_online_type = online_type;
403 return count;
404}
405
406static DEVICE_ATTR_RW(auto_online_blocks);
407
408/*
409 * Some architectures will have custom drivers to do this, and
410 * will not need to do it from userspace. The fake hot-add code
411 * as well as ppc64 will do all of their discovery in userspace
412 * and will require this interface.
413 */
414#ifdef CONFIG_ARCH_MEMORY_PROBE
415static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
416 const char *buf, size_t count)
417{
418 u64 phys_addr;
419 int nid, ret;
420 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
421
422 ret = kstrtoull(buf, 0, &phys_addr);
423 if (ret)
424 return ret;
425
426 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
427 return -EINVAL;
428
429 ret = lock_device_hotplug_sysfs();
430 if (ret)
431 return ret;
432
433 nid = memory_add_physaddr_to_nid(phys_addr);
434 ret = __add_memory(nid, phys_addr,
435 MIN_MEMORY_BLOCK_SIZE * sections_per_block);
436
437 if (ret)
438 goto out;
439
440 ret = count;
441out:
442 unlock_device_hotplug();
443 return ret;
444}
445
446static DEVICE_ATTR_WO(probe);
447#endif
448
449#ifdef CONFIG_MEMORY_FAILURE
450/*
451 * Support for offlining pages of memory
452 */
453
454/* Soft offline a page */
455static ssize_t soft_offline_page_store(struct device *dev,
456 struct device_attribute *attr,
457 const char *buf, size_t count)
458{
459 int ret;
460 u64 pfn;
461 if (!capable(CAP_SYS_ADMIN))
462 return -EPERM;
463 if (kstrtoull(buf, 0, &pfn) < 0)
464 return -EINVAL;
465 pfn >>= PAGE_SHIFT;
466 ret = soft_offline_page(pfn, 0);
467 return ret == 0 ? count : ret;
468}
469
470/* Forcibly offline a page, including killing processes. */
471static ssize_t hard_offline_page_store(struct device *dev,
472 struct device_attribute *attr,
473 const char *buf, size_t count)
474{
475 int ret;
476 u64 pfn;
477 if (!capable(CAP_SYS_ADMIN))
478 return -EPERM;
479 if (kstrtoull(buf, 0, &pfn) < 0)
480 return -EINVAL;
481 pfn >>= PAGE_SHIFT;
482 ret = memory_failure(pfn, 0);
483 return ret ? ret : count;
484}
485
486static DEVICE_ATTR_WO(soft_offline_page);
487static DEVICE_ATTR_WO(hard_offline_page);
488#endif
489
490/*
491 * Note that phys_device is optional. It is here to allow for
492 * differentiation between which *physical* devices each
493 * section belongs to...
494 */
495int __weak arch_get_memory_phys_device(unsigned long start_pfn)
496{
497 return 0;
498}
499
500/*
501 * A reference for the returned memory block device is acquired.
502 *
503 * Called under device_hotplug_lock.
504 */
505static struct memory_block *find_memory_block_by_id(unsigned long block_id)
506{
507 struct memory_block *mem;
508
509 mem = xa_load(&memory_blocks, block_id);
510 if (mem)
511 get_device(&mem->dev);
512 return mem;
513}
514
515/*
516 * Called under device_hotplug_lock.
517 */
518struct memory_block *find_memory_block(struct mem_section *section)
519{
520 unsigned long block_id = memory_block_id(__section_nr(section));
521
522 return find_memory_block_by_id(block_id);
523}
524
525static struct attribute *memory_memblk_attrs[] = {
526 &dev_attr_phys_index.attr,
527 &dev_attr_state.attr,
528 &dev_attr_phys_device.attr,
529 &dev_attr_removable.attr,
530#ifdef CONFIG_MEMORY_HOTREMOVE
531 &dev_attr_valid_zones.attr,
532#endif
533 NULL
534};
535
536static struct attribute_group memory_memblk_attr_group = {
537 .attrs = memory_memblk_attrs,
538};
539
540static const struct attribute_group *memory_memblk_attr_groups[] = {
541 &memory_memblk_attr_group,
542 NULL,
543};
544
545/*
546 * register_memory - Setup a sysfs device for a memory block
547 */
548static
549int register_memory(struct memory_block *memory)
550{
551 int ret;
552
553 memory->dev.bus = &memory_subsys;
554 memory->dev.id = memory->start_section_nr / sections_per_block;
555 memory->dev.release = memory_block_release;
556 memory->dev.groups = memory_memblk_attr_groups;
557 memory->dev.offline = memory->state == MEM_OFFLINE;
558
559 ret = device_register(&memory->dev);
560 if (ret) {
561 put_device(&memory->dev);
562 return ret;
563 }
564 ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
565 GFP_KERNEL));
566 if (ret) {
567 put_device(&memory->dev);
568 device_unregister(&memory->dev);
569 }
570 return ret;
571}
572
573static int init_memory_block(unsigned long block_id, unsigned long state)
574{
575 struct memory_block *mem;
576 unsigned long start_pfn;
577 int ret = 0;
578
579 mem = find_memory_block_by_id(block_id);
580 if (mem) {
581 put_device(&mem->dev);
582 return -EEXIST;
583 }
584 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
585 if (!mem)
586 return -ENOMEM;
587
588 mem->start_section_nr = block_id * sections_per_block;
589 mem->state = state;
590 start_pfn = section_nr_to_pfn(mem->start_section_nr);
591 mem->phys_device = arch_get_memory_phys_device(start_pfn);
592 mem->nid = NUMA_NO_NODE;
593
594 ret = register_memory(mem);
595
596 return ret;
597}
598
599static int add_memory_block(unsigned long base_section_nr)
600{
601 int section_count = 0;
602 unsigned long nr;
603
604 for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
605 nr++)
606 if (present_section_nr(nr))
607 section_count++;
608
609 if (section_count == 0)
610 return 0;
611 return init_memory_block(memory_block_id(base_section_nr),
612 MEM_ONLINE);
613}
614
615static void unregister_memory(struct memory_block *memory)
616{
617 if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
618 return;
619
620 WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
621
622 /* drop the ref. we got via find_memory_block() */
623 put_device(&memory->dev);
624 device_unregister(&memory->dev);
625}
626
627/*
628 * Create memory block devices for the given memory area. Start and size
629 * have to be aligned to memory block granularity. Memory block devices
630 * will be initialized as offline.
631 *
632 * Called under device_hotplug_lock.
633 */
634int create_memory_block_devices(unsigned long start, unsigned long size)
635{
636 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
637 unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
638 struct memory_block *mem;
639 unsigned long block_id;
640 int ret = 0;
641
642 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
643 !IS_ALIGNED(size, memory_block_size_bytes())))
644 return -EINVAL;
645
646 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
647 ret = init_memory_block(block_id, MEM_OFFLINE);
648 if (ret)
649 break;
650 }
651 if (ret) {
652 end_block_id = block_id;
653 for (block_id = start_block_id; block_id != end_block_id;
654 block_id++) {
655 mem = find_memory_block_by_id(block_id);
656 if (WARN_ON_ONCE(!mem))
657 continue;
658 unregister_memory(mem);
659 }
660 }
661 return ret;
662}
663
664/*
665 * Remove memory block devices for the given memory area. Start and size
666 * have to be aligned to memory block granularity. Memory block devices
667 * have to be offline.
668 *
669 * Called under device_hotplug_lock.
670 */
671void remove_memory_block_devices(unsigned long start, unsigned long size)
672{
673 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
674 const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
675 struct memory_block *mem;
676 unsigned long block_id;
677
678 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
679 !IS_ALIGNED(size, memory_block_size_bytes())))
680 return;
681
682 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
683 mem = find_memory_block_by_id(block_id);
684 if (WARN_ON_ONCE(!mem))
685 continue;
686 unregister_memory_block_under_nodes(mem);
687 unregister_memory(mem);
688 }
689}
690
691/* return true if the memory block is offlined, otherwise, return false */
692bool is_memblock_offlined(struct memory_block *mem)
693{
694 return mem->state == MEM_OFFLINE;
695}
696
697static struct attribute *memory_root_attrs[] = {
698#ifdef CONFIG_ARCH_MEMORY_PROBE
699 &dev_attr_probe.attr,
700#endif
701
702#ifdef CONFIG_MEMORY_FAILURE
703 &dev_attr_soft_offline_page.attr,
704 &dev_attr_hard_offline_page.attr,
705#endif
706
707 &dev_attr_block_size_bytes.attr,
708 &dev_attr_auto_online_blocks.attr,
709 NULL
710};
711
712static struct attribute_group memory_root_attr_group = {
713 .attrs = memory_root_attrs,
714};
715
716static const struct attribute_group *memory_root_attr_groups[] = {
717 &memory_root_attr_group,
718 NULL,
719};
720
721/*
722 * Initialize the sysfs support for memory devices. At the time this function
723 * is called, we cannot have concurrent creation/deletion of memory block
724 * devices, the device_hotplug_lock is not needed.
725 */
726void __init memory_dev_init(void)
727{
728 int ret;
729 unsigned long block_sz, nr;
730
731 /* Validate the configured memory block size */
732 block_sz = memory_block_size_bytes();
733 if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
734 panic("Memory block size not suitable: 0x%lx\n", block_sz);
735 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
736
737 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
738 if (ret)
739 panic("%s() failed to register subsystem: %d\n", __func__, ret);
740
741 /*
742 * Create entries for memory sections that were found
743 * during boot and have been initialized
744 */
745 for (nr = 0; nr <= __highest_present_section_nr;
746 nr += sections_per_block) {
747 ret = add_memory_block(nr);
748 if (ret)
749 panic("%s() failed to add memory block: %d\n", __func__,
750 ret);
751 }
752}
753
754/**
755 * walk_memory_blocks - walk through all present memory blocks overlapped
756 * by the range [start, start + size)
757 *
758 * @start: start address of the memory range
759 * @size: size of the memory range
760 * @arg: argument passed to func
761 * @func: callback for each memory section walked
762 *
763 * This function walks through all present memory blocks overlapped by the
764 * range [start, start + size), calling func on each memory block.
765 *
766 * In case func() returns an error, walking is aborted and the error is
767 * returned.
768 *
769 * Called under device_hotplug_lock.
770 */
771int walk_memory_blocks(unsigned long start, unsigned long size,
772 void *arg, walk_memory_blocks_func_t func)
773{
774 const unsigned long start_block_id = phys_to_block_id(start);
775 const unsigned long end_block_id = phys_to_block_id(start + size - 1);
776 struct memory_block *mem;
777 unsigned long block_id;
778 int ret = 0;
779
780 if (!size)
781 return 0;
782
783 for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
784 mem = find_memory_block_by_id(block_id);
785 if (!mem)
786 continue;
787
788 ret = func(mem, arg);
789 put_device(&mem->dev);
790 if (ret)
791 break;
792 }
793 return ret;
794}
795
796struct for_each_memory_block_cb_data {
797 walk_memory_blocks_func_t func;
798 void *arg;
799};
800
801static int for_each_memory_block_cb(struct device *dev, void *data)
802{
803 struct memory_block *mem = to_memory_block(dev);
804 struct for_each_memory_block_cb_data *cb_data = data;
805
806 return cb_data->func(mem, cb_data->arg);
807}
808
809/**
810 * for_each_memory_block - walk through all present memory blocks
811 *
812 * @arg: argument passed to func
813 * @func: callback for each memory block walked
814 *
815 * This function walks through all present memory blocks, calling func on
816 * each memory block.
817 *
818 * In case func() returns an error, walking is aborted and the error is
819 * returned.
820 */
821int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
822{
823 struct for_each_memory_block_cb_data cb_data = {
824 .func = func,
825 .arg = arg,
826 };
827
828 return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
829 for_each_memory_block_cb);
830}
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 const 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 /* Verify that the altmap is freed */
109 WARN_ON(mem->altmap);
110 kfree(mem);
111}
112
113unsigned long __weak memory_block_size_bytes(void)
114{
115 return MIN_MEMORY_BLOCK_SIZE;
116}
117EXPORT_SYMBOL_GPL(memory_block_size_bytes);
118
119/* Show the memory block ID, relative to the memory block size */
120static ssize_t phys_index_show(struct device *dev,
121 struct device_attribute *attr, char *buf)
122{
123 struct memory_block *mem = to_memory_block(dev);
124
125 return sysfs_emit(buf, "%08lx\n", memory_block_id(mem->start_section_nr));
126}
127
128/*
129 * Legacy interface that we cannot remove. Always indicate "removable"
130 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
131 */
132static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
133 char *buf)
134{
135 return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
136}
137
138/*
139 * online, offline, going offline, etc.
140 */
141static ssize_t state_show(struct device *dev, struct device_attribute *attr,
142 char *buf)
143{
144 struct memory_block *mem = to_memory_block(dev);
145 const char *output;
146
147 /*
148 * We can probably put these states in a nice little array
149 * so that they're not open-coded
150 */
151 switch (mem->state) {
152 case MEM_ONLINE:
153 output = "online";
154 break;
155 case MEM_OFFLINE:
156 output = "offline";
157 break;
158 case MEM_GOING_OFFLINE:
159 output = "going-offline";
160 break;
161 default:
162 WARN_ON(1);
163 return sysfs_emit(buf, "ERROR-UNKNOWN-%ld\n", mem->state);
164 }
165
166 return sysfs_emit(buf, "%s\n", output);
167}
168
169int memory_notify(unsigned long val, void *v)
170{
171 return blocking_notifier_call_chain(&memory_chain, val, v);
172}
173
174#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
175static unsigned long memblk_nr_poison(struct memory_block *mem);
176#else
177static inline unsigned long memblk_nr_poison(struct memory_block *mem)
178{
179 return 0;
180}
181#endif
182
183/*
184 * Must acquire mem_hotplug_lock in write mode.
185 */
186static int memory_block_online(struct memory_block *mem)
187{
188 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
189 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
190 unsigned long nr_vmemmap_pages = 0;
191 struct memory_notify arg;
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 (mem->altmap)
209 nr_vmemmap_pages = mem->altmap->free;
210
211 arg.altmap_start_pfn = start_pfn;
212 arg.altmap_nr_pages = nr_vmemmap_pages;
213 arg.start_pfn = start_pfn + nr_vmemmap_pages;
214 arg.nr_pages = nr_pages - nr_vmemmap_pages;
215 mem_hotplug_begin();
216 ret = memory_notify(MEM_PREPARE_ONLINE, &arg);
217 ret = notifier_to_errno(ret);
218 if (ret)
219 goto out_notifier;
220
221 if (nr_vmemmap_pages) {
222 ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages,
223 zone, mem->altmap->inaccessible);
224 if (ret)
225 goto out;
226 }
227
228 ret = online_pages(start_pfn + nr_vmemmap_pages,
229 nr_pages - nr_vmemmap_pages, zone, mem->group);
230 if (ret) {
231 if (nr_vmemmap_pages)
232 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
233 goto out;
234 }
235
236 /*
237 * Account once onlining succeeded. If the zone was unpopulated, it is
238 * now already properly populated.
239 */
240 if (nr_vmemmap_pages)
241 adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
242 nr_vmemmap_pages);
243
244 mem->zone = zone;
245 mem_hotplug_done();
246 return ret;
247out:
248 memory_notify(MEM_FINISH_OFFLINE, &arg);
249out_notifier:
250 mem_hotplug_done();
251 return ret;
252}
253
254/*
255 * Must acquire mem_hotplug_lock in write mode.
256 */
257static int memory_block_offline(struct memory_block *mem)
258{
259 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
260 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
261 unsigned long nr_vmemmap_pages = 0;
262 struct memory_notify arg;
263 int ret;
264
265 if (!mem->zone)
266 return -EINVAL;
267
268 /*
269 * Unaccount before offlining, such that unpopulated zone and kthreads
270 * can properly be torn down in offline_pages().
271 */
272 if (mem->altmap)
273 nr_vmemmap_pages = mem->altmap->free;
274
275 mem_hotplug_begin();
276 if (nr_vmemmap_pages)
277 adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
278 -nr_vmemmap_pages);
279
280 ret = offline_pages(start_pfn + nr_vmemmap_pages,
281 nr_pages - nr_vmemmap_pages, mem->zone, mem->group);
282 if (ret) {
283 /* offline_pages() failed. Account back. */
284 if (nr_vmemmap_pages)
285 adjust_present_page_count(pfn_to_page(start_pfn),
286 mem->group, nr_vmemmap_pages);
287 goto out;
288 }
289
290 if (nr_vmemmap_pages)
291 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
292
293 mem->zone = NULL;
294 arg.altmap_start_pfn = start_pfn;
295 arg.altmap_nr_pages = nr_vmemmap_pages;
296 arg.start_pfn = start_pfn + nr_vmemmap_pages;
297 arg.nr_pages = nr_pages - nr_vmemmap_pages;
298 memory_notify(MEM_FINISH_OFFLINE, &arg);
299out:
300 mem_hotplug_done();
301 return ret;
302}
303
304/*
305 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
306 * OK to have direct references to sparsemem variables in here.
307 */
308static int
309memory_block_action(struct memory_block *mem, unsigned long action)
310{
311 int ret;
312
313 switch (action) {
314 case MEM_ONLINE:
315 ret = memory_block_online(mem);
316 break;
317 case MEM_OFFLINE:
318 ret = memory_block_offline(mem);
319 break;
320 default:
321 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
322 "%ld\n", __func__, mem->start_section_nr, action, action);
323 ret = -EINVAL;
324 }
325
326 return ret;
327}
328
329static int memory_block_change_state(struct memory_block *mem,
330 unsigned long to_state, unsigned long from_state_req)
331{
332 int ret = 0;
333
334 if (mem->state != from_state_req)
335 return -EINVAL;
336
337 if (to_state == MEM_OFFLINE)
338 mem->state = MEM_GOING_OFFLINE;
339
340 ret = memory_block_action(mem, to_state);
341 mem->state = ret ? from_state_req : to_state;
342
343 return ret;
344}
345
346/* The device lock serializes operations on memory_subsys_[online|offline] */
347static int memory_subsys_online(struct device *dev)
348{
349 struct memory_block *mem = to_memory_block(dev);
350 int ret;
351
352 if (mem->state == MEM_ONLINE)
353 return 0;
354
355 /*
356 * When called via device_online() without configuring the online_type,
357 * we want to default to MMOP_ONLINE.
358 */
359 if (mem->online_type == MMOP_OFFLINE)
360 mem->online_type = MMOP_ONLINE;
361
362 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
363 mem->online_type = MMOP_OFFLINE;
364
365 return ret;
366}
367
368static int memory_subsys_offline(struct device *dev)
369{
370 struct memory_block *mem = to_memory_block(dev);
371
372 if (mem->state == MEM_OFFLINE)
373 return 0;
374
375 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
376}
377
378static ssize_t state_store(struct device *dev, struct device_attribute *attr,
379 const char *buf, size_t count)
380{
381 const int online_type = mhp_online_type_from_str(buf);
382 struct memory_block *mem = to_memory_block(dev);
383 int ret;
384
385 if (online_type < 0)
386 return -EINVAL;
387
388 ret = lock_device_hotplug_sysfs();
389 if (ret)
390 return ret;
391
392 switch (online_type) {
393 case MMOP_ONLINE_KERNEL:
394 case MMOP_ONLINE_MOVABLE:
395 case MMOP_ONLINE:
396 /* mem->online_type is protected by device_hotplug_lock */
397 mem->online_type = online_type;
398 ret = device_online(&mem->dev);
399 break;
400 case MMOP_OFFLINE:
401 ret = device_offline(&mem->dev);
402 break;
403 default:
404 ret = -EINVAL; /* should never happen */
405 }
406
407 unlock_device_hotplug();
408
409 if (ret < 0)
410 return ret;
411 if (ret)
412 return -EINVAL;
413
414 return count;
415}
416
417/*
418 * Legacy interface that we cannot remove: s390x exposes the storage increment
419 * covered by a memory block, allowing for identifying which memory blocks
420 * comprise a storage increment. Since a memory block spans complete
421 * storage increments nowadays, this interface is basically unused. Other
422 * archs never exposed != 0.
423 */
424static ssize_t phys_device_show(struct device *dev,
425 struct device_attribute *attr, char *buf)
426{
427 struct memory_block *mem = to_memory_block(dev);
428 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
429
430 return sysfs_emit(buf, "%d\n",
431 arch_get_memory_phys_device(start_pfn));
432}
433
434#ifdef CONFIG_MEMORY_HOTREMOVE
435static int print_allowed_zone(char *buf, int len, int nid,
436 struct memory_group *group,
437 unsigned long start_pfn, unsigned long nr_pages,
438 int online_type, struct zone *default_zone)
439{
440 struct zone *zone;
441
442 zone = zone_for_pfn_range(online_type, nid, group, start_pfn, nr_pages);
443 if (zone == default_zone)
444 return 0;
445
446 return sysfs_emit_at(buf, len, " %s", zone->name);
447}
448
449static ssize_t valid_zones_show(struct device *dev,
450 struct device_attribute *attr, char *buf)
451{
452 struct memory_block *mem = to_memory_block(dev);
453 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
454 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
455 struct memory_group *group = mem->group;
456 struct zone *default_zone;
457 int nid = mem->nid;
458 int len = 0;
459
460 /*
461 * Check the existing zone. Make sure that we do that only on the
462 * online nodes otherwise the page_zone is not reliable
463 */
464 if (mem->state == MEM_ONLINE) {
465 /*
466 * If !mem->zone, the memory block spans multiple zones and
467 * cannot get offlined.
468 */
469 default_zone = mem->zone;
470 if (!default_zone)
471 return sysfs_emit(buf, "%s\n", "none");
472 len += sysfs_emit_at(buf, len, "%s", default_zone->name);
473 goto out;
474 }
475
476 default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, group,
477 start_pfn, nr_pages);
478
479 len += sysfs_emit_at(buf, len, "%s", default_zone->name);
480 len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
481 MMOP_ONLINE_KERNEL, default_zone);
482 len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
483 MMOP_ONLINE_MOVABLE, default_zone);
484out:
485 len += sysfs_emit_at(buf, len, "\n");
486 return len;
487}
488static DEVICE_ATTR_RO(valid_zones);
489#endif
490
491static DEVICE_ATTR_RO(phys_index);
492static DEVICE_ATTR_RW(state);
493static DEVICE_ATTR_RO(phys_device);
494static DEVICE_ATTR_RO(removable);
495
496/*
497 * Show the memory block size (shared by all memory blocks).
498 */
499static ssize_t block_size_bytes_show(struct device *dev,
500 struct device_attribute *attr, char *buf)
501{
502 return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
503}
504
505static DEVICE_ATTR_RO(block_size_bytes);
506
507/*
508 * Memory auto online policy.
509 */
510
511static ssize_t auto_online_blocks_show(struct device *dev,
512 struct device_attribute *attr, char *buf)
513{
514 return sysfs_emit(buf, "%s\n",
515 online_type_to_str[mhp_default_online_type]);
516}
517
518static ssize_t auto_online_blocks_store(struct device *dev,
519 struct device_attribute *attr,
520 const char *buf, size_t count)
521{
522 const int online_type = mhp_online_type_from_str(buf);
523
524 if (online_type < 0)
525 return -EINVAL;
526
527 mhp_default_online_type = online_type;
528 return count;
529}
530
531static DEVICE_ATTR_RW(auto_online_blocks);
532
533#ifdef CONFIG_CRASH_HOTPLUG
534#include <linux/kexec.h>
535static ssize_t crash_hotplug_show(struct device *dev,
536 struct device_attribute *attr, char *buf)
537{
538 return sysfs_emit(buf, "%d\n", crash_hotplug_memory_support());
539}
540static DEVICE_ATTR_RO(crash_hotplug);
541#endif
542
543/*
544 * Some architectures will have custom drivers to do this, and
545 * will not need to do it from userspace. The fake hot-add code
546 * as well as ppc64 will do all of their discovery in userspace
547 * and will require this interface.
548 */
549#ifdef CONFIG_ARCH_MEMORY_PROBE
550static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
551 const char *buf, size_t count)
552{
553 u64 phys_addr;
554 int nid, ret;
555 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
556
557 ret = kstrtoull(buf, 0, &phys_addr);
558 if (ret)
559 return ret;
560
561 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
562 return -EINVAL;
563
564 ret = lock_device_hotplug_sysfs();
565 if (ret)
566 return ret;
567
568 nid = memory_add_physaddr_to_nid(phys_addr);
569 ret = __add_memory(nid, phys_addr,
570 MIN_MEMORY_BLOCK_SIZE * sections_per_block,
571 MHP_NONE);
572
573 if (ret)
574 goto out;
575
576 ret = count;
577out:
578 unlock_device_hotplug();
579 return ret;
580}
581
582static DEVICE_ATTR_WO(probe);
583#endif
584
585#ifdef CONFIG_MEMORY_FAILURE
586/*
587 * Support for offlining pages of memory
588 */
589
590/* Soft offline a page */
591static ssize_t soft_offline_page_store(struct device *dev,
592 struct device_attribute *attr,
593 const char *buf, size_t count)
594{
595 int ret;
596 u64 pfn;
597 if (!capable(CAP_SYS_ADMIN))
598 return -EPERM;
599 if (kstrtoull(buf, 0, &pfn) < 0)
600 return -EINVAL;
601 pfn >>= PAGE_SHIFT;
602 ret = soft_offline_page(pfn, 0);
603 return ret == 0 ? count : ret;
604}
605
606/* Forcibly offline a page, including killing processes. */
607static ssize_t hard_offline_page_store(struct device *dev,
608 struct device_attribute *attr,
609 const char *buf, size_t count)
610{
611 int ret;
612 u64 pfn;
613 if (!capable(CAP_SYS_ADMIN))
614 return -EPERM;
615 if (kstrtoull(buf, 0, &pfn) < 0)
616 return -EINVAL;
617 pfn >>= PAGE_SHIFT;
618 ret = memory_failure(pfn, MF_SW_SIMULATED);
619 if (ret == -EOPNOTSUPP)
620 ret = 0;
621 return ret ? ret : count;
622}
623
624static DEVICE_ATTR_WO(soft_offline_page);
625static DEVICE_ATTR_WO(hard_offline_page);
626#endif
627
628/* See phys_device_show(). */
629int __weak arch_get_memory_phys_device(unsigned long start_pfn)
630{
631 return 0;
632}
633
634/*
635 * A reference for the returned memory block device is acquired.
636 *
637 * Called under device_hotplug_lock.
638 */
639static struct memory_block *find_memory_block_by_id(unsigned long block_id)
640{
641 struct memory_block *mem;
642
643 mem = xa_load(&memory_blocks, block_id);
644 if (mem)
645 get_device(&mem->dev);
646 return mem;
647}
648
649/*
650 * Called under device_hotplug_lock.
651 */
652struct memory_block *find_memory_block(unsigned long section_nr)
653{
654 unsigned long block_id = memory_block_id(section_nr);
655
656 return find_memory_block_by_id(block_id);
657}
658
659static struct attribute *memory_memblk_attrs[] = {
660 &dev_attr_phys_index.attr,
661 &dev_attr_state.attr,
662 &dev_attr_phys_device.attr,
663 &dev_attr_removable.attr,
664#ifdef CONFIG_MEMORY_HOTREMOVE
665 &dev_attr_valid_zones.attr,
666#endif
667 NULL
668};
669
670static const struct attribute_group memory_memblk_attr_group = {
671 .attrs = memory_memblk_attrs,
672};
673
674static const struct attribute_group *memory_memblk_attr_groups[] = {
675 &memory_memblk_attr_group,
676 NULL,
677};
678
679static int __add_memory_block(struct memory_block *memory)
680{
681 int ret;
682
683 memory->dev.bus = &memory_subsys;
684 memory->dev.id = memory->start_section_nr / sections_per_block;
685 memory->dev.release = memory_block_release;
686 memory->dev.groups = memory_memblk_attr_groups;
687 memory->dev.offline = memory->state == MEM_OFFLINE;
688
689 ret = device_register(&memory->dev);
690 if (ret) {
691 put_device(&memory->dev);
692 return ret;
693 }
694 ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
695 GFP_KERNEL));
696 if (ret)
697 device_unregister(&memory->dev);
698
699 return ret;
700}
701
702static struct zone *early_node_zone_for_memory_block(struct memory_block *mem,
703 int nid)
704{
705 const unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
706 const unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
707 struct zone *zone, *matching_zone = NULL;
708 pg_data_t *pgdat = NODE_DATA(nid);
709 int i;
710
711 /*
712 * This logic only works for early memory, when the applicable zones
713 * already span the memory block. We don't expect overlapping zones on
714 * a single node for early memory. So if we're told that some PFNs
715 * of a node fall into this memory block, we can assume that all node
716 * zones that intersect with the memory block are actually applicable.
717 * No need to look at the memmap.
718 */
719 for (i = 0; i < MAX_NR_ZONES; i++) {
720 zone = pgdat->node_zones + i;
721 if (!populated_zone(zone))
722 continue;
723 if (!zone_intersects(zone, start_pfn, nr_pages))
724 continue;
725 if (!matching_zone) {
726 matching_zone = zone;
727 continue;
728 }
729 /* Spans multiple zones ... */
730 matching_zone = NULL;
731 break;
732 }
733 return matching_zone;
734}
735
736#ifdef CONFIG_NUMA
737/**
738 * memory_block_add_nid() - Indicate that system RAM falling into this memory
739 * block device (partially) belongs to the given node.
740 * @mem: The memory block device.
741 * @nid: The node id.
742 * @context: The memory initialization context.
743 *
744 * Indicate that system RAM falling into this memory block (partially) belongs
745 * to the given node. If the context indicates ("early") that we are adding the
746 * node during node device subsystem initialization, this will also properly
747 * set/adjust mem->zone based on the zone ranges of the given node.
748 */
749void memory_block_add_nid(struct memory_block *mem, int nid,
750 enum meminit_context context)
751{
752 if (context == MEMINIT_EARLY && mem->nid != nid) {
753 /*
754 * For early memory we have to determine the zone when setting
755 * the node id and handle multiple nodes spanning a single
756 * memory block by indicate via zone == NULL that we're not
757 * dealing with a single zone. So if we're setting the node id
758 * the first time, determine if there is a single zone. If we're
759 * setting the node id a second time to a different node,
760 * invalidate the single detected zone.
761 */
762 if (mem->nid == NUMA_NO_NODE)
763 mem->zone = early_node_zone_for_memory_block(mem, nid);
764 else
765 mem->zone = NULL;
766 }
767
768 /*
769 * If this memory block spans multiple nodes, we only indicate
770 * the last processed node. If we span multiple nodes (not applicable
771 * to hotplugged memory), zone == NULL will prohibit memory offlining
772 * and consequently unplug.
773 */
774 mem->nid = nid;
775}
776#endif
777
778static int add_memory_block(unsigned long block_id, unsigned long state,
779 struct vmem_altmap *altmap,
780 struct memory_group *group)
781{
782 struct memory_block *mem;
783 int ret = 0;
784
785 mem = find_memory_block_by_id(block_id);
786 if (mem) {
787 put_device(&mem->dev);
788 return -EEXIST;
789 }
790 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
791 if (!mem)
792 return -ENOMEM;
793
794 mem->start_section_nr = block_id * sections_per_block;
795 mem->state = state;
796 mem->nid = NUMA_NO_NODE;
797 mem->altmap = altmap;
798 INIT_LIST_HEAD(&mem->group_next);
799
800#ifndef CONFIG_NUMA
801 if (state == MEM_ONLINE)
802 /*
803 * MEM_ONLINE at this point implies early memory. With NUMA,
804 * we'll determine the zone when setting the node id via
805 * memory_block_add_nid(). Memory hotplug updated the zone
806 * manually when memory onlining/offlining succeeds.
807 */
808 mem->zone = early_node_zone_for_memory_block(mem, NUMA_NO_NODE);
809#endif /* CONFIG_NUMA */
810
811 ret = __add_memory_block(mem);
812 if (ret)
813 return ret;
814
815 if (group) {
816 mem->group = group;
817 list_add(&mem->group_next, &group->memory_blocks);
818 }
819
820 return 0;
821}
822
823static int __init add_boot_memory_block(unsigned long base_section_nr)
824{
825 int section_count = 0;
826 unsigned long nr;
827
828 for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
829 nr++)
830 if (present_section_nr(nr))
831 section_count++;
832
833 if (section_count == 0)
834 return 0;
835 return add_memory_block(memory_block_id(base_section_nr),
836 MEM_ONLINE, NULL, NULL);
837}
838
839static int add_hotplug_memory_block(unsigned long block_id,
840 struct vmem_altmap *altmap,
841 struct memory_group *group)
842{
843 return add_memory_block(block_id, MEM_OFFLINE, altmap, group);
844}
845
846static void remove_memory_block(struct memory_block *memory)
847{
848 if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
849 return;
850
851 WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
852
853 if (memory->group) {
854 list_del(&memory->group_next);
855 memory->group = NULL;
856 }
857
858 /* drop the ref. we got via find_memory_block() */
859 put_device(&memory->dev);
860 device_unregister(&memory->dev);
861}
862
863/*
864 * Create memory block devices for the given memory area. Start and size
865 * have to be aligned to memory block granularity. Memory block devices
866 * will be initialized as offline.
867 *
868 * Called under device_hotplug_lock.
869 */
870int create_memory_block_devices(unsigned long start, unsigned long size,
871 struct vmem_altmap *altmap,
872 struct memory_group *group)
873{
874 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
875 unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
876 struct memory_block *mem;
877 unsigned long block_id;
878 int ret = 0;
879
880 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
881 !IS_ALIGNED(size, memory_block_size_bytes())))
882 return -EINVAL;
883
884 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
885 ret = add_hotplug_memory_block(block_id, altmap, group);
886 if (ret)
887 break;
888 }
889 if (ret) {
890 end_block_id = block_id;
891 for (block_id = start_block_id; block_id != end_block_id;
892 block_id++) {
893 mem = find_memory_block_by_id(block_id);
894 if (WARN_ON_ONCE(!mem))
895 continue;
896 remove_memory_block(mem);
897 }
898 }
899 return ret;
900}
901
902/*
903 * Remove memory block devices for the given memory area. Start and size
904 * have to be aligned to memory block granularity. Memory block devices
905 * have to be offline.
906 *
907 * Called under device_hotplug_lock.
908 */
909void remove_memory_block_devices(unsigned long start, unsigned long size)
910{
911 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
912 const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
913 struct memory_block *mem;
914 unsigned long block_id;
915
916 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
917 !IS_ALIGNED(size, memory_block_size_bytes())))
918 return;
919
920 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
921 mem = find_memory_block_by_id(block_id);
922 if (WARN_ON_ONCE(!mem))
923 continue;
924 num_poisoned_pages_sub(-1UL, memblk_nr_poison(mem));
925 unregister_memory_block_under_nodes(mem);
926 remove_memory_block(mem);
927 }
928}
929
930static struct attribute *memory_root_attrs[] = {
931#ifdef CONFIG_ARCH_MEMORY_PROBE
932 &dev_attr_probe.attr,
933#endif
934
935#ifdef CONFIG_MEMORY_FAILURE
936 &dev_attr_soft_offline_page.attr,
937 &dev_attr_hard_offline_page.attr,
938#endif
939
940 &dev_attr_block_size_bytes.attr,
941 &dev_attr_auto_online_blocks.attr,
942#ifdef CONFIG_CRASH_HOTPLUG
943 &dev_attr_crash_hotplug.attr,
944#endif
945 NULL
946};
947
948static const struct attribute_group memory_root_attr_group = {
949 .attrs = memory_root_attrs,
950};
951
952static const struct attribute_group *memory_root_attr_groups[] = {
953 &memory_root_attr_group,
954 NULL,
955};
956
957/*
958 * Initialize the sysfs support for memory devices. At the time this function
959 * is called, we cannot have concurrent creation/deletion of memory block
960 * devices, the device_hotplug_lock is not needed.
961 */
962void __init memory_dev_init(void)
963{
964 int ret;
965 unsigned long block_sz, nr;
966
967 /* Validate the configured memory block size */
968 block_sz = memory_block_size_bytes();
969 if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
970 panic("Memory block size not suitable: 0x%lx\n", block_sz);
971 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
972
973 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
974 if (ret)
975 panic("%s() failed to register subsystem: %d\n", __func__, ret);
976
977 /*
978 * Create entries for memory sections that were found
979 * during boot and have been initialized
980 */
981 for (nr = 0; nr <= __highest_present_section_nr;
982 nr += sections_per_block) {
983 ret = add_boot_memory_block(nr);
984 if (ret)
985 panic("%s() failed to add memory block: %d\n", __func__,
986 ret);
987 }
988}
989
990/**
991 * walk_memory_blocks - walk through all present memory blocks overlapped
992 * by the range [start, start + size)
993 *
994 * @start: start address of the memory range
995 * @size: size of the memory range
996 * @arg: argument passed to func
997 * @func: callback for each memory section walked
998 *
999 * This function walks through all present memory blocks overlapped by the
1000 * range [start, start + size), calling func on each memory block.
1001 *
1002 * In case func() returns an error, walking is aborted and the error is
1003 * returned.
1004 *
1005 * Called under device_hotplug_lock.
1006 */
1007int walk_memory_blocks(unsigned long start, unsigned long size,
1008 void *arg, walk_memory_blocks_func_t func)
1009{
1010 const unsigned long start_block_id = phys_to_block_id(start);
1011 const unsigned long end_block_id = phys_to_block_id(start + size - 1);
1012 struct memory_block *mem;
1013 unsigned long block_id;
1014 int ret = 0;
1015
1016 if (!size)
1017 return 0;
1018
1019 for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
1020 mem = find_memory_block_by_id(block_id);
1021 if (!mem)
1022 continue;
1023
1024 ret = func(mem, arg);
1025 put_device(&mem->dev);
1026 if (ret)
1027 break;
1028 }
1029 return ret;
1030}
1031
1032struct for_each_memory_block_cb_data {
1033 walk_memory_blocks_func_t func;
1034 void *arg;
1035};
1036
1037static int for_each_memory_block_cb(struct device *dev, void *data)
1038{
1039 struct memory_block *mem = to_memory_block(dev);
1040 struct for_each_memory_block_cb_data *cb_data = data;
1041
1042 return cb_data->func(mem, cb_data->arg);
1043}
1044
1045/**
1046 * for_each_memory_block - walk through all present memory blocks
1047 *
1048 * @arg: argument passed to func
1049 * @func: callback for each memory block walked
1050 *
1051 * This function walks through all present memory blocks, calling func on
1052 * each memory block.
1053 *
1054 * In case func() returns an error, walking is aborted and the error is
1055 * returned.
1056 */
1057int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
1058{
1059 struct for_each_memory_block_cb_data cb_data = {
1060 .func = func,
1061 .arg = arg,
1062 };
1063
1064 return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
1065 for_each_memory_block_cb);
1066}
1067
1068/*
1069 * This is an internal helper to unify allocation and initialization of
1070 * memory groups. Note that the passed memory group will be copied to a
1071 * dynamically allocated memory group. After this call, the passed
1072 * memory group should no longer be used.
1073 */
1074static int memory_group_register(struct memory_group group)
1075{
1076 struct memory_group *new_group;
1077 uint32_t mgid;
1078 int ret;
1079
1080 if (!node_possible(group.nid))
1081 return -EINVAL;
1082
1083 new_group = kzalloc(sizeof(group), GFP_KERNEL);
1084 if (!new_group)
1085 return -ENOMEM;
1086 *new_group = group;
1087 INIT_LIST_HEAD(&new_group->memory_blocks);
1088
1089 ret = xa_alloc(&memory_groups, &mgid, new_group, xa_limit_31b,
1090 GFP_KERNEL);
1091 if (ret) {
1092 kfree(new_group);
1093 return ret;
1094 } else if (group.is_dynamic) {
1095 xa_set_mark(&memory_groups, mgid, MEMORY_GROUP_MARK_DYNAMIC);
1096 }
1097 return mgid;
1098}
1099
1100/**
1101 * memory_group_register_static() - Register a static memory group.
1102 * @nid: The node id.
1103 * @max_pages: The maximum number of pages we'll have in this static memory
1104 * group.
1105 *
1106 * Register a new static memory group and return the memory group id.
1107 * All memory in the group belongs to a single unit, such as a DIMM. All
1108 * memory belonging to a static memory group is added in one go to be removed
1109 * in one go -- it's static.
1110 *
1111 * Returns an error if out of memory, if the node id is invalid, if no new
1112 * memory groups can be registered, or if max_pages is invalid (0). Otherwise,
1113 * returns the new memory group id.
1114 */
1115int memory_group_register_static(int nid, unsigned long max_pages)
1116{
1117 struct memory_group group = {
1118 .nid = nid,
1119 .s = {
1120 .max_pages = max_pages,
1121 },
1122 };
1123
1124 if (!max_pages)
1125 return -EINVAL;
1126 return memory_group_register(group);
1127}
1128EXPORT_SYMBOL_GPL(memory_group_register_static);
1129
1130/**
1131 * memory_group_register_dynamic() - Register a dynamic memory group.
1132 * @nid: The node id.
1133 * @unit_pages: Unit in pages in which is memory added/removed in this dynamic
1134 * memory group.
1135 *
1136 * Register a new dynamic memory group and return the memory group id.
1137 * Memory within a dynamic memory group is added/removed dynamically
1138 * in unit_pages.
1139 *
1140 * Returns an error if out of memory, if the node id is invalid, if no new
1141 * memory groups can be registered, or if unit_pages is invalid (0, not a
1142 * power of two, smaller than a single memory block). Otherwise, returns the
1143 * new memory group id.
1144 */
1145int memory_group_register_dynamic(int nid, unsigned long unit_pages)
1146{
1147 struct memory_group group = {
1148 .nid = nid,
1149 .is_dynamic = true,
1150 .d = {
1151 .unit_pages = unit_pages,
1152 },
1153 };
1154
1155 if (!unit_pages || !is_power_of_2(unit_pages) ||
1156 unit_pages < PHYS_PFN(memory_block_size_bytes()))
1157 return -EINVAL;
1158 return memory_group_register(group);
1159}
1160EXPORT_SYMBOL_GPL(memory_group_register_dynamic);
1161
1162/**
1163 * memory_group_unregister() - Unregister a memory group.
1164 * @mgid: the memory group id
1165 *
1166 * Unregister a memory group. If any memory block still belongs to this
1167 * memory group, unregistering will fail.
1168 *
1169 * Returns -EINVAL if the memory group id is invalid, returns -EBUSY if some
1170 * memory blocks still belong to this memory group and returns 0 if
1171 * unregistering succeeded.
1172 */
1173int memory_group_unregister(int mgid)
1174{
1175 struct memory_group *group;
1176
1177 if (mgid < 0)
1178 return -EINVAL;
1179
1180 group = xa_load(&memory_groups, mgid);
1181 if (!group)
1182 return -EINVAL;
1183 if (!list_empty(&group->memory_blocks))
1184 return -EBUSY;
1185 xa_erase(&memory_groups, mgid);
1186 kfree(group);
1187 return 0;
1188}
1189EXPORT_SYMBOL_GPL(memory_group_unregister);
1190
1191/*
1192 * This is an internal helper only to be used in core memory hotplug code to
1193 * lookup a memory group. We don't care about locking, as we don't expect a
1194 * memory group to get unregistered while adding memory to it -- because
1195 * the group and the memory is managed by the same driver.
1196 */
1197struct memory_group *memory_group_find_by_id(int mgid)
1198{
1199 return xa_load(&memory_groups, mgid);
1200}
1201
1202/*
1203 * This is an internal helper only to be used in core memory hotplug code to
1204 * walk all dynamic memory groups excluding a given memory group, either
1205 * belonging to a specific node, or belonging to any node.
1206 */
1207int walk_dynamic_memory_groups(int nid, walk_memory_groups_func_t func,
1208 struct memory_group *excluded, void *arg)
1209{
1210 struct memory_group *group;
1211 unsigned long index;
1212 int ret = 0;
1213
1214 xa_for_each_marked(&memory_groups, index, group,
1215 MEMORY_GROUP_MARK_DYNAMIC) {
1216 if (group == excluded)
1217 continue;
1218#ifdef CONFIG_NUMA
1219 if (nid != NUMA_NO_NODE && group->nid != nid)
1220 continue;
1221#endif /* CONFIG_NUMA */
1222 ret = func(group, arg);
1223 if (ret)
1224 break;
1225 }
1226 return ret;
1227}
1228
1229#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
1230void memblk_nr_poison_inc(unsigned long pfn)
1231{
1232 const unsigned long block_id = pfn_to_block_id(pfn);
1233 struct memory_block *mem = find_memory_block_by_id(block_id);
1234
1235 if (mem)
1236 atomic_long_inc(&mem->nr_hwpoison);
1237}
1238
1239void memblk_nr_poison_sub(unsigned long pfn, long i)
1240{
1241 const unsigned long block_id = pfn_to_block_id(pfn);
1242 struct memory_block *mem = find_memory_block_by_id(block_id);
1243
1244 if (mem)
1245 atomic_long_sub(i, &mem->nr_hwpoison);
1246}
1247
1248static unsigned long memblk_nr_poison(struct memory_block *mem)
1249{
1250 return atomic_long_read(&mem->nr_hwpoison);
1251}
1252#endif