Linux Audio

Check our new training course

Loading...
v5.9
  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}
v5.14.15
  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
 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
123	return sysfs_emit(buf, "%08lx\n", phys_index);
124}
125
126/*
127 * Legacy interface that we cannot remove. Always indicate "removable"
128 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
129 */
130static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
131			      char *buf)
132{
133	return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
134}
135
136/*
137 * online, offline, going offline, etc.
138 */
139static ssize_t state_show(struct device *dev, struct device_attribute *attr,
140			  char *buf)
141{
142	struct memory_block *mem = to_memory_block(dev);
143	const char *output;
144
145	/*
146	 * We can probably put these states in a nice little array
147	 * so that they're not open-coded
148	 */
149	switch (mem->state) {
150	case MEM_ONLINE:
151		output = "online";
152		break;
153	case MEM_OFFLINE:
154		output = "offline";
155		break;
156	case MEM_GOING_OFFLINE:
157		output = "going-offline";
158		break;
159	default:
 
 
160		WARN_ON(1);
161		return sysfs_emit(buf, "ERROR-UNKNOWN-%ld\n", mem->state);
162	}
163
164	return sysfs_emit(buf, "%s\n", output);
165}
166
167int memory_notify(unsigned long val, void *v)
168{
169	return blocking_notifier_call_chain(&memory_chain, val, v);
170}
171
172static int memory_block_online(struct memory_block *mem)
173{
174	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
175	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
176	unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
177	struct zone *zone;
178	int ret;
179
180	zone = zone_for_pfn_range(mem->online_type, mem->nid, start_pfn, nr_pages);
181
182	/*
183	 * Although vmemmap pages have a different lifecycle than the pages
184	 * they describe (they remain until the memory is unplugged), doing
185	 * their initialization and accounting at memory onlining/offlining
186	 * stage helps to keep accounting easier to follow - e.g vmemmaps
187	 * belong to the same zone as the memory they backed.
188	 */
189	if (nr_vmemmap_pages) {
190		ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
191		if (ret)
192			return ret;
193	}
194
195	ret = online_pages(start_pfn + nr_vmemmap_pages,
196			   nr_pages - nr_vmemmap_pages, zone);
197	if (ret) {
198		if (nr_vmemmap_pages)
199			mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
200		return ret;
201	}
202
203	/*
204	 * Account once onlining succeeded. If the zone was unpopulated, it is
205	 * now already properly populated.
206	 */
207	if (nr_vmemmap_pages)
208		adjust_present_page_count(zone, nr_vmemmap_pages);
209
210	return ret;
211}
212
213static int memory_block_offline(struct memory_block *mem)
214{
215	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
216	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
217	unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
218	struct zone *zone;
219	int ret;
220
221	/*
222	 * Unaccount before offlining, such that unpopulated zone and kthreads
223	 * can properly be torn down in offline_pages().
224	 */
225	if (nr_vmemmap_pages) {
226		zone = page_zone(pfn_to_page(start_pfn));
227		adjust_present_page_count(zone, -nr_vmemmap_pages);
228	}
229
230	ret = offline_pages(start_pfn + nr_vmemmap_pages,
231			    nr_pages - nr_vmemmap_pages);
232	if (ret) {
233		/* offline_pages() failed. Account back. */
234		if (nr_vmemmap_pages)
235			adjust_present_page_count(zone, nr_vmemmap_pages);
236		return ret;
237	}
238
239	if (nr_vmemmap_pages)
240		mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
241
242	return ret;
243}
244
245/*
246 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
247 * OK to have direct references to sparsemem variables in here.
248 */
249static int
250memory_block_action(struct memory_block *mem, unsigned long action)
 
251{
 
 
252	int ret;
253
 
 
254	switch (action) {
255	case MEM_ONLINE:
256		ret = memory_block_online(mem);
257		break;
258	case MEM_OFFLINE:
259		ret = memory_block_offline(mem);
260		break;
261	default:
262		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
263		     "%ld\n", __func__, mem->start_section_nr, action, action);
264		ret = -EINVAL;
265	}
266
267	return ret;
268}
269
270static int memory_block_change_state(struct memory_block *mem,
271		unsigned long to_state, unsigned long from_state_req)
272{
273	int ret = 0;
274
275	if (mem->state != from_state_req)
276		return -EINVAL;
277
278	if (to_state == MEM_OFFLINE)
279		mem->state = MEM_GOING_OFFLINE;
280
281	ret = memory_block_action(mem, to_state);
 
 
282	mem->state = ret ? from_state_req : to_state;
283
284	return ret;
285}
286
287/* The device lock serializes operations on memory_subsys_[online|offline] */
288static int memory_subsys_online(struct device *dev)
289{
290	struct memory_block *mem = to_memory_block(dev);
291	int ret;
292
293	if (mem->state == MEM_ONLINE)
294		return 0;
295
296	/*
297	 * When called via device_online() without configuring the online_type,
298	 * we want to default to MMOP_ONLINE.
299	 */
300	if (mem->online_type == MMOP_OFFLINE)
301		mem->online_type = MMOP_ONLINE;
302
303	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
304	mem->online_type = MMOP_OFFLINE;
305
306	return ret;
307}
308
309static int memory_subsys_offline(struct device *dev)
310{
311	struct memory_block *mem = to_memory_block(dev);
312
313	if (mem->state == MEM_OFFLINE)
314		return 0;
315
316	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
317}
318
319static ssize_t state_store(struct device *dev, struct device_attribute *attr,
320			   const char *buf, size_t count)
321{
322	const int online_type = mhp_online_type_from_str(buf);
323	struct memory_block *mem = to_memory_block(dev);
324	int ret;
325
326	if (online_type < 0)
327		return -EINVAL;
328
329	ret = lock_device_hotplug_sysfs();
330	if (ret)
331		return ret;
332
333	switch (online_type) {
334	case MMOP_ONLINE_KERNEL:
335	case MMOP_ONLINE_MOVABLE:
336	case MMOP_ONLINE:
337		/* mem->online_type is protected by device_hotplug_lock */
338		mem->online_type = online_type;
339		ret = device_online(&mem->dev);
340		break;
341	case MMOP_OFFLINE:
342		ret = device_offline(&mem->dev);
343		break;
344	default:
345		ret = -EINVAL; /* should never happen */
346	}
347
348	unlock_device_hotplug();
349
350	if (ret < 0)
351		return ret;
352	if (ret)
353		return -EINVAL;
354
355	return count;
356}
357
358/*
359 * Legacy interface that we cannot remove: s390x exposes the storage increment
360 * covered by a memory block, allowing for identifying which memory blocks
361 * comprise a storage increment. Since a memory block spans complete
362 * storage increments nowadays, this interface is basically unused. Other
363 * archs never exposed != 0.
 
 
364 */
365static ssize_t phys_device_show(struct device *dev,
366				struct device_attribute *attr, char *buf)
367{
368	struct memory_block *mem = to_memory_block(dev);
369	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
370
371	return sysfs_emit(buf, "%d\n",
372			  arch_get_memory_phys_device(start_pfn));
373}
374
375#ifdef CONFIG_MEMORY_HOTREMOVE
376static int print_allowed_zone(char *buf, int len, int nid,
377			      unsigned long start_pfn, unsigned long nr_pages,
378			      int online_type, struct zone *default_zone)
379{
380	struct zone *zone;
381
382	zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
383	if (zone == default_zone)
384		return 0;
385
386	return sysfs_emit_at(buf, len, " %s", zone->name);
387}
388
389static ssize_t valid_zones_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	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
395	struct zone *default_zone;
396	int len = 0;
397	int nid;
398
399	/*
400	 * Check the existing zone. Make sure that we do that only on the
401	 * online nodes otherwise the page_zone is not reliable
402	 */
403	if (mem->state == MEM_ONLINE) {
404		/*
405		 * The block contains more than one zone can not be offlined.
406		 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
407		 */
408		default_zone = test_pages_in_a_zone(start_pfn,
409						    start_pfn + nr_pages);
410		if (!default_zone)
411			return sysfs_emit(buf, "%s\n", "none");
412		len += sysfs_emit_at(buf, len, "%s", default_zone->name);
413		goto out;
414	}
415
416	nid = mem->nid;
417	default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, start_pfn,
418					  nr_pages);
 
419
420	len += sysfs_emit_at(buf, len, "%s", default_zone->name);
421	len += print_allowed_zone(buf, len, nid, start_pfn, nr_pages,
422				  MMOP_ONLINE_KERNEL, default_zone);
423	len += print_allowed_zone(buf, len, nid, start_pfn, nr_pages,
424				  MMOP_ONLINE_MOVABLE, default_zone);
425out:
426	len += sysfs_emit_at(buf, len, "\n");
427	return len;
 
428}
429static DEVICE_ATTR_RO(valid_zones);
430#endif
431
432static DEVICE_ATTR_RO(phys_index);
433static DEVICE_ATTR_RW(state);
434static DEVICE_ATTR_RO(phys_device);
435static DEVICE_ATTR_RO(removable);
436
437/*
438 * Show the memory block size (shared by all memory blocks).
439 */
440static ssize_t block_size_bytes_show(struct device *dev,
441				     struct device_attribute *attr, char *buf)
442{
443	return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
444}
445
446static DEVICE_ATTR_RO(block_size_bytes);
447
448/*
449 * Memory auto online policy.
450 */
451
452static ssize_t auto_online_blocks_show(struct device *dev,
453				       struct device_attribute *attr, char *buf)
454{
455	return sysfs_emit(buf, "%s\n",
456			  online_type_to_str[mhp_default_online_type]);
457}
458
459static ssize_t auto_online_blocks_store(struct device *dev,
460					struct device_attribute *attr,
461					const char *buf, size_t count)
462{
463	const int online_type = mhp_online_type_from_str(buf);
464
465	if (online_type < 0)
466		return -EINVAL;
467
468	mhp_default_online_type = online_type;
469	return count;
470}
471
472static DEVICE_ATTR_RW(auto_online_blocks);
473
474/*
475 * Some architectures will have custom drivers to do this, and
476 * will not need to do it from userspace.  The fake hot-add code
477 * as well as ppc64 will do all of their discovery in userspace
478 * and will require this interface.
479 */
480#ifdef CONFIG_ARCH_MEMORY_PROBE
481static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
482			   const char *buf, size_t count)
483{
484	u64 phys_addr;
485	int nid, ret;
486	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
487
488	ret = kstrtoull(buf, 0, &phys_addr);
489	if (ret)
490		return ret;
491
492	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
493		return -EINVAL;
494
495	ret = lock_device_hotplug_sysfs();
496	if (ret)
497		return ret;
498
499	nid = memory_add_physaddr_to_nid(phys_addr);
500	ret = __add_memory(nid, phys_addr,
501			   MIN_MEMORY_BLOCK_SIZE * sections_per_block,
502			   MHP_NONE);
503
504	if (ret)
505		goto out;
506
507	ret = count;
508out:
509	unlock_device_hotplug();
510	return ret;
511}
512
513static DEVICE_ATTR_WO(probe);
514#endif
515
516#ifdef CONFIG_MEMORY_FAILURE
517/*
518 * Support for offlining pages of memory
519 */
520
521/* Soft offline a page */
522static ssize_t soft_offline_page_store(struct device *dev,
523				       struct device_attribute *attr,
524				       const char *buf, size_t count)
525{
526	int ret;
527	u64 pfn;
528	if (!capable(CAP_SYS_ADMIN))
529		return -EPERM;
530	if (kstrtoull(buf, 0, &pfn) < 0)
531		return -EINVAL;
532	pfn >>= PAGE_SHIFT;
533	ret = soft_offline_page(pfn, 0);
534	return ret == 0 ? count : ret;
535}
536
537/* Forcibly offline a page, including killing processes. */
538static ssize_t hard_offline_page_store(struct device *dev,
539				       struct device_attribute *attr,
540				       const char *buf, size_t count)
541{
542	int ret;
543	u64 pfn;
544	if (!capable(CAP_SYS_ADMIN))
545		return -EPERM;
546	if (kstrtoull(buf, 0, &pfn) < 0)
547		return -EINVAL;
548	pfn >>= PAGE_SHIFT;
549	ret = memory_failure(pfn, 0);
550	return ret ? ret : count;
551}
552
553static DEVICE_ATTR_WO(soft_offline_page);
554static DEVICE_ATTR_WO(hard_offline_page);
555#endif
556
557/* See phys_device_show(). */
 
 
 
 
558int __weak arch_get_memory_phys_device(unsigned long start_pfn)
559{
560	return 0;
561}
562
563/*
564 * A reference for the returned memory block device is acquired.
565 *
566 * Called under device_hotplug_lock.
567 */
568static struct memory_block *find_memory_block_by_id(unsigned long block_id)
569{
570	struct memory_block *mem;
571
572	mem = xa_load(&memory_blocks, block_id);
573	if (mem)
574		get_device(&mem->dev);
575	return mem;
576}
577
578/*
579 * Called under device_hotplug_lock.
580 */
581struct memory_block *find_memory_block(struct mem_section *section)
582{
583	unsigned long block_id = memory_block_id(__section_nr(section));
584
585	return find_memory_block_by_id(block_id);
586}
587
588static struct attribute *memory_memblk_attrs[] = {
589	&dev_attr_phys_index.attr,
590	&dev_attr_state.attr,
591	&dev_attr_phys_device.attr,
592	&dev_attr_removable.attr,
593#ifdef CONFIG_MEMORY_HOTREMOVE
594	&dev_attr_valid_zones.attr,
595#endif
596	NULL
597};
598
599static const struct attribute_group memory_memblk_attr_group = {
600	.attrs = memory_memblk_attrs,
601};
602
603static const struct attribute_group *memory_memblk_attr_groups[] = {
604	&memory_memblk_attr_group,
605	NULL,
606};
607
608/*
609 * register_memory - Setup a sysfs device for a memory block
610 */
611static
612int register_memory(struct memory_block *memory)
613{
614	int ret;
615
616	memory->dev.bus = &memory_subsys;
617	memory->dev.id = memory->start_section_nr / sections_per_block;
618	memory->dev.release = memory_block_release;
619	memory->dev.groups = memory_memblk_attr_groups;
620	memory->dev.offline = memory->state == MEM_OFFLINE;
621
622	ret = device_register(&memory->dev);
623	if (ret) {
624		put_device(&memory->dev);
625		return ret;
626	}
627	ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
628			      GFP_KERNEL));
629	if (ret) {
630		put_device(&memory->dev);
631		device_unregister(&memory->dev);
632	}
633	return ret;
634}
635
636static int init_memory_block(unsigned long block_id, unsigned long state,
637			     unsigned long nr_vmemmap_pages)
638{
639	struct memory_block *mem;
 
640	int ret = 0;
641
642	mem = find_memory_block_by_id(block_id);
643	if (mem) {
644		put_device(&mem->dev);
645		return -EEXIST;
646	}
647	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
648	if (!mem)
649		return -ENOMEM;
650
651	mem->start_section_nr = block_id * sections_per_block;
652	mem->state = state;
 
 
653	mem->nid = NUMA_NO_NODE;
654	mem->nr_vmemmap_pages = nr_vmemmap_pages;
655
656	ret = register_memory(mem);
657
658	return ret;
659}
660
661static int add_memory_block(unsigned long base_section_nr)
662{
663	int section_count = 0;
664	unsigned long nr;
665
666	for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
667	     nr++)
668		if (present_section_nr(nr))
669			section_count++;
670
671	if (section_count == 0)
672		return 0;
673	return init_memory_block(memory_block_id(base_section_nr),
674				 MEM_ONLINE, 0);
675}
676
677static void unregister_memory(struct memory_block *memory)
678{
679	if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
680		return;
681
682	WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
683
684	/* drop the ref. we got via find_memory_block() */
685	put_device(&memory->dev);
686	device_unregister(&memory->dev);
687}
688
689/*
690 * Create memory block devices for the given memory area. Start and size
691 * have to be aligned to memory block granularity. Memory block devices
692 * will be initialized as offline.
693 *
694 * Called under device_hotplug_lock.
695 */
696int create_memory_block_devices(unsigned long start, unsigned long size,
697				unsigned long vmemmap_pages)
698{
699	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
700	unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
701	struct memory_block *mem;
702	unsigned long block_id;
703	int ret = 0;
704
705	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
706			 !IS_ALIGNED(size, memory_block_size_bytes())))
707		return -EINVAL;
708
709	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
710		ret = init_memory_block(block_id, MEM_OFFLINE, vmemmap_pages);
711		if (ret)
712			break;
713	}
714	if (ret) {
715		end_block_id = block_id;
716		for (block_id = start_block_id; block_id != end_block_id;
717		     block_id++) {
718			mem = find_memory_block_by_id(block_id);
719			if (WARN_ON_ONCE(!mem))
720				continue;
721			unregister_memory(mem);
722		}
723	}
724	return ret;
725}
726
727/*
728 * Remove memory block devices for the given memory area. Start and size
729 * have to be aligned to memory block granularity. Memory block devices
730 * have to be offline.
731 *
732 * Called under device_hotplug_lock.
733 */
734void remove_memory_block_devices(unsigned long start, unsigned long size)
735{
736	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
737	const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
738	struct memory_block *mem;
739	unsigned long block_id;
740
741	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
742			 !IS_ALIGNED(size, memory_block_size_bytes())))
743		return;
744
745	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
746		mem = find_memory_block_by_id(block_id);
747		if (WARN_ON_ONCE(!mem))
748			continue;
749		unregister_memory_block_under_nodes(mem);
750		unregister_memory(mem);
751	}
752}
753
754/* return true if the memory block is offlined, otherwise, return false */
755bool is_memblock_offlined(struct memory_block *mem)
756{
757	return mem->state == MEM_OFFLINE;
758}
759
760static struct attribute *memory_root_attrs[] = {
761#ifdef CONFIG_ARCH_MEMORY_PROBE
762	&dev_attr_probe.attr,
763#endif
764
765#ifdef CONFIG_MEMORY_FAILURE
766	&dev_attr_soft_offline_page.attr,
767	&dev_attr_hard_offline_page.attr,
768#endif
769
770	&dev_attr_block_size_bytes.attr,
771	&dev_attr_auto_online_blocks.attr,
772	NULL
773};
774
775static const struct attribute_group memory_root_attr_group = {
776	.attrs = memory_root_attrs,
777};
778
779static const struct attribute_group *memory_root_attr_groups[] = {
780	&memory_root_attr_group,
781	NULL,
782};
783
784/*
785 * Initialize the sysfs support for memory devices. At the time this function
786 * is called, we cannot have concurrent creation/deletion of memory block
787 * devices, the device_hotplug_lock is not needed.
788 */
789void __init memory_dev_init(void)
790{
791	int ret;
792	unsigned long block_sz, nr;
793
794	/* Validate the configured memory block size */
795	block_sz = memory_block_size_bytes();
796	if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
797		panic("Memory block size not suitable: 0x%lx\n", block_sz);
798	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
799
800	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
801	if (ret)
802		panic("%s() failed to register subsystem: %d\n", __func__, ret);
803
804	/*
805	 * Create entries for memory sections that were found
806	 * during boot and have been initialized
807	 */
808	for (nr = 0; nr <= __highest_present_section_nr;
809	     nr += sections_per_block) {
810		ret = add_memory_block(nr);
811		if (ret)
812			panic("%s() failed to add memory block: %d\n", __func__,
813			      ret);
814	}
815}
816
817/**
818 * walk_memory_blocks - walk through all present memory blocks overlapped
819 *			by the range [start, start + size)
820 *
821 * @start: start address of the memory range
822 * @size: size of the memory range
823 * @arg: argument passed to func
824 * @func: callback for each memory section walked
825 *
826 * This function walks through all present memory blocks overlapped by the
827 * range [start, start + size), calling func on each memory block.
828 *
829 * In case func() returns an error, walking is aborted and the error is
830 * returned.
831 *
832 * Called under device_hotplug_lock.
833 */
834int walk_memory_blocks(unsigned long start, unsigned long size,
835		       void *arg, walk_memory_blocks_func_t func)
836{
837	const unsigned long start_block_id = phys_to_block_id(start);
838	const unsigned long end_block_id = phys_to_block_id(start + size - 1);
839	struct memory_block *mem;
840	unsigned long block_id;
841	int ret = 0;
842
843	if (!size)
844		return 0;
845
846	for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
847		mem = find_memory_block_by_id(block_id);
848		if (!mem)
849			continue;
850
851		ret = func(mem, arg);
852		put_device(&mem->dev);
853		if (ret)
854			break;
855	}
856	return ret;
857}
858
859struct for_each_memory_block_cb_data {
860	walk_memory_blocks_func_t func;
861	void *arg;
862};
863
864static int for_each_memory_block_cb(struct device *dev, void *data)
865{
866	struct memory_block *mem = to_memory_block(dev);
867	struct for_each_memory_block_cb_data *cb_data = data;
868
869	return cb_data->func(mem, cb_data->arg);
870}
871
872/**
873 * for_each_memory_block - walk through all present memory blocks
874 *
875 * @arg: argument passed to func
876 * @func: callback for each memory block walked
877 *
878 * This function walks through all present memory blocks, calling func on
879 * each memory block.
880 *
881 * In case func() returns an error, walking is aborted and the error is
882 * returned.
883 */
884int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
885{
886	struct for_each_memory_block_cb_data cb_data = {
887		.func = func,
888		.arg = arg,
889	};
890
891	return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
892				for_each_memory_block_cb);
893}