Linux Audio

Check our new training course

Loading...
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}
v5.4
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Memory subsystem support
  4 *
  5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
  6 *            Dave Hansen <haveblue@us.ibm.com>
  7 *
  8 * This file provides the necessary infrastructure to represent
  9 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
 10 * All arch-independent code that assumes MEMORY_HOTPLUG requires
 11 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
 12 */
 13
 14#include <linux/module.h>
 15#include <linux/init.h>
 16#include <linux/topology.h>
 17#include <linux/capability.h>
 18#include <linux/device.h>
 19#include <linux/memory.h>
 20#include <linux/memory_hotplug.h>
 21#include <linux/mm.h>
 22#include <linux/mutex.h>
 23#include <linux/stat.h>
 24#include <linux/slab.h>
 
 25
 26#include <linux/atomic.h>
 27#include <linux/uaccess.h>
 28
 29static DEFINE_MUTEX(mem_sysfs_mutex);
 
 
 
 
 
 
 
 30
 31#define MEMORY_CLASS_NAME	"memory"
 
 
 
 
 
 
 
 
 
 32
 33#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
 34
 35static int sections_per_block;
 36
 37static inline unsigned long base_memory_block_id(unsigned long section_nr)
 38{
 39	return section_nr / sections_per_block;
 40}
 41
 42static inline unsigned long pfn_to_block_id(unsigned long pfn)
 43{
 44	return base_memory_block_id(pfn_to_section_nr(pfn));
 45}
 46
 47static inline unsigned long phys_to_block_id(unsigned long phys)
 48{
 49	return pfn_to_block_id(PFN_DOWN(phys));
 50}
 51
 52static int memory_subsys_online(struct device *dev);
 53static int memory_subsys_offline(struct device *dev);
 54
 55static struct bus_type memory_subsys = {
 56	.name = MEMORY_CLASS_NAME,
 57	.dev_name = MEMORY_CLASS_NAME,
 58	.online = memory_subsys_online,
 59	.offline = memory_subsys_offline,
 60};
 61
 
 
 
 
 
 
 
 62static BLOCKING_NOTIFIER_HEAD(memory_chain);
 63
 64int register_memory_notifier(struct notifier_block *nb)
 65{
 66	return blocking_notifier_chain_register(&memory_chain, nb);
 67}
 68EXPORT_SYMBOL(register_memory_notifier);
 69
 70void unregister_memory_notifier(struct notifier_block *nb)
 71{
 72	blocking_notifier_chain_unregister(&memory_chain, nb);
 73}
 74EXPORT_SYMBOL(unregister_memory_notifier);
 75
 76static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
 77
 78int register_memory_isolate_notifier(struct notifier_block *nb)
 79{
 80	return atomic_notifier_chain_register(&memory_isolate_chain, nb);
 81}
 82EXPORT_SYMBOL(register_memory_isolate_notifier);
 83
 84void unregister_memory_isolate_notifier(struct notifier_block *nb)
 85{
 86	atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
 87}
 88EXPORT_SYMBOL(unregister_memory_isolate_notifier);
 89
 90static void memory_block_release(struct device *dev)
 91{
 92	struct memory_block *mem = to_memory_block(dev);
 93
 94	kfree(mem);
 95}
 96
 97unsigned long __weak memory_block_size_bytes(void)
 98{
 99	return MIN_MEMORY_BLOCK_SIZE;
100}
101EXPORT_SYMBOL_GPL(memory_block_size_bytes);
102
103/*
104 * Show the first physical section index (number) of this memory block.
105 */
106static ssize_t phys_index_show(struct device *dev,
107			       struct device_attribute *attr, char *buf)
108{
109	struct memory_block *mem = to_memory_block(dev);
110	unsigned long phys_index;
111
112	phys_index = mem->start_section_nr / sections_per_block;
113	return sprintf(buf, "%08lx\n", phys_index);
 
114}
115
116/*
117 * Show whether the memory block is likely to be offlineable (or is already
118 * offline). Once offline, the memory block could be removed. The return
119 * value does, however, not indicate that there is a way to remove the
120 * memory block.
121 */
122static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
123			      char *buf)
124{
125	struct memory_block *mem = to_memory_block(dev);
126	unsigned long pfn;
127	int ret = 1, i;
128
129	if (mem->state != MEM_ONLINE)
130		goto out;
131
132	for (i = 0; i < sections_per_block; i++) {
133		if (!present_section_nr(mem->start_section_nr + i))
134			continue;
135		pfn = section_nr_to_pfn(mem->start_section_nr + i);
136		ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
137	}
138
139out:
140	return sprintf(buf, "%d\n", ret);
141}
142
143/*
144 * online, offline, going offline, etc.
145 */
146static ssize_t state_show(struct device *dev, struct device_attribute *attr,
147			  char *buf)
148{
149	struct memory_block *mem = to_memory_block(dev);
150	ssize_t len = 0;
151
152	/*
153	 * We can probably put these states in a nice little array
154	 * so that they're not open-coded
155	 */
156	switch (mem->state) {
157	case MEM_ONLINE:
158		len = sprintf(buf, "online\n");
159		break;
160	case MEM_OFFLINE:
161		len = sprintf(buf, "offline\n");
162		break;
163	case MEM_GOING_OFFLINE:
164		len = sprintf(buf, "going-offline\n");
165		break;
166	default:
167		len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
168				mem->state);
169		WARN_ON(1);
170		break;
171	}
172
173	return len;
174}
175
176int memory_notify(unsigned long val, void *v)
177{
178	return blocking_notifier_call_chain(&memory_chain, val, v);
179}
180
181int memory_isolate_notify(unsigned long val, void *v)
182{
183	return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
184}
185
186/*
187 * The probe routines leave the pages uninitialized, just as the bootmem code
188 * does. Make sure we do not access them, but instead use only information from
189 * within sections.
190 */
191static bool pages_correctly_probed(unsigned long start_pfn)
192{
193	unsigned long section_nr = pfn_to_section_nr(start_pfn);
194	unsigned long section_nr_end = section_nr + sections_per_block;
195	unsigned long pfn = start_pfn;
 
 
196
197	/*
198	 * memmap between sections is not contiguous except with
199	 * SPARSEMEM_VMEMMAP. We lookup the page once per section
200	 * and assume memmap is contiguous within each section
201	 */
202	for (; section_nr < section_nr_end; section_nr++) {
203		if (WARN_ON_ONCE(!pfn_valid(pfn)))
204			return false;
205
206		if (!present_section_nr(section_nr)) {
207			pr_warn("section %ld pfn[%lx, %lx) not present\n",
208				section_nr, pfn, pfn + PAGES_PER_SECTION);
209			return false;
210		} else if (!valid_section_nr(section_nr)) {
211			pr_warn("section %ld pfn[%lx, %lx) no valid memmap\n",
212				section_nr, pfn, pfn + PAGES_PER_SECTION);
213			return false;
214		} else if (online_section_nr(section_nr)) {
215			pr_warn("section %ld pfn[%lx, %lx) is already online\n",
216				section_nr, pfn, pfn + PAGES_PER_SECTION);
217			return false;
218		}
219		pfn += PAGES_PER_SECTION;
220	}
221
222	return true;
 
 
 
223}
224
225/*
226 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
227 * OK to have direct references to sparsemem variables in here.
228 */
229static int
230memory_block_action(unsigned long start_section_nr, unsigned long action,
231		    int online_type)
232{
233	unsigned long start_pfn;
234	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
235	int ret;
236
237	start_pfn = section_nr_to_pfn(start_section_nr);
238
239	switch (action) {
240	case MEM_ONLINE:
241		if (!pages_correctly_probed(start_pfn))
242			return -EBUSY;
243
244		ret = online_pages(start_pfn, nr_pages, online_type);
245		break;
246	case MEM_OFFLINE:
247		ret = offline_pages(start_pfn, nr_pages);
248		break;
249	default:
250		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
251		     "%ld\n", __func__, start_section_nr, action, action);
252		ret = -EINVAL;
253	}
254
255	return ret;
256}
257
258static int memory_block_change_state(struct memory_block *mem,
259		unsigned long to_state, unsigned long from_state_req)
260{
261	int ret = 0;
262
263	if (mem->state != from_state_req)
264		return -EINVAL;
265
266	if (to_state == MEM_OFFLINE)
267		mem->state = MEM_GOING_OFFLINE;
268
269	ret = memory_block_action(mem->start_section_nr, to_state,
270				mem->online_type);
271
272	mem->state = ret ? from_state_req : to_state;
273
274	return ret;
275}
276
277/* The device lock serializes operations on memory_subsys_[online|offline] */
278static int memory_subsys_online(struct device *dev)
279{
280	struct memory_block *mem = to_memory_block(dev);
281	int ret;
282
283	if (mem->state == MEM_ONLINE)
284		return 0;
285
286	/*
287	 * If we are called from state_store(), online_type will be
288	 * set >= 0 Otherwise we were called from the device online
289	 * attribute and need to set the online_type.
290	 */
291	if (mem->online_type < 0)
292		mem->online_type = MMOP_ONLINE_KEEP;
293
294	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
295
296	/* clear online_type */
297	mem->online_type = -1;
298
299	return ret;
300}
301
302static int memory_subsys_offline(struct device *dev)
303{
304	struct memory_block *mem = to_memory_block(dev);
305
306	if (mem->state == MEM_OFFLINE)
307		return 0;
308
309	/* Can't offline block with non-present sections */
310	if (mem->section_count != sections_per_block)
311		return -EINVAL;
312
313	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
314}
315
316static ssize_t state_store(struct device *dev, struct device_attribute *attr,
317			   const char *buf, size_t count)
318{
 
319	struct memory_block *mem = to_memory_block(dev);
320	int ret, online_type;
 
 
 
321
322	ret = lock_device_hotplug_sysfs();
323	if (ret)
324		return ret;
325
326	if (sysfs_streq(buf, "online_kernel"))
327		online_type = MMOP_ONLINE_KERNEL;
328	else if (sysfs_streq(buf, "online_movable"))
329		online_type = MMOP_ONLINE_MOVABLE;
330	else if (sysfs_streq(buf, "online"))
331		online_type = MMOP_ONLINE_KEEP;
332	else if (sysfs_streq(buf, "offline"))
333		online_type = MMOP_OFFLINE;
334	else {
335		ret = -EINVAL;
336		goto err;
337	}
338
339	switch (online_type) {
340	case MMOP_ONLINE_KERNEL:
341	case MMOP_ONLINE_MOVABLE:
342	case MMOP_ONLINE_KEEP:
343		/* mem->online_type is protected by device_hotplug_lock */
344		mem->online_type = online_type;
345		ret = device_online(&mem->dev);
346		break;
347	case MMOP_OFFLINE:
348		ret = device_offline(&mem->dev);
349		break;
350	default:
351		ret = -EINVAL; /* should never happen */
352	}
353
354err:
355	unlock_device_hotplug();
356
357	if (ret < 0)
358		return ret;
359	if (ret)
360		return -EINVAL;
361
362	return count;
363}
364
365/*
366 * phys_device is a bad name for this.  What I really want
367 * is a way to differentiate between memory ranges that
368 * are part of physical devices that constitute
369 * a complete removable unit or fru.
370 * i.e. do these ranges belong to the same physical device,
371 * s.t. if I offline all of these sections I can then
372 * remove the physical device?
373 */
374static ssize_t phys_device_show(struct device *dev,
375				struct device_attribute *attr, char *buf)
376{
377	struct memory_block *mem = to_memory_block(dev);
378	return sprintf(buf, "%d\n", mem->phys_device);
 
 
 
379}
380
381#ifdef CONFIG_MEMORY_HOTREMOVE
382static void print_allowed_zone(char *buf, int nid, unsigned long start_pfn,
383		unsigned long nr_pages, int online_type,
384		struct zone *default_zone)
385{
386	struct zone *zone;
387
388	zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
389	if (zone != default_zone) {
390		strcat(buf, " ");
391		strcat(buf, zone->name);
392	}
393}
394
395static ssize_t valid_zones_show(struct device *dev,
396				struct device_attribute *attr, char *buf)
397{
398	struct memory_block *mem = to_memory_block(dev);
399	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
400	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
401	unsigned long valid_start_pfn, valid_end_pfn;
402	struct zone *default_zone;
 
403	int nid;
404
405	/*
406	 * Check the existing zone. Make sure that we do that only on the
407	 * online nodes otherwise the page_zone is not reliable
408	 */
409	if (mem->state == MEM_ONLINE) {
410		/*
411		 * The block contains more than one zone can not be offlined.
412		 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
413		 */
414		if (!test_pages_in_a_zone(start_pfn, start_pfn + nr_pages,
415					  &valid_start_pfn, &valid_end_pfn))
416			return sprintf(buf, "none\n");
417		start_pfn = valid_start_pfn;
418		strcat(buf, page_zone(pfn_to_page(start_pfn))->name);
419		goto out;
420	}
421
422	nid = mem->nid;
423	default_zone = zone_for_pfn_range(MMOP_ONLINE_KEEP, nid, start_pfn, nr_pages);
424	strcat(buf, default_zone->name);
425
426	print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_KERNEL,
427			default_zone);
428	print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_MOVABLE,
429			default_zone);
 
430out:
431	strcat(buf, "\n");
432
433	return strlen(buf);
434}
435static DEVICE_ATTR_RO(valid_zones);
436#endif
437
438static DEVICE_ATTR_RO(phys_index);
439static DEVICE_ATTR_RW(state);
440static DEVICE_ATTR_RO(phys_device);
441static DEVICE_ATTR_RO(removable);
442
443/*
444 * Show the memory block size (shared by all memory blocks).
445 */
446static ssize_t block_size_bytes_show(struct device *dev,
447				     struct device_attribute *attr, char *buf)
448{
449	return sprintf(buf, "%lx\n", memory_block_size_bytes());
450}
451
452static DEVICE_ATTR_RO(block_size_bytes);
453
454/*
455 * Memory auto online policy.
456 */
457
458static ssize_t auto_online_blocks_show(struct device *dev,
459				       struct device_attribute *attr, char *buf)
460{
461	if (memhp_auto_online)
462		return sprintf(buf, "online\n");
463	else
464		return sprintf(buf, "offline\n");
465}
466
467static ssize_t auto_online_blocks_store(struct device *dev,
468					struct device_attribute *attr,
469					const char *buf, size_t count)
470{
471	if (sysfs_streq(buf, "online"))
472		memhp_auto_online = true;
473	else if (sysfs_streq(buf, "offline"))
474		memhp_auto_online = false;
475	else
476		return -EINVAL;
477
 
478	return count;
479}
480
481static DEVICE_ATTR_RW(auto_online_blocks);
482
483/*
484 * Some architectures will have custom drivers to do this, and
485 * will not need to do it from userspace.  The fake hot-add code
486 * as well as ppc64 will do all of their discovery in userspace
487 * and will require this interface.
488 */
489#ifdef CONFIG_ARCH_MEMORY_PROBE
490static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
491			   const char *buf, size_t count)
492{
493	u64 phys_addr;
494	int nid, ret;
495	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
496
497	ret = kstrtoull(buf, 0, &phys_addr);
498	if (ret)
499		return ret;
500
501	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
502		return -EINVAL;
503
504	ret = lock_device_hotplug_sysfs();
505	if (ret)
506		return ret;
507
508	nid = memory_add_physaddr_to_nid(phys_addr);
509	ret = __add_memory(nid, phys_addr,
510			   MIN_MEMORY_BLOCK_SIZE * sections_per_block);
 
511
512	if (ret)
513		goto out;
514
515	ret = count;
516out:
517	unlock_device_hotplug();
518	return ret;
519}
520
521static DEVICE_ATTR_WO(probe);
522#endif
523
524#ifdef CONFIG_MEMORY_FAILURE
525/*
526 * Support for offlining pages of memory
527 */
528
529/* Soft offline a page */
530static ssize_t soft_offline_page_store(struct device *dev,
531				       struct device_attribute *attr,
532				       const char *buf, size_t count)
533{
534	int ret;
535	u64 pfn;
536	if (!capable(CAP_SYS_ADMIN))
537		return -EPERM;
538	if (kstrtoull(buf, 0, &pfn) < 0)
539		return -EINVAL;
540	pfn >>= PAGE_SHIFT;
541	if (!pfn_valid(pfn))
542		return -ENXIO;
543	/* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
544	if (!pfn_to_online_page(pfn))
545		return -EIO;
546	ret = soft_offline_page(pfn_to_page(pfn), 0);
547	return ret == 0 ? count : ret;
548}
549
550/* Forcibly offline a page, including killing processes. */
551static ssize_t hard_offline_page_store(struct device *dev,
552				       struct device_attribute *attr,
553				       const char *buf, size_t count)
554{
555	int ret;
556	u64 pfn;
557	if (!capable(CAP_SYS_ADMIN))
558		return -EPERM;
559	if (kstrtoull(buf, 0, &pfn) < 0)
560		return -EINVAL;
561	pfn >>= PAGE_SHIFT;
562	ret = memory_failure(pfn, 0);
563	return ret ? ret : count;
564}
565
566static DEVICE_ATTR_WO(soft_offline_page);
567static DEVICE_ATTR_WO(hard_offline_page);
568#endif
569
570/*
571 * Note that phys_device is optional.  It is here to allow for
572 * differentiation between which *physical* devices each
573 * section belongs to...
574 */
575int __weak arch_get_memory_phys_device(unsigned long start_pfn)
576{
577	return 0;
578}
579
580/* A reference for the returned memory block device is acquired. */
 
 
 
 
581static struct memory_block *find_memory_block_by_id(unsigned long block_id)
582{
583	struct device *dev;
584
585	dev = subsys_find_device_by_id(&memory_subsys, block_id, NULL);
586	return dev ? to_memory_block(dev) : NULL;
 
 
587}
588
589/*
590 * For now, we have a linear search to go find the appropriate
591 * memory_block corresponding to a particular phys_index. If
592 * this gets to be a real problem, we can always use a radix
593 * tree or something here.
594 *
595 * This could be made generic for all device subsystems.
596 */
597struct memory_block *find_memory_block(struct mem_section *section)
598{
599	unsigned long block_id = base_memory_block_id(__section_nr(section));
600
601	return find_memory_block_by_id(block_id);
602}
603
604static struct attribute *memory_memblk_attrs[] = {
605	&dev_attr_phys_index.attr,
606	&dev_attr_state.attr,
607	&dev_attr_phys_device.attr,
608	&dev_attr_removable.attr,
609#ifdef CONFIG_MEMORY_HOTREMOVE
610	&dev_attr_valid_zones.attr,
611#endif
612	NULL
613};
614
615static struct attribute_group memory_memblk_attr_group = {
616	.attrs = memory_memblk_attrs,
617};
618
619static const struct attribute_group *memory_memblk_attr_groups[] = {
620	&memory_memblk_attr_group,
621	NULL,
622};
623
624/*
625 * register_memory - Setup a sysfs device for a memory block
626 */
627static
628int register_memory(struct memory_block *memory)
629{
630	int ret;
631
632	memory->dev.bus = &memory_subsys;
633	memory->dev.id = memory->start_section_nr / sections_per_block;
634	memory->dev.release = memory_block_release;
635	memory->dev.groups = memory_memblk_attr_groups;
636	memory->dev.offline = memory->state == MEM_OFFLINE;
637
638	ret = device_register(&memory->dev);
639	if (ret)
 
 
 
 
 
 
640		put_device(&memory->dev);
641
 
642	return ret;
643}
644
645static int init_memory_block(struct memory_block **memory,
646			     unsigned long block_id, unsigned long state)
647{
648	struct memory_block *mem;
649	unsigned long start_pfn;
650	int ret = 0;
651
652	mem = find_memory_block_by_id(block_id);
653	if (mem) {
654		put_device(&mem->dev);
655		return -EEXIST;
656	}
657	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
658	if (!mem)
659		return -ENOMEM;
660
661	mem->start_section_nr = block_id * sections_per_block;
662	mem->state = state;
663	start_pfn = section_nr_to_pfn(mem->start_section_nr);
664	mem->phys_device = arch_get_memory_phys_device(start_pfn);
665	mem->nid = NUMA_NO_NODE;
 
666
667	ret = register_memory(mem);
668
669	*memory = mem;
670	return ret;
671}
672
673static int add_memory_block(unsigned long base_section_nr)
674{
675	int ret, section_count = 0;
676	struct memory_block *mem;
677	unsigned long nr;
678
679	for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
680	     nr++)
681		if (present_section_nr(nr))
682			section_count++;
683
684	if (section_count == 0)
685		return 0;
686	ret = init_memory_block(&mem, base_memory_block_id(base_section_nr),
687				MEM_ONLINE);
688	if (ret)
689		return ret;
690	mem->section_count = section_count;
691	return 0;
692}
693
694static void unregister_memory(struct memory_block *memory)
695{
696	if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
697		return;
698
 
 
699	/* drop the ref. we got via find_memory_block() */
700	put_device(&memory->dev);
701	device_unregister(&memory->dev);
702}
703
704/*
705 * Create memory block devices for the given memory area. Start and size
706 * have to be aligned to memory block granularity. Memory block devices
707 * will be initialized as offline.
 
 
708 */
709int create_memory_block_devices(unsigned long start, unsigned long size)
 
710{
711	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
712	unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
713	struct memory_block *mem;
714	unsigned long block_id;
715	int ret = 0;
716
717	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
718			 !IS_ALIGNED(size, memory_block_size_bytes())))
719		return -EINVAL;
720
721	mutex_lock(&mem_sysfs_mutex);
722	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
723		ret = init_memory_block(&mem, block_id, MEM_OFFLINE);
724		if (ret)
725			break;
726		mem->section_count = sections_per_block;
727	}
728	if (ret) {
729		end_block_id = block_id;
730		for (block_id = start_block_id; block_id != end_block_id;
731		     block_id++) {
732			mem = find_memory_block_by_id(block_id);
733			mem->section_count = 0;
 
734			unregister_memory(mem);
735		}
736	}
737	mutex_unlock(&mem_sysfs_mutex);
738	return ret;
739}
740
741/*
742 * Remove memory block devices for the given memory area. Start and size
743 * have to be aligned to memory block granularity. Memory block devices
744 * have to be offline.
 
 
745 */
746void remove_memory_block_devices(unsigned long start, unsigned long size)
747{
748	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
749	const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
750	struct memory_block *mem;
751	unsigned long block_id;
752
753	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
754			 !IS_ALIGNED(size, memory_block_size_bytes())))
755		return;
756
757	mutex_lock(&mem_sysfs_mutex);
758	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
759		mem = find_memory_block_by_id(block_id);
760		if (WARN_ON_ONCE(!mem))
761			continue;
762		mem->section_count = 0;
763		unregister_memory_block_under_nodes(mem);
764		unregister_memory(mem);
765	}
766	mutex_unlock(&mem_sysfs_mutex);
767}
768
769/* return true if the memory block is offlined, otherwise, return false */
770bool is_memblock_offlined(struct memory_block *mem)
771{
772	return mem->state == MEM_OFFLINE;
773}
774
775static struct attribute *memory_root_attrs[] = {
776#ifdef CONFIG_ARCH_MEMORY_PROBE
777	&dev_attr_probe.attr,
778#endif
779
780#ifdef CONFIG_MEMORY_FAILURE
781	&dev_attr_soft_offline_page.attr,
782	&dev_attr_hard_offline_page.attr,
783#endif
784
785	&dev_attr_block_size_bytes.attr,
786	&dev_attr_auto_online_blocks.attr,
787	NULL
788};
789
790static struct attribute_group memory_root_attr_group = {
791	.attrs = memory_root_attrs,
792};
793
794static const struct attribute_group *memory_root_attr_groups[] = {
795	&memory_root_attr_group,
796	NULL,
797};
798
799/*
800 * Initialize the sysfs support for memory devices...
 
 
801 */
802void __init memory_dev_init(void)
803{
804	int ret;
805	int err;
806	unsigned long block_sz, nr;
807
808	/* Validate the configured memory block size */
809	block_sz = memory_block_size_bytes();
810	if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
811		panic("Memory block size not suitable: 0x%lx\n", block_sz);
812	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
813
814	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
815	if (ret)
816		goto out;
817
818	/*
819	 * Create entries for memory sections that were found
820	 * during boot and have been initialized
821	 */
822	mutex_lock(&mem_sysfs_mutex);
823	for (nr = 0; nr <= __highest_present_section_nr;
824	     nr += sections_per_block) {
825		err = add_memory_block(nr);
826		if (!ret)
827			ret = err;
 
828	}
829	mutex_unlock(&mem_sysfs_mutex);
830
831out:
832	if (ret)
833		panic("%s() failed: %d\n", __func__, ret);
834}
835
836/**
837 * walk_memory_blocks - walk through all present memory blocks overlapped
838 *			by the range [start, start + size)
839 *
840 * @start: start address of the memory range
841 * @size: size of the memory range
842 * @arg: argument passed to func
843 * @func: callback for each memory section walked
844 *
845 * This function walks through all present memory blocks overlapped by the
846 * range [start, start + size), calling func on each memory block.
847 *
848 * In case func() returns an error, walking is aborted and the error is
849 * returned.
 
 
850 */
851int walk_memory_blocks(unsigned long start, unsigned long size,
852		       void *arg, walk_memory_blocks_func_t func)
853{
854	const unsigned long start_block_id = phys_to_block_id(start);
855	const unsigned long end_block_id = phys_to_block_id(start + size - 1);
856	struct memory_block *mem;
857	unsigned long block_id;
858	int ret = 0;
859
860	if (!size)
861		return 0;
862
863	for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
864		mem = find_memory_block_by_id(block_id);
865		if (!mem)
866			continue;
867
868		ret = func(mem, arg);
869		put_device(&mem->dev);
870		if (ret)
871			break;
872	}
873	return ret;
874}
875
876struct for_each_memory_block_cb_data {
877	walk_memory_blocks_func_t func;
878	void *arg;
879};
880
881static int for_each_memory_block_cb(struct device *dev, void *data)
882{
883	struct memory_block *mem = to_memory_block(dev);
884	struct for_each_memory_block_cb_data *cb_data = data;
885
886	return cb_data->func(mem, cb_data->arg);
887}
888
889/**
890 * for_each_memory_block - walk through all present memory blocks
891 *
892 * @arg: argument passed to func
893 * @func: callback for each memory block walked
894 *
895 * This function walks through all present memory blocks, calling func on
896 * each memory block.
897 *
898 * In case func() returns an error, walking is aborted and the error is
899 * returned.
900 */
901int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
902{
903	struct for_each_memory_block_cb_data cb_data = {
904		.func = func,
905		.arg = arg,
906	};
907
908	return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
909				for_each_memory_block_cb);
910}