1/*
2 * Memory subsystem support
3 *
4 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
5 *            Dave Hansen <haveblue@us.ibm.com>
6 *
7 * This file provides the necessary infrastructure to represent
8 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
9 * All arch-independent code that assumes MEMORY_HOTPLUG requires
10 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
11 */
12
13#include <linux/module.h>
14#include <linux/init.h>
15#include <linux/topology.h>
16#include <linux/capability.h>
17#include <linux/device.h>
18#include <linux/memory.h>
19#include <linux/memory_hotplug.h>
20#include <linux/mm.h>
21#include <linux/mutex.h>
22#include <linux/stat.h>
23#include <linux/slab.h>
24
25#include <linux/atomic.h>
26#include <asm/uaccess.h>
27
28static DEFINE_MUTEX(mem_sysfs_mutex);
29
30#define MEMORY_CLASS_NAME	"memory"
31
32#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
33
34static int sections_per_block;
35
36static inline int base_memory_block_id(int section_nr)
37{
38	return section_nr / sections_per_block;
39}
40
41static int memory_subsys_online(struct device *dev);
42static int memory_subsys_offline(struct device *dev);
43
44static struct bus_type memory_subsys = {
45	.name = MEMORY_CLASS_NAME,
46	.dev_name = MEMORY_CLASS_NAME,
47	.online = memory_subsys_online,
48	.offline = memory_subsys_offline,
49};
50
51static BLOCKING_NOTIFIER_HEAD(memory_chain);
52
53int register_memory_notifier(struct notifier_block *nb)
54{
55	return blocking_notifier_chain_register(&memory_chain, nb);
56}
57EXPORT_SYMBOL(register_memory_notifier);
58
59void unregister_memory_notifier(struct notifier_block *nb)
60{
61	blocking_notifier_chain_unregister(&memory_chain, nb);
62}
63EXPORT_SYMBOL(unregister_memory_notifier);
64
65static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
66
67int register_memory_isolate_notifier(struct notifier_block *nb)
68{
69	return atomic_notifier_chain_register(&memory_isolate_chain, nb);
70}
71EXPORT_SYMBOL(register_memory_isolate_notifier);
72
73void unregister_memory_isolate_notifier(struct notifier_block *nb)
74{
75	atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
76}
77EXPORT_SYMBOL(unregister_memory_isolate_notifier);
78
79static void memory_block_release(struct device *dev)
80{
81	struct memory_block *mem = to_memory_block(dev);
82
83	kfree(mem);
84}
85
86unsigned long __weak memory_block_size_bytes(void)
87{
88	return MIN_MEMORY_BLOCK_SIZE;
89}
90
91static unsigned long get_memory_block_size(void)
92{
93	unsigned long block_sz;
94
95	block_sz = memory_block_size_bytes();
96
97	/* Validate blk_sz is a power of 2 and not less than section size */
98	if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
99		WARN_ON(1);
100		block_sz = MIN_MEMORY_BLOCK_SIZE;
101	}
102
103	return block_sz;
104}
105
106/*
107 * use this as the physical section index that this memsection
108 * uses.
109 */
110
111static ssize_t show_mem_start_phys_index(struct device *dev,
112			struct device_attribute *attr, char *buf)
113{
114	struct memory_block *mem = to_memory_block(dev);
115	unsigned long phys_index;
116
117	phys_index = mem->start_section_nr / sections_per_block;
118	return sprintf(buf, "%08lx\n", phys_index);
119}
120
121/*
122 * Show whether the section of memory is likely to be hot-removable
123 */
124static ssize_t show_mem_removable(struct device *dev,
125			struct device_attribute *attr, char *buf)
126{
127	unsigned long i, pfn;
128	int ret = 1;
129	struct memory_block *mem = to_memory_block(dev);
130
131	for (i = 0; i < sections_per_block; i++) {
132		if (!present_section_nr(mem->start_section_nr + i))
133			continue;
134		pfn = section_nr_to_pfn(mem->start_section_nr + i);
135		ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
136	}
137
138	return sprintf(buf, "%d\n", ret);
139}
140
141/*
142 * online, offline, going offline, etc.
143 */
144static ssize_t show_mem_state(struct device *dev,
145			struct device_attribute *attr, char *buf)
146{
147	struct memory_block *mem = to_memory_block(dev);
148	ssize_t len = 0;
149
150	/*
151	 * We can probably put these states in a nice little array
152	 * so that they're not open-coded
153	 */
154	switch (mem->state) {
155	case MEM_ONLINE:
156		len = sprintf(buf, "online\n");
157		break;
158	case MEM_OFFLINE:
159		len = sprintf(buf, "offline\n");
160		break;
161	case MEM_GOING_OFFLINE:
162		len = sprintf(buf, "going-offline\n");
163		break;
164	default:
165		len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
166				mem->state);
167		WARN_ON(1);
168		break;
169	}
170
171	return len;
172}
173
174int memory_notify(unsigned long val, void *v)
175{
176	return blocking_notifier_call_chain(&memory_chain, val, v);
177}
178
179int memory_isolate_notify(unsigned long val, void *v)
180{
181	return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
182}
183
184/*
185 * The probe routines leave the pages reserved, just as the bootmem code does.
186 * Make sure they're still that way.
187 */
188static bool pages_correctly_reserved(unsigned long start_pfn)
189{
190	int i, j;
191	struct page *page;
192	unsigned long pfn = start_pfn;
193
194	/*
195	 * memmap between sections is not contiguous except with
196	 * SPARSEMEM_VMEMMAP. We lookup the page once per section
197	 * and assume memmap is contiguous within each section
198	 */
199	for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
200		if (WARN_ON_ONCE(!pfn_valid(pfn)))
201			return false;
202		page = pfn_to_page(pfn);
203
204		for (j = 0; j < PAGES_PER_SECTION; j++) {
205			if (PageReserved(page + j))
206				continue;
207
208			printk(KERN_WARNING "section number %ld page number %d "
209				"not reserved, was it already online?\n",
210				pfn_to_section_nr(pfn), j);
211
212			return false;
213		}
214	}
215
216	return true;
217}
218
219/*
220 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
221 * OK to have direct references to sparsemem variables in here.
222 * Must already be protected by mem_hotplug_begin().
223 */
224static int
225memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
226{
227	unsigned long start_pfn;
228	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
229	struct page *first_page;
230	int ret;
231
232	start_pfn = section_nr_to_pfn(phys_index);
233	first_page = pfn_to_page(start_pfn);
234
235	switch (action) {
236	case MEM_ONLINE:
237		if (!pages_correctly_reserved(start_pfn))
238			return -EBUSY;
239
240		ret = online_pages(start_pfn, nr_pages, online_type);
241		break;
242	case MEM_OFFLINE:
243		ret = offline_pages(start_pfn, nr_pages);
244		break;
245	default:
246		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
247		     "%ld\n", __func__, phys_index, action, action);
248		ret = -EINVAL;
249	}
250
251	return ret;
252}
253
254static int memory_block_change_state(struct memory_block *mem,
255		unsigned long to_state, unsigned long from_state_req)
256{
257	int ret = 0;
258
259	if (mem->state != from_state_req)
260		return -EINVAL;
261
262	if (to_state == MEM_OFFLINE)
263		mem->state = MEM_GOING_OFFLINE;
264
265	ret = memory_block_action(mem->start_section_nr, to_state,
266				mem->online_type);
267
268	mem->state = ret ? from_state_req : to_state;
269
270	return ret;
271}
272
273/* The device lock serializes operations on memory_subsys_[online|offline] */
274static int memory_subsys_online(struct device *dev)
275{
276	struct memory_block *mem = to_memory_block(dev);
277	int ret;
278
279	if (mem->state == MEM_ONLINE)
280		return 0;
281
282	/*
283	 * If we are called from store_mem_state(), online_type will be
284	 * set >= 0 Otherwise we were called from the device online
285	 * attribute and need to set the online_type.
286	 */
287	if (mem->online_type < 0)
288		mem->online_type = MMOP_ONLINE_KEEP;
289
290	/* Already under protection of mem_hotplug_begin() */
291	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
292
293	/* clear online_type */
294	mem->online_type = -1;
295
296	return ret;
297}
298
299static int memory_subsys_offline(struct device *dev)
300{
301	struct memory_block *mem = to_memory_block(dev);
302
303	if (mem->state == MEM_OFFLINE)
304		return 0;
305
306	/* Can't offline block with non-present sections */
307	if (mem->section_count != sections_per_block)
308		return -EINVAL;
309
310	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
311}
312
313static ssize_t
314store_mem_state(struct device *dev,
315		struct device_attribute *attr, const char *buf, size_t count)
316{
317	struct memory_block *mem = to_memory_block(dev);
318	int ret, online_type;
319
320	ret = lock_device_hotplug_sysfs();
321	if (ret)
322		return ret;
323
324	if (sysfs_streq(buf, "online_kernel"))
325		online_type = MMOP_ONLINE_KERNEL;
326	else if (sysfs_streq(buf, "online_movable"))
327		online_type = MMOP_ONLINE_MOVABLE;
328	else if (sysfs_streq(buf, "online"))
329		online_type = MMOP_ONLINE_KEEP;
330	else if (sysfs_streq(buf, "offline"))
331		online_type = MMOP_OFFLINE;
332	else {
333		ret = -EINVAL;
334		goto err;
335	}
336
337	/*
338	 * Memory hotplug needs to hold mem_hotplug_begin() for probe to find
339	 * the correct memory block to online before doing device_online(dev),
340	 * which will take dev->mutex.  Take the lock early to prevent an
341	 * inversion, memory_subsys_online() callbacks will be implemented by
342	 * assuming it's already protected.
343	 */
344	mem_hotplug_begin();
345
346	switch (online_type) {
347	case MMOP_ONLINE_KERNEL:
348	case MMOP_ONLINE_MOVABLE:
349	case MMOP_ONLINE_KEEP:
350		mem->online_type = online_type;
351		ret = device_online(&mem->dev);
352		break;
353	case MMOP_OFFLINE:
354		ret = device_offline(&mem->dev);
355		break;
356	default:
357		ret = -EINVAL; /* should never happen */
358	}
359
360	mem_hotplug_done();
361err:
362	unlock_device_hotplug();
363
364	if (ret)
365		return ret;
366	return count;
367}
368
369/*
370 * phys_device is a bad name for this.  What I really want
371 * is a way to differentiate between memory ranges that
372 * are part of physical devices that constitute
373 * a complete removable unit or fru.
374 * i.e. do these ranges belong to the same physical device,
375 * s.t. if I offline all of these sections I can then
376 * remove the physical device?
377 */
378static ssize_t show_phys_device(struct device *dev,
379				struct device_attribute *attr, char *buf)
380{
381	struct memory_block *mem = to_memory_block(dev);
382	return sprintf(buf, "%d\n", mem->phys_device);
383}
384
385#ifdef CONFIG_MEMORY_HOTREMOVE
386static ssize_t show_valid_zones(struct device *dev,
387				struct device_attribute *attr, char *buf)
388{
389	struct memory_block *mem = to_memory_block(dev);
390	unsigned long start_pfn, end_pfn;
391	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
392	struct page *first_page;
393	struct zone *zone;
394
395	start_pfn = section_nr_to_pfn(mem->start_section_nr);
396	end_pfn = start_pfn + nr_pages;
397	first_page = pfn_to_page(start_pfn);
398
399	/* The block contains more than one zone can not be offlined. */
400	if (!test_pages_in_a_zone(start_pfn, end_pfn))
401		return sprintf(buf, "none\n");
402
403	zone = page_zone(first_page);
404
405	if (zone_idx(zone) == ZONE_MOVABLE - 1) {
406		/*The mem block is the last memoryblock of this zone.*/
407		if (end_pfn == zone_end_pfn(zone))
408			return sprintf(buf, "%s %s\n",
409					zone->name, (zone + 1)->name);
410	}
411
412	if (zone_idx(zone) == ZONE_MOVABLE) {
413		/*The mem block is the first memoryblock of ZONE_MOVABLE.*/
414		if (start_pfn == zone->zone_start_pfn)
415			return sprintf(buf, "%s %s\n",
416					zone->name, (zone - 1)->name);
417	}
418
419	return sprintf(buf, "%s\n", zone->name);
420}
421static DEVICE_ATTR(valid_zones, 0444, show_valid_zones, NULL);
422#endif
423
424static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
425static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
426static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
427static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
428
429/*
430 * Block size attribute stuff
431 */
432static ssize_t
433print_block_size(struct device *dev, struct device_attribute *attr,
434		 char *buf)
435{
436	return sprintf(buf, "%lx\n", get_memory_block_size());
437}
438
439static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
440
441/*
442 * Some architectures will have custom drivers to do this, and
443 * will not need to do it from userspace.  The fake hot-add code
444 * as well as ppc64 will do all of their discovery in userspace
445 * and will require this interface.
446 */
447#ifdef CONFIG_ARCH_MEMORY_PROBE
448static ssize_t
449memory_probe_store(struct device *dev, struct device_attribute *attr,
450		   const char *buf, size_t count)
451{
452	u64 phys_addr;
453	int nid;
454	int i, ret;
455	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
456
457	ret = kstrtoull(buf, 0, &phys_addr);
458	if (ret)
459		return ret;
460
461	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
462		return -EINVAL;
463
464	for (i = 0; i < sections_per_block; i++) {
465		nid = memory_add_physaddr_to_nid(phys_addr);
466		ret = add_memory(nid, phys_addr,
467				 PAGES_PER_SECTION << PAGE_SHIFT);
468		if (ret)
469			goto out;
470
471		phys_addr += MIN_MEMORY_BLOCK_SIZE;
472	}
473
474	ret = count;
475out:
476	return ret;
477}
478
479static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
480#endif
481
482#ifdef CONFIG_MEMORY_FAILURE
483/*
484 * Support for offlining pages of memory
485 */
486
487/* Soft offline a page */
488static ssize_t
489store_soft_offline_page(struct device *dev,
490			struct device_attribute *attr,
491			const char *buf, size_t count)
492{
493	int ret;
494	u64 pfn;
495	if (!capable(CAP_SYS_ADMIN))
496		return -EPERM;
497	if (kstrtoull(buf, 0, &pfn) < 0)
498		return -EINVAL;
499	pfn >>= PAGE_SHIFT;
500	if (!pfn_valid(pfn))
501		return -ENXIO;
502	ret = soft_offline_page(pfn_to_page(pfn), 0);
503	return ret == 0 ? count : ret;
504}
505
506/* Forcibly offline a page, including killing processes. */
507static ssize_t
508store_hard_offline_page(struct device *dev,
509			struct device_attribute *attr,
510			const char *buf, size_t count)
511{
512	int ret;
513	u64 pfn;
514	if (!capable(CAP_SYS_ADMIN))
515		return -EPERM;
516	if (kstrtoull(buf, 0, &pfn) < 0)
517		return -EINVAL;
518	pfn >>= PAGE_SHIFT;
519	ret = memory_failure(pfn, 0, 0);
520	return ret ? ret : count;
521}
522
523static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
524static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
525#endif
526
527/*
528 * Note that phys_device is optional.  It is here to allow for
529 * differentiation between which *physical* devices each
530 * section belongs to...
531 */
532int __weak arch_get_memory_phys_device(unsigned long start_pfn)
533{
534	return 0;
535}
536
537/*
538 * A reference for the returned object is held and the reference for the
539 * hinted object is released.
540 */
541struct memory_block *find_memory_block_hinted(struct mem_section *section,
542					      struct memory_block *hint)
543{
544	int block_id = base_memory_block_id(__section_nr(section));
545	struct device *hintdev = hint ? &hint->dev : NULL;
546	struct device *dev;
547
548	dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
549	if (hint)
550		put_device(&hint->dev);
551	if (!dev)
552		return NULL;
553	return to_memory_block(dev);
554}
555
556/*
557 * For now, we have a linear search to go find the appropriate
558 * memory_block corresponding to a particular phys_index. If
559 * this gets to be a real problem, we can always use a radix
560 * tree or something here.
561 *
562 * This could be made generic for all device subsystems.
563 */
564struct memory_block *find_memory_block(struct mem_section *section)
565{
566	return find_memory_block_hinted(section, NULL);
567}
568
569static struct attribute *memory_memblk_attrs[] = {
570	&dev_attr_phys_index.attr,
571	&dev_attr_state.attr,
572	&dev_attr_phys_device.attr,
573	&dev_attr_removable.attr,
574#ifdef CONFIG_MEMORY_HOTREMOVE
575	&dev_attr_valid_zones.attr,
576#endif
577	NULL
578};
579
580static struct attribute_group memory_memblk_attr_group = {
581	.attrs = memory_memblk_attrs,
582};
583
584static const struct attribute_group *memory_memblk_attr_groups[] = {
585	&memory_memblk_attr_group,
586	NULL,
587};
588
589/*
590 * register_memory - Setup a sysfs device for a memory block
591 */
592static
593int register_memory(struct memory_block *memory)
594{
595	memory->dev.bus = &memory_subsys;
596	memory->dev.id = memory->start_section_nr / sections_per_block;
597	memory->dev.release = memory_block_release;
598	memory->dev.groups = memory_memblk_attr_groups;
599	memory->dev.offline = memory->state == MEM_OFFLINE;
600
601	return device_register(&memory->dev);
602}
603
604static int init_memory_block(struct memory_block **memory,
605			     struct mem_section *section, unsigned long state)
606{
607	struct memory_block *mem;
608	unsigned long start_pfn;
609	int scn_nr;
610	int ret = 0;
611
612	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
613	if (!mem)
614		return -ENOMEM;
615
616	scn_nr = __section_nr(section);
617	mem->start_section_nr =
618			base_memory_block_id(scn_nr) * sections_per_block;
619	mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
620	mem->state = state;
621	mem->section_count++;
622	start_pfn = section_nr_to_pfn(mem->start_section_nr);
623	mem->phys_device = arch_get_memory_phys_device(start_pfn);
624
625	ret = register_memory(mem);
626
627	*memory = mem;
628	return ret;
629}
630
631static int add_memory_block(int base_section_nr)
632{
633	struct memory_block *mem;
634	int i, ret, section_count = 0, section_nr;
635
636	for (i = base_section_nr;
637	     (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
638	     i++) {
639		if (!present_section_nr(i))
640			continue;
641		if (section_count == 0)
642			section_nr = i;
643		section_count++;
644	}
645
646	if (section_count == 0)
647		return 0;
648	ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
649	if (ret)
650		return ret;
651	mem->section_count = section_count;
652	return 0;
653}
654
655
656/*
657 * need an interface for the VM to add new memory regions,
658 * but without onlining it.
659 */
660int register_new_memory(int nid, struct mem_section *section)
661{
662	int ret = 0;
663	struct memory_block *mem;
664
665	mutex_lock(&mem_sysfs_mutex);
666
667	mem = find_memory_block(section);
668	if (mem) {
669		mem->section_count++;
670		put_device(&mem->dev);
671	} else {
672		ret = init_memory_block(&mem, section, MEM_OFFLINE);
673		if (ret)
674			goto out;
675	}
676
677	if (mem->section_count == sections_per_block)
678		ret = register_mem_sect_under_node(mem, nid);
679out:
680	mutex_unlock(&mem_sysfs_mutex);
681	return ret;
682}
683
684#ifdef CONFIG_MEMORY_HOTREMOVE
685static void
686unregister_memory(struct memory_block *memory)
687{
688	BUG_ON(memory->dev.bus != &memory_subsys);
689
690	/* drop the ref. we got in remove_memory_block() */
691	put_device(&memory->dev);
692	device_unregister(&memory->dev);
693}
694
695static int remove_memory_block(unsigned long node_id,
696			       struct mem_section *section, int phys_device)
697{
698	struct memory_block *mem;
699
700	mutex_lock(&mem_sysfs_mutex);
701	mem = find_memory_block(section);
702	unregister_mem_sect_under_nodes(mem, __section_nr(section));
703
704	mem->section_count--;
705	if (mem->section_count == 0)
706		unregister_memory(mem);
707	else
708		put_device(&mem->dev);
709
710	mutex_unlock(&mem_sysfs_mutex);
711	return 0;
712}
713
714int unregister_memory_section(struct mem_section *section)
715{
716	if (!present_section(section))
717		return -EINVAL;
718
719	return remove_memory_block(0, section, 0);
720}
721#endif /* CONFIG_MEMORY_HOTREMOVE */
722
723/* return true if the memory block is offlined, otherwise, return false */
724bool is_memblock_offlined(struct memory_block *mem)
725{
726	return mem->state == MEM_OFFLINE;
727}
728
729static struct attribute *memory_root_attrs[] = {
730#ifdef CONFIG_ARCH_MEMORY_PROBE
731	&dev_attr_probe.attr,
732#endif
733
734#ifdef CONFIG_MEMORY_FAILURE
735	&dev_attr_soft_offline_page.attr,
736	&dev_attr_hard_offline_page.attr,
737#endif
738
739	&dev_attr_block_size_bytes.attr,
740	NULL
741};
742
743static struct attribute_group memory_root_attr_group = {
744	.attrs = memory_root_attrs,
745};
746
747static const struct attribute_group *memory_root_attr_groups[] = {
748	&memory_root_attr_group,
749	NULL,
750};
751
752/*
753 * Initialize the sysfs support for memory devices...
754 */
755int __init memory_dev_init(void)
756{
757	unsigned int i;
758	int ret;
759	int err;
760	unsigned long block_sz;
761
762	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
763	if (ret)
764		goto out;
765
766	block_sz = get_memory_block_size();
767	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
768
769	/*
770	 * Create entries for memory sections that were found
771	 * during boot and have been initialized
772	 */
773	mutex_lock(&mem_sysfs_mutex);
774	for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
775		err = add_memory_block(i);
776		if (!ret)
777			ret = err;
778	}
779	mutex_unlock(&mem_sysfs_mutex);
780
781out:
782	if (ret)
783		printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
784	return ret;
785}
786