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/kobject.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 <asm/uaccess.h>
28
29static DEFINE_MUTEX(mem_sysfs_mutex);
30
31#define MEMORY_CLASS_NAME	"memory"
32
33static int sections_per_block;
34
35static inline int base_memory_block_id(int section_nr)
36{
37	return section_nr / sections_per_block;
38}
39
40static struct bus_type memory_subsys = {
41	.name = MEMORY_CLASS_NAME,
42	.dev_name = MEMORY_CLASS_NAME,
43};
44
45static BLOCKING_NOTIFIER_HEAD(memory_chain);
46
47int register_memory_notifier(struct notifier_block *nb)
48{
49        return blocking_notifier_chain_register(&memory_chain, nb);
50}
51EXPORT_SYMBOL(register_memory_notifier);
52
53void unregister_memory_notifier(struct notifier_block *nb)
54{
55        blocking_notifier_chain_unregister(&memory_chain, nb);
56}
57EXPORT_SYMBOL(unregister_memory_notifier);
58
59static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
60
61int register_memory_isolate_notifier(struct notifier_block *nb)
62{
63	return atomic_notifier_chain_register(&memory_isolate_chain, nb);
64}
65EXPORT_SYMBOL(register_memory_isolate_notifier);
66
67void unregister_memory_isolate_notifier(struct notifier_block *nb)
68{
69	atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
70}
71EXPORT_SYMBOL(unregister_memory_isolate_notifier);
72
73/*
74 * register_memory - Setup a sysfs device for a memory block
75 */
76static
77int register_memory(struct memory_block *memory)
78{
79	int error;
80
81	memory->dev.bus = &memory_subsys;
82	memory->dev.id = memory->start_section_nr / sections_per_block;
83
84	error = device_register(&memory->dev);
85	return error;
86}
87
88static void
89unregister_memory(struct memory_block *memory)
90{
91	BUG_ON(memory->dev.bus != &memory_subsys);
92
93	/* drop the ref. we got in remove_memory_block() */
94	kobject_put(&memory->dev.kobj);
95	device_unregister(&memory->dev);
96}
97
98unsigned long __weak memory_block_size_bytes(void)
99{
100	return MIN_MEMORY_BLOCK_SIZE;
101}
102
103static unsigned long get_memory_block_size(void)
104{
105	unsigned long block_sz;
106
107	block_sz = memory_block_size_bytes();
108
109	/* Validate blk_sz is a power of 2 and not less than section size */
110	if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
111		WARN_ON(1);
112		block_sz = MIN_MEMORY_BLOCK_SIZE;
113	}
114
115	return block_sz;
116}
117
118/*
119 * use this as the physical section index that this memsection
120 * uses.
121 */
122
123static ssize_t show_mem_start_phys_index(struct device *dev,
124			struct device_attribute *attr, char *buf)
125{
126	struct memory_block *mem =
127		container_of(dev, struct memory_block, dev);
128	unsigned long phys_index;
129
130	phys_index = mem->start_section_nr / sections_per_block;
131	return sprintf(buf, "%08lx\n", phys_index);
132}
133
134static ssize_t show_mem_end_phys_index(struct device *dev,
135			struct device_attribute *attr, char *buf)
136{
137	struct memory_block *mem =
138		container_of(dev, struct memory_block, dev);
139	unsigned long phys_index;
140
141	phys_index = mem->end_section_nr / sections_per_block;
142	return sprintf(buf, "%08lx\n", phys_index);
143}
144
145/*
146 * Show whether the section of memory is likely to be hot-removable
147 */
148static ssize_t show_mem_removable(struct device *dev,
149			struct device_attribute *attr, char *buf)
150{
151	unsigned long i, pfn;
152	int ret = 1;
153	struct memory_block *mem =
154		container_of(dev, struct memory_block, dev);
155
156	for (i = 0; i < sections_per_block; i++) {
157		pfn = section_nr_to_pfn(mem->start_section_nr + i);
158		ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
159	}
160
161	return sprintf(buf, "%d\n", ret);
162}
163
164/*
165 * online, offline, going offline, etc.
166 */
167static ssize_t show_mem_state(struct device *dev,
168			struct device_attribute *attr, char *buf)
169{
170	struct memory_block *mem =
171		container_of(dev, struct memory_block, dev);
172	ssize_t len = 0;
173
174	/*
175	 * We can probably put these states in a nice little array
176	 * so that they're not open-coded
177	 */
178	switch (mem->state) {
179		case MEM_ONLINE:
180			len = sprintf(buf, "online\n");
181			break;
182		case MEM_OFFLINE:
183			len = sprintf(buf, "offline\n");
184			break;
185		case MEM_GOING_OFFLINE:
186			len = sprintf(buf, "going-offline\n");
187			break;
188		default:
189			len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
190					mem->state);
191			WARN_ON(1);
192			break;
193	}
194
195	return len;
196}
197
198int memory_notify(unsigned long val, void *v)
199{
200	return blocking_notifier_call_chain(&memory_chain, val, v);
201}
202
203int memory_isolate_notify(unsigned long val, void *v)
204{
205	return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
206}
207
208/*
209 * The probe routines leave the pages reserved, just as the bootmem code does.
210 * Make sure they're still that way.
211 */
212static bool pages_correctly_reserved(unsigned long start_pfn,
213					unsigned long nr_pages)
214{
215	int i, j;
216	struct page *page;
217	unsigned long pfn = start_pfn;
218
219	/*
220	 * memmap between sections is not contiguous except with
221	 * SPARSEMEM_VMEMMAP. We lookup the page once per section
222	 * and assume memmap is contiguous within each section
223	 */
224	for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
225		if (WARN_ON_ONCE(!pfn_valid(pfn)))
226			return false;
227		page = pfn_to_page(pfn);
228
229		for (j = 0; j < PAGES_PER_SECTION; j++) {
230			if (PageReserved(page + j))
231				continue;
232
233			printk(KERN_WARNING "section number %ld page number %d "
234				"not reserved, was it already online?\n",
235				pfn_to_section_nr(pfn), j);
236
237			return false;
238		}
239	}
240
241	return true;
242}
243
244/*
245 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
246 * OK to have direct references to sparsemem variables in here.
247 */
248static int
249memory_block_action(unsigned long phys_index, unsigned long action)
250{
251	unsigned long start_pfn, start_paddr;
252	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
253	struct page *first_page;
254	int ret;
255
256	first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
257
258	switch (action) {
259		case MEM_ONLINE:
260			start_pfn = page_to_pfn(first_page);
261
262			if (!pages_correctly_reserved(start_pfn, nr_pages))
263				return -EBUSY;
264
265			ret = online_pages(start_pfn, nr_pages);
266			break;
267		case MEM_OFFLINE:
268			start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
269			ret = remove_memory(start_paddr,
270					    nr_pages << PAGE_SHIFT);
271			break;
272		default:
273			WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
274			     "%ld\n", __func__, phys_index, action, action);
275			ret = -EINVAL;
276	}
277
278	return ret;
279}
280
281static int memory_block_change_state(struct memory_block *mem,
282		unsigned long to_state, unsigned long from_state_req)
283{
284	int ret = 0;
285
286	mutex_lock(&mem->state_mutex);
287
288	if (mem->state != from_state_req) {
289		ret = -EINVAL;
290		goto out;
291	}
292
293	if (to_state == MEM_OFFLINE)
294		mem->state = MEM_GOING_OFFLINE;
295
296	ret = memory_block_action(mem->start_section_nr, to_state);
297
298	if (ret) {
299		mem->state = from_state_req;
300		goto out;
301	}
302
303	mem->state = to_state;
304	switch (mem->state) {
305	case MEM_OFFLINE:
306		kobject_uevent(&mem->dev.kobj, KOBJ_OFFLINE);
307		break;
308	case MEM_ONLINE:
309		kobject_uevent(&mem->dev.kobj, KOBJ_ONLINE);
310		break;
311	default:
312		break;
313	}
314out:
315	mutex_unlock(&mem->state_mutex);
316	return ret;
317}
318
319static ssize_t
320store_mem_state(struct device *dev,
321		struct device_attribute *attr, const char *buf, size_t count)
322{
323	struct memory_block *mem;
324	int ret = -EINVAL;
325
326	mem = container_of(dev, struct memory_block, dev);
327
328	if (!strncmp(buf, "online", min((int)count, 6)))
329		ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
330	else if(!strncmp(buf, "offline", min((int)count, 7)))
331		ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
332
333	if (ret)
334		return ret;
335	return count;
336}
337
338/*
339 * phys_device is a bad name for this.  What I really want
340 * is a way to differentiate between memory ranges that
341 * are part of physical devices that constitute
342 * a complete removable unit or fru.
343 * i.e. do these ranges belong to the same physical device,
344 * s.t. if I offline all of these sections I can then
345 * remove the physical device?
346 */
347static ssize_t show_phys_device(struct device *dev,
348				struct device_attribute *attr, char *buf)
349{
350	struct memory_block *mem =
351		container_of(dev, struct memory_block, dev);
352	return sprintf(buf, "%d\n", mem->phys_device);
353}
354
355static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
356static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
357static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
358static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
359static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
360
361#define mem_create_simple_file(mem, attr_name)	\
362	device_create_file(&mem->dev, &dev_attr_##attr_name)
363#define mem_remove_simple_file(mem, attr_name)	\
364	device_remove_file(&mem->dev, &dev_attr_##attr_name)
365
366/*
367 * Block size attribute stuff
368 */
369static ssize_t
370print_block_size(struct device *dev, struct device_attribute *attr,
371		 char *buf)
372{
373	return sprintf(buf, "%lx\n", get_memory_block_size());
374}
375
376static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
377
378static int block_size_init(void)
379{
380	return device_create_file(memory_subsys.dev_root,
381				  &dev_attr_block_size_bytes);
382}
383
384/*
385 * Some architectures will have custom drivers to do this, and
386 * will not need to do it from userspace.  The fake hot-add code
387 * as well as ppc64 will do all of their discovery in userspace
388 * and will require this interface.
389 */
390#ifdef CONFIG_ARCH_MEMORY_PROBE
391static ssize_t
392memory_probe_store(struct device *dev, struct device_attribute *attr,
393		   const char *buf, size_t count)
394{
395	u64 phys_addr;
396	int nid;
397	int i, ret;
398	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
399
400	phys_addr = simple_strtoull(buf, NULL, 0);
401
402	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
403		return -EINVAL;
404
405	for (i = 0; i < sections_per_block; i++) {
406		nid = memory_add_physaddr_to_nid(phys_addr);
407		ret = add_memory(nid, phys_addr,
408				 PAGES_PER_SECTION << PAGE_SHIFT);
409		if (ret)
410			goto out;
411
412		phys_addr += MIN_MEMORY_BLOCK_SIZE;
413	}
414
415	ret = count;
416out:
417	return ret;
418}
419static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
420
421static int memory_probe_init(void)
422{
423	return device_create_file(memory_subsys.dev_root, &dev_attr_probe);
424}
425#else
426static inline int memory_probe_init(void)
427{
428	return 0;
429}
430#endif
431
432#ifdef CONFIG_MEMORY_FAILURE
433/*
434 * Support for offlining pages of memory
435 */
436
437/* Soft offline a page */
438static ssize_t
439store_soft_offline_page(struct device *dev,
440			struct device_attribute *attr,
441			const char *buf, size_t count)
442{
443	int ret;
444	u64 pfn;
445	if (!capable(CAP_SYS_ADMIN))
446		return -EPERM;
447	if (strict_strtoull(buf, 0, &pfn) < 0)
448		return -EINVAL;
449	pfn >>= PAGE_SHIFT;
450	if (!pfn_valid(pfn))
451		return -ENXIO;
452	ret = soft_offline_page(pfn_to_page(pfn), 0);
453	return ret == 0 ? count : ret;
454}
455
456/* Forcibly offline a page, including killing processes. */
457static ssize_t
458store_hard_offline_page(struct device *dev,
459			struct device_attribute *attr,
460			const char *buf, size_t count)
461{
462	int ret;
463	u64 pfn;
464	if (!capable(CAP_SYS_ADMIN))
465		return -EPERM;
466	if (strict_strtoull(buf, 0, &pfn) < 0)
467		return -EINVAL;
468	pfn >>= PAGE_SHIFT;
469	ret = memory_failure(pfn, 0, 0);
470	return ret ? ret : count;
471}
472
473static DEVICE_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
474static DEVICE_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
475
476static __init int memory_fail_init(void)
477{
478	int err;
479
480	err = device_create_file(memory_subsys.dev_root,
481				&dev_attr_soft_offline_page);
482	if (!err)
483		err = device_create_file(memory_subsys.dev_root,
484				&dev_attr_hard_offline_page);
485	return err;
486}
487#else
488static inline int memory_fail_init(void)
489{
490	return 0;
491}
492#endif
493
494/*
495 * Note that phys_device is optional.  It is here to allow for
496 * differentiation between which *physical* devices each
497 * section belongs to...
498 */
499int __weak arch_get_memory_phys_device(unsigned long start_pfn)
500{
501	return 0;
502}
503
504/*
505 * A reference for the returned object is held and the reference for the
506 * hinted object is released.
507 */
508struct memory_block *find_memory_block_hinted(struct mem_section *section,
509					      struct memory_block *hint)
510{
511	int block_id = base_memory_block_id(__section_nr(section));
512	struct device *hintdev = hint ? &hint->dev : NULL;
513	struct device *dev;
514
515	dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
516	if (hint)
517		put_device(&hint->dev);
518	if (!dev)
519		return NULL;
520	return container_of(dev, struct memory_block, dev);
521}
522
523/*
524 * For now, we have a linear search to go find the appropriate
525 * memory_block corresponding to a particular phys_index. If
526 * this gets to be a real problem, we can always use a radix
527 * tree or something here.
528 *
529 * This could be made generic for all device subsystems.
530 */
531struct memory_block *find_memory_block(struct mem_section *section)
532{
533	return find_memory_block_hinted(section, NULL);
534}
535
536static int init_memory_block(struct memory_block **memory,
537			     struct mem_section *section, unsigned long state)
538{
539	struct memory_block *mem;
540	unsigned long start_pfn;
541	int scn_nr;
542	int ret = 0;
543
544	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
545	if (!mem)
546		return -ENOMEM;
547
548	scn_nr = __section_nr(section);
549	mem->start_section_nr =
550			base_memory_block_id(scn_nr) * sections_per_block;
551	mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
552	mem->state = state;
553	mem->section_count++;
554	mutex_init(&mem->state_mutex);
555	start_pfn = section_nr_to_pfn(mem->start_section_nr);
556	mem->phys_device = arch_get_memory_phys_device(start_pfn);
557
558	ret = register_memory(mem);
559	if (!ret)
560		ret = mem_create_simple_file(mem, phys_index);
561	if (!ret)
562		ret = mem_create_simple_file(mem, end_phys_index);
563	if (!ret)
564		ret = mem_create_simple_file(mem, state);
565	if (!ret)
566		ret = mem_create_simple_file(mem, phys_device);
567	if (!ret)
568		ret = mem_create_simple_file(mem, removable);
569
570	*memory = mem;
571	return ret;
572}
573
574static int add_memory_section(int nid, struct mem_section *section,
575			struct memory_block **mem_p,
576			unsigned long state, enum mem_add_context context)
577{
578	struct memory_block *mem = NULL;
579	int scn_nr = __section_nr(section);
580	int ret = 0;
581
582	mutex_lock(&mem_sysfs_mutex);
583
584	if (context == BOOT) {
585		/* same memory block ? */
586		if (mem_p && *mem_p)
587			if (scn_nr >= (*mem_p)->start_section_nr &&
588			    scn_nr <= (*mem_p)->end_section_nr) {
589				mem = *mem_p;
590				kobject_get(&mem->dev.kobj);
591			}
592	} else
593		mem = find_memory_block(section);
594
595	if (mem) {
596		mem->section_count++;
597		kobject_put(&mem->dev.kobj);
598	} else {
599		ret = init_memory_block(&mem, section, state);
600		/* store memory_block pointer for next loop */
601		if (!ret && context == BOOT)
602			if (mem_p)
603				*mem_p = mem;
604	}
605
606	if (!ret) {
607		if (context == HOTPLUG &&
608		    mem->section_count == sections_per_block)
609			ret = register_mem_sect_under_node(mem, nid);
610	}
611
612	mutex_unlock(&mem_sysfs_mutex);
613	return ret;
614}
615
616int remove_memory_block(unsigned long node_id, struct mem_section *section,
617		int phys_device)
618{
619	struct memory_block *mem;
620
621	mutex_lock(&mem_sysfs_mutex);
622	mem = find_memory_block(section);
623	unregister_mem_sect_under_nodes(mem, __section_nr(section));
624
625	mem->section_count--;
626	if (mem->section_count == 0) {
627		mem_remove_simple_file(mem, phys_index);
628		mem_remove_simple_file(mem, end_phys_index);
629		mem_remove_simple_file(mem, state);
630		mem_remove_simple_file(mem, phys_device);
631		mem_remove_simple_file(mem, removable);
632		unregister_memory(mem);
633		kfree(mem);
634	} else
635		kobject_put(&mem->dev.kobj);
636
637	mutex_unlock(&mem_sysfs_mutex);
638	return 0;
639}
640
641/*
642 * need an interface for the VM to add new memory regions,
643 * but without onlining it.
644 */
645int register_new_memory(int nid, struct mem_section *section)
646{
647	return add_memory_section(nid, section, NULL, MEM_OFFLINE, HOTPLUG);
648}
649
650int unregister_memory_section(struct mem_section *section)
651{
652	if (!present_section(section))
653		return -EINVAL;
654
655	return remove_memory_block(0, section, 0);
656}
657
658/*
659 * Initialize the sysfs support for memory devices...
660 */
661int __init memory_dev_init(void)
662{
663	unsigned int i;
664	int ret;
665	int err;
666	unsigned long block_sz;
667	struct memory_block *mem = NULL;
668
669	ret = subsys_system_register(&memory_subsys, NULL);
670	if (ret)
671		goto out;
672
673	block_sz = get_memory_block_size();
674	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
675
676	/*
677	 * Create entries for memory sections that were found
678	 * during boot and have been initialized
679	 */
680	for (i = 0; i < NR_MEM_SECTIONS; i++) {
681		if (!present_section_nr(i))
682			continue;
683		/* don't need to reuse memory_block if only one per block */
684		err = add_memory_section(0, __nr_to_section(i),
685				 (sections_per_block == 1) ? NULL : &mem,
686					 MEM_ONLINE,
687					 BOOT);
688		if (!ret)
689			ret = err;
690	}
691
692	err = memory_probe_init();
693	if (!ret)
694		ret = err;
695	err = memory_fail_init();
696	if (!ret)
697		ret = err;
698	err = block_size_init();
699	if (!ret)
700		ret = err;
701out:
702	if (ret)
703		printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
704	return ret;
705}
706