Loading...
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 <linux/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 int ret;
230
231 start_pfn = section_nr_to_pfn(phys_index);
232
233 switch (action) {
234 case MEM_ONLINE:
235 if (!pages_correctly_reserved(start_pfn))
236 return -EBUSY;
237
238 ret = online_pages(start_pfn, nr_pages, online_type);
239 break;
240 case MEM_OFFLINE:
241 ret = offline_pages(start_pfn, nr_pages);
242 break;
243 default:
244 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
245 "%ld\n", __func__, phys_index, action, action);
246 ret = -EINVAL;
247 }
248
249 return ret;
250}
251
252int memory_block_change_state(struct memory_block *mem,
253 unsigned long to_state, unsigned long from_state_req)
254{
255 int ret = 0;
256
257 if (mem->state != from_state_req)
258 return -EINVAL;
259
260 if (to_state == MEM_OFFLINE)
261 mem->state = MEM_GOING_OFFLINE;
262
263 ret = memory_block_action(mem->start_section_nr, to_state,
264 mem->online_type);
265
266 mem->state = ret ? from_state_req : to_state;
267
268 return ret;
269}
270
271/* The device lock serializes operations on memory_subsys_[online|offline] */
272static int memory_subsys_online(struct device *dev)
273{
274 struct memory_block *mem = to_memory_block(dev);
275 int ret;
276
277 if (mem->state == MEM_ONLINE)
278 return 0;
279
280 /*
281 * If we are called from store_mem_state(), online_type will be
282 * set >= 0 Otherwise we were called from the device online
283 * attribute and need to set the online_type.
284 */
285 if (mem->online_type < 0)
286 mem->online_type = MMOP_ONLINE_KEEP;
287
288 /* Already under protection of mem_hotplug_begin() */
289 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
290
291 /* clear online_type */
292 mem->online_type = -1;
293
294 return ret;
295}
296
297static int memory_subsys_offline(struct device *dev)
298{
299 struct memory_block *mem = to_memory_block(dev);
300
301 if (mem->state == MEM_OFFLINE)
302 return 0;
303
304 /* Can't offline block with non-present sections */
305 if (mem->section_count != sections_per_block)
306 return -EINVAL;
307
308 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
309}
310
311static ssize_t
312store_mem_state(struct device *dev,
313 struct device_attribute *attr, const char *buf, size_t count)
314{
315 struct memory_block *mem = to_memory_block(dev);
316 int ret, online_type;
317
318 ret = lock_device_hotplug_sysfs();
319 if (ret)
320 return ret;
321
322 if (sysfs_streq(buf, "online_kernel"))
323 online_type = MMOP_ONLINE_KERNEL;
324 else if (sysfs_streq(buf, "online_movable"))
325 online_type = MMOP_ONLINE_MOVABLE;
326 else if (sysfs_streq(buf, "online"))
327 online_type = MMOP_ONLINE_KEEP;
328 else if (sysfs_streq(buf, "offline"))
329 online_type = MMOP_OFFLINE;
330 else {
331 ret = -EINVAL;
332 goto err;
333 }
334
335 /*
336 * Memory hotplug needs to hold mem_hotplug_begin() for probe to find
337 * the correct memory block to online before doing device_online(dev),
338 * which will take dev->mutex. Take the lock early to prevent an
339 * inversion, memory_subsys_online() callbacks will be implemented by
340 * assuming it's already protected.
341 */
342 mem_hotplug_begin();
343
344 switch (online_type) {
345 case MMOP_ONLINE_KERNEL:
346 case MMOP_ONLINE_MOVABLE:
347 case MMOP_ONLINE_KEEP:
348 mem->online_type = online_type;
349 ret = device_online(&mem->dev);
350 break;
351 case MMOP_OFFLINE:
352 ret = device_offline(&mem->dev);
353 break;
354 default:
355 ret = -EINVAL; /* should never happen */
356 }
357
358 mem_hotplug_done();
359err:
360 unlock_device_hotplug();
361
362 if (ret < 0)
363 return ret;
364 if (ret)
365 return -EINVAL;
366
367 return count;
368}
369
370/*
371 * phys_device is a bad name for this. What I really want
372 * is a way to differentiate between memory ranges that
373 * are part of physical devices that constitute
374 * a complete removable unit or fru.
375 * i.e. do these ranges belong to the same physical device,
376 * s.t. if I offline all of these sections I can then
377 * remove the physical device?
378 */
379static ssize_t show_phys_device(struct device *dev,
380 struct device_attribute *attr, char *buf)
381{
382 struct memory_block *mem = to_memory_block(dev);
383 return sprintf(buf, "%d\n", mem->phys_device);
384}
385
386#ifdef CONFIG_MEMORY_HOTREMOVE
387static ssize_t show_valid_zones(struct device *dev,
388 struct device_attribute *attr, char *buf)
389{
390 struct memory_block *mem = to_memory_block(dev);
391 unsigned long start_pfn, end_pfn;
392 unsigned long valid_start, valid_end, valid_pages;
393 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
394 struct zone *zone;
395 int zone_shift = 0;
396
397 start_pfn = section_nr_to_pfn(mem->start_section_nr);
398 end_pfn = start_pfn + nr_pages;
399
400 /* The block contains more than one zone can not be offlined. */
401 if (!test_pages_in_a_zone(start_pfn, end_pfn, &valid_start, &valid_end))
402 return sprintf(buf, "none\n");
403
404 zone = page_zone(pfn_to_page(valid_start));
405 valid_pages = valid_end - valid_start;
406
407 /* MMOP_ONLINE_KEEP */
408 sprintf(buf, "%s", zone->name);
409
410 /* MMOP_ONLINE_KERNEL */
411 zone_can_shift(valid_start, valid_pages, ZONE_NORMAL, &zone_shift);
412 if (zone_shift) {
413 strcat(buf, " ");
414 strcat(buf, (zone + zone_shift)->name);
415 }
416
417 /* MMOP_ONLINE_MOVABLE */
418 zone_can_shift(valid_start, valid_pages, ZONE_MOVABLE, &zone_shift);
419 if (zone_shift) {
420 strcat(buf, " ");
421 strcat(buf, (zone + zone_shift)->name);
422 }
423
424 strcat(buf, "\n");
425
426 return strlen(buf);
427}
428static DEVICE_ATTR(valid_zones, 0444, show_valid_zones, NULL);
429#endif
430
431static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
432static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
433static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
434static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
435
436/*
437 * Block size attribute stuff
438 */
439static ssize_t
440print_block_size(struct device *dev, struct device_attribute *attr,
441 char *buf)
442{
443 return sprintf(buf, "%lx\n", get_memory_block_size());
444}
445
446static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
447
448/*
449 * Memory auto online policy.
450 */
451
452static ssize_t
453show_auto_online_blocks(struct device *dev, struct device_attribute *attr,
454 char *buf)
455{
456 if (memhp_auto_online)
457 return sprintf(buf, "online\n");
458 else
459 return sprintf(buf, "offline\n");
460}
461
462static ssize_t
463store_auto_online_blocks(struct device *dev, struct device_attribute *attr,
464 const char *buf, size_t count)
465{
466 if (sysfs_streq(buf, "online"))
467 memhp_auto_online = true;
468 else if (sysfs_streq(buf, "offline"))
469 memhp_auto_online = false;
470 else
471 return -EINVAL;
472
473 return count;
474}
475
476static DEVICE_ATTR(auto_online_blocks, 0644, show_auto_online_blocks,
477 store_auto_online_blocks);
478
479/*
480 * Some architectures will have custom drivers to do this, and
481 * will not need to do it from userspace. The fake hot-add code
482 * as well as ppc64 will do all of their discovery in userspace
483 * and will require this interface.
484 */
485#ifdef CONFIG_ARCH_MEMORY_PROBE
486static ssize_t
487memory_probe_store(struct device *dev, struct device_attribute *attr,
488 const char *buf, size_t count)
489{
490 u64 phys_addr;
491 int nid, ret;
492 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
493
494 ret = kstrtoull(buf, 0, &phys_addr);
495 if (ret)
496 return ret;
497
498 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
499 return -EINVAL;
500
501 nid = memory_add_physaddr_to_nid(phys_addr);
502 ret = add_memory(nid, phys_addr,
503 MIN_MEMORY_BLOCK_SIZE * sections_per_block);
504
505 if (ret)
506 goto out;
507
508 ret = count;
509out:
510 return ret;
511}
512
513static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
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
523store_soft_offline_page(struct device *dev,
524 struct device_attribute *attr,
525 const char *buf, size_t count)
526{
527 int ret;
528 u64 pfn;
529 if (!capable(CAP_SYS_ADMIN))
530 return -EPERM;
531 if (kstrtoull(buf, 0, &pfn) < 0)
532 return -EINVAL;
533 pfn >>= PAGE_SHIFT;
534 if (!pfn_valid(pfn))
535 return -ENXIO;
536 ret = soft_offline_page(pfn_to_page(pfn), 0);
537 return ret == 0 ? count : ret;
538}
539
540/* Forcibly offline a page, including killing processes. */
541static ssize_t
542store_hard_offline_page(struct device *dev,
543 struct device_attribute *attr,
544 const char *buf, size_t count)
545{
546 int ret;
547 u64 pfn;
548 if (!capable(CAP_SYS_ADMIN))
549 return -EPERM;
550 if (kstrtoull(buf, 0, &pfn) < 0)
551 return -EINVAL;
552 pfn >>= PAGE_SHIFT;
553 ret = memory_failure(pfn, 0, 0);
554 return ret ? ret : count;
555}
556
557static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
558static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
559#endif
560
561/*
562 * Note that phys_device is optional. It is here to allow for
563 * differentiation between which *physical* devices each
564 * section belongs to...
565 */
566int __weak arch_get_memory_phys_device(unsigned long start_pfn)
567{
568 return 0;
569}
570
571/*
572 * A reference for the returned object is held and the reference for the
573 * hinted object is released.
574 */
575struct memory_block *find_memory_block_hinted(struct mem_section *section,
576 struct memory_block *hint)
577{
578 int block_id = base_memory_block_id(__section_nr(section));
579 struct device *hintdev = hint ? &hint->dev : NULL;
580 struct device *dev;
581
582 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
583 if (hint)
584 put_device(&hint->dev);
585 if (!dev)
586 return NULL;
587 return to_memory_block(dev);
588}
589
590/*
591 * For now, we have a linear search to go find the appropriate
592 * memory_block corresponding to a particular phys_index. If
593 * this gets to be a real problem, we can always use a radix
594 * tree or something here.
595 *
596 * This could be made generic for all device subsystems.
597 */
598struct memory_block *find_memory_block(struct mem_section *section)
599{
600 return find_memory_block_hinted(section, NULL);
601}
602
603static struct attribute *memory_memblk_attrs[] = {
604 &dev_attr_phys_index.attr,
605 &dev_attr_state.attr,
606 &dev_attr_phys_device.attr,
607 &dev_attr_removable.attr,
608#ifdef CONFIG_MEMORY_HOTREMOVE
609 &dev_attr_valid_zones.attr,
610#endif
611 NULL
612};
613
614static struct attribute_group memory_memblk_attr_group = {
615 .attrs = memory_memblk_attrs,
616};
617
618static const struct attribute_group *memory_memblk_attr_groups[] = {
619 &memory_memblk_attr_group,
620 NULL,
621};
622
623/*
624 * register_memory - Setup a sysfs device for a memory block
625 */
626static
627int register_memory(struct memory_block *memory)
628{
629 memory->dev.bus = &memory_subsys;
630 memory->dev.id = memory->start_section_nr / sections_per_block;
631 memory->dev.release = memory_block_release;
632 memory->dev.groups = memory_memblk_attr_groups;
633 memory->dev.offline = memory->state == MEM_OFFLINE;
634
635 return device_register(&memory->dev);
636}
637
638static int init_memory_block(struct memory_block **memory,
639 struct mem_section *section, unsigned long state)
640{
641 struct memory_block *mem;
642 unsigned long start_pfn;
643 int scn_nr;
644 int ret = 0;
645
646 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
647 if (!mem)
648 return -ENOMEM;
649
650 scn_nr = __section_nr(section);
651 mem->start_section_nr =
652 base_memory_block_id(scn_nr) * sections_per_block;
653 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
654 mem->state = state;
655 start_pfn = section_nr_to_pfn(mem->start_section_nr);
656 mem->phys_device = arch_get_memory_phys_device(start_pfn);
657
658 ret = register_memory(mem);
659
660 *memory = mem;
661 return ret;
662}
663
664static int add_memory_block(int base_section_nr)
665{
666 struct memory_block *mem;
667 int i, ret, section_count = 0, section_nr;
668
669 for (i = base_section_nr;
670 (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
671 i++) {
672 if (!present_section_nr(i))
673 continue;
674 if (section_count == 0)
675 section_nr = i;
676 section_count++;
677 }
678
679 if (section_count == 0)
680 return 0;
681 ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
682 if (ret)
683 return ret;
684 mem->section_count = section_count;
685 return 0;
686}
687
688static bool is_zone_device_section(struct mem_section *ms)
689{
690 struct page *page;
691
692 page = sparse_decode_mem_map(ms->section_mem_map, __section_nr(ms));
693 return is_zone_device_page(page);
694}
695
696/*
697 * need an interface for the VM to add new memory regions,
698 * but without onlining it.
699 */
700int register_new_memory(int nid, struct mem_section *section)
701{
702 int ret = 0;
703 struct memory_block *mem;
704
705 if (is_zone_device_section(section))
706 return 0;
707
708 mutex_lock(&mem_sysfs_mutex);
709
710 mem = find_memory_block(section);
711 if (mem) {
712 mem->section_count++;
713 put_device(&mem->dev);
714 } else {
715 ret = init_memory_block(&mem, section, MEM_OFFLINE);
716 if (ret)
717 goto out;
718 mem->section_count++;
719 }
720
721 if (mem->section_count == sections_per_block)
722 ret = register_mem_sect_under_node(mem, nid);
723out:
724 mutex_unlock(&mem_sysfs_mutex);
725 return ret;
726}
727
728#ifdef CONFIG_MEMORY_HOTREMOVE
729static void
730unregister_memory(struct memory_block *memory)
731{
732 BUG_ON(memory->dev.bus != &memory_subsys);
733
734 /* drop the ref. we got in remove_memory_block() */
735 put_device(&memory->dev);
736 device_unregister(&memory->dev);
737}
738
739static int remove_memory_section(unsigned long node_id,
740 struct mem_section *section, int phys_device)
741{
742 struct memory_block *mem;
743
744 if (is_zone_device_section(section))
745 return 0;
746
747 mutex_lock(&mem_sysfs_mutex);
748 mem = find_memory_block(section);
749 unregister_mem_sect_under_nodes(mem, __section_nr(section));
750
751 mem->section_count--;
752 if (mem->section_count == 0)
753 unregister_memory(mem);
754 else
755 put_device(&mem->dev);
756
757 mutex_unlock(&mem_sysfs_mutex);
758 return 0;
759}
760
761int unregister_memory_section(struct mem_section *section)
762{
763 if (!present_section(section))
764 return -EINVAL;
765
766 return remove_memory_section(0, section, 0);
767}
768#endif /* CONFIG_MEMORY_HOTREMOVE */
769
770/* return true if the memory block is offlined, otherwise, return false */
771bool is_memblock_offlined(struct memory_block *mem)
772{
773 return mem->state == MEM_OFFLINE;
774}
775
776static struct attribute *memory_root_attrs[] = {
777#ifdef CONFIG_ARCH_MEMORY_PROBE
778 &dev_attr_probe.attr,
779#endif
780
781#ifdef CONFIG_MEMORY_FAILURE
782 &dev_attr_soft_offline_page.attr,
783 &dev_attr_hard_offline_page.attr,
784#endif
785
786 &dev_attr_block_size_bytes.attr,
787 &dev_attr_auto_online_blocks.attr,
788 NULL
789};
790
791static struct attribute_group memory_root_attr_group = {
792 .attrs = memory_root_attrs,
793};
794
795static const struct attribute_group *memory_root_attr_groups[] = {
796 &memory_root_attr_group,
797 NULL,
798};
799
800/*
801 * Initialize the sysfs support for memory devices...
802 */
803int __init memory_dev_init(void)
804{
805 unsigned int i;
806 int ret;
807 int err;
808 unsigned long block_sz;
809
810 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
811 if (ret)
812 goto out;
813
814 block_sz = get_memory_block_size();
815 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
816
817 /*
818 * Create entries for memory sections that were found
819 * during boot and have been initialized
820 */
821 mutex_lock(&mem_sysfs_mutex);
822 for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
823 err = add_memory_block(i);
824 if (!ret)
825 ret = err;
826 }
827 mutex_unlock(&mem_sysfs_mutex);
828
829out:
830 if (ret)
831 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
832 return ret;
833}
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}