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 <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
121static ssize_t show_mem_end_phys_index(struct device *dev,
122 struct device_attribute *attr, char *buf)
123{
124 struct memory_block *mem = to_memory_block(dev);
125 unsigned long phys_index;
126
127 phys_index = mem->end_section_nr / sections_per_block;
128 return sprintf(buf, "%08lx\n", phys_index);
129}
130
131/*
132 * Show whether the section of memory is likely to be hot-removable
133 */
134static ssize_t show_mem_removable(struct device *dev,
135 struct device_attribute *attr, char *buf)
136{
137 unsigned long i, pfn;
138 int ret = 1;
139 struct memory_block *mem = to_memory_block(dev);
140
141 for (i = 0; i < sections_per_block; i++) {
142 if (!present_section_nr(mem->start_section_nr + i))
143 continue;
144 pfn = section_nr_to_pfn(mem->start_section_nr + i);
145 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
146 }
147
148 return sprintf(buf, "%d\n", ret);
149}
150
151/*
152 * online, offline, going offline, etc.
153 */
154static ssize_t show_mem_state(struct device *dev,
155 struct device_attribute *attr, char *buf)
156{
157 struct memory_block *mem = to_memory_block(dev);
158 ssize_t len = 0;
159
160 /*
161 * We can probably put these states in a nice little array
162 * so that they're not open-coded
163 */
164 switch (mem->state) {
165 case MEM_ONLINE:
166 len = sprintf(buf, "online\n");
167 break;
168 case MEM_OFFLINE:
169 len = sprintf(buf, "offline\n");
170 break;
171 case MEM_GOING_OFFLINE:
172 len = sprintf(buf, "going-offline\n");
173 break;
174 default:
175 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
176 mem->state);
177 WARN_ON(1);
178 break;
179 }
180
181 return len;
182}
183
184int memory_notify(unsigned long val, void *v)
185{
186 return blocking_notifier_call_chain(&memory_chain, val, v);
187}
188
189int memory_isolate_notify(unsigned long val, void *v)
190{
191 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
192}
193
194/*
195 * The probe routines leave the pages reserved, just as the bootmem code does.
196 * Make sure they're still that way.
197 */
198static bool pages_correctly_reserved(unsigned long start_pfn)
199{
200 int i, j;
201 struct page *page;
202 unsigned long pfn = start_pfn;
203
204 /*
205 * memmap between sections is not contiguous except with
206 * SPARSEMEM_VMEMMAP. We lookup the page once per section
207 * and assume memmap is contiguous within each section
208 */
209 for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
210 if (WARN_ON_ONCE(!pfn_valid(pfn)))
211 return false;
212 page = pfn_to_page(pfn);
213
214 for (j = 0; j < PAGES_PER_SECTION; j++) {
215 if (PageReserved(page + j))
216 continue;
217
218 printk(KERN_WARNING "section number %ld page number %d "
219 "not reserved, was it already online?\n",
220 pfn_to_section_nr(pfn), j);
221
222 return false;
223 }
224 }
225
226 return true;
227}
228
229/*
230 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
231 * OK to have direct references to sparsemem variables in here.
232 */
233static int
234memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
235{
236 unsigned long start_pfn;
237 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
238 struct page *first_page;
239 int ret;
240
241 first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
242 start_pfn = page_to_pfn(first_page);
243
244 switch (action) {
245 case MEM_ONLINE:
246 if (!pages_correctly_reserved(start_pfn))
247 return -EBUSY;
248
249 ret = online_pages(start_pfn, nr_pages, online_type);
250 break;
251 case MEM_OFFLINE:
252 ret = offline_pages(start_pfn, nr_pages);
253 break;
254 default:
255 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
256 "%ld\n", __func__, phys_index, action, action);
257 ret = -EINVAL;
258 }
259
260 return ret;
261}
262
263static int memory_block_change_state(struct memory_block *mem,
264 unsigned long to_state, unsigned long from_state_req)
265{
266 int ret = 0;
267
268 if (mem->state != from_state_req)
269 return -EINVAL;
270
271 if (to_state == MEM_OFFLINE)
272 mem->state = MEM_GOING_OFFLINE;
273
274 ret = memory_block_action(mem->start_section_nr, to_state,
275 mem->online_type);
276
277 mem->state = ret ? from_state_req : to_state;
278
279 return ret;
280}
281
282/* The device lock serializes operations on memory_subsys_[online|offline] */
283static int memory_subsys_online(struct device *dev)
284{
285 struct memory_block *mem = to_memory_block(dev);
286 int ret;
287
288 if (mem->state == MEM_ONLINE)
289 return 0;
290
291 /*
292 * If we are called from store_mem_state(), online_type will be
293 * set >= 0 Otherwise we were called from the device online
294 * attribute and need to set the online_type.
295 */
296 if (mem->online_type < 0)
297 mem->online_type = ONLINE_KEEP;
298
299 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
300
301 /* clear online_type */
302 mem->online_type = -1;
303
304 return ret;
305}
306
307static int memory_subsys_offline(struct device *dev)
308{
309 struct memory_block *mem = to_memory_block(dev);
310
311 if (mem->state == MEM_OFFLINE)
312 return 0;
313
314 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
315}
316
317static ssize_t
318store_mem_state(struct device *dev,
319 struct device_attribute *attr, const char *buf, size_t count)
320{
321 struct memory_block *mem = to_memory_block(dev);
322 int ret, online_type;
323
324 ret = lock_device_hotplug_sysfs();
325 if (ret)
326 return ret;
327
328 if (!strncmp(buf, "online_kernel", min_t(int, count, 13)))
329 online_type = ONLINE_KERNEL;
330 else if (!strncmp(buf, "online_movable", min_t(int, count, 14)))
331 online_type = ONLINE_MOVABLE;
332 else if (!strncmp(buf, "online", min_t(int, count, 6)))
333 online_type = ONLINE_KEEP;
334 else if (!strncmp(buf, "offline", min_t(int, count, 7)))
335 online_type = -1;
336 else {
337 ret = -EINVAL;
338 goto err;
339 }
340
341 switch (online_type) {
342 case ONLINE_KERNEL:
343 case ONLINE_MOVABLE:
344 case ONLINE_KEEP:
345 /*
346 * mem->online_type is not protected so there can be a
347 * race here. However, when racing online, the first
348 * will succeed and the second will just return as the
349 * block will already be online. The online type
350 * could be either one, but that is expected.
351 */
352 mem->online_type = online_type;
353 ret = device_online(&mem->dev);
354 break;
355 case -1:
356 ret = device_offline(&mem->dev);
357 break;
358 default:
359 ret = -EINVAL; /* should never happen */
360 }
361
362err:
363 unlock_device_hotplug();
364
365 if (ret)
366 return ret;
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
386static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
387static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
388static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
389static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
390static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
391
392/*
393 * Block size attribute stuff
394 */
395static ssize_t
396print_block_size(struct device *dev, struct device_attribute *attr,
397 char *buf)
398{
399 return sprintf(buf, "%lx\n", get_memory_block_size());
400}
401
402static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
403
404/*
405 * Some architectures will have custom drivers to do this, and
406 * will not need to do it from userspace. The fake hot-add code
407 * as well as ppc64 will do all of their discovery in userspace
408 * and will require this interface.
409 */
410#ifdef CONFIG_ARCH_MEMORY_PROBE
411static ssize_t
412memory_probe_store(struct device *dev, struct device_attribute *attr,
413 const char *buf, size_t count)
414{
415 u64 phys_addr;
416 int nid;
417 int i, ret;
418 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
419
420 phys_addr = simple_strtoull(buf, NULL, 0);
421
422 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
423 return -EINVAL;
424
425 for (i = 0; i < sections_per_block; i++) {
426 nid = memory_add_physaddr_to_nid(phys_addr);
427 ret = add_memory(nid, phys_addr,
428 PAGES_PER_SECTION << PAGE_SHIFT);
429 if (ret)
430 goto out;
431
432 phys_addr += MIN_MEMORY_BLOCK_SIZE;
433 }
434
435 ret = count;
436out:
437 return ret;
438}
439
440static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
441#endif
442
443#ifdef CONFIG_MEMORY_FAILURE
444/*
445 * Support for offlining pages of memory
446 */
447
448/* Soft offline a page */
449static ssize_t
450store_soft_offline_page(struct device *dev,
451 struct device_attribute *attr,
452 const char *buf, size_t count)
453{
454 int ret;
455 u64 pfn;
456 if (!capable(CAP_SYS_ADMIN))
457 return -EPERM;
458 if (kstrtoull(buf, 0, &pfn) < 0)
459 return -EINVAL;
460 pfn >>= PAGE_SHIFT;
461 if (!pfn_valid(pfn))
462 return -ENXIO;
463 ret = soft_offline_page(pfn_to_page(pfn), 0);
464 return ret == 0 ? count : ret;
465}
466
467/* Forcibly offline a page, including killing processes. */
468static ssize_t
469store_hard_offline_page(struct device *dev,
470 struct device_attribute *attr,
471 const char *buf, size_t count)
472{
473 int ret;
474 u64 pfn;
475 if (!capable(CAP_SYS_ADMIN))
476 return -EPERM;
477 if (kstrtoull(buf, 0, &pfn) < 0)
478 return -EINVAL;
479 pfn >>= PAGE_SHIFT;
480 ret = memory_failure(pfn, 0, 0);
481 return ret ? ret : count;
482}
483
484static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
485static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
486#endif
487
488/*
489 * Note that phys_device is optional. It is here to allow for
490 * differentiation between which *physical* devices each
491 * section belongs to...
492 */
493int __weak arch_get_memory_phys_device(unsigned long start_pfn)
494{
495 return 0;
496}
497
498/*
499 * A reference for the returned object is held and the reference for the
500 * hinted object is released.
501 */
502struct memory_block *find_memory_block_hinted(struct mem_section *section,
503 struct memory_block *hint)
504{
505 int block_id = base_memory_block_id(__section_nr(section));
506 struct device *hintdev = hint ? &hint->dev : NULL;
507 struct device *dev;
508
509 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
510 if (hint)
511 put_device(&hint->dev);
512 if (!dev)
513 return NULL;
514 return to_memory_block(dev);
515}
516
517/*
518 * For now, we have a linear search to go find the appropriate
519 * memory_block corresponding to a particular phys_index. If
520 * this gets to be a real problem, we can always use a radix
521 * tree or something here.
522 *
523 * This could be made generic for all device subsystems.
524 */
525struct memory_block *find_memory_block(struct mem_section *section)
526{
527 return find_memory_block_hinted(section, NULL);
528}
529
530static struct attribute *memory_memblk_attrs[] = {
531 &dev_attr_phys_index.attr,
532 &dev_attr_end_phys_index.attr,
533 &dev_attr_state.attr,
534 &dev_attr_phys_device.attr,
535 &dev_attr_removable.attr,
536 NULL
537};
538
539static struct attribute_group memory_memblk_attr_group = {
540 .attrs = memory_memblk_attrs,
541};
542
543static const struct attribute_group *memory_memblk_attr_groups[] = {
544 &memory_memblk_attr_group,
545 NULL,
546};
547
548/*
549 * register_memory - Setup a sysfs device for a memory block
550 */
551static
552int register_memory(struct memory_block *memory)
553{
554 memory->dev.bus = &memory_subsys;
555 memory->dev.id = memory->start_section_nr / sections_per_block;
556 memory->dev.release = memory_block_release;
557 memory->dev.groups = memory_memblk_attr_groups;
558 memory->dev.offline = memory->state == MEM_OFFLINE;
559
560 return device_register(&memory->dev);
561}
562
563static int init_memory_block(struct memory_block **memory,
564 struct mem_section *section, unsigned long state)
565{
566 struct memory_block *mem;
567 unsigned long start_pfn;
568 int scn_nr;
569 int ret = 0;
570
571 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
572 if (!mem)
573 return -ENOMEM;
574
575 scn_nr = __section_nr(section);
576 mem->start_section_nr =
577 base_memory_block_id(scn_nr) * sections_per_block;
578 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
579 mem->state = state;
580 mem->section_count++;
581 start_pfn = section_nr_to_pfn(mem->start_section_nr);
582 mem->phys_device = arch_get_memory_phys_device(start_pfn);
583
584 ret = register_memory(mem);
585
586 *memory = mem;
587 return ret;
588}
589
590static int add_memory_block(int base_section_nr)
591{
592 struct memory_block *mem;
593 int i, ret, section_count = 0, section_nr;
594
595 for (i = base_section_nr;
596 (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
597 i++) {
598 if (!present_section_nr(i))
599 continue;
600 if (section_count == 0)
601 section_nr = i;
602 section_count++;
603 }
604
605 if (section_count == 0)
606 return 0;
607 ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
608 if (ret)
609 return ret;
610 mem->section_count = section_count;
611 return 0;
612}
613
614
615/*
616 * need an interface for the VM to add new memory regions,
617 * but without onlining it.
618 */
619int register_new_memory(int nid, struct mem_section *section)
620{
621 int ret = 0;
622 struct memory_block *mem;
623
624 mutex_lock(&mem_sysfs_mutex);
625
626 mem = find_memory_block(section);
627 if (mem) {
628 mem->section_count++;
629 put_device(&mem->dev);
630 } else {
631 ret = init_memory_block(&mem, section, MEM_OFFLINE);
632 if (ret)
633 goto out;
634 }
635
636 if (mem->section_count == sections_per_block)
637 ret = register_mem_sect_under_node(mem, nid);
638out:
639 mutex_unlock(&mem_sysfs_mutex);
640 return ret;
641}
642
643#ifdef CONFIG_MEMORY_HOTREMOVE
644static void
645unregister_memory(struct memory_block *memory)
646{
647 BUG_ON(memory->dev.bus != &memory_subsys);
648
649 /* drop the ref. we got in remove_memory_block() */
650 put_device(&memory->dev);
651 device_unregister(&memory->dev);
652}
653
654static int remove_memory_block(unsigned long node_id,
655 struct mem_section *section, int phys_device)
656{
657 struct memory_block *mem;
658
659 mutex_lock(&mem_sysfs_mutex);
660 mem = find_memory_block(section);
661 unregister_mem_sect_under_nodes(mem, __section_nr(section));
662
663 mem->section_count--;
664 if (mem->section_count == 0)
665 unregister_memory(mem);
666 else
667 put_device(&mem->dev);
668
669 mutex_unlock(&mem_sysfs_mutex);
670 return 0;
671}
672
673int unregister_memory_section(struct mem_section *section)
674{
675 if (!present_section(section))
676 return -EINVAL;
677
678 return remove_memory_block(0, section, 0);
679}
680#endif /* CONFIG_MEMORY_HOTREMOVE */
681
682/* return true if the memory block is offlined, otherwise, return false */
683bool is_memblock_offlined(struct memory_block *mem)
684{
685 return mem->state == MEM_OFFLINE;
686}
687
688static struct attribute *memory_root_attrs[] = {
689#ifdef CONFIG_ARCH_MEMORY_PROBE
690 &dev_attr_probe.attr,
691#endif
692
693#ifdef CONFIG_MEMORY_FAILURE
694 &dev_attr_soft_offline_page.attr,
695 &dev_attr_hard_offline_page.attr,
696#endif
697
698 &dev_attr_block_size_bytes.attr,
699 NULL
700};
701
702static struct attribute_group memory_root_attr_group = {
703 .attrs = memory_root_attrs,
704};
705
706static const struct attribute_group *memory_root_attr_groups[] = {
707 &memory_root_attr_group,
708 NULL,
709};
710
711/*
712 * Initialize the sysfs support for memory devices...
713 */
714int __init memory_dev_init(void)
715{
716 unsigned int i;
717 int ret;
718 int err;
719 unsigned long block_sz;
720
721 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
722 if (ret)
723 goto out;
724
725 block_sz = get_memory_block_size();
726 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
727
728 /*
729 * Create entries for memory sections that were found
730 * during boot and have been initialized
731 */
732 mutex_lock(&mem_sysfs_mutex);
733 for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
734 err = add_memory_block(i);
735 if (!ret)
736 ret = err;
737 }
738 mutex_unlock(&mem_sysfs_mutex);
739
740out:
741 if (ret)
742 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
743 return ret;
744}
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
254int 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 * Memory auto online policy.
443 */
444
445static ssize_t
446show_auto_online_blocks(struct device *dev, struct device_attribute *attr,
447 char *buf)
448{
449 if (memhp_auto_online)
450 return sprintf(buf, "online\n");
451 else
452 return sprintf(buf, "offline\n");
453}
454
455static ssize_t
456store_auto_online_blocks(struct device *dev, struct device_attribute *attr,
457 const char *buf, size_t count)
458{
459 if (sysfs_streq(buf, "online"))
460 memhp_auto_online = true;
461 else if (sysfs_streq(buf, "offline"))
462 memhp_auto_online = false;
463 else
464 return -EINVAL;
465
466 return count;
467}
468
469static DEVICE_ATTR(auto_online_blocks, 0644, show_auto_online_blocks,
470 store_auto_online_blocks);
471
472/*
473 * Some architectures will have custom drivers to do this, and
474 * will not need to do it from userspace. The fake hot-add code
475 * as well as ppc64 will do all of their discovery in userspace
476 * and will require this interface.
477 */
478#ifdef CONFIG_ARCH_MEMORY_PROBE
479static ssize_t
480memory_probe_store(struct device *dev, struct device_attribute *attr,
481 const char *buf, size_t count)
482{
483 u64 phys_addr;
484 int nid, ret;
485 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
486
487 ret = kstrtoull(buf, 0, &phys_addr);
488 if (ret)
489 return ret;
490
491 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
492 return -EINVAL;
493
494 nid = memory_add_physaddr_to_nid(phys_addr);
495 ret = add_memory(nid, phys_addr,
496 MIN_MEMORY_BLOCK_SIZE * sections_per_block);
497
498 if (ret)
499 goto out;
500
501 ret = count;
502out:
503 return ret;
504}
505
506static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
507#endif
508
509#ifdef CONFIG_MEMORY_FAILURE
510/*
511 * Support for offlining pages of memory
512 */
513
514/* Soft offline a page */
515static ssize_t
516store_soft_offline_page(struct device *dev,
517 struct device_attribute *attr,
518 const char *buf, size_t count)
519{
520 int ret;
521 u64 pfn;
522 if (!capable(CAP_SYS_ADMIN))
523 return -EPERM;
524 if (kstrtoull(buf, 0, &pfn) < 0)
525 return -EINVAL;
526 pfn >>= PAGE_SHIFT;
527 if (!pfn_valid(pfn))
528 return -ENXIO;
529 ret = soft_offline_page(pfn_to_page(pfn), 0);
530 return ret == 0 ? count : ret;
531}
532
533/* Forcibly offline a page, including killing processes. */
534static ssize_t
535store_hard_offline_page(struct device *dev,
536 struct device_attribute *attr,
537 const char *buf, size_t count)
538{
539 int ret;
540 u64 pfn;
541 if (!capable(CAP_SYS_ADMIN))
542 return -EPERM;
543 if (kstrtoull(buf, 0, &pfn) < 0)
544 return -EINVAL;
545 pfn >>= PAGE_SHIFT;
546 ret = memory_failure(pfn, 0, 0);
547 return ret ? ret : count;
548}
549
550static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
551static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
552#endif
553
554/*
555 * Note that phys_device is optional. It is here to allow for
556 * differentiation between which *physical* devices each
557 * section belongs to...
558 */
559int __weak arch_get_memory_phys_device(unsigned long start_pfn)
560{
561 return 0;
562}
563
564/*
565 * A reference for the returned object is held and the reference for the
566 * hinted object is released.
567 */
568struct memory_block *find_memory_block_hinted(struct mem_section *section,
569 struct memory_block *hint)
570{
571 int block_id = base_memory_block_id(__section_nr(section));
572 struct device *hintdev = hint ? &hint->dev : NULL;
573 struct device *dev;
574
575 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
576 if (hint)
577 put_device(&hint->dev);
578 if (!dev)
579 return NULL;
580 return to_memory_block(dev);
581}
582
583/*
584 * For now, we have a linear search to go find the appropriate
585 * memory_block corresponding to a particular phys_index. If
586 * this gets to be a real problem, we can always use a radix
587 * tree or something here.
588 *
589 * This could be made generic for all device subsystems.
590 */
591struct memory_block *find_memory_block(struct mem_section *section)
592{
593 return find_memory_block_hinted(section, NULL);
594}
595
596static struct attribute *memory_memblk_attrs[] = {
597 &dev_attr_phys_index.attr,
598 &dev_attr_state.attr,
599 &dev_attr_phys_device.attr,
600 &dev_attr_removable.attr,
601#ifdef CONFIG_MEMORY_HOTREMOVE
602 &dev_attr_valid_zones.attr,
603#endif
604 NULL
605};
606
607static struct attribute_group memory_memblk_attr_group = {
608 .attrs = memory_memblk_attrs,
609};
610
611static const struct attribute_group *memory_memblk_attr_groups[] = {
612 &memory_memblk_attr_group,
613 NULL,
614};
615
616/*
617 * register_memory - Setup a sysfs device for a memory block
618 */
619static
620int register_memory(struct memory_block *memory)
621{
622 memory->dev.bus = &memory_subsys;
623 memory->dev.id = memory->start_section_nr / sections_per_block;
624 memory->dev.release = memory_block_release;
625 memory->dev.groups = memory_memblk_attr_groups;
626 memory->dev.offline = memory->state == MEM_OFFLINE;
627
628 return device_register(&memory->dev);
629}
630
631static int init_memory_block(struct memory_block **memory,
632 struct mem_section *section, unsigned long state)
633{
634 struct memory_block *mem;
635 unsigned long start_pfn;
636 int scn_nr;
637 int ret = 0;
638
639 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
640 if (!mem)
641 return -ENOMEM;
642
643 scn_nr = __section_nr(section);
644 mem->start_section_nr =
645 base_memory_block_id(scn_nr) * sections_per_block;
646 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
647 mem->state = state;
648 start_pfn = section_nr_to_pfn(mem->start_section_nr);
649 mem->phys_device = arch_get_memory_phys_device(start_pfn);
650
651 ret = register_memory(mem);
652
653 *memory = mem;
654 return ret;
655}
656
657static int add_memory_block(int base_section_nr)
658{
659 struct memory_block *mem;
660 int i, ret, section_count = 0, section_nr;
661
662 for (i = base_section_nr;
663 (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
664 i++) {
665 if (!present_section_nr(i))
666 continue;
667 if (section_count == 0)
668 section_nr = i;
669 section_count++;
670 }
671
672 if (section_count == 0)
673 return 0;
674 ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
675 if (ret)
676 return ret;
677 mem->section_count = section_count;
678 return 0;
679}
680
681static bool is_zone_device_section(struct mem_section *ms)
682{
683 struct page *page;
684
685 page = sparse_decode_mem_map(ms->section_mem_map, __section_nr(ms));
686 return is_zone_device_page(page);
687}
688
689/*
690 * need an interface for the VM to add new memory regions,
691 * but without onlining it.
692 */
693int register_new_memory(int nid, struct mem_section *section)
694{
695 int ret = 0;
696 struct memory_block *mem;
697
698 if (is_zone_device_section(section))
699 return 0;
700
701 mutex_lock(&mem_sysfs_mutex);
702
703 mem = find_memory_block(section);
704 if (mem) {
705 mem->section_count++;
706 put_device(&mem->dev);
707 } else {
708 ret = init_memory_block(&mem, section, MEM_OFFLINE);
709 if (ret)
710 goto out;
711 mem->section_count++;
712 }
713
714 if (mem->section_count == sections_per_block)
715 ret = register_mem_sect_under_node(mem, nid);
716out:
717 mutex_unlock(&mem_sysfs_mutex);
718 return ret;
719}
720
721#ifdef CONFIG_MEMORY_HOTREMOVE
722static void
723unregister_memory(struct memory_block *memory)
724{
725 BUG_ON(memory->dev.bus != &memory_subsys);
726
727 /* drop the ref. we got in remove_memory_block() */
728 put_device(&memory->dev);
729 device_unregister(&memory->dev);
730}
731
732static int remove_memory_section(unsigned long node_id,
733 struct mem_section *section, int phys_device)
734{
735 struct memory_block *mem;
736
737 if (is_zone_device_section(section))
738 return 0;
739
740 mutex_lock(&mem_sysfs_mutex);
741 mem = find_memory_block(section);
742 unregister_mem_sect_under_nodes(mem, __section_nr(section));
743
744 mem->section_count--;
745 if (mem->section_count == 0)
746 unregister_memory(mem);
747 else
748 put_device(&mem->dev);
749
750 mutex_unlock(&mem_sysfs_mutex);
751 return 0;
752}
753
754int unregister_memory_section(struct mem_section *section)
755{
756 if (!present_section(section))
757 return -EINVAL;
758
759 return remove_memory_section(0, section, 0);
760}
761#endif /* CONFIG_MEMORY_HOTREMOVE */
762
763/* return true if the memory block is offlined, otherwise, return false */
764bool is_memblock_offlined(struct memory_block *mem)
765{
766 return mem->state == MEM_OFFLINE;
767}
768
769static struct attribute *memory_root_attrs[] = {
770#ifdef CONFIG_ARCH_MEMORY_PROBE
771 &dev_attr_probe.attr,
772#endif
773
774#ifdef CONFIG_MEMORY_FAILURE
775 &dev_attr_soft_offline_page.attr,
776 &dev_attr_hard_offline_page.attr,
777#endif
778
779 &dev_attr_block_size_bytes.attr,
780 &dev_attr_auto_online_blocks.attr,
781 NULL
782};
783
784static struct attribute_group memory_root_attr_group = {
785 .attrs = memory_root_attrs,
786};
787
788static const struct attribute_group *memory_root_attr_groups[] = {
789 &memory_root_attr_group,
790 NULL,
791};
792
793/*
794 * Initialize the sysfs support for memory devices...
795 */
796int __init memory_dev_init(void)
797{
798 unsigned int i;
799 int ret;
800 int err;
801 unsigned long block_sz;
802
803 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
804 if (ret)
805 goto out;
806
807 block_sz = get_memory_block_size();
808 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
809
810 /*
811 * Create entries for memory sections that were found
812 * during boot and have been initialized
813 */
814 mutex_lock(&mem_sysfs_mutex);
815 for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
816 err = add_memory_block(i);
817 if (!ret)
818 ret = err;
819 }
820 mutex_unlock(&mem_sysfs_mutex);
821
822out:
823 if (ret)
824 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
825 return ret;
826}