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 <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}