Loading...
1/*
2 * drivers/base/memory.c - basic Memory class 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/sysdev.h>
14#include <linux/module.h>
15#include <linux/init.h>
16#include <linux/topology.h>
17#include <linux/capability.h>
18#include <linux/device.h>
19#include <linux/memory.h>
20#include <linux/kobject.h>
21#include <linux/memory_hotplug.h>
22#include <linux/mm.h>
23#include <linux/mutex.h>
24#include <linux/stat.h>
25#include <linux/slab.h>
26
27#include <linux/atomic.h>
28#include <asm/uaccess.h>
29
30static DEFINE_MUTEX(mem_sysfs_mutex);
31
32#define MEMORY_CLASS_NAME "memory"
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 struct sysdev_class memory_sysdev_class = {
42 .name = MEMORY_CLASS_NAME,
43};
44
45static const char *memory_uevent_name(struct kset *kset, struct kobject *kobj)
46{
47 return MEMORY_CLASS_NAME;
48}
49
50static int memory_uevent(struct kset *kset, struct kobject *obj,
51 struct kobj_uevent_env *env)
52{
53 int retval = 0;
54
55 return retval;
56}
57
58static const struct kset_uevent_ops memory_uevent_ops = {
59 .name = memory_uevent_name,
60 .uevent = memory_uevent,
61};
62
63static BLOCKING_NOTIFIER_HEAD(memory_chain);
64
65int register_memory_notifier(struct notifier_block *nb)
66{
67 return blocking_notifier_chain_register(&memory_chain, nb);
68}
69EXPORT_SYMBOL(register_memory_notifier);
70
71void unregister_memory_notifier(struct notifier_block *nb)
72{
73 blocking_notifier_chain_unregister(&memory_chain, nb);
74}
75EXPORT_SYMBOL(unregister_memory_notifier);
76
77static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
78
79int register_memory_isolate_notifier(struct notifier_block *nb)
80{
81 return atomic_notifier_chain_register(&memory_isolate_chain, nb);
82}
83EXPORT_SYMBOL(register_memory_isolate_notifier);
84
85void unregister_memory_isolate_notifier(struct notifier_block *nb)
86{
87 atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
88}
89EXPORT_SYMBOL(unregister_memory_isolate_notifier);
90
91/*
92 * register_memory - Setup a sysfs device for a memory block
93 */
94static
95int register_memory(struct memory_block *memory)
96{
97 int error;
98
99 memory->sysdev.cls = &memory_sysdev_class;
100 memory->sysdev.id = memory->start_section_nr / sections_per_block;
101
102 error = sysdev_register(&memory->sysdev);
103 return error;
104}
105
106static void
107unregister_memory(struct memory_block *memory)
108{
109 BUG_ON(memory->sysdev.cls != &memory_sysdev_class);
110
111 /* drop the ref. we got in remove_memory_block() */
112 kobject_put(&memory->sysdev.kobj);
113 sysdev_unregister(&memory->sysdev);
114}
115
116unsigned long __weak memory_block_size_bytes(void)
117{
118 return MIN_MEMORY_BLOCK_SIZE;
119}
120
121static unsigned long get_memory_block_size(void)
122{
123 unsigned long block_sz;
124
125 block_sz = memory_block_size_bytes();
126
127 /* Validate blk_sz is a power of 2 and not less than section size */
128 if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
129 WARN_ON(1);
130 block_sz = MIN_MEMORY_BLOCK_SIZE;
131 }
132
133 return block_sz;
134}
135
136/*
137 * use this as the physical section index that this memsection
138 * uses.
139 */
140
141static ssize_t show_mem_start_phys_index(struct sys_device *dev,
142 struct sysdev_attribute *attr, char *buf)
143{
144 struct memory_block *mem =
145 container_of(dev, struct memory_block, sysdev);
146 unsigned long phys_index;
147
148 phys_index = mem->start_section_nr / sections_per_block;
149 return sprintf(buf, "%08lx\n", phys_index);
150}
151
152static ssize_t show_mem_end_phys_index(struct sys_device *dev,
153 struct sysdev_attribute *attr, char *buf)
154{
155 struct memory_block *mem =
156 container_of(dev, struct memory_block, sysdev);
157 unsigned long phys_index;
158
159 phys_index = mem->end_section_nr / sections_per_block;
160 return sprintf(buf, "%08lx\n", phys_index);
161}
162
163/*
164 * Show whether the section of memory is likely to be hot-removable
165 */
166static ssize_t show_mem_removable(struct sys_device *dev,
167 struct sysdev_attribute *attr, char *buf)
168{
169 unsigned long i, pfn;
170 int ret = 1;
171 struct memory_block *mem =
172 container_of(dev, struct memory_block, sysdev);
173
174 for (i = 0; i < sections_per_block; i++) {
175 pfn = section_nr_to_pfn(mem->start_section_nr + i);
176 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
177 }
178
179 return sprintf(buf, "%d\n", ret);
180}
181
182/*
183 * online, offline, going offline, etc.
184 */
185static ssize_t show_mem_state(struct sys_device *dev,
186 struct sysdev_attribute *attr, char *buf)
187{
188 struct memory_block *mem =
189 container_of(dev, struct memory_block, sysdev);
190 ssize_t len = 0;
191
192 /*
193 * We can probably put these states in a nice little array
194 * so that they're not open-coded
195 */
196 switch (mem->state) {
197 case MEM_ONLINE:
198 len = sprintf(buf, "online\n");
199 break;
200 case MEM_OFFLINE:
201 len = sprintf(buf, "offline\n");
202 break;
203 case MEM_GOING_OFFLINE:
204 len = sprintf(buf, "going-offline\n");
205 break;
206 default:
207 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
208 mem->state);
209 WARN_ON(1);
210 break;
211 }
212
213 return len;
214}
215
216int memory_notify(unsigned long val, void *v)
217{
218 return blocking_notifier_call_chain(&memory_chain, val, v);
219}
220
221int memory_isolate_notify(unsigned long val, void *v)
222{
223 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
224}
225
226/*
227 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
228 * OK to have direct references to sparsemem variables in here.
229 */
230static int
231memory_block_action(unsigned long phys_index, unsigned long action)
232{
233 int i;
234 unsigned long start_pfn, start_paddr;
235 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
236 struct page *first_page;
237 int ret;
238
239 first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
240
241 /*
242 * The probe routines leave the pages reserved, just
243 * as the bootmem code does. Make sure they're still
244 * that way.
245 */
246 if (action == MEM_ONLINE) {
247 for (i = 0; i < nr_pages; i++) {
248 if (PageReserved(first_page+i))
249 continue;
250
251 printk(KERN_WARNING "section number %ld page number %d "
252 "not reserved, was it already online?\n",
253 phys_index, i);
254 return -EBUSY;
255 }
256 }
257
258 switch (action) {
259 case MEM_ONLINE:
260 start_pfn = page_to_pfn(first_page);
261 ret = online_pages(start_pfn, nr_pages);
262 break;
263 case MEM_OFFLINE:
264 start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
265 ret = remove_memory(start_paddr,
266 nr_pages << PAGE_SHIFT);
267 break;
268 default:
269 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
270 "%ld\n", __func__, phys_index, action, action);
271 ret = -EINVAL;
272 }
273
274 return ret;
275}
276
277static int memory_block_change_state(struct memory_block *mem,
278 unsigned long to_state, unsigned long from_state_req)
279{
280 int ret = 0;
281
282 mutex_lock(&mem->state_mutex);
283
284 if (mem->state != from_state_req) {
285 ret = -EINVAL;
286 goto out;
287 }
288
289 if (to_state == MEM_OFFLINE)
290 mem->state = MEM_GOING_OFFLINE;
291
292 ret = memory_block_action(mem->start_section_nr, to_state);
293
294 if (ret)
295 mem->state = from_state_req;
296 else
297 mem->state = to_state;
298
299out:
300 mutex_unlock(&mem->state_mutex);
301 return ret;
302}
303
304static ssize_t
305store_mem_state(struct sys_device *dev,
306 struct sysdev_attribute *attr, const char *buf, size_t count)
307{
308 struct memory_block *mem;
309 int ret = -EINVAL;
310
311 mem = container_of(dev, struct memory_block, sysdev);
312
313 if (!strncmp(buf, "online", min((int)count, 6)))
314 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
315 else if(!strncmp(buf, "offline", min((int)count, 7)))
316 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
317
318 if (ret)
319 return ret;
320 return count;
321}
322
323/*
324 * phys_device is a bad name for this. What I really want
325 * is a way to differentiate between memory ranges that
326 * are part of physical devices that constitute
327 * a complete removable unit or fru.
328 * i.e. do these ranges belong to the same physical device,
329 * s.t. if I offline all of these sections I can then
330 * remove the physical device?
331 */
332static ssize_t show_phys_device(struct sys_device *dev,
333 struct sysdev_attribute *attr, char *buf)
334{
335 struct memory_block *mem =
336 container_of(dev, struct memory_block, sysdev);
337 return sprintf(buf, "%d\n", mem->phys_device);
338}
339
340static SYSDEV_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
341static SYSDEV_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
342static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
343static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
344static SYSDEV_ATTR(removable, 0444, show_mem_removable, NULL);
345
346#define mem_create_simple_file(mem, attr_name) \
347 sysdev_create_file(&mem->sysdev, &attr_##attr_name)
348#define mem_remove_simple_file(mem, attr_name) \
349 sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
350
351/*
352 * Block size attribute stuff
353 */
354static ssize_t
355print_block_size(struct sysdev_class *class, struct sysdev_class_attribute *attr,
356 char *buf)
357{
358 return sprintf(buf, "%lx\n", get_memory_block_size());
359}
360
361static SYSDEV_CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
362
363static int block_size_init(void)
364{
365 return sysfs_create_file(&memory_sysdev_class.kset.kobj,
366 &attr_block_size_bytes.attr);
367}
368
369/*
370 * Some architectures will have custom drivers to do this, and
371 * will not need to do it from userspace. The fake hot-add code
372 * as well as ppc64 will do all of their discovery in userspace
373 * and will require this interface.
374 */
375#ifdef CONFIG_ARCH_MEMORY_PROBE
376static ssize_t
377memory_probe_store(struct class *class, struct class_attribute *attr,
378 const char *buf, size_t count)
379{
380 u64 phys_addr;
381 int nid;
382 int i, ret;
383
384 phys_addr = simple_strtoull(buf, NULL, 0);
385
386 for (i = 0; i < sections_per_block; i++) {
387 nid = memory_add_physaddr_to_nid(phys_addr);
388 ret = add_memory(nid, phys_addr,
389 PAGES_PER_SECTION << PAGE_SHIFT);
390 if (ret)
391 goto out;
392
393 phys_addr += MIN_MEMORY_BLOCK_SIZE;
394 }
395
396 ret = count;
397out:
398 return ret;
399}
400static CLASS_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
401
402static int memory_probe_init(void)
403{
404 return sysfs_create_file(&memory_sysdev_class.kset.kobj,
405 &class_attr_probe.attr);
406}
407#else
408static inline int memory_probe_init(void)
409{
410 return 0;
411}
412#endif
413
414#ifdef CONFIG_MEMORY_FAILURE
415/*
416 * Support for offlining pages of memory
417 */
418
419/* Soft offline a page */
420static ssize_t
421store_soft_offline_page(struct class *class,
422 struct class_attribute *attr,
423 const char *buf, size_t count)
424{
425 int ret;
426 u64 pfn;
427 if (!capable(CAP_SYS_ADMIN))
428 return -EPERM;
429 if (strict_strtoull(buf, 0, &pfn) < 0)
430 return -EINVAL;
431 pfn >>= PAGE_SHIFT;
432 if (!pfn_valid(pfn))
433 return -ENXIO;
434 ret = soft_offline_page(pfn_to_page(pfn), 0);
435 return ret == 0 ? count : ret;
436}
437
438/* Forcibly offline a page, including killing processes. */
439static ssize_t
440store_hard_offline_page(struct class *class,
441 struct class_attribute *attr,
442 const char *buf, size_t count)
443{
444 int ret;
445 u64 pfn;
446 if (!capable(CAP_SYS_ADMIN))
447 return -EPERM;
448 if (strict_strtoull(buf, 0, &pfn) < 0)
449 return -EINVAL;
450 pfn >>= PAGE_SHIFT;
451 ret = __memory_failure(pfn, 0, 0);
452 return ret ? ret : count;
453}
454
455static CLASS_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
456static CLASS_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
457
458static __init int memory_fail_init(void)
459{
460 int err;
461
462 err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
463 &class_attr_soft_offline_page.attr);
464 if (!err)
465 err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
466 &class_attr_hard_offline_page.attr);
467 return err;
468}
469#else
470static inline int memory_fail_init(void)
471{
472 return 0;
473}
474#endif
475
476/*
477 * Note that phys_device is optional. It is here to allow for
478 * differentiation between which *physical* devices each
479 * section belongs to...
480 */
481int __weak arch_get_memory_phys_device(unsigned long start_pfn)
482{
483 return 0;
484}
485
486struct memory_block *find_memory_block_hinted(struct mem_section *section,
487 struct memory_block *hint)
488{
489 struct kobject *kobj;
490 struct sys_device *sysdev;
491 struct memory_block *mem;
492 char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
493 int block_id = base_memory_block_id(__section_nr(section));
494
495 kobj = hint ? &hint->sysdev.kobj : NULL;
496
497 /*
498 * This only works because we know that section == sysdev->id
499 * slightly redundant with sysdev_register()
500 */
501 sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, block_id);
502
503 kobj = kset_find_obj_hinted(&memory_sysdev_class.kset, name, kobj);
504 if (!kobj)
505 return NULL;
506
507 sysdev = container_of(kobj, struct sys_device, kobj);
508 mem = container_of(sysdev, struct memory_block, sysdev);
509
510 return mem;
511}
512
513/*
514 * For now, we have a linear search to go find the appropriate
515 * memory_block corresponding to a particular phys_index. If
516 * this gets to be a real problem, we can always use a radix
517 * tree or something here.
518 *
519 * This could be made generic for all sysdev classes.
520 */
521struct memory_block *find_memory_block(struct mem_section *section)
522{
523 return find_memory_block_hinted(section, NULL);
524}
525
526static int init_memory_block(struct memory_block **memory,
527 struct mem_section *section, unsigned long state)
528{
529 struct memory_block *mem;
530 unsigned long start_pfn;
531 int scn_nr;
532 int ret = 0;
533
534 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
535 if (!mem)
536 return -ENOMEM;
537
538 scn_nr = __section_nr(section);
539 mem->start_section_nr =
540 base_memory_block_id(scn_nr) * sections_per_block;
541 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
542 mem->state = state;
543 mem->section_count++;
544 mutex_init(&mem->state_mutex);
545 start_pfn = section_nr_to_pfn(mem->start_section_nr);
546 mem->phys_device = arch_get_memory_phys_device(start_pfn);
547
548 ret = register_memory(mem);
549 if (!ret)
550 ret = mem_create_simple_file(mem, phys_index);
551 if (!ret)
552 ret = mem_create_simple_file(mem, end_phys_index);
553 if (!ret)
554 ret = mem_create_simple_file(mem, state);
555 if (!ret)
556 ret = mem_create_simple_file(mem, phys_device);
557 if (!ret)
558 ret = mem_create_simple_file(mem, removable);
559
560 *memory = mem;
561 return ret;
562}
563
564static int add_memory_section(int nid, struct mem_section *section,
565 unsigned long state, enum mem_add_context context)
566{
567 struct memory_block *mem;
568 int ret = 0;
569
570 mutex_lock(&mem_sysfs_mutex);
571
572 mem = find_memory_block(section);
573 if (mem) {
574 mem->section_count++;
575 kobject_put(&mem->sysdev.kobj);
576 } else
577 ret = init_memory_block(&mem, section, state);
578
579 if (!ret) {
580 if (context == HOTPLUG &&
581 mem->section_count == sections_per_block)
582 ret = register_mem_sect_under_node(mem, nid);
583 }
584
585 mutex_unlock(&mem_sysfs_mutex);
586 return ret;
587}
588
589int remove_memory_block(unsigned long node_id, struct mem_section *section,
590 int phys_device)
591{
592 struct memory_block *mem;
593
594 mutex_lock(&mem_sysfs_mutex);
595 mem = find_memory_block(section);
596 unregister_mem_sect_under_nodes(mem, __section_nr(section));
597
598 mem->section_count--;
599 if (mem->section_count == 0) {
600 mem_remove_simple_file(mem, phys_index);
601 mem_remove_simple_file(mem, end_phys_index);
602 mem_remove_simple_file(mem, state);
603 mem_remove_simple_file(mem, phys_device);
604 mem_remove_simple_file(mem, removable);
605 unregister_memory(mem);
606 kfree(mem);
607 } else
608 kobject_put(&mem->sysdev.kobj);
609
610 mutex_unlock(&mem_sysfs_mutex);
611 return 0;
612}
613
614/*
615 * need an interface for the VM to add new memory regions,
616 * but without onlining it.
617 */
618int register_new_memory(int nid, struct mem_section *section)
619{
620 return add_memory_section(nid, section, MEM_OFFLINE, HOTPLUG);
621}
622
623int unregister_memory_section(struct mem_section *section)
624{
625 if (!present_section(section))
626 return -EINVAL;
627
628 return remove_memory_block(0, section, 0);
629}
630
631/*
632 * Initialize the sysfs support for memory devices...
633 */
634int __init memory_dev_init(void)
635{
636 unsigned int i;
637 int ret;
638 int err;
639 unsigned long block_sz;
640
641 memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
642 ret = sysdev_class_register(&memory_sysdev_class);
643 if (ret)
644 goto out;
645
646 block_sz = get_memory_block_size();
647 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
648
649 /*
650 * Create entries for memory sections that were found
651 * during boot and have been initialized
652 */
653 for (i = 0; i < NR_MEM_SECTIONS; i++) {
654 if (!present_section_nr(i))
655 continue;
656 err = add_memory_section(0, __nr_to_section(i), MEM_ONLINE,
657 BOOT);
658 if (!ret)
659 ret = err;
660 }
661
662 err = memory_probe_init();
663 if (!ret)
664 ret = err;
665 err = memory_fail_init();
666 if (!ret)
667 ret = err;
668 err = block_size_init();
669 if (!ret)
670 ret = err;
671out:
672 if (ret)
673 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
674 return ret;
675}
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}