Merge tag 'driver-core-3.11-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git...
[cascardo/linux.git] / drivers / base / memory.c
1 /*
2  * Memory subsystem support
3  *
4  * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
5  *            Dave Hansen <haveblue@us.ibm.com>
6  *
7  * This file provides the necessary infrastructure to represent
8  * a SPARSEMEM-memory-model system's physical memory in /sysfs.
9  * All arch-independent code that assumes MEMORY_HOTPLUG requires
10  * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
11  */
12
13 #include <linux/module.h>
14 #include <linux/init.h>
15 #include <linux/topology.h>
16 #include <linux/capability.h>
17 #include <linux/device.h>
18 #include <linux/memory.h>
19 #include <linux/kobject.h>
20 #include <linux/memory_hotplug.h>
21 #include <linux/mm.h>
22 #include <linux/mutex.h>
23 #include <linux/stat.h>
24 #include <linux/slab.h>
25
26 #include <linux/atomic.h>
27 #include <asm/uaccess.h>
28
29 static DEFINE_MUTEX(mem_sysfs_mutex);
30
31 #define MEMORY_CLASS_NAME       "memory"
32
33 static int sections_per_block;
34
35 static inline int base_memory_block_id(int section_nr)
36 {
37         return section_nr / sections_per_block;
38 }
39
40 static struct bus_type memory_subsys = {
41         .name = MEMORY_CLASS_NAME,
42         .dev_name = MEMORY_CLASS_NAME,
43 };
44
45 static BLOCKING_NOTIFIER_HEAD(memory_chain);
46
47 int register_memory_notifier(struct notifier_block *nb)
48 {
49         return blocking_notifier_chain_register(&memory_chain, nb);
50 }
51 EXPORT_SYMBOL(register_memory_notifier);
52
53 void unregister_memory_notifier(struct notifier_block *nb)
54 {
55         blocking_notifier_chain_unregister(&memory_chain, nb);
56 }
57 EXPORT_SYMBOL(unregister_memory_notifier);
58
59 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
60
61 int register_memory_isolate_notifier(struct notifier_block *nb)
62 {
63         return atomic_notifier_chain_register(&memory_isolate_chain, nb);
64 }
65 EXPORT_SYMBOL(register_memory_isolate_notifier);
66
67 void unregister_memory_isolate_notifier(struct notifier_block *nb)
68 {
69         atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
70 }
71 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
72
73 static void memory_block_release(struct device *dev)
74 {
75         struct memory_block *mem = container_of(dev, struct memory_block, dev);
76
77         kfree(mem);
78 }
79
80 unsigned long __weak memory_block_size_bytes(void)
81 {
82         return MIN_MEMORY_BLOCK_SIZE;
83 }
84
85 static unsigned long get_memory_block_size(void)
86 {
87         unsigned long block_sz;
88
89         block_sz = memory_block_size_bytes();
90
91         /* Validate blk_sz is a power of 2 and not less than section size */
92         if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
93                 WARN_ON(1);
94                 block_sz = MIN_MEMORY_BLOCK_SIZE;
95         }
96
97         return block_sz;
98 }
99
100 /*
101  * use this as the physical section index that this memsection
102  * uses.
103  */
104
105 static ssize_t show_mem_start_phys_index(struct device *dev,
106                         struct device_attribute *attr, char *buf)
107 {
108         struct memory_block *mem =
109                 container_of(dev, struct memory_block, dev);
110         unsigned long phys_index;
111
112         phys_index = mem->start_section_nr / sections_per_block;
113         return sprintf(buf, "%08lx\n", phys_index);
114 }
115
116 static ssize_t show_mem_end_phys_index(struct device *dev,
117                         struct device_attribute *attr, char *buf)
118 {
119         struct memory_block *mem =
120                 container_of(dev, struct memory_block, dev);
121         unsigned long phys_index;
122
123         phys_index = mem->end_section_nr / sections_per_block;
124         return sprintf(buf, "%08lx\n", phys_index);
125 }
126
127 /*
128  * Show whether the section of memory is likely to be hot-removable
129  */
130 static ssize_t show_mem_removable(struct device *dev,
131                         struct device_attribute *attr, char *buf)
132 {
133         unsigned long i, pfn;
134         int ret = 1;
135         struct memory_block *mem =
136                 container_of(dev, struct memory_block, dev);
137
138         for (i = 0; i < sections_per_block; i++) {
139                 pfn = section_nr_to_pfn(mem->start_section_nr + i);
140                 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
141         }
142
143         return sprintf(buf, "%d\n", ret);
144 }
145
146 /*
147  * online, offline, going offline, etc.
148  */
149 static ssize_t show_mem_state(struct device *dev,
150                         struct device_attribute *attr, char *buf)
151 {
152         struct memory_block *mem =
153                 container_of(dev, struct memory_block, dev);
154         ssize_t len = 0;
155
156         /*
157          * We can probably put these states in a nice little array
158          * so that they're not open-coded
159          */
160         switch (mem->state) {
161                 case MEM_ONLINE:
162                         len = sprintf(buf, "online\n");
163                         break;
164                 case MEM_OFFLINE:
165                         len = sprintf(buf, "offline\n");
166                         break;
167                 case MEM_GOING_OFFLINE:
168                         len = sprintf(buf, "going-offline\n");
169                         break;
170                 default:
171                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
172                                         mem->state);
173                         WARN_ON(1);
174                         break;
175         }
176
177         return len;
178 }
179
180 int memory_notify(unsigned long val, void *v)
181 {
182         return blocking_notifier_call_chain(&memory_chain, val, v);
183 }
184
185 int memory_isolate_notify(unsigned long val, void *v)
186 {
187         return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
188 }
189
190 /*
191  * The probe routines leave the pages reserved, just as the bootmem code does.
192  * Make sure they're still that way.
193  */
194 static bool pages_correctly_reserved(unsigned long start_pfn)
195 {
196         int i, j;
197         struct page *page;
198         unsigned long pfn = start_pfn;
199
200         /*
201          * memmap between sections is not contiguous except with
202          * SPARSEMEM_VMEMMAP. We lookup the page once per section
203          * and assume memmap is contiguous within each section
204          */
205         for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
206                 if (WARN_ON_ONCE(!pfn_valid(pfn)))
207                         return false;
208                 page = pfn_to_page(pfn);
209
210                 for (j = 0; j < PAGES_PER_SECTION; j++) {
211                         if (PageReserved(page + j))
212                                 continue;
213
214                         printk(KERN_WARNING "section number %ld page number %d "
215                                 "not reserved, was it already online?\n",
216                                 pfn_to_section_nr(pfn), j);
217
218                         return false;
219                 }
220         }
221
222         return true;
223 }
224
225 /*
226  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
227  * OK to have direct references to sparsemem variables in here.
228  */
229 static int
230 memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
231 {
232         unsigned long start_pfn;
233         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
234         struct page *first_page;
235         int ret;
236
237         first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
238         start_pfn = page_to_pfn(first_page);
239
240         switch (action) {
241                 case MEM_ONLINE:
242                         if (!pages_correctly_reserved(start_pfn))
243                                 return -EBUSY;
244
245                         ret = online_pages(start_pfn, nr_pages, online_type);
246                         break;
247                 case MEM_OFFLINE:
248                         ret = offline_pages(start_pfn, nr_pages);
249                         break;
250                 default:
251                         WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
252                              "%ld\n", __func__, phys_index, action, action);
253                         ret = -EINVAL;
254         }
255
256         return ret;
257 }
258
259 static int __memory_block_change_state(struct memory_block *mem,
260                 unsigned long to_state, unsigned long from_state_req,
261                 int online_type)
262 {
263         int ret = 0;
264
265         if (mem->state != from_state_req) {
266                 ret = -EINVAL;
267                 goto out;
268         }
269
270         if (to_state == MEM_OFFLINE)
271                 mem->state = MEM_GOING_OFFLINE;
272
273         ret = memory_block_action(mem->start_section_nr, to_state, online_type);
274
275         if (ret) {
276                 mem->state = from_state_req;
277                 goto out;
278         }
279
280         mem->state = to_state;
281         switch (mem->state) {
282         case MEM_OFFLINE:
283                 kobject_uevent(&mem->dev.kobj, KOBJ_OFFLINE);
284                 break;
285         case MEM_ONLINE:
286                 kobject_uevent(&mem->dev.kobj, KOBJ_ONLINE);
287                 break;
288         default:
289                 break;
290         }
291 out:
292         return ret;
293 }
294
295 static int memory_block_change_state(struct memory_block *mem,
296                 unsigned long to_state, unsigned long from_state_req,
297                 int online_type)
298 {
299         int ret;
300
301         mutex_lock(&mem->state_mutex);
302         ret = __memory_block_change_state(mem, to_state, from_state_req,
303                                           online_type);
304         mutex_unlock(&mem->state_mutex);
305
306         return ret;
307 }
308 static ssize_t
309 store_mem_state(struct device *dev,
310                 struct device_attribute *attr, const char *buf, size_t count)
311 {
312         struct memory_block *mem;
313         int ret = -EINVAL;
314
315         mem = container_of(dev, struct memory_block, dev);
316
317         if (!strncmp(buf, "online_kernel", min_t(int, count, 13)))
318                 ret = memory_block_change_state(mem, MEM_ONLINE,
319                                                 MEM_OFFLINE, ONLINE_KERNEL);
320         else if (!strncmp(buf, "online_movable", min_t(int, count, 14)))
321                 ret = memory_block_change_state(mem, MEM_ONLINE,
322                                                 MEM_OFFLINE, ONLINE_MOVABLE);
323         else if (!strncmp(buf, "online", min_t(int, count, 6)))
324                 ret = memory_block_change_state(mem, MEM_ONLINE,
325                                                 MEM_OFFLINE, ONLINE_KEEP);
326         else if(!strncmp(buf, "offline", min_t(int, count, 7)))
327                 ret = memory_block_change_state(mem, MEM_OFFLINE,
328                                                 MEM_ONLINE, -1);
329
330         if (ret)
331                 return ret;
332         return count;
333 }
334
335 /*
336  * phys_device is a bad name for this.  What I really want
337  * is a way to differentiate between memory ranges that
338  * are part of physical devices that constitute
339  * a complete removable unit or fru.
340  * i.e. do these ranges belong to the same physical device,
341  * s.t. if I offline all of these sections I can then
342  * remove the physical device?
343  */
344 static ssize_t show_phys_device(struct device *dev,
345                                 struct device_attribute *attr, char *buf)
346 {
347         struct memory_block *mem =
348                 container_of(dev, struct memory_block, dev);
349         return sprintf(buf, "%d\n", mem->phys_device);
350 }
351
352 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
353 static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
354 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
355 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
356 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
357
358 /*
359  * Block size attribute stuff
360  */
361 static ssize_t
362 print_block_size(struct device *dev, struct device_attribute *attr,
363                  char *buf)
364 {
365         return sprintf(buf, "%lx\n", get_memory_block_size());
366 }
367
368 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
369
370 /*
371  * Some architectures will have custom drivers to do this, and
372  * will not need to do it from userspace.  The fake hot-add code
373  * as well as ppc64 will do all of their discovery in userspace
374  * and will require this interface.
375  */
376 #ifdef CONFIG_ARCH_MEMORY_PROBE
377 static ssize_t
378 memory_probe_store(struct device *dev, struct device_attribute *attr,
379                    const char *buf, size_t count)
380 {
381         u64 phys_addr;
382         int nid;
383         int i, ret;
384         unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
385
386         phys_addr = simple_strtoull(buf, NULL, 0);
387
388         if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
389                 return -EINVAL;
390
391         for (i = 0; i < sections_per_block; i++) {
392                 nid = memory_add_physaddr_to_nid(phys_addr);
393                 ret = add_memory(nid, phys_addr,
394                                  PAGES_PER_SECTION << PAGE_SHIFT);
395                 if (ret)
396                         goto out;
397
398                 phys_addr += MIN_MEMORY_BLOCK_SIZE;
399         }
400
401         ret = count;
402 out:
403         return ret;
404 }
405
406 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
407 #endif
408
409 #ifdef CONFIG_MEMORY_FAILURE
410 /*
411  * Support for offlining pages of memory
412  */
413
414 /* Soft offline a page */
415 static ssize_t
416 store_soft_offline_page(struct device *dev,
417                         struct device_attribute *attr,
418                         const char *buf, size_t count)
419 {
420         int ret;
421         u64 pfn;
422         if (!capable(CAP_SYS_ADMIN))
423                 return -EPERM;
424         if (strict_strtoull(buf, 0, &pfn) < 0)
425                 return -EINVAL;
426         pfn >>= PAGE_SHIFT;
427         if (!pfn_valid(pfn))
428                 return -ENXIO;
429         ret = soft_offline_page(pfn_to_page(pfn), 0);
430         return ret == 0 ? count : ret;
431 }
432
433 /* Forcibly offline a page, including killing processes. */
434 static ssize_t
435 store_hard_offline_page(struct device *dev,
436                         struct device_attribute *attr,
437                         const char *buf, size_t count)
438 {
439         int ret;
440         u64 pfn;
441         if (!capable(CAP_SYS_ADMIN))
442                 return -EPERM;
443         if (strict_strtoull(buf, 0, &pfn) < 0)
444                 return -EINVAL;
445         pfn >>= PAGE_SHIFT;
446         ret = memory_failure(pfn, 0, 0);
447         return ret ? ret : count;
448 }
449
450 static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
451 static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
452 #endif
453
454 /*
455  * Note that phys_device is optional.  It is here to allow for
456  * differentiation between which *physical* devices each
457  * section belongs to...
458  */
459 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
460 {
461         return 0;
462 }
463
464 /*
465  * A reference for the returned object is held and the reference for the
466  * hinted object is released.
467  */
468 struct memory_block *find_memory_block_hinted(struct mem_section *section,
469                                               struct memory_block *hint)
470 {
471         int block_id = base_memory_block_id(__section_nr(section));
472         struct device *hintdev = hint ? &hint->dev : NULL;
473         struct device *dev;
474
475         dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
476         if (hint)
477                 put_device(&hint->dev);
478         if (!dev)
479                 return NULL;
480         return container_of(dev, struct memory_block, dev);
481 }
482
483 /*
484  * For now, we have a linear search to go find the appropriate
485  * memory_block corresponding to a particular phys_index. If
486  * this gets to be a real problem, we can always use a radix
487  * tree or something here.
488  *
489  * This could be made generic for all device subsystems.
490  */
491 struct memory_block *find_memory_block(struct mem_section *section)
492 {
493         return find_memory_block_hinted(section, NULL);
494 }
495
496 static struct attribute *memory_memblk_attrs[] = {
497         &dev_attr_phys_index.attr,
498         &dev_attr_end_phys_index.attr,
499         &dev_attr_state.attr,
500         &dev_attr_phys_device.attr,
501         &dev_attr_removable.attr,
502         NULL
503 };
504
505 static struct attribute_group memory_memblk_attr_group = {
506         .attrs = memory_memblk_attrs,
507 };
508
509 static const struct attribute_group *memory_memblk_attr_groups[] = {
510         &memory_memblk_attr_group,
511         NULL,
512 };
513
514 /*
515  * register_memory - Setup a sysfs device for a memory block
516  */
517 static
518 int register_memory(struct memory_block *memory)
519 {
520         int error;
521
522         memory->dev.bus = &memory_subsys;
523         memory->dev.id = memory->start_section_nr / sections_per_block;
524         memory->dev.release = memory_block_release;
525         memory->dev.groups = memory_memblk_attr_groups;
526
527         error = device_register(&memory->dev);
528         return error;
529 }
530
531 static int init_memory_block(struct memory_block **memory,
532                              struct mem_section *section, unsigned long state)
533 {
534         struct memory_block *mem;
535         unsigned long start_pfn;
536         int scn_nr;
537         int ret = 0;
538
539         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
540         if (!mem)
541                 return -ENOMEM;
542
543         scn_nr = __section_nr(section);
544         mem->start_section_nr =
545                         base_memory_block_id(scn_nr) * sections_per_block;
546         mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
547         mem->state = state;
548         mem->section_count++;
549         mutex_init(&mem->state_mutex);
550         start_pfn = section_nr_to_pfn(mem->start_section_nr);
551         mem->phys_device = arch_get_memory_phys_device(start_pfn);
552
553         ret = register_memory(mem);
554
555         *memory = mem;
556         return ret;
557 }
558
559 static int add_memory_section(int nid, struct mem_section *section,
560                         struct memory_block **mem_p,
561                         unsigned long state, enum mem_add_context context)
562 {
563         struct memory_block *mem = NULL;
564         int scn_nr = __section_nr(section);
565         int ret = 0;
566
567         mutex_lock(&mem_sysfs_mutex);
568
569         if (context == BOOT) {
570                 /* same memory block ? */
571                 if (mem_p && *mem_p)
572                         if (scn_nr >= (*mem_p)->start_section_nr &&
573                             scn_nr <= (*mem_p)->end_section_nr) {
574                                 mem = *mem_p;
575                                 kobject_get(&mem->dev.kobj);
576                         }
577         } else
578                 mem = find_memory_block(section);
579
580         if (mem) {
581                 mem->section_count++;
582                 kobject_put(&mem->dev.kobj);
583         } else {
584                 ret = init_memory_block(&mem, section, state);
585                 /* store memory_block pointer for next loop */
586                 if (!ret && context == BOOT)
587                         if (mem_p)
588                                 *mem_p = mem;
589         }
590
591         if (!ret) {
592                 if (context == HOTPLUG &&
593                     mem->section_count == sections_per_block)
594                         ret = register_mem_sect_under_node(mem, nid);
595         }
596
597         mutex_unlock(&mem_sysfs_mutex);
598         return ret;
599 }
600
601 /*
602  * need an interface for the VM to add new memory regions,
603  * but without onlining it.
604  */
605 int register_new_memory(int nid, struct mem_section *section)
606 {
607         return add_memory_section(nid, section, NULL, MEM_OFFLINE, HOTPLUG);
608 }
609
610 #ifdef CONFIG_MEMORY_HOTREMOVE
611 static void
612 unregister_memory(struct memory_block *memory)
613 {
614         BUG_ON(memory->dev.bus != &memory_subsys);
615
616         /* drop the ref. we got in remove_memory_block() */
617         kobject_put(&memory->dev.kobj);
618         device_unregister(&memory->dev);
619 }
620
621 static int remove_memory_block(unsigned long node_id,
622                                struct mem_section *section, int phys_device)
623 {
624         struct memory_block *mem;
625
626         mutex_lock(&mem_sysfs_mutex);
627         mem = find_memory_block(section);
628         unregister_mem_sect_under_nodes(mem, __section_nr(section));
629
630         mem->section_count--;
631         if (mem->section_count == 0)
632                 unregister_memory(mem);
633         else
634                 kobject_put(&mem->dev.kobj);
635
636         mutex_unlock(&mem_sysfs_mutex);
637         return 0;
638 }
639
640 int unregister_memory_section(struct mem_section *section)
641 {
642         if (!present_section(section))
643                 return -EINVAL;
644
645         return remove_memory_block(0, section, 0);
646 }
647 #endif /* CONFIG_MEMORY_HOTREMOVE */
648
649 /*
650  * offline one memory block. If the memory block has been offlined, do nothing.
651  */
652 int offline_memory_block(struct memory_block *mem)
653 {
654         int ret = 0;
655
656         mutex_lock(&mem->state_mutex);
657         if (mem->state != MEM_OFFLINE)
658                 ret = __memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE, -1);
659         mutex_unlock(&mem->state_mutex);
660
661         return ret;
662 }
663
664 /* return true if the memory block is offlined, otherwise, return false */
665 bool is_memblock_offlined(struct memory_block *mem)
666 {
667         return mem->state == MEM_OFFLINE;
668 }
669
670 static struct attribute *memory_root_attrs[] = {
671 #ifdef CONFIG_ARCH_MEMORY_PROBE
672         &dev_attr_probe.attr,
673 #endif
674
675 #ifdef CONFIG_MEMORY_FAILURE
676         &dev_attr_soft_offline_page.attr,
677         &dev_attr_hard_offline_page.attr,
678 #endif
679
680         &dev_attr_block_size_bytes.attr,
681         NULL
682 };
683
684 static struct attribute_group memory_root_attr_group = {
685         .attrs = memory_root_attrs,
686 };
687
688 static const struct attribute_group *memory_root_attr_groups[] = {
689         &memory_root_attr_group,
690         NULL,
691 };
692
693 /*
694  * Initialize the sysfs support for memory devices...
695  */
696 int __init memory_dev_init(void)
697 {
698         unsigned int i;
699         int ret;
700         int err;
701         unsigned long block_sz;
702         struct memory_block *mem = NULL;
703
704         ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
705         if (ret)
706                 goto out;
707
708         block_sz = get_memory_block_size();
709         sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
710
711         /*
712          * Create entries for memory sections that were found
713          * during boot and have been initialized
714          */
715         for (i = 0; i < NR_MEM_SECTIONS; i++) {
716                 if (!present_section_nr(i))
717                         continue;
718                 /* don't need to reuse memory_block if only one per block */
719                 err = add_memory_section(0, __nr_to_section(i),
720                                  (sections_per_block == 1) ? NULL : &mem,
721                                          MEM_ONLINE,
722                                          BOOT);
723                 if (!ret)
724                         ret = err;
725         }
726
727 out:
728         if (ret)
729                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
730         return ret;
731 }