]> Pileus Git - ~andy/linux/blob - drivers/base/memory.c
memory-hotplug: preparation to notify memory block's state at memory hot remove
[~andy/linux] / 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 /*
74  * register_memory - Setup a sysfs device for a memory block
75  */
76 static
77 int register_memory(struct memory_block *memory)
78 {
79         int error;
80
81         memory->dev.bus = &memory_subsys;
82         memory->dev.id = memory->start_section_nr / sections_per_block;
83
84         error = device_register(&memory->dev);
85         return error;
86 }
87
88 static void
89 unregister_memory(struct memory_block *memory)
90 {
91         BUG_ON(memory->dev.bus != &memory_subsys);
92
93         /* drop the ref. we got in remove_memory_block() */
94         kobject_put(&memory->dev.kobj);
95         device_unregister(&memory->dev);
96 }
97
98 unsigned long __weak memory_block_size_bytes(void)
99 {
100         return MIN_MEMORY_BLOCK_SIZE;
101 }
102
103 static unsigned long get_memory_block_size(void)
104 {
105         unsigned long block_sz;
106
107         block_sz = memory_block_size_bytes();
108
109         /* Validate blk_sz is a power of 2 and not less than section size */
110         if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
111                 WARN_ON(1);
112                 block_sz = MIN_MEMORY_BLOCK_SIZE;
113         }
114
115         return block_sz;
116 }
117
118 /*
119  * use this as the physical section index that this memsection
120  * uses.
121  */
122
123 static ssize_t show_mem_start_phys_index(struct device *dev,
124                         struct device_attribute *attr, char *buf)
125 {
126         struct memory_block *mem =
127                 container_of(dev, struct memory_block, dev);
128         unsigned long phys_index;
129
130         phys_index = mem->start_section_nr / sections_per_block;
131         return sprintf(buf, "%08lx\n", phys_index);
132 }
133
134 static ssize_t show_mem_end_phys_index(struct device *dev,
135                         struct device_attribute *attr, char *buf)
136 {
137         struct memory_block *mem =
138                 container_of(dev, struct memory_block, dev);
139         unsigned long phys_index;
140
141         phys_index = mem->end_section_nr / sections_per_block;
142         return sprintf(buf, "%08lx\n", phys_index);
143 }
144
145 /*
146  * Show whether the section of memory is likely to be hot-removable
147  */
148 static ssize_t show_mem_removable(struct device *dev,
149                         struct device_attribute *attr, char *buf)
150 {
151         unsigned long i, pfn;
152         int ret = 1;
153         struct memory_block *mem =
154                 container_of(dev, struct memory_block, dev);
155
156         for (i = 0; i < sections_per_block; i++) {
157                 pfn = section_nr_to_pfn(mem->start_section_nr + i);
158                 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
159         }
160
161         return sprintf(buf, "%d\n", ret);
162 }
163
164 /*
165  * online, offline, going offline, etc.
166  */
167 static ssize_t show_mem_state(struct device *dev,
168                         struct device_attribute *attr, char *buf)
169 {
170         struct memory_block *mem =
171                 container_of(dev, struct memory_block, dev);
172         ssize_t len = 0;
173
174         /*
175          * We can probably put these states in a nice little array
176          * so that they're not open-coded
177          */
178         switch (mem->state) {
179                 case MEM_ONLINE:
180                         len = sprintf(buf, "online\n");
181                         break;
182                 case MEM_OFFLINE:
183                         len = sprintf(buf, "offline\n");
184                         break;
185                 case MEM_GOING_OFFLINE:
186                         len = sprintf(buf, "going-offline\n");
187                         break;
188                 default:
189                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
190                                         mem->state);
191                         WARN_ON(1);
192                         break;
193         }
194
195         return len;
196 }
197
198 int memory_notify(unsigned long val, void *v)
199 {
200         return blocking_notifier_call_chain(&memory_chain, val, v);
201 }
202
203 int memory_isolate_notify(unsigned long val, void *v)
204 {
205         return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
206 }
207
208 /*
209  * The probe routines leave the pages reserved, just as the bootmem code does.
210  * Make sure they're still that way.
211  */
212 static bool pages_correctly_reserved(unsigned long start_pfn,
213                                         unsigned long nr_pages)
214 {
215         int i, j;
216         struct page *page;
217         unsigned long pfn = start_pfn;
218
219         /*
220          * memmap between sections is not contiguous except with
221          * SPARSEMEM_VMEMMAP. We lookup the page once per section
222          * and assume memmap is contiguous within each section
223          */
224         for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
225                 if (WARN_ON_ONCE(!pfn_valid(pfn)))
226                         return false;
227                 page = pfn_to_page(pfn);
228
229                 for (j = 0; j < PAGES_PER_SECTION; j++) {
230                         if (PageReserved(page + j))
231                                 continue;
232
233                         printk(KERN_WARNING "section number %ld page number %d "
234                                 "not reserved, was it already online?\n",
235                                 pfn_to_section_nr(pfn), j);
236
237                         return false;
238                 }
239         }
240
241         return true;
242 }
243
244 /*
245  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
246  * OK to have direct references to sparsemem variables in here.
247  */
248 static int
249 memory_block_action(unsigned long phys_index, unsigned long action)
250 {
251         unsigned long start_pfn;
252         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
253         struct page *first_page;
254         int ret;
255
256         first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
257         start_pfn = page_to_pfn(first_page);
258
259         switch (action) {
260                 case MEM_ONLINE:
261                         if (!pages_correctly_reserved(start_pfn, nr_pages))
262                                 return -EBUSY;
263
264                         ret = online_pages(start_pfn, nr_pages);
265                         break;
266                 case MEM_OFFLINE:
267                         ret = offline_pages(start_pfn, nr_pages);
268                         break;
269                 default:
270                         WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
271                              "%ld\n", __func__, phys_index, action, action);
272                         ret = -EINVAL;
273         }
274
275         return ret;
276 }
277
278 static int memory_block_change_state(struct memory_block *mem,
279                 unsigned long to_state, unsigned long from_state_req)
280 {
281         int ret = 0;
282
283         mutex_lock(&mem->state_mutex);
284
285         if (mem->state != from_state_req) {
286                 ret = -EINVAL;
287                 goto out;
288         }
289
290         if (to_state == MEM_OFFLINE)
291                 mem->state = MEM_GOING_OFFLINE;
292
293         ret = memory_block_action(mem->start_section_nr, to_state);
294
295         if (ret) {
296                 mem->state = from_state_req;
297                 goto out;
298         }
299
300         mem->state = to_state;
301         switch (mem->state) {
302         case MEM_OFFLINE:
303                 kobject_uevent(&mem->dev.kobj, KOBJ_OFFLINE);
304                 break;
305         case MEM_ONLINE:
306                 kobject_uevent(&mem->dev.kobj, KOBJ_ONLINE);
307                 break;
308         default:
309                 break;
310         }
311 out:
312         mutex_unlock(&mem->state_mutex);
313         return ret;
314 }
315
316 static ssize_t
317 store_mem_state(struct device *dev,
318                 struct device_attribute *attr, const char *buf, size_t count)
319 {
320         struct memory_block *mem;
321         int ret = -EINVAL;
322
323         mem = container_of(dev, struct memory_block, dev);
324
325         if (!strncmp(buf, "online", min((int)count, 6)))
326                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
327         else if(!strncmp(buf, "offline", min((int)count, 7)))
328                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
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 #define mem_create_simple_file(mem, attr_name)  \
359         device_create_file(&mem->dev, &dev_attr_##attr_name)
360 #define mem_remove_simple_file(mem, attr_name)  \
361         device_remove_file(&mem->dev, &dev_attr_##attr_name)
362
363 /*
364  * Block size attribute stuff
365  */
366 static ssize_t
367 print_block_size(struct device *dev, struct device_attribute *attr,
368                  char *buf)
369 {
370         return sprintf(buf, "%lx\n", get_memory_block_size());
371 }
372
373 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
374
375 static int block_size_init(void)
376 {
377         return device_create_file(memory_subsys.dev_root,
378                                   &dev_attr_block_size_bytes);
379 }
380
381 /*
382  * Some architectures will have custom drivers to do this, and
383  * will not need to do it from userspace.  The fake hot-add code
384  * as well as ppc64 will do all of their discovery in userspace
385  * and will require this interface.
386  */
387 #ifdef CONFIG_ARCH_MEMORY_PROBE
388 static ssize_t
389 memory_probe_store(struct device *dev, struct device_attribute *attr,
390                    const char *buf, size_t count)
391 {
392         u64 phys_addr;
393         int nid;
394         int i, ret;
395         unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
396
397         phys_addr = simple_strtoull(buf, NULL, 0);
398
399         if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
400                 return -EINVAL;
401
402         for (i = 0; i < sections_per_block; i++) {
403                 nid = memory_add_physaddr_to_nid(phys_addr);
404                 ret = add_memory(nid, phys_addr,
405                                  PAGES_PER_SECTION << PAGE_SHIFT);
406                 if (ret)
407                         goto out;
408
409                 phys_addr += MIN_MEMORY_BLOCK_SIZE;
410         }
411
412         ret = count;
413 out:
414         return ret;
415 }
416 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
417
418 static int memory_probe_init(void)
419 {
420         return device_create_file(memory_subsys.dev_root, &dev_attr_probe);
421 }
422 #else
423 static inline int memory_probe_init(void)
424 {
425         return 0;
426 }
427 #endif
428
429 #ifdef CONFIG_MEMORY_FAILURE
430 /*
431  * Support for offlining pages of memory
432  */
433
434 /* Soft offline a page */
435 static ssize_t
436 store_soft_offline_page(struct device *dev,
437                         struct device_attribute *attr,
438                         const char *buf, size_t count)
439 {
440         int ret;
441         u64 pfn;
442         if (!capable(CAP_SYS_ADMIN))
443                 return -EPERM;
444         if (strict_strtoull(buf, 0, &pfn) < 0)
445                 return -EINVAL;
446         pfn >>= PAGE_SHIFT;
447         if (!pfn_valid(pfn))
448                 return -ENXIO;
449         ret = soft_offline_page(pfn_to_page(pfn), 0);
450         return ret == 0 ? count : ret;
451 }
452
453 /* Forcibly offline a page, including killing processes. */
454 static ssize_t
455 store_hard_offline_page(struct device *dev,
456                         struct device_attribute *attr,
457                         const char *buf, size_t count)
458 {
459         int ret;
460         u64 pfn;
461         if (!capable(CAP_SYS_ADMIN))
462                 return -EPERM;
463         if (strict_strtoull(buf, 0, &pfn) < 0)
464                 return -EINVAL;
465         pfn >>= PAGE_SHIFT;
466         ret = memory_failure(pfn, 0, 0);
467         return ret ? ret : count;
468 }
469
470 static DEVICE_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
471 static DEVICE_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
472
473 static __init int memory_fail_init(void)
474 {
475         int err;
476
477         err = device_create_file(memory_subsys.dev_root,
478                                 &dev_attr_soft_offline_page);
479         if (!err)
480                 err = device_create_file(memory_subsys.dev_root,
481                                 &dev_attr_hard_offline_page);
482         return err;
483 }
484 #else
485 static inline int memory_fail_init(void)
486 {
487         return 0;
488 }
489 #endif
490
491 /*
492  * Note that phys_device is optional.  It is here to allow for
493  * differentiation between which *physical* devices each
494  * section belongs to...
495  */
496 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
497 {
498         return 0;
499 }
500
501 /*
502  * A reference for the returned object is held and the reference for the
503  * hinted object is released.
504  */
505 struct memory_block *find_memory_block_hinted(struct mem_section *section,
506                                               struct memory_block *hint)
507 {
508         int block_id = base_memory_block_id(__section_nr(section));
509         struct device *hintdev = hint ? &hint->dev : NULL;
510         struct device *dev;
511
512         dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
513         if (hint)
514                 put_device(&hint->dev);
515         if (!dev)
516                 return NULL;
517         return container_of(dev, struct memory_block, dev);
518 }
519
520 /*
521  * For now, we have a linear search to go find the appropriate
522  * memory_block corresponding to a particular phys_index. If
523  * this gets to be a real problem, we can always use a radix
524  * tree or something here.
525  *
526  * This could be made generic for all device subsystems.
527  */
528 struct memory_block *find_memory_block(struct mem_section *section)
529 {
530         return find_memory_block_hinted(section, NULL);
531 }
532
533 static int init_memory_block(struct memory_block **memory,
534                              struct mem_section *section, unsigned long state)
535 {
536         struct memory_block *mem;
537         unsigned long start_pfn;
538         int scn_nr;
539         int ret = 0;
540
541         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
542         if (!mem)
543                 return -ENOMEM;
544
545         scn_nr = __section_nr(section);
546         mem->start_section_nr =
547                         base_memory_block_id(scn_nr) * sections_per_block;
548         mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
549         mem->state = state;
550         mem->section_count++;
551         mutex_init(&mem->state_mutex);
552         start_pfn = section_nr_to_pfn(mem->start_section_nr);
553         mem->phys_device = arch_get_memory_phys_device(start_pfn);
554
555         ret = register_memory(mem);
556         if (!ret)
557                 ret = mem_create_simple_file(mem, phys_index);
558         if (!ret)
559                 ret = mem_create_simple_file(mem, end_phys_index);
560         if (!ret)
561                 ret = mem_create_simple_file(mem, state);
562         if (!ret)
563                 ret = mem_create_simple_file(mem, phys_device);
564         if (!ret)
565                 ret = mem_create_simple_file(mem, removable);
566
567         *memory = mem;
568         return ret;
569 }
570
571 static int add_memory_section(int nid, struct mem_section *section,
572                         struct memory_block **mem_p,
573                         unsigned long state, enum mem_add_context context)
574 {
575         struct memory_block *mem = NULL;
576         int scn_nr = __section_nr(section);
577         int ret = 0;
578
579         mutex_lock(&mem_sysfs_mutex);
580
581         if (context == BOOT) {
582                 /* same memory block ? */
583                 if (mem_p && *mem_p)
584                         if (scn_nr >= (*mem_p)->start_section_nr &&
585                             scn_nr <= (*mem_p)->end_section_nr) {
586                                 mem = *mem_p;
587                                 kobject_get(&mem->dev.kobj);
588                         }
589         } else
590                 mem = find_memory_block(section);
591
592         if (mem) {
593                 mem->section_count++;
594                 kobject_put(&mem->dev.kobj);
595         } else {
596                 ret = init_memory_block(&mem, section, state);
597                 /* store memory_block pointer for next loop */
598                 if (!ret && context == BOOT)
599                         if (mem_p)
600                                 *mem_p = mem;
601         }
602
603         if (!ret) {
604                 if (context == HOTPLUG &&
605                     mem->section_count == sections_per_block)
606                         ret = register_mem_sect_under_node(mem, nid);
607         }
608
609         mutex_unlock(&mem_sysfs_mutex);
610         return ret;
611 }
612
613 int remove_memory_block(unsigned long node_id, struct mem_section *section,
614                 int phys_device)
615 {
616         struct memory_block *mem;
617
618         mutex_lock(&mem_sysfs_mutex);
619         mem = find_memory_block(section);
620         unregister_mem_sect_under_nodes(mem, __section_nr(section));
621
622         mem->section_count--;
623         if (mem->section_count == 0) {
624                 mem_remove_simple_file(mem, phys_index);
625                 mem_remove_simple_file(mem, end_phys_index);
626                 mem_remove_simple_file(mem, state);
627                 mem_remove_simple_file(mem, phys_device);
628                 mem_remove_simple_file(mem, removable);
629                 unregister_memory(mem);
630                 kfree(mem);
631         } else
632                 kobject_put(&mem->dev.kobj);
633
634         mutex_unlock(&mem_sysfs_mutex);
635         return 0;
636 }
637
638 /*
639  * need an interface for the VM to add new memory regions,
640  * but without onlining it.
641  */
642 int register_new_memory(int nid, struct mem_section *section)
643 {
644         return add_memory_section(nid, section, NULL, MEM_OFFLINE, HOTPLUG);
645 }
646
647 int unregister_memory_section(struct mem_section *section)
648 {
649         if (!present_section(section))
650                 return -EINVAL;
651
652         return remove_memory_block(0, section, 0);
653 }
654
655 /*
656  * Initialize the sysfs support for memory devices...
657  */
658 int __init memory_dev_init(void)
659 {
660         unsigned int i;
661         int ret;
662         int err;
663         unsigned long block_sz;
664         struct memory_block *mem = NULL;
665
666         ret = subsys_system_register(&memory_subsys, NULL);
667         if (ret)
668                 goto out;
669
670         block_sz = get_memory_block_size();
671         sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
672
673         /*
674          * Create entries for memory sections that were found
675          * during boot and have been initialized
676          */
677         for (i = 0; i < NR_MEM_SECTIONS; i++) {
678                 if (!present_section_nr(i))
679                         continue;
680                 /* don't need to reuse memory_block if only one per block */
681                 err = add_memory_section(0, __nr_to_section(i),
682                                  (sections_per_block == 1) ? NULL : &mem,
683                                          MEM_ONLINE,
684                                          BOOT);
685                 if (!ret)
686                         ret = err;
687         }
688
689         err = memory_probe_init();
690         if (!ret)
691                 ret = err;
692         err = memory_fail_init();
693         if (!ret)
694                 ret = err;
695         err = block_size_init();
696         if (!ret)
697                 ret = err;
698 out:
699         if (ret)
700                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
701         return ret;
702 }