]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - drivers/base/memory.c
memory hotplug: suppress "Device memoryX does not have a release() function" warning
[~shefty/rdma-dev.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 /*
81  * register_memory - Setup a sysfs device for a memory block
82  */
83 static
84 int register_memory(struct memory_block *memory)
85 {
86         int error;
87
88         memory->dev.bus = &memory_subsys;
89         memory->dev.id = memory->start_section_nr / sections_per_block;
90         memory->dev.release = memory_block_release;
91
92         error = device_register(&memory->dev);
93         return error;
94 }
95
96 static void
97 unregister_memory(struct memory_block *memory)
98 {
99         BUG_ON(memory->dev.bus != &memory_subsys);
100
101         /* drop the ref. we got in remove_memory_block() */
102         kobject_put(&memory->dev.kobj);
103         device_unregister(&memory->dev);
104 }
105
106 unsigned long __weak memory_block_size_bytes(void)
107 {
108         return MIN_MEMORY_BLOCK_SIZE;
109 }
110
111 static unsigned long get_memory_block_size(void)
112 {
113         unsigned long block_sz;
114
115         block_sz = memory_block_size_bytes();
116
117         /* Validate blk_sz is a power of 2 and not less than section size */
118         if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
119                 WARN_ON(1);
120                 block_sz = MIN_MEMORY_BLOCK_SIZE;
121         }
122
123         return block_sz;
124 }
125
126 /*
127  * use this as the physical section index that this memsection
128  * uses.
129  */
130
131 static ssize_t show_mem_start_phys_index(struct device *dev,
132                         struct device_attribute *attr, char *buf)
133 {
134         struct memory_block *mem =
135                 container_of(dev, struct memory_block, dev);
136         unsigned long phys_index;
137
138         phys_index = mem->start_section_nr / sections_per_block;
139         return sprintf(buf, "%08lx\n", phys_index);
140 }
141
142 static ssize_t show_mem_end_phys_index(struct device *dev,
143                         struct device_attribute *attr, char *buf)
144 {
145         struct memory_block *mem =
146                 container_of(dev, struct memory_block, dev);
147         unsigned long phys_index;
148
149         phys_index = mem->end_section_nr / sections_per_block;
150         return sprintf(buf, "%08lx\n", phys_index);
151 }
152
153 /*
154  * Show whether the section of memory is likely to be hot-removable
155  */
156 static ssize_t show_mem_removable(struct device *dev,
157                         struct device_attribute *attr, char *buf)
158 {
159         unsigned long i, pfn;
160         int ret = 1;
161         struct memory_block *mem =
162                 container_of(dev, struct memory_block, dev);
163
164         for (i = 0; i < sections_per_block; i++) {
165                 pfn = section_nr_to_pfn(mem->start_section_nr + i);
166                 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
167         }
168
169         return sprintf(buf, "%d\n", ret);
170 }
171
172 /*
173  * online, offline, going offline, etc.
174  */
175 static ssize_t show_mem_state(struct device *dev,
176                         struct device_attribute *attr, char *buf)
177 {
178         struct memory_block *mem =
179                 container_of(dev, struct memory_block, dev);
180         ssize_t len = 0;
181
182         /*
183          * We can probably put these states in a nice little array
184          * so that they're not open-coded
185          */
186         switch (mem->state) {
187                 case MEM_ONLINE:
188                         len = sprintf(buf, "online\n");
189                         break;
190                 case MEM_OFFLINE:
191                         len = sprintf(buf, "offline\n");
192                         break;
193                 case MEM_GOING_OFFLINE:
194                         len = sprintf(buf, "going-offline\n");
195                         break;
196                 default:
197                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
198                                         mem->state);
199                         WARN_ON(1);
200                         break;
201         }
202
203         return len;
204 }
205
206 int memory_notify(unsigned long val, void *v)
207 {
208         return blocking_notifier_call_chain(&memory_chain, val, v);
209 }
210
211 int memory_isolate_notify(unsigned long val, void *v)
212 {
213         return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
214 }
215
216 /*
217  * The probe routines leave the pages reserved, just as the bootmem code does.
218  * Make sure they're still that way.
219  */
220 static bool pages_correctly_reserved(unsigned long start_pfn,
221                                         unsigned long nr_pages)
222 {
223         int i, j;
224         struct page *page;
225         unsigned long pfn = start_pfn;
226
227         /*
228          * memmap between sections is not contiguous except with
229          * SPARSEMEM_VMEMMAP. We lookup the page once per section
230          * and assume memmap is contiguous within each section
231          */
232         for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
233                 if (WARN_ON_ONCE(!pfn_valid(pfn)))
234                         return false;
235                 page = pfn_to_page(pfn);
236
237                 for (j = 0; j < PAGES_PER_SECTION; j++) {
238                         if (PageReserved(page + j))
239                                 continue;
240
241                         printk(KERN_WARNING "section number %ld page number %d "
242                                 "not reserved, was it already online?\n",
243                                 pfn_to_section_nr(pfn), j);
244
245                         return false;
246                 }
247         }
248
249         return true;
250 }
251
252 /*
253  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
254  * OK to have direct references to sparsemem variables in here.
255  */
256 static int
257 memory_block_action(unsigned long phys_index, unsigned long action)
258 {
259         unsigned long start_pfn;
260         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
261         struct page *first_page;
262         int ret;
263
264         first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
265         start_pfn = page_to_pfn(first_page);
266
267         switch (action) {
268                 case MEM_ONLINE:
269                         if (!pages_correctly_reserved(start_pfn, nr_pages))
270                                 return -EBUSY;
271
272                         ret = online_pages(start_pfn, nr_pages);
273                         break;
274                 case MEM_OFFLINE:
275                         ret = offline_pages(start_pfn, nr_pages);
276                         break;
277                 default:
278                         WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
279                              "%ld\n", __func__, phys_index, action, action);
280                         ret = -EINVAL;
281         }
282
283         return ret;
284 }
285
286 static int __memory_block_change_state(struct memory_block *mem,
287                 unsigned long to_state, unsigned long from_state_req)
288 {
289         int ret = 0;
290
291         if (mem->state != from_state_req) {
292                 ret = -EINVAL;
293                 goto out;
294         }
295
296         if (to_state == MEM_OFFLINE)
297                 mem->state = MEM_GOING_OFFLINE;
298
299         ret = memory_block_action(mem->start_section_nr, to_state);
300
301         if (ret) {
302                 mem->state = from_state_req;
303                 goto out;
304         }
305
306         mem->state = to_state;
307         switch (mem->state) {
308         case MEM_OFFLINE:
309                 kobject_uevent(&mem->dev.kobj, KOBJ_OFFLINE);
310                 break;
311         case MEM_ONLINE:
312                 kobject_uevent(&mem->dev.kobj, KOBJ_ONLINE);
313                 break;
314         default:
315                 break;
316         }
317 out:
318         return ret;
319 }
320
321 static int memory_block_change_state(struct memory_block *mem,
322                 unsigned long to_state, unsigned long from_state_req)
323 {
324         int ret;
325
326         mutex_lock(&mem->state_mutex);
327         ret = __memory_block_change_state(mem, to_state, from_state_req);
328         mutex_unlock(&mem->state_mutex);
329
330         return ret;
331 }
332 static ssize_t
333 store_mem_state(struct device *dev,
334                 struct device_attribute *attr, const char *buf, size_t count)
335 {
336         struct memory_block *mem;
337         int ret = -EINVAL;
338
339         mem = container_of(dev, struct memory_block, dev);
340
341         if (!strncmp(buf, "online", min((int)count, 6)))
342                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
343         else if(!strncmp(buf, "offline", min((int)count, 7)))
344                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
345
346         if (ret)
347                 return ret;
348         return count;
349 }
350
351 /*
352  * phys_device is a bad name for this.  What I really want
353  * is a way to differentiate between memory ranges that
354  * are part of physical devices that constitute
355  * a complete removable unit or fru.
356  * i.e. do these ranges belong to the same physical device,
357  * s.t. if I offline all of these sections I can then
358  * remove the physical device?
359  */
360 static ssize_t show_phys_device(struct device *dev,
361                                 struct device_attribute *attr, char *buf)
362 {
363         struct memory_block *mem =
364                 container_of(dev, struct memory_block, dev);
365         return sprintf(buf, "%d\n", mem->phys_device);
366 }
367
368 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
369 static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
370 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
371 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
372 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
373
374 #define mem_create_simple_file(mem, attr_name)  \
375         device_create_file(&mem->dev, &dev_attr_##attr_name)
376 #define mem_remove_simple_file(mem, attr_name)  \
377         device_remove_file(&mem->dev, &dev_attr_##attr_name)
378
379 /*
380  * Block size attribute stuff
381  */
382 static ssize_t
383 print_block_size(struct device *dev, struct device_attribute *attr,
384                  char *buf)
385 {
386         return sprintf(buf, "%lx\n", get_memory_block_size());
387 }
388
389 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
390
391 static int block_size_init(void)
392 {
393         return device_create_file(memory_subsys.dev_root,
394                                   &dev_attr_block_size_bytes);
395 }
396
397 /*
398  * Some architectures will have custom drivers to do this, and
399  * will not need to do it from userspace.  The fake hot-add code
400  * as well as ppc64 will do all of their discovery in userspace
401  * and will require this interface.
402  */
403 #ifdef CONFIG_ARCH_MEMORY_PROBE
404 static ssize_t
405 memory_probe_store(struct device *dev, struct device_attribute *attr,
406                    const char *buf, size_t count)
407 {
408         u64 phys_addr;
409         int nid;
410         int i, ret;
411         unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
412
413         phys_addr = simple_strtoull(buf, NULL, 0);
414
415         if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
416                 return -EINVAL;
417
418         for (i = 0; i < sections_per_block; i++) {
419                 nid = memory_add_physaddr_to_nid(phys_addr);
420                 ret = add_memory(nid, phys_addr,
421                                  PAGES_PER_SECTION << PAGE_SHIFT);
422                 if (ret)
423                         goto out;
424
425                 phys_addr += MIN_MEMORY_BLOCK_SIZE;
426         }
427
428         ret = count;
429 out:
430         return ret;
431 }
432 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
433
434 static int memory_probe_init(void)
435 {
436         return device_create_file(memory_subsys.dev_root, &dev_attr_probe);
437 }
438 #else
439 static inline int memory_probe_init(void)
440 {
441         return 0;
442 }
443 #endif
444
445 #ifdef CONFIG_MEMORY_FAILURE
446 /*
447  * Support for offlining pages of memory
448  */
449
450 /* Soft offline a page */
451 static ssize_t
452 store_soft_offline_page(struct device *dev,
453                         struct device_attribute *attr,
454                         const char *buf, size_t count)
455 {
456         int ret;
457         u64 pfn;
458         if (!capable(CAP_SYS_ADMIN))
459                 return -EPERM;
460         if (strict_strtoull(buf, 0, &pfn) < 0)
461                 return -EINVAL;
462         pfn >>= PAGE_SHIFT;
463         if (!pfn_valid(pfn))
464                 return -ENXIO;
465         ret = soft_offline_page(pfn_to_page(pfn), 0);
466         return ret == 0 ? count : ret;
467 }
468
469 /* Forcibly offline a page, including killing processes. */
470 static ssize_t
471 store_hard_offline_page(struct device *dev,
472                         struct device_attribute *attr,
473                         const char *buf, size_t count)
474 {
475         int ret;
476         u64 pfn;
477         if (!capable(CAP_SYS_ADMIN))
478                 return -EPERM;
479         if (strict_strtoull(buf, 0, &pfn) < 0)
480                 return -EINVAL;
481         pfn >>= PAGE_SHIFT;
482         ret = memory_failure(pfn, 0, 0);
483         return ret ? ret : count;
484 }
485
486 static DEVICE_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
487 static DEVICE_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
488
489 static __init int memory_fail_init(void)
490 {
491         int err;
492
493         err = device_create_file(memory_subsys.dev_root,
494                                 &dev_attr_soft_offline_page);
495         if (!err)
496                 err = device_create_file(memory_subsys.dev_root,
497                                 &dev_attr_hard_offline_page);
498         return err;
499 }
500 #else
501 static inline int memory_fail_init(void)
502 {
503         return 0;
504 }
505 #endif
506
507 /*
508  * Note that phys_device is optional.  It is here to allow for
509  * differentiation between which *physical* devices each
510  * section belongs to...
511  */
512 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
513 {
514         return 0;
515 }
516
517 /*
518  * A reference for the returned object is held and the reference for the
519  * hinted object is released.
520  */
521 struct memory_block *find_memory_block_hinted(struct mem_section *section,
522                                               struct memory_block *hint)
523 {
524         int block_id = base_memory_block_id(__section_nr(section));
525         struct device *hintdev = hint ? &hint->dev : NULL;
526         struct device *dev;
527
528         dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
529         if (hint)
530                 put_device(&hint->dev);
531         if (!dev)
532                 return NULL;
533         return container_of(dev, struct memory_block, dev);
534 }
535
536 /*
537  * For now, we have a linear search to go find the appropriate
538  * memory_block corresponding to a particular phys_index. If
539  * this gets to be a real problem, we can always use a radix
540  * tree or something here.
541  *
542  * This could be made generic for all device subsystems.
543  */
544 struct memory_block *find_memory_block(struct mem_section *section)
545 {
546         return find_memory_block_hinted(section, NULL);
547 }
548
549 static int init_memory_block(struct memory_block **memory,
550                              struct mem_section *section, unsigned long state)
551 {
552         struct memory_block *mem;
553         unsigned long start_pfn;
554         int scn_nr;
555         int ret = 0;
556
557         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
558         if (!mem)
559                 return -ENOMEM;
560
561         scn_nr = __section_nr(section);
562         mem->start_section_nr =
563                         base_memory_block_id(scn_nr) * sections_per_block;
564         mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
565         mem->state = state;
566         mem->section_count++;
567         mutex_init(&mem->state_mutex);
568         start_pfn = section_nr_to_pfn(mem->start_section_nr);
569         mem->phys_device = arch_get_memory_phys_device(start_pfn);
570
571         ret = register_memory(mem);
572         if (!ret)
573                 ret = mem_create_simple_file(mem, phys_index);
574         if (!ret)
575                 ret = mem_create_simple_file(mem, end_phys_index);
576         if (!ret)
577                 ret = mem_create_simple_file(mem, state);
578         if (!ret)
579                 ret = mem_create_simple_file(mem, phys_device);
580         if (!ret)
581                 ret = mem_create_simple_file(mem, removable);
582
583         *memory = mem;
584         return ret;
585 }
586
587 static int add_memory_section(int nid, struct mem_section *section,
588                         struct memory_block **mem_p,
589                         unsigned long state, enum mem_add_context context)
590 {
591         struct memory_block *mem = NULL;
592         int scn_nr = __section_nr(section);
593         int ret = 0;
594
595         mutex_lock(&mem_sysfs_mutex);
596
597         if (context == BOOT) {
598                 /* same memory block ? */
599                 if (mem_p && *mem_p)
600                         if (scn_nr >= (*mem_p)->start_section_nr &&
601                             scn_nr <= (*mem_p)->end_section_nr) {
602                                 mem = *mem_p;
603                                 kobject_get(&mem->dev.kobj);
604                         }
605         } else
606                 mem = find_memory_block(section);
607
608         if (mem) {
609                 mem->section_count++;
610                 kobject_put(&mem->dev.kobj);
611         } else {
612                 ret = init_memory_block(&mem, section, state);
613                 /* store memory_block pointer for next loop */
614                 if (!ret && context == BOOT)
615                         if (mem_p)
616                                 *mem_p = mem;
617         }
618
619         if (!ret) {
620                 if (context == HOTPLUG &&
621                     mem->section_count == sections_per_block)
622                         ret = register_mem_sect_under_node(mem, nid);
623         }
624
625         mutex_unlock(&mem_sysfs_mutex);
626         return ret;
627 }
628
629 int remove_memory_block(unsigned long node_id, struct mem_section *section,
630                 int phys_device)
631 {
632         struct memory_block *mem;
633
634         mutex_lock(&mem_sysfs_mutex);
635         mem = find_memory_block(section);
636         unregister_mem_sect_under_nodes(mem, __section_nr(section));
637
638         mem->section_count--;
639         if (mem->section_count == 0) {
640                 mem_remove_simple_file(mem, phys_index);
641                 mem_remove_simple_file(mem, end_phys_index);
642                 mem_remove_simple_file(mem, state);
643                 mem_remove_simple_file(mem, phys_device);
644                 mem_remove_simple_file(mem, removable);
645                 unregister_memory(mem);
646         } else
647                 kobject_put(&mem->dev.kobj);
648
649         mutex_unlock(&mem_sysfs_mutex);
650         return 0;
651 }
652
653 /*
654  * need an interface for the VM to add new memory regions,
655  * but without onlining it.
656  */
657 int register_new_memory(int nid, struct mem_section *section)
658 {
659         return add_memory_section(nid, section, NULL, MEM_OFFLINE, HOTPLUG);
660 }
661
662 int unregister_memory_section(struct mem_section *section)
663 {
664         if (!present_section(section))
665                 return -EINVAL;
666
667         return remove_memory_block(0, section, 0);
668 }
669
670 /*
671  * offline one memory block. If the memory block has been offlined, do nothing.
672  */
673 int offline_memory_block(struct memory_block *mem)
674 {
675         int ret = 0;
676
677         mutex_lock(&mem->state_mutex);
678         if (mem->state != MEM_OFFLINE)
679                 ret = __memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
680         mutex_unlock(&mem->state_mutex);
681
682         return ret;
683 }
684
685 /*
686  * Initialize the sysfs support for memory devices...
687  */
688 int __init memory_dev_init(void)
689 {
690         unsigned int i;
691         int ret;
692         int err;
693         unsigned long block_sz;
694         struct memory_block *mem = NULL;
695
696         ret = subsys_system_register(&memory_subsys, NULL);
697         if (ret)
698                 goto out;
699
700         block_sz = get_memory_block_size();
701         sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
702
703         /*
704          * Create entries for memory sections that were found
705          * during boot and have been initialized
706          */
707         for (i = 0; i < NR_MEM_SECTIONS; i++) {
708                 if (!present_section_nr(i))
709                         continue;
710                 /* don't need to reuse memory_block if only one per block */
711                 err = add_memory_section(0, __nr_to_section(i),
712                                  (sections_per_block == 1) ? NULL : &mem,
713                                          MEM_ONLINE,
714                                          BOOT);
715                 if (!ret)
716                         ret = err;
717         }
718
719         err = memory_probe_init();
720         if (!ret)
721                 ret = err;
722         err = memory_fail_init();
723         if (!ret)
724                 ret = err;
725         err = block_size_init();
726         if (!ret)
727                 ret = err;
728 out:
729         if (ret)
730                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
731         return ret;
732 }