Merge tag 'virtio-next-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[deliverable/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/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
28 static 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
34 static int sections_per_block;
35
36 static inline int base_memory_block_id(int section_nr)
37 {
38 return section_nr / sections_per_block;
39 }
40
41 static int memory_subsys_online(struct device *dev);
42 static int memory_subsys_offline(struct device *dev);
43
44 static 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
51 static BLOCKING_NOTIFIER_HEAD(memory_chain);
52
53 int register_memory_notifier(struct notifier_block *nb)
54 {
55 return blocking_notifier_chain_register(&memory_chain, nb);
56 }
57 EXPORT_SYMBOL(register_memory_notifier);
58
59 void unregister_memory_notifier(struct notifier_block *nb)
60 {
61 blocking_notifier_chain_unregister(&memory_chain, nb);
62 }
63 EXPORT_SYMBOL(unregister_memory_notifier);
64
65 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
66
67 int register_memory_isolate_notifier(struct notifier_block *nb)
68 {
69 return atomic_notifier_chain_register(&memory_isolate_chain, nb);
70 }
71 EXPORT_SYMBOL(register_memory_isolate_notifier);
72
73 void unregister_memory_isolate_notifier(struct notifier_block *nb)
74 {
75 atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
76 }
77 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
78
79 static void memory_block_release(struct device *dev)
80 {
81 struct memory_block *mem = to_memory_block(dev);
82
83 kfree(mem);
84 }
85
86 unsigned long __weak memory_block_size_bytes(void)
87 {
88 return MIN_MEMORY_BLOCK_SIZE;
89 }
90
91 static 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
111 static 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 */
124 static 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 */
144 static 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
174 int memory_notify(unsigned long val, void *v)
175 {
176 return blocking_notifier_call_chain(&memory_chain, val, v);
177 }
178
179 int 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 */
188 static 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 */
223 static int
224 memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
225 {
226 unsigned long start_pfn;
227 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
228 struct page *first_page;
229 int ret;
230
231 first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
232 start_pfn = page_to_pfn(first_page);
233
234 switch (action) {
235 case MEM_ONLINE:
236 if (!pages_correctly_reserved(start_pfn))
237 return -EBUSY;
238
239 ret = online_pages(start_pfn, nr_pages, online_type);
240 break;
241 case MEM_OFFLINE:
242 ret = offline_pages(start_pfn, nr_pages);
243 break;
244 default:
245 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
246 "%ld\n", __func__, phys_index, action, action);
247 ret = -EINVAL;
248 }
249
250 return ret;
251 }
252
253 static int memory_block_change_state(struct memory_block *mem,
254 unsigned long to_state, unsigned long from_state_req)
255 {
256 int ret = 0;
257
258 if (mem->state != from_state_req)
259 return -EINVAL;
260
261 if (to_state == MEM_OFFLINE)
262 mem->state = MEM_GOING_OFFLINE;
263
264 ret = memory_block_action(mem->start_section_nr, to_state,
265 mem->online_type);
266
267 mem->state = ret ? from_state_req : to_state;
268
269 return ret;
270 }
271
272 /* The device lock serializes operations on memory_subsys_[online|offline] */
273 static int memory_subsys_online(struct device *dev)
274 {
275 struct memory_block *mem = to_memory_block(dev);
276 int ret;
277
278 if (mem->state == MEM_ONLINE)
279 return 0;
280
281 /*
282 * If we are called from store_mem_state(), online_type will be
283 * set >= 0 Otherwise we were called from the device online
284 * attribute and need to set the online_type.
285 */
286 if (mem->online_type < 0)
287 mem->online_type = MMOP_ONLINE_KEEP;
288
289 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
290
291 /* clear online_type */
292 mem->online_type = -1;
293
294 return ret;
295 }
296
297 static int memory_subsys_offline(struct device *dev)
298 {
299 struct memory_block *mem = to_memory_block(dev);
300
301 if (mem->state == MEM_OFFLINE)
302 return 0;
303
304 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
305 }
306
307 static ssize_t
308 store_mem_state(struct device *dev,
309 struct device_attribute *attr, const char *buf, size_t count)
310 {
311 struct memory_block *mem = to_memory_block(dev);
312 int ret, online_type;
313
314 ret = lock_device_hotplug_sysfs();
315 if (ret)
316 return ret;
317
318 if (sysfs_streq(buf, "online_kernel"))
319 online_type = MMOP_ONLINE_KERNEL;
320 else if (sysfs_streq(buf, "online_movable"))
321 online_type = MMOP_ONLINE_MOVABLE;
322 else if (sysfs_streq(buf, "online"))
323 online_type = MMOP_ONLINE_KEEP;
324 else if (sysfs_streq(buf, "offline"))
325 online_type = MMOP_OFFLINE;
326 else {
327 ret = -EINVAL;
328 goto err;
329 }
330
331 switch (online_type) {
332 case MMOP_ONLINE_KERNEL:
333 case MMOP_ONLINE_MOVABLE:
334 case MMOP_ONLINE_KEEP:
335 /*
336 * mem->online_type is not protected so there can be a
337 * race here. However, when racing online, the first
338 * will succeed and the second will just return as the
339 * block will already be online. The online type
340 * could be either one, but that is expected.
341 */
342 mem->online_type = online_type;
343 ret = device_online(&mem->dev);
344 break;
345 case MMOP_OFFLINE:
346 ret = device_offline(&mem->dev);
347 break;
348 default:
349 ret = -EINVAL; /* should never happen */
350 }
351
352 err:
353 unlock_device_hotplug();
354
355 if (ret)
356 return ret;
357 return count;
358 }
359
360 /*
361 * phys_device is a bad name for this. What I really want
362 * is a way to differentiate between memory ranges that
363 * are part of physical devices that constitute
364 * a complete removable unit or fru.
365 * i.e. do these ranges belong to the same physical device,
366 * s.t. if I offline all of these sections I can then
367 * remove the physical device?
368 */
369 static ssize_t show_phys_device(struct device *dev,
370 struct device_attribute *attr, char *buf)
371 {
372 struct memory_block *mem = to_memory_block(dev);
373 return sprintf(buf, "%d\n", mem->phys_device);
374 }
375
376 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
377 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
378 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
379 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
380
381 /*
382 * Block size attribute stuff
383 */
384 static ssize_t
385 print_block_size(struct device *dev, struct device_attribute *attr,
386 char *buf)
387 {
388 return sprintf(buf, "%lx\n", get_memory_block_size());
389 }
390
391 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
392
393 /*
394 * Some architectures will have custom drivers to do this, and
395 * will not need to do it from userspace. The fake hot-add code
396 * as well as ppc64 will do all of their discovery in userspace
397 * and will require this interface.
398 */
399 #ifdef CONFIG_ARCH_MEMORY_PROBE
400 static ssize_t
401 memory_probe_store(struct device *dev, struct device_attribute *attr,
402 const char *buf, size_t count)
403 {
404 u64 phys_addr;
405 int nid;
406 int i, ret;
407 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
408
409 ret = kstrtoull(buf, 0, &phys_addr);
410 if (ret)
411 return ret;
412
413 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
414 return -EINVAL;
415
416 for (i = 0; i < sections_per_block; i++) {
417 nid = memory_add_physaddr_to_nid(phys_addr);
418 ret = add_memory(nid, phys_addr,
419 PAGES_PER_SECTION << PAGE_SHIFT);
420 if (ret)
421 goto out;
422
423 phys_addr += MIN_MEMORY_BLOCK_SIZE;
424 }
425
426 ret = count;
427 out:
428 return ret;
429 }
430
431 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
432 #endif
433
434 #ifdef CONFIG_MEMORY_FAILURE
435 /*
436 * Support for offlining pages of memory
437 */
438
439 /* Soft offline a page */
440 static ssize_t
441 store_soft_offline_page(struct device *dev,
442 struct device_attribute *attr,
443 const char *buf, size_t count)
444 {
445 int ret;
446 u64 pfn;
447 if (!capable(CAP_SYS_ADMIN))
448 return -EPERM;
449 if (kstrtoull(buf, 0, &pfn) < 0)
450 return -EINVAL;
451 pfn >>= PAGE_SHIFT;
452 if (!pfn_valid(pfn))
453 return -ENXIO;
454 ret = soft_offline_page(pfn_to_page(pfn), 0);
455 return ret == 0 ? count : ret;
456 }
457
458 /* Forcibly offline a page, including killing processes. */
459 static ssize_t
460 store_hard_offline_page(struct device *dev,
461 struct device_attribute *attr,
462 const char *buf, size_t count)
463 {
464 int ret;
465 u64 pfn;
466 if (!capable(CAP_SYS_ADMIN))
467 return -EPERM;
468 if (kstrtoull(buf, 0, &pfn) < 0)
469 return -EINVAL;
470 pfn >>= PAGE_SHIFT;
471 ret = memory_failure(pfn, 0, 0);
472 return ret ? ret : count;
473 }
474
475 static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
476 static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
477 #endif
478
479 /*
480 * Note that phys_device is optional. It is here to allow for
481 * differentiation between which *physical* devices each
482 * section belongs to...
483 */
484 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
485 {
486 return 0;
487 }
488
489 /*
490 * A reference for the returned object is held and the reference for the
491 * hinted object is released.
492 */
493 struct memory_block *find_memory_block_hinted(struct mem_section *section,
494 struct memory_block *hint)
495 {
496 int block_id = base_memory_block_id(__section_nr(section));
497 struct device *hintdev = hint ? &hint->dev : NULL;
498 struct device *dev;
499
500 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
501 if (hint)
502 put_device(&hint->dev);
503 if (!dev)
504 return NULL;
505 return to_memory_block(dev);
506 }
507
508 /*
509 * For now, we have a linear search to go find the appropriate
510 * memory_block corresponding to a particular phys_index. If
511 * this gets to be a real problem, we can always use a radix
512 * tree or something here.
513 *
514 * This could be made generic for all device subsystems.
515 */
516 struct memory_block *find_memory_block(struct mem_section *section)
517 {
518 return find_memory_block_hinted(section, NULL);
519 }
520
521 static struct attribute *memory_memblk_attrs[] = {
522 &dev_attr_phys_index.attr,
523 &dev_attr_state.attr,
524 &dev_attr_phys_device.attr,
525 &dev_attr_removable.attr,
526 NULL
527 };
528
529 static struct attribute_group memory_memblk_attr_group = {
530 .attrs = memory_memblk_attrs,
531 };
532
533 static const struct attribute_group *memory_memblk_attr_groups[] = {
534 &memory_memblk_attr_group,
535 NULL,
536 };
537
538 /*
539 * register_memory - Setup a sysfs device for a memory block
540 */
541 static
542 int register_memory(struct memory_block *memory)
543 {
544 memory->dev.bus = &memory_subsys;
545 memory->dev.id = memory->start_section_nr / sections_per_block;
546 memory->dev.release = memory_block_release;
547 memory->dev.groups = memory_memblk_attr_groups;
548 memory->dev.offline = memory->state == MEM_OFFLINE;
549
550 return device_register(&memory->dev);
551 }
552
553 static int init_memory_block(struct memory_block **memory,
554 struct mem_section *section, unsigned long state)
555 {
556 struct memory_block *mem;
557 unsigned long start_pfn;
558 int scn_nr;
559 int ret = 0;
560
561 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
562 if (!mem)
563 return -ENOMEM;
564
565 scn_nr = __section_nr(section);
566 mem->start_section_nr =
567 base_memory_block_id(scn_nr) * sections_per_block;
568 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
569 mem->state = state;
570 mem->section_count++;
571 start_pfn = section_nr_to_pfn(mem->start_section_nr);
572 mem->phys_device = arch_get_memory_phys_device(start_pfn);
573
574 ret = register_memory(mem);
575
576 *memory = mem;
577 return ret;
578 }
579
580 static int add_memory_block(int base_section_nr)
581 {
582 struct memory_block *mem;
583 int i, ret, section_count = 0, section_nr;
584
585 for (i = base_section_nr;
586 (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
587 i++) {
588 if (!present_section_nr(i))
589 continue;
590 if (section_count == 0)
591 section_nr = i;
592 section_count++;
593 }
594
595 if (section_count == 0)
596 return 0;
597 ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
598 if (ret)
599 return ret;
600 mem->section_count = section_count;
601 return 0;
602 }
603
604
605 /*
606 * need an interface for the VM to add new memory regions,
607 * but without onlining it.
608 */
609 int register_new_memory(int nid, struct mem_section *section)
610 {
611 int ret = 0;
612 struct memory_block *mem;
613
614 mutex_lock(&mem_sysfs_mutex);
615
616 mem = find_memory_block(section);
617 if (mem) {
618 mem->section_count++;
619 put_device(&mem->dev);
620 } else {
621 ret = init_memory_block(&mem, section, MEM_OFFLINE);
622 if (ret)
623 goto out;
624 }
625
626 if (mem->section_count == sections_per_block)
627 ret = register_mem_sect_under_node(mem, nid);
628 out:
629 mutex_unlock(&mem_sysfs_mutex);
630 return ret;
631 }
632
633 #ifdef CONFIG_MEMORY_HOTREMOVE
634 static void
635 unregister_memory(struct memory_block *memory)
636 {
637 BUG_ON(memory->dev.bus != &memory_subsys);
638
639 /* drop the ref. we got in remove_memory_block() */
640 put_device(&memory->dev);
641 device_unregister(&memory->dev);
642 }
643
644 static int remove_memory_block(unsigned long node_id,
645 struct mem_section *section, int phys_device)
646 {
647 struct memory_block *mem;
648
649 mutex_lock(&mem_sysfs_mutex);
650 mem = find_memory_block(section);
651 unregister_mem_sect_under_nodes(mem, __section_nr(section));
652
653 mem->section_count--;
654 if (mem->section_count == 0)
655 unregister_memory(mem);
656 else
657 put_device(&mem->dev);
658
659 mutex_unlock(&mem_sysfs_mutex);
660 return 0;
661 }
662
663 int unregister_memory_section(struct mem_section *section)
664 {
665 if (!present_section(section))
666 return -EINVAL;
667
668 return remove_memory_block(0, section, 0);
669 }
670 #endif /* CONFIG_MEMORY_HOTREMOVE */
671
672 /* return true if the memory block is offlined, otherwise, return false */
673 bool is_memblock_offlined(struct memory_block *mem)
674 {
675 return mem->state == MEM_OFFLINE;
676 }
677
678 static struct attribute *memory_root_attrs[] = {
679 #ifdef CONFIG_ARCH_MEMORY_PROBE
680 &dev_attr_probe.attr,
681 #endif
682
683 #ifdef CONFIG_MEMORY_FAILURE
684 &dev_attr_soft_offline_page.attr,
685 &dev_attr_hard_offline_page.attr,
686 #endif
687
688 &dev_attr_block_size_bytes.attr,
689 NULL
690 };
691
692 static struct attribute_group memory_root_attr_group = {
693 .attrs = memory_root_attrs,
694 };
695
696 static const struct attribute_group *memory_root_attr_groups[] = {
697 &memory_root_attr_group,
698 NULL,
699 };
700
701 /*
702 * Initialize the sysfs support for memory devices...
703 */
704 int __init memory_dev_init(void)
705 {
706 unsigned int i;
707 int ret;
708 int err;
709 unsigned long block_sz;
710
711 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
712 if (ret)
713 goto out;
714
715 block_sz = get_memory_block_size();
716 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
717
718 /*
719 * Create entries for memory sections that were found
720 * during boot and have been initialized
721 */
722 mutex_lock(&mem_sysfs_mutex);
723 for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
724 err = add_memory_block(i);
725 if (!ret)
726 ret = err;
727 }
728 mutex_unlock(&mem_sysfs_mutex);
729
730 out:
731 if (ret)
732 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
733 return ret;
734 }
This page took 0.109277 seconds and 5 git commands to generate.