]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - drivers/base/memory.c
Merge git://git.kvack.org/~bcrl/aio-next
[mirror_ubuntu-bionic-kernel.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 static ssize_t show_mem_end_phys_index(struct device *dev,
122 struct device_attribute *attr, char *buf)
123 {
124 struct memory_block *mem = to_memory_block(dev);
125 unsigned long phys_index;
126
127 phys_index = mem->end_section_nr / sections_per_block;
128 return sprintf(buf, "%08lx\n", phys_index);
129 }
130
131 /*
132 * Show whether the section of memory is likely to be hot-removable
133 */
134 static ssize_t show_mem_removable(struct device *dev,
135 struct device_attribute *attr, char *buf)
136 {
137 unsigned long i, pfn;
138 int ret = 1;
139 struct memory_block *mem = to_memory_block(dev);
140
141 for (i = 0; i < sections_per_block; i++) {
142 if (!present_section_nr(mem->start_section_nr + i))
143 continue;
144 pfn = section_nr_to_pfn(mem->start_section_nr + i);
145 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
146 }
147
148 return sprintf(buf, "%d\n", ret);
149 }
150
151 /*
152 * online, offline, going offline, etc.
153 */
154 static ssize_t show_mem_state(struct device *dev,
155 struct device_attribute *attr, char *buf)
156 {
157 struct memory_block *mem = to_memory_block(dev);
158 ssize_t len = 0;
159
160 /*
161 * We can probably put these states in a nice little array
162 * so that they're not open-coded
163 */
164 switch (mem->state) {
165 case MEM_ONLINE:
166 len = sprintf(buf, "online\n");
167 break;
168 case MEM_OFFLINE:
169 len = sprintf(buf, "offline\n");
170 break;
171 case MEM_GOING_OFFLINE:
172 len = sprintf(buf, "going-offline\n");
173 break;
174 default:
175 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
176 mem->state);
177 WARN_ON(1);
178 break;
179 }
180
181 return len;
182 }
183
184 int memory_notify(unsigned long val, void *v)
185 {
186 return blocking_notifier_call_chain(&memory_chain, val, v);
187 }
188
189 int memory_isolate_notify(unsigned long val, void *v)
190 {
191 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
192 }
193
194 /*
195 * The probe routines leave the pages reserved, just as the bootmem code does.
196 * Make sure they're still that way.
197 */
198 static bool pages_correctly_reserved(unsigned long start_pfn)
199 {
200 int i, j;
201 struct page *page;
202 unsigned long pfn = start_pfn;
203
204 /*
205 * memmap between sections is not contiguous except with
206 * SPARSEMEM_VMEMMAP. We lookup the page once per section
207 * and assume memmap is contiguous within each section
208 */
209 for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
210 if (WARN_ON_ONCE(!pfn_valid(pfn)))
211 return false;
212 page = pfn_to_page(pfn);
213
214 for (j = 0; j < PAGES_PER_SECTION; j++) {
215 if (PageReserved(page + j))
216 continue;
217
218 printk(KERN_WARNING "section number %ld page number %d "
219 "not reserved, was it already online?\n",
220 pfn_to_section_nr(pfn), j);
221
222 return false;
223 }
224 }
225
226 return true;
227 }
228
229 /*
230 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
231 * OK to have direct references to sparsemem variables in here.
232 */
233 static int
234 memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
235 {
236 unsigned long start_pfn;
237 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
238 struct page *first_page;
239 int ret;
240
241 first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
242 start_pfn = page_to_pfn(first_page);
243
244 switch (action) {
245 case MEM_ONLINE:
246 if (!pages_correctly_reserved(start_pfn))
247 return -EBUSY;
248
249 ret = online_pages(start_pfn, nr_pages, online_type);
250 break;
251 case MEM_OFFLINE:
252 ret = offline_pages(start_pfn, nr_pages);
253 break;
254 default:
255 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
256 "%ld\n", __func__, phys_index, action, action);
257 ret = -EINVAL;
258 }
259
260 return ret;
261 }
262
263 static int memory_block_change_state(struct memory_block *mem,
264 unsigned long to_state, unsigned long from_state_req)
265 {
266 int ret = 0;
267
268 if (mem->state != from_state_req)
269 return -EINVAL;
270
271 if (to_state == MEM_OFFLINE)
272 mem->state = MEM_GOING_OFFLINE;
273
274 ret = memory_block_action(mem->start_section_nr, to_state,
275 mem->online_type);
276
277 mem->state = ret ? from_state_req : to_state;
278
279 return ret;
280 }
281
282 /* The device lock serializes operations on memory_subsys_[online|offline] */
283 static int memory_subsys_online(struct device *dev)
284 {
285 struct memory_block *mem = to_memory_block(dev);
286 int ret;
287
288 if (mem->state == MEM_ONLINE)
289 return 0;
290
291 /*
292 * If we are called from store_mem_state(), online_type will be
293 * set >= 0 Otherwise we were called from the device online
294 * attribute and need to set the online_type.
295 */
296 if (mem->online_type < 0)
297 mem->online_type = ONLINE_KEEP;
298
299 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
300
301 /* clear online_type */
302 mem->online_type = -1;
303
304 return ret;
305 }
306
307 static int memory_subsys_offline(struct device *dev)
308 {
309 struct memory_block *mem = to_memory_block(dev);
310
311 if (mem->state == MEM_OFFLINE)
312 return 0;
313
314 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
315 }
316
317 static ssize_t
318 store_mem_state(struct device *dev,
319 struct device_attribute *attr, const char *buf, size_t count)
320 {
321 struct memory_block *mem = to_memory_block(dev);
322 int ret, online_type;
323
324 ret = lock_device_hotplug_sysfs();
325 if (ret)
326 return ret;
327
328 if (!strncmp(buf, "online_kernel", min_t(int, count, 13)))
329 online_type = ONLINE_KERNEL;
330 else if (!strncmp(buf, "online_movable", min_t(int, count, 14)))
331 online_type = ONLINE_MOVABLE;
332 else if (!strncmp(buf, "online", min_t(int, count, 6)))
333 online_type = ONLINE_KEEP;
334 else if (!strncmp(buf, "offline", min_t(int, count, 7)))
335 online_type = -1;
336 else
337 return -EINVAL;
338
339 switch (online_type) {
340 case ONLINE_KERNEL:
341 case ONLINE_MOVABLE:
342 case ONLINE_KEEP:
343 /*
344 * mem->online_type is not protected so there can be a
345 * race here. However, when racing online, the first
346 * will succeed and the second will just return as the
347 * block will already be online. The online type
348 * could be either one, but that is expected.
349 */
350 mem->online_type = online_type;
351 ret = device_online(&mem->dev);
352 break;
353 case -1:
354 ret = device_offline(&mem->dev);
355 break;
356 default:
357 ret = -EINVAL; /* should never happen */
358 }
359
360 unlock_device_hotplug();
361
362 if (ret)
363 return ret;
364 return count;
365 }
366
367 /*
368 * phys_device is a bad name for this. What I really want
369 * is a way to differentiate between memory ranges that
370 * are part of physical devices that constitute
371 * a complete removable unit or fru.
372 * i.e. do these ranges belong to the same physical device,
373 * s.t. if I offline all of these sections I can then
374 * remove the physical device?
375 */
376 static ssize_t show_phys_device(struct device *dev,
377 struct device_attribute *attr, char *buf)
378 {
379 struct memory_block *mem = to_memory_block(dev);
380 return sprintf(buf, "%d\n", mem->phys_device);
381 }
382
383 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
384 static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
385 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
386 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
387 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
388
389 /*
390 * Block size attribute stuff
391 */
392 static ssize_t
393 print_block_size(struct device *dev, struct device_attribute *attr,
394 char *buf)
395 {
396 return sprintf(buf, "%lx\n", get_memory_block_size());
397 }
398
399 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
400
401 /*
402 * Some architectures will have custom drivers to do this, and
403 * will not need to do it from userspace. The fake hot-add code
404 * as well as ppc64 will do all of their discovery in userspace
405 * and will require this interface.
406 */
407 #ifdef CONFIG_ARCH_MEMORY_PROBE
408 static ssize_t
409 memory_probe_store(struct device *dev, struct device_attribute *attr,
410 const char *buf, size_t count)
411 {
412 u64 phys_addr;
413 int nid;
414 int i, ret;
415 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
416
417 phys_addr = simple_strtoull(buf, NULL, 0);
418
419 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
420 return -EINVAL;
421
422 for (i = 0; i < sections_per_block; i++) {
423 nid = memory_add_physaddr_to_nid(phys_addr);
424 ret = add_memory(nid, phys_addr,
425 PAGES_PER_SECTION << PAGE_SHIFT);
426 if (ret)
427 goto out;
428
429 phys_addr += MIN_MEMORY_BLOCK_SIZE;
430 }
431
432 ret = count;
433 out:
434 return ret;
435 }
436
437 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
438 #endif
439
440 #ifdef CONFIG_MEMORY_FAILURE
441 /*
442 * Support for offlining pages of memory
443 */
444
445 /* Soft offline a page */
446 static ssize_t
447 store_soft_offline_page(struct device *dev,
448 struct device_attribute *attr,
449 const char *buf, size_t count)
450 {
451 int ret;
452 u64 pfn;
453 if (!capable(CAP_SYS_ADMIN))
454 return -EPERM;
455 if (kstrtoull(buf, 0, &pfn) < 0)
456 return -EINVAL;
457 pfn >>= PAGE_SHIFT;
458 if (!pfn_valid(pfn))
459 return -ENXIO;
460 ret = soft_offline_page(pfn_to_page(pfn), 0);
461 return ret == 0 ? count : ret;
462 }
463
464 /* Forcibly offline a page, including killing processes. */
465 static ssize_t
466 store_hard_offline_page(struct device *dev,
467 struct device_attribute *attr,
468 const char *buf, size_t count)
469 {
470 int ret;
471 u64 pfn;
472 if (!capable(CAP_SYS_ADMIN))
473 return -EPERM;
474 if (kstrtoull(buf, 0, &pfn) < 0)
475 return -EINVAL;
476 pfn >>= PAGE_SHIFT;
477 ret = memory_failure(pfn, 0, 0);
478 return ret ? ret : count;
479 }
480
481 static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
482 static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
483 #endif
484
485 /*
486 * Note that phys_device is optional. It is here to allow for
487 * differentiation between which *physical* devices each
488 * section belongs to...
489 */
490 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
491 {
492 return 0;
493 }
494
495 /*
496 * A reference for the returned object is held and the reference for the
497 * hinted object is released.
498 */
499 struct memory_block *find_memory_block_hinted(struct mem_section *section,
500 struct memory_block *hint)
501 {
502 int block_id = base_memory_block_id(__section_nr(section));
503 struct device *hintdev = hint ? &hint->dev : NULL;
504 struct device *dev;
505
506 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
507 if (hint)
508 put_device(&hint->dev);
509 if (!dev)
510 return NULL;
511 return to_memory_block(dev);
512 }
513
514 /*
515 * For now, we have a linear search to go find the appropriate
516 * memory_block corresponding to a particular phys_index. If
517 * this gets to be a real problem, we can always use a radix
518 * tree or something here.
519 *
520 * This could be made generic for all device subsystems.
521 */
522 struct memory_block *find_memory_block(struct mem_section *section)
523 {
524 return find_memory_block_hinted(section, NULL);
525 }
526
527 static struct attribute *memory_memblk_attrs[] = {
528 &dev_attr_phys_index.attr,
529 &dev_attr_end_phys_index.attr,
530 &dev_attr_state.attr,
531 &dev_attr_phys_device.attr,
532 &dev_attr_removable.attr,
533 NULL
534 };
535
536 static struct attribute_group memory_memblk_attr_group = {
537 .attrs = memory_memblk_attrs,
538 };
539
540 static const struct attribute_group *memory_memblk_attr_groups[] = {
541 &memory_memblk_attr_group,
542 NULL,
543 };
544
545 /*
546 * register_memory - Setup a sysfs device for a memory block
547 */
548 static
549 int register_memory(struct memory_block *memory)
550 {
551 memory->dev.bus = &memory_subsys;
552 memory->dev.id = memory->start_section_nr / sections_per_block;
553 memory->dev.release = memory_block_release;
554 memory->dev.groups = memory_memblk_attr_groups;
555 memory->dev.offline = memory->state == MEM_OFFLINE;
556
557 return device_register(&memory->dev);
558 }
559
560 static int init_memory_block(struct memory_block **memory,
561 struct mem_section *section, unsigned long state)
562 {
563 struct memory_block *mem;
564 unsigned long start_pfn;
565 int scn_nr;
566 int ret = 0;
567
568 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
569 if (!mem)
570 return -ENOMEM;
571
572 scn_nr = __section_nr(section);
573 mem->start_section_nr =
574 base_memory_block_id(scn_nr) * sections_per_block;
575 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
576 mem->state = state;
577 mem->section_count++;
578 start_pfn = section_nr_to_pfn(mem->start_section_nr);
579 mem->phys_device = arch_get_memory_phys_device(start_pfn);
580
581 ret = register_memory(mem);
582
583 *memory = mem;
584 return ret;
585 }
586
587 static int add_memory_block(int base_section_nr)
588 {
589 struct memory_block *mem;
590 int i, ret, section_count = 0, section_nr;
591
592 for (i = base_section_nr;
593 (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
594 i++) {
595 if (!present_section_nr(i))
596 continue;
597 if (section_count == 0)
598 section_nr = i;
599 section_count++;
600 }
601
602 if (section_count == 0)
603 return 0;
604 ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
605 if (ret)
606 return ret;
607 mem->section_count = section_count;
608 return 0;
609 }
610
611
612 /*
613 * need an interface for the VM to add new memory regions,
614 * but without onlining it.
615 */
616 int register_new_memory(int nid, struct mem_section *section)
617 {
618 int ret = 0;
619 struct memory_block *mem;
620
621 mutex_lock(&mem_sysfs_mutex);
622
623 mem = find_memory_block(section);
624 if (mem) {
625 mem->section_count++;
626 put_device(&mem->dev);
627 } else {
628 ret = init_memory_block(&mem, section, MEM_OFFLINE);
629 if (ret)
630 goto out;
631 }
632
633 if (mem->section_count == sections_per_block)
634 ret = register_mem_sect_under_node(mem, nid);
635 out:
636 mutex_unlock(&mem_sysfs_mutex);
637 return ret;
638 }
639
640 #ifdef CONFIG_MEMORY_HOTREMOVE
641 static void
642 unregister_memory(struct memory_block *memory)
643 {
644 BUG_ON(memory->dev.bus != &memory_subsys);
645
646 /* drop the ref. we got in remove_memory_block() */
647 put_device(&memory->dev);
648 device_unregister(&memory->dev);
649 }
650
651 static int remove_memory_block(unsigned long node_id,
652 struct mem_section *section, int phys_device)
653 {
654 struct memory_block *mem;
655
656 mutex_lock(&mem_sysfs_mutex);
657 mem = find_memory_block(section);
658 unregister_mem_sect_under_nodes(mem, __section_nr(section));
659
660 mem->section_count--;
661 if (mem->section_count == 0)
662 unregister_memory(mem);
663 else
664 put_device(&mem->dev);
665
666 mutex_unlock(&mem_sysfs_mutex);
667 return 0;
668 }
669
670 int unregister_memory_section(struct mem_section *section)
671 {
672 if (!present_section(section))
673 return -EINVAL;
674
675 return remove_memory_block(0, section, 0);
676 }
677 #endif /* CONFIG_MEMORY_HOTREMOVE */
678
679 /* return true if the memory block is offlined, otherwise, return false */
680 bool is_memblock_offlined(struct memory_block *mem)
681 {
682 return mem->state == MEM_OFFLINE;
683 }
684
685 static struct attribute *memory_root_attrs[] = {
686 #ifdef CONFIG_ARCH_MEMORY_PROBE
687 &dev_attr_probe.attr,
688 #endif
689
690 #ifdef CONFIG_MEMORY_FAILURE
691 &dev_attr_soft_offline_page.attr,
692 &dev_attr_hard_offline_page.attr,
693 #endif
694
695 &dev_attr_block_size_bytes.attr,
696 NULL
697 };
698
699 static struct attribute_group memory_root_attr_group = {
700 .attrs = memory_root_attrs,
701 };
702
703 static const struct attribute_group *memory_root_attr_groups[] = {
704 &memory_root_attr_group,
705 NULL,
706 };
707
708 /*
709 * Initialize the sysfs support for memory devices...
710 */
711 int __init memory_dev_init(void)
712 {
713 unsigned int i;
714 int ret;
715 int err;
716 unsigned long block_sz;
717
718 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
719 if (ret)
720 goto out;
721
722 block_sz = get_memory_block_size();
723 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
724
725 /*
726 * Create entries for memory sections that were found
727 * during boot and have been initialized
728 */
729 mutex_lock(&mem_sysfs_mutex);
730 for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
731 err = add_memory_block(i);
732 if (!ret)
733 ret = err;
734 }
735 mutex_unlock(&mem_sysfs_mutex);
736
737 out:
738 if (ret)
739 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
740 return ret;
741 }