Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Memory subsystem support
4 *
5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
6 * Dave Hansen <haveblue@us.ibm.com>
7 *
8 * This file provides the necessary infrastructure to represent
9 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
10 * All arch-independent code that assumes MEMORY_HOTPLUG requires
11 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
12 */
13
14#include <linux/module.h>
15#include <linux/init.h>
16#include <linux/topology.h>
17#include <linux/capability.h>
18#include <linux/device.h>
19#include <linux/memory.h>
20#include <linux/memory_hotplug.h>
21#include <linux/mm.h>
22#include <linux/stat.h>
23#include <linux/slab.h>
24#include <linux/xarray.h>
25
26#include <linux/atomic.h>
27#include <linux/uaccess.h>
28
29#define MEMORY_CLASS_NAME "memory"
30
31static const char *const online_type_to_str[] = {
32 [MMOP_OFFLINE] = "offline",
33 [MMOP_ONLINE] = "online",
34 [MMOP_ONLINE_KERNEL] = "online_kernel",
35 [MMOP_ONLINE_MOVABLE] = "online_movable",
36};
37
38int memhp_online_type_from_str(const char *str)
39{
40 int i;
41
42 for (i = 0; i < ARRAY_SIZE(online_type_to_str); i++) {
43 if (sysfs_streq(str, online_type_to_str[i]))
44 return i;
45 }
46 return -EINVAL;
47}
48
49#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
50
51static int sections_per_block;
52
53static inline unsigned long memory_block_id(unsigned long section_nr)
54{
55 return section_nr / sections_per_block;
56}
57
58static inline unsigned long pfn_to_block_id(unsigned long pfn)
59{
60 return memory_block_id(pfn_to_section_nr(pfn));
61}
62
63static inline unsigned long phys_to_block_id(unsigned long phys)
64{
65 return pfn_to_block_id(PFN_DOWN(phys));
66}
67
68static int memory_subsys_online(struct device *dev);
69static int memory_subsys_offline(struct device *dev);
70
71static struct bus_type memory_subsys = {
72 .name = MEMORY_CLASS_NAME,
73 .dev_name = MEMORY_CLASS_NAME,
74 .online = memory_subsys_online,
75 .offline = memory_subsys_offline,
76};
77
78/*
79 * Memory blocks are cached in a local radix tree to avoid
80 * a costly linear search for the corresponding device on
81 * the subsystem bus.
82 */
83static DEFINE_XARRAY(memory_blocks);
84
85static BLOCKING_NOTIFIER_HEAD(memory_chain);
86
87int register_memory_notifier(struct notifier_block *nb)
88{
89 return blocking_notifier_chain_register(&memory_chain, nb);
90}
91EXPORT_SYMBOL(register_memory_notifier);
92
93void unregister_memory_notifier(struct notifier_block *nb)
94{
95 blocking_notifier_chain_unregister(&memory_chain, nb);
96}
97EXPORT_SYMBOL(unregister_memory_notifier);
98
99static void memory_block_release(struct device *dev)
100{
101 struct memory_block *mem = to_memory_block(dev);
102
103 kfree(mem);
104}
105
106unsigned long __weak memory_block_size_bytes(void)
107{
108 return MIN_MEMORY_BLOCK_SIZE;
109}
110EXPORT_SYMBOL_GPL(memory_block_size_bytes);
111
112/*
113 * Show the first physical section index (number) of this memory block.
114 */
115static ssize_t phys_index_show(struct device *dev,
116 struct device_attribute *attr, char *buf)
117{
118 struct memory_block *mem = to_memory_block(dev);
119 unsigned long phys_index;
120
121 phys_index = mem->start_section_nr / sections_per_block;
122 return sprintf(buf, "%08lx\n", phys_index);
123}
124
125/*
126 * Legacy interface that we cannot remove. Always indicate "removable"
127 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
128 */
129static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
130 char *buf)
131{
132 return sprintf(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
133}
134
135/*
136 * online, offline, going offline, etc.
137 */
138static ssize_t state_show(struct device *dev, struct device_attribute *attr,
139 char *buf)
140{
141 struct memory_block *mem = to_memory_block(dev);
142 ssize_t len = 0;
143
144 /*
145 * We can probably put these states in a nice little array
146 * so that they're not open-coded
147 */
148 switch (mem->state) {
149 case MEM_ONLINE:
150 len = sprintf(buf, "online\n");
151 break;
152 case MEM_OFFLINE:
153 len = sprintf(buf, "offline\n");
154 break;
155 case MEM_GOING_OFFLINE:
156 len = sprintf(buf, "going-offline\n");
157 break;
158 default:
159 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
160 mem->state);
161 WARN_ON(1);
162 break;
163 }
164
165 return len;
166}
167
168int memory_notify(unsigned long val, void *v)
169{
170 return blocking_notifier_call_chain(&memory_chain, val, v);
171}
172
173/*
174 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
175 * OK to have direct references to sparsemem variables in here.
176 */
177static int
178memory_block_action(unsigned long start_section_nr, unsigned long action,
179 int online_type, int nid)
180{
181 unsigned long start_pfn;
182 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
183 int ret;
184
185 start_pfn = section_nr_to_pfn(start_section_nr);
186
187 switch (action) {
188 case MEM_ONLINE:
189 ret = online_pages(start_pfn, nr_pages, online_type, nid);
190 break;
191 case MEM_OFFLINE:
192 ret = offline_pages(start_pfn, nr_pages);
193 break;
194 default:
195 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
196 "%ld\n", __func__, start_section_nr, action, action);
197 ret = -EINVAL;
198 }
199
200 return ret;
201}
202
203static int memory_block_change_state(struct memory_block *mem,
204 unsigned long to_state, unsigned long from_state_req)
205{
206 int ret = 0;
207
208 if (mem->state != from_state_req)
209 return -EINVAL;
210
211 if (to_state == MEM_OFFLINE)
212 mem->state = MEM_GOING_OFFLINE;
213
214 ret = memory_block_action(mem->start_section_nr, to_state,
215 mem->online_type, mem->nid);
216
217 mem->state = ret ? from_state_req : to_state;
218
219 return ret;
220}
221
222/* The device lock serializes operations on memory_subsys_[online|offline] */
223static int memory_subsys_online(struct device *dev)
224{
225 struct memory_block *mem = to_memory_block(dev);
226 int ret;
227
228 if (mem->state == MEM_ONLINE)
229 return 0;
230
231 /*
232 * When called via device_online() without configuring the online_type,
233 * we want to default to MMOP_ONLINE.
234 */
235 if (mem->online_type == MMOP_OFFLINE)
236 mem->online_type = MMOP_ONLINE;
237
238 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
239 mem->online_type = MMOP_OFFLINE;
240
241 return ret;
242}
243
244static int memory_subsys_offline(struct device *dev)
245{
246 struct memory_block *mem = to_memory_block(dev);
247
248 if (mem->state == MEM_OFFLINE)
249 return 0;
250
251 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
252}
253
254static ssize_t state_store(struct device *dev, struct device_attribute *attr,
255 const char *buf, size_t count)
256{
257 const int online_type = memhp_online_type_from_str(buf);
258 struct memory_block *mem = to_memory_block(dev);
259 int ret;
260
261 if (online_type < 0)
262 return -EINVAL;
263
264 ret = lock_device_hotplug_sysfs();
265 if (ret)
266 return ret;
267
268 switch (online_type) {
269 case MMOP_ONLINE_KERNEL:
270 case MMOP_ONLINE_MOVABLE:
271 case MMOP_ONLINE:
272 /* mem->online_type is protected by device_hotplug_lock */
273 mem->online_type = online_type;
274 ret = device_online(&mem->dev);
275 break;
276 case MMOP_OFFLINE:
277 ret = device_offline(&mem->dev);
278 break;
279 default:
280 ret = -EINVAL; /* should never happen */
281 }
282
283 unlock_device_hotplug();
284
285 if (ret < 0)
286 return ret;
287 if (ret)
288 return -EINVAL;
289
290 return count;
291}
292
293/*
294 * phys_device is a bad name for this. What I really want
295 * is a way to differentiate between memory ranges that
296 * are part of physical devices that constitute
297 * a complete removable unit or fru.
298 * i.e. do these ranges belong to the same physical device,
299 * s.t. if I offline all of these sections I can then
300 * remove the physical device?
301 */
302static ssize_t phys_device_show(struct device *dev,
303 struct device_attribute *attr, char *buf)
304{
305 struct memory_block *mem = to_memory_block(dev);
306 return sprintf(buf, "%d\n", mem->phys_device);
307}
308
309#ifdef CONFIG_MEMORY_HOTREMOVE
310static void print_allowed_zone(char *buf, int nid, unsigned long start_pfn,
311 unsigned long nr_pages, int online_type,
312 struct zone *default_zone)
313{
314 struct zone *zone;
315
316 zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
317 if (zone != default_zone) {
318 strcat(buf, " ");
319 strcat(buf, zone->name);
320 }
321}
322
323static ssize_t valid_zones_show(struct device *dev,
324 struct device_attribute *attr, char *buf)
325{
326 struct memory_block *mem = to_memory_block(dev);
327 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
328 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
329 struct zone *default_zone;
330 int nid;
331
332 /*
333 * Check the existing zone. Make sure that we do that only on the
334 * online nodes otherwise the page_zone is not reliable
335 */
336 if (mem->state == MEM_ONLINE) {
337 /*
338 * The block contains more than one zone can not be offlined.
339 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
340 */
341 default_zone = test_pages_in_a_zone(start_pfn,
342 start_pfn + nr_pages);
343 if (!default_zone)
344 return sprintf(buf, "none\n");
345 strcat(buf, default_zone->name);
346 goto out;
347 }
348
349 nid = mem->nid;
350 default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, start_pfn,
351 nr_pages);
352 strcat(buf, default_zone->name);
353
354 print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_KERNEL,
355 default_zone);
356 print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_MOVABLE,
357 default_zone);
358out:
359 strcat(buf, "\n");
360
361 return strlen(buf);
362}
363static DEVICE_ATTR_RO(valid_zones);
364#endif
365
366static DEVICE_ATTR_RO(phys_index);
367static DEVICE_ATTR_RW(state);
368static DEVICE_ATTR_RO(phys_device);
369static DEVICE_ATTR_RO(removable);
370
371/*
372 * Show the memory block size (shared by all memory blocks).
373 */
374static ssize_t block_size_bytes_show(struct device *dev,
375 struct device_attribute *attr, char *buf)
376{
377 return sprintf(buf, "%lx\n", memory_block_size_bytes());
378}
379
380static DEVICE_ATTR_RO(block_size_bytes);
381
382/*
383 * Memory auto online policy.
384 */
385
386static ssize_t auto_online_blocks_show(struct device *dev,
387 struct device_attribute *attr, char *buf)
388{
389 return sprintf(buf, "%s\n",
390 online_type_to_str[memhp_default_online_type]);
391}
392
393static ssize_t auto_online_blocks_store(struct device *dev,
394 struct device_attribute *attr,
395 const char *buf, size_t count)
396{
397 const int online_type = memhp_online_type_from_str(buf);
398
399 if (online_type < 0)
400 return -EINVAL;
401
402 memhp_default_online_type = online_type;
403 return count;
404}
405
406static DEVICE_ATTR_RW(auto_online_blocks);
407
408/*
409 * Some architectures will have custom drivers to do this, and
410 * will not need to do it from userspace. The fake hot-add code
411 * as well as ppc64 will do all of their discovery in userspace
412 * and will require this interface.
413 */
414#ifdef CONFIG_ARCH_MEMORY_PROBE
415static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
416 const char *buf, size_t count)
417{
418 u64 phys_addr;
419 int nid, ret;
420 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
421
422 ret = kstrtoull(buf, 0, &phys_addr);
423 if (ret)
424 return ret;
425
426 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
427 return -EINVAL;
428
429 ret = lock_device_hotplug_sysfs();
430 if (ret)
431 return ret;
432
433 nid = memory_add_physaddr_to_nid(phys_addr);
434 ret = __add_memory(nid, phys_addr,
435 MIN_MEMORY_BLOCK_SIZE * sections_per_block);
436
437 if (ret)
438 goto out;
439
440 ret = count;
441out:
442 unlock_device_hotplug();
443 return ret;
444}
445
446static DEVICE_ATTR_WO(probe);
447#endif
448
449#ifdef CONFIG_MEMORY_FAILURE
450/*
451 * Support for offlining pages of memory
452 */
453
454/* Soft offline a page */
455static ssize_t soft_offline_page_store(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 (kstrtoull(buf, 0, &pfn) < 0)
464 return -EINVAL;
465 pfn >>= PAGE_SHIFT;
466 ret = soft_offline_page(pfn, 0);
467 return ret == 0 ? count : ret;
468}
469
470/* Forcibly offline a page, including killing processes. */
471static ssize_t hard_offline_page_store(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 (kstrtoull(buf, 0, &pfn) < 0)
480 return -EINVAL;
481 pfn >>= PAGE_SHIFT;
482 ret = memory_failure(pfn, 0);
483 return ret ? ret : count;
484}
485
486static DEVICE_ATTR_WO(soft_offline_page);
487static DEVICE_ATTR_WO(hard_offline_page);
488#endif
489
490/*
491 * Note that phys_device is optional. It is here to allow for
492 * differentiation between which *physical* devices each
493 * section belongs to...
494 */
495int __weak arch_get_memory_phys_device(unsigned long start_pfn)
496{
497 return 0;
498}
499
500/*
501 * A reference for the returned memory block device is acquired.
502 *
503 * Called under device_hotplug_lock.
504 */
505static struct memory_block *find_memory_block_by_id(unsigned long block_id)
506{
507 struct memory_block *mem;
508
509 mem = xa_load(&memory_blocks, block_id);
510 if (mem)
511 get_device(&mem->dev);
512 return mem;
513}
514
515/*
516 * Called under device_hotplug_lock.
517 */
518struct memory_block *find_memory_block(struct mem_section *section)
519{
520 unsigned long block_id = memory_block_id(__section_nr(section));
521
522 return find_memory_block_by_id(block_id);
523}
524
525static struct attribute *memory_memblk_attrs[] = {
526 &dev_attr_phys_index.attr,
527 &dev_attr_state.attr,
528 &dev_attr_phys_device.attr,
529 &dev_attr_removable.attr,
530#ifdef CONFIG_MEMORY_HOTREMOVE
531 &dev_attr_valid_zones.attr,
532#endif
533 NULL
534};
535
536static struct attribute_group memory_memblk_attr_group = {
537 .attrs = memory_memblk_attrs,
538};
539
540static 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 */
548static
549int register_memory(struct memory_block *memory)
550{
551 int ret;
552
553 memory->dev.bus = &memory_subsys;
554 memory->dev.id = memory->start_section_nr / sections_per_block;
555 memory->dev.release = memory_block_release;
556 memory->dev.groups = memory_memblk_attr_groups;
557 memory->dev.offline = memory->state == MEM_OFFLINE;
558
559 ret = device_register(&memory->dev);
560 if (ret) {
561 put_device(&memory->dev);
562 return ret;
563 }
564 ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
565 GFP_KERNEL));
566 if (ret) {
567 put_device(&memory->dev);
568 device_unregister(&memory->dev);
569 }
570 return ret;
571}
572
573static int init_memory_block(unsigned long block_id, unsigned long state)
574{
575 struct memory_block *mem;
576 unsigned long start_pfn;
577 int ret = 0;
578
579 mem = find_memory_block_by_id(block_id);
580 if (mem) {
581 put_device(&mem->dev);
582 return -EEXIST;
583 }
584 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
585 if (!mem)
586 return -ENOMEM;
587
588 mem->start_section_nr = block_id * sections_per_block;
589 mem->state = state;
590 start_pfn = section_nr_to_pfn(mem->start_section_nr);
591 mem->phys_device = arch_get_memory_phys_device(start_pfn);
592 mem->nid = NUMA_NO_NODE;
593
594 ret = register_memory(mem);
595
596 return ret;
597}
598
599static int add_memory_block(unsigned long base_section_nr)
600{
601 int section_count = 0;
602 unsigned long nr;
603
604 for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
605 nr++)
606 if (present_section_nr(nr))
607 section_count++;
608
609 if (section_count == 0)
610 return 0;
611 return init_memory_block(memory_block_id(base_section_nr),
612 MEM_ONLINE);
613}
614
615static void unregister_memory(struct memory_block *memory)
616{
617 if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
618 return;
619
620 WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
621
622 /* drop the ref. we got via find_memory_block() */
623 put_device(&memory->dev);
624 device_unregister(&memory->dev);
625}
626
627/*
628 * Create memory block devices for the given memory area. Start and size
629 * have to be aligned to memory block granularity. Memory block devices
630 * will be initialized as offline.
631 *
632 * Called under device_hotplug_lock.
633 */
634int create_memory_block_devices(unsigned long start, unsigned long size)
635{
636 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
637 unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
638 struct memory_block *mem;
639 unsigned long block_id;
640 int ret = 0;
641
642 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
643 !IS_ALIGNED(size, memory_block_size_bytes())))
644 return -EINVAL;
645
646 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
647 ret = init_memory_block(block_id, MEM_OFFLINE);
648 if (ret)
649 break;
650 }
651 if (ret) {
652 end_block_id = block_id;
653 for (block_id = start_block_id; block_id != end_block_id;
654 block_id++) {
655 mem = find_memory_block_by_id(block_id);
656 if (WARN_ON_ONCE(!mem))
657 continue;
658 unregister_memory(mem);
659 }
660 }
661 return ret;
662}
663
664/*
665 * Remove memory block devices for the given memory area. Start and size
666 * have to be aligned to memory block granularity. Memory block devices
667 * have to be offline.
668 *
669 * Called under device_hotplug_lock.
670 */
671void remove_memory_block_devices(unsigned long start, unsigned long size)
672{
673 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
674 const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
675 struct memory_block *mem;
676 unsigned long block_id;
677
678 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
679 !IS_ALIGNED(size, memory_block_size_bytes())))
680 return;
681
682 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
683 mem = find_memory_block_by_id(block_id);
684 if (WARN_ON_ONCE(!mem))
685 continue;
686 unregister_memory_block_under_nodes(mem);
687 unregister_memory(mem);
688 }
689}
690
691/* return true if the memory block is offlined, otherwise, return false */
692bool is_memblock_offlined(struct memory_block *mem)
693{
694 return mem->state == MEM_OFFLINE;
695}
696
697static struct attribute *memory_root_attrs[] = {
698#ifdef CONFIG_ARCH_MEMORY_PROBE
699 &dev_attr_probe.attr,
700#endif
701
702#ifdef CONFIG_MEMORY_FAILURE
703 &dev_attr_soft_offline_page.attr,
704 &dev_attr_hard_offline_page.attr,
705#endif
706
707 &dev_attr_block_size_bytes.attr,
708 &dev_attr_auto_online_blocks.attr,
709 NULL
710};
711
712static struct attribute_group memory_root_attr_group = {
713 .attrs = memory_root_attrs,
714};
715
716static const struct attribute_group *memory_root_attr_groups[] = {
717 &memory_root_attr_group,
718 NULL,
719};
720
721/*
722 * Initialize the sysfs support for memory devices. At the time this function
723 * is called, we cannot have concurrent creation/deletion of memory block
724 * devices, the device_hotplug_lock is not needed.
725 */
726void __init memory_dev_init(void)
727{
728 int ret;
729 unsigned long block_sz, nr;
730
731 /* Validate the configured memory block size */
732 block_sz = memory_block_size_bytes();
733 if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
734 panic("Memory block size not suitable: 0x%lx\n", block_sz);
735 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
736
737 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
738 if (ret)
739 panic("%s() failed to register subsystem: %d\n", __func__, ret);
740
741 /*
742 * Create entries for memory sections that were found
743 * during boot and have been initialized
744 */
745 for (nr = 0; nr <= __highest_present_section_nr;
746 nr += sections_per_block) {
747 ret = add_memory_block(nr);
748 if (ret)
749 panic("%s() failed to add memory block: %d\n", __func__,
750 ret);
751 }
752}
753
754/**
755 * walk_memory_blocks - walk through all present memory blocks overlapped
756 * by the range [start, start + size)
757 *
758 * @start: start address of the memory range
759 * @size: size of the memory range
760 * @arg: argument passed to func
761 * @func: callback for each memory section walked
762 *
763 * This function walks through all present memory blocks overlapped by the
764 * range [start, start + size), calling func on each memory block.
765 *
766 * In case func() returns an error, walking is aborted and the error is
767 * returned.
768 *
769 * Called under device_hotplug_lock.
770 */
771int walk_memory_blocks(unsigned long start, unsigned long size,
772 void *arg, walk_memory_blocks_func_t func)
773{
774 const unsigned long start_block_id = phys_to_block_id(start);
775 const unsigned long end_block_id = phys_to_block_id(start + size - 1);
776 struct memory_block *mem;
777 unsigned long block_id;
778 int ret = 0;
779
780 if (!size)
781 return 0;
782
783 for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
784 mem = find_memory_block_by_id(block_id);
785 if (!mem)
786 continue;
787
788 ret = func(mem, arg);
789 put_device(&mem->dev);
790 if (ret)
791 break;
792 }
793 return ret;
794}
795
796struct for_each_memory_block_cb_data {
797 walk_memory_blocks_func_t func;
798 void *arg;
799};
800
801static int for_each_memory_block_cb(struct device *dev, void *data)
802{
803 struct memory_block *mem = to_memory_block(dev);
804 struct for_each_memory_block_cb_data *cb_data = data;
805
806 return cb_data->func(mem, cb_data->arg);
807}
808
809/**
810 * for_each_memory_block - walk through all present memory blocks
811 *
812 * @arg: argument passed to func
813 * @func: callback for each memory block walked
814 *
815 * This function walks through all present memory blocks, calling func on
816 * each memory block.
817 *
818 * In case func() returns an error, walking is aborted and the error is
819 * returned.
820 */
821int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
822{
823 struct for_each_memory_block_cb_data cb_data = {
824 .func = func,
825 .arg = arg,
826 };
827
828 return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
829 for_each_memory_block_cb);
830}
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
28static 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
34static int sections_per_block;
35
36static inline int base_memory_block_id(int section_nr)
37{
38 return section_nr / sections_per_block;
39}
40
41static int memory_subsys_online(struct device *dev);
42static int memory_subsys_offline(struct device *dev);
43
44static 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
51static BLOCKING_NOTIFIER_HEAD(memory_chain);
52
53int register_memory_notifier(struct notifier_block *nb)
54{
55 return blocking_notifier_chain_register(&memory_chain, nb);
56}
57EXPORT_SYMBOL(register_memory_notifier);
58
59void unregister_memory_notifier(struct notifier_block *nb)
60{
61 blocking_notifier_chain_unregister(&memory_chain, nb);
62}
63EXPORT_SYMBOL(unregister_memory_notifier);
64
65static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
66
67int register_memory_isolate_notifier(struct notifier_block *nb)
68{
69 return atomic_notifier_chain_register(&memory_isolate_chain, nb);
70}
71EXPORT_SYMBOL(register_memory_isolate_notifier);
72
73void unregister_memory_isolate_notifier(struct notifier_block *nb)
74{
75 atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
76}
77EXPORT_SYMBOL(unregister_memory_isolate_notifier);
78
79static void memory_block_release(struct device *dev)
80{
81 struct memory_block *mem = to_memory_block(dev);
82
83 kfree(mem);
84}
85
86unsigned long __weak memory_block_size_bytes(void)
87{
88 return MIN_MEMORY_BLOCK_SIZE;
89}
90
91static 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
111static 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
121static 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 */
134static 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 */
154static 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
184int memory_notify(unsigned long val, void *v)
185{
186 return blocking_notifier_call_chain(&memory_chain, val, v);
187}
188
189int 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 */
198static 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 */
233static int
234memory_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
263static 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] */
283static 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
307static 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
317static ssize_t
318store_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 ret = -EINVAL;
338 goto err;
339 }
340
341 switch (online_type) {
342 case ONLINE_KERNEL:
343 case ONLINE_MOVABLE:
344 case ONLINE_KEEP:
345 /*
346 * mem->online_type is not protected so there can be a
347 * race here. However, when racing online, the first
348 * will succeed and the second will just return as the
349 * block will already be online. The online type
350 * could be either one, but that is expected.
351 */
352 mem->online_type = online_type;
353 ret = device_online(&mem->dev);
354 break;
355 case -1:
356 ret = device_offline(&mem->dev);
357 break;
358 default:
359 ret = -EINVAL; /* should never happen */
360 }
361
362err:
363 unlock_device_hotplug();
364
365 if (ret)
366 return ret;
367 return count;
368}
369
370/*
371 * phys_device is a bad name for this. What I really want
372 * is a way to differentiate between memory ranges that
373 * are part of physical devices that constitute
374 * a complete removable unit or fru.
375 * i.e. do these ranges belong to the same physical device,
376 * s.t. if I offline all of these sections I can then
377 * remove the physical device?
378 */
379static ssize_t show_phys_device(struct device *dev,
380 struct device_attribute *attr, char *buf)
381{
382 struct memory_block *mem = to_memory_block(dev);
383 return sprintf(buf, "%d\n", mem->phys_device);
384}
385
386static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
387static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
388static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
389static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
390static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
391
392/*
393 * Block size attribute stuff
394 */
395static ssize_t
396print_block_size(struct device *dev, struct device_attribute *attr,
397 char *buf)
398{
399 return sprintf(buf, "%lx\n", get_memory_block_size());
400}
401
402static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
403
404/*
405 * Some architectures will have custom drivers to do this, and
406 * will not need to do it from userspace. The fake hot-add code
407 * as well as ppc64 will do all of their discovery in userspace
408 * and will require this interface.
409 */
410#ifdef CONFIG_ARCH_MEMORY_PROBE
411static ssize_t
412memory_probe_store(struct device *dev, struct device_attribute *attr,
413 const char *buf, size_t count)
414{
415 u64 phys_addr;
416 int nid;
417 int i, ret;
418 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
419
420 phys_addr = simple_strtoull(buf, NULL, 0);
421
422 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
423 return -EINVAL;
424
425 for (i = 0; i < sections_per_block; i++) {
426 nid = memory_add_physaddr_to_nid(phys_addr);
427 ret = add_memory(nid, phys_addr,
428 PAGES_PER_SECTION << PAGE_SHIFT);
429 if (ret)
430 goto out;
431
432 phys_addr += MIN_MEMORY_BLOCK_SIZE;
433 }
434
435 ret = count;
436out:
437 return ret;
438}
439
440static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
441#endif
442
443#ifdef CONFIG_MEMORY_FAILURE
444/*
445 * Support for offlining pages of memory
446 */
447
448/* Soft offline a page */
449static ssize_t
450store_soft_offline_page(struct device *dev,
451 struct device_attribute *attr,
452 const char *buf, size_t count)
453{
454 int ret;
455 u64 pfn;
456 if (!capable(CAP_SYS_ADMIN))
457 return -EPERM;
458 if (kstrtoull(buf, 0, &pfn) < 0)
459 return -EINVAL;
460 pfn >>= PAGE_SHIFT;
461 if (!pfn_valid(pfn))
462 return -ENXIO;
463 ret = soft_offline_page(pfn_to_page(pfn), 0);
464 return ret == 0 ? count : ret;
465}
466
467/* Forcibly offline a page, including killing processes. */
468static ssize_t
469store_hard_offline_page(struct device *dev,
470 struct device_attribute *attr,
471 const char *buf, size_t count)
472{
473 int ret;
474 u64 pfn;
475 if (!capable(CAP_SYS_ADMIN))
476 return -EPERM;
477 if (kstrtoull(buf, 0, &pfn) < 0)
478 return -EINVAL;
479 pfn >>= PAGE_SHIFT;
480 ret = memory_failure(pfn, 0, 0);
481 return ret ? ret : count;
482}
483
484static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
485static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
486#endif
487
488/*
489 * Note that phys_device is optional. It is here to allow for
490 * differentiation between which *physical* devices each
491 * section belongs to...
492 */
493int __weak arch_get_memory_phys_device(unsigned long start_pfn)
494{
495 return 0;
496}
497
498/*
499 * A reference for the returned object is held and the reference for the
500 * hinted object is released.
501 */
502struct memory_block *find_memory_block_hinted(struct mem_section *section,
503 struct memory_block *hint)
504{
505 int block_id = base_memory_block_id(__section_nr(section));
506 struct device *hintdev = hint ? &hint->dev : NULL;
507 struct device *dev;
508
509 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
510 if (hint)
511 put_device(&hint->dev);
512 if (!dev)
513 return NULL;
514 return to_memory_block(dev);
515}
516
517/*
518 * For now, we have a linear search to go find the appropriate
519 * memory_block corresponding to a particular phys_index. If
520 * this gets to be a real problem, we can always use a radix
521 * tree or something here.
522 *
523 * This could be made generic for all device subsystems.
524 */
525struct memory_block *find_memory_block(struct mem_section *section)
526{
527 return find_memory_block_hinted(section, NULL);
528}
529
530static struct attribute *memory_memblk_attrs[] = {
531 &dev_attr_phys_index.attr,
532 &dev_attr_end_phys_index.attr,
533 &dev_attr_state.attr,
534 &dev_attr_phys_device.attr,
535 &dev_attr_removable.attr,
536 NULL
537};
538
539static struct attribute_group memory_memblk_attr_group = {
540 .attrs = memory_memblk_attrs,
541};
542
543static const struct attribute_group *memory_memblk_attr_groups[] = {
544 &memory_memblk_attr_group,
545 NULL,
546};
547
548/*
549 * register_memory - Setup a sysfs device for a memory block
550 */
551static
552int register_memory(struct memory_block *memory)
553{
554 memory->dev.bus = &memory_subsys;
555 memory->dev.id = memory->start_section_nr / sections_per_block;
556 memory->dev.release = memory_block_release;
557 memory->dev.groups = memory_memblk_attr_groups;
558 memory->dev.offline = memory->state == MEM_OFFLINE;
559
560 return device_register(&memory->dev);
561}
562
563static int init_memory_block(struct memory_block **memory,
564 struct mem_section *section, unsigned long state)
565{
566 struct memory_block *mem;
567 unsigned long start_pfn;
568 int scn_nr;
569 int ret = 0;
570
571 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
572 if (!mem)
573 return -ENOMEM;
574
575 scn_nr = __section_nr(section);
576 mem->start_section_nr =
577 base_memory_block_id(scn_nr) * sections_per_block;
578 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
579 mem->state = state;
580 mem->section_count++;
581 start_pfn = section_nr_to_pfn(mem->start_section_nr);
582 mem->phys_device = arch_get_memory_phys_device(start_pfn);
583
584 ret = register_memory(mem);
585
586 *memory = mem;
587 return ret;
588}
589
590static int add_memory_block(int base_section_nr)
591{
592 struct memory_block *mem;
593 int i, ret, section_count = 0, section_nr;
594
595 for (i = base_section_nr;
596 (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
597 i++) {
598 if (!present_section_nr(i))
599 continue;
600 if (section_count == 0)
601 section_nr = i;
602 section_count++;
603 }
604
605 if (section_count == 0)
606 return 0;
607 ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
608 if (ret)
609 return ret;
610 mem->section_count = section_count;
611 return 0;
612}
613
614
615/*
616 * need an interface for the VM to add new memory regions,
617 * but without onlining it.
618 */
619int register_new_memory(int nid, struct mem_section *section)
620{
621 int ret = 0;
622 struct memory_block *mem;
623
624 mutex_lock(&mem_sysfs_mutex);
625
626 mem = find_memory_block(section);
627 if (mem) {
628 mem->section_count++;
629 put_device(&mem->dev);
630 } else {
631 ret = init_memory_block(&mem, section, MEM_OFFLINE);
632 if (ret)
633 goto out;
634 }
635
636 if (mem->section_count == sections_per_block)
637 ret = register_mem_sect_under_node(mem, nid);
638out:
639 mutex_unlock(&mem_sysfs_mutex);
640 return ret;
641}
642
643#ifdef CONFIG_MEMORY_HOTREMOVE
644static void
645unregister_memory(struct memory_block *memory)
646{
647 BUG_ON(memory->dev.bus != &memory_subsys);
648
649 /* drop the ref. we got in remove_memory_block() */
650 put_device(&memory->dev);
651 device_unregister(&memory->dev);
652}
653
654static int remove_memory_block(unsigned long node_id,
655 struct mem_section *section, int phys_device)
656{
657 struct memory_block *mem;
658
659 mutex_lock(&mem_sysfs_mutex);
660 mem = find_memory_block(section);
661 unregister_mem_sect_under_nodes(mem, __section_nr(section));
662
663 mem->section_count--;
664 if (mem->section_count == 0)
665 unregister_memory(mem);
666 else
667 put_device(&mem->dev);
668
669 mutex_unlock(&mem_sysfs_mutex);
670 return 0;
671}
672
673int unregister_memory_section(struct mem_section *section)
674{
675 if (!present_section(section))
676 return -EINVAL;
677
678 return remove_memory_block(0, section, 0);
679}
680#endif /* CONFIG_MEMORY_HOTREMOVE */
681
682/* return true if the memory block is offlined, otherwise, return false */
683bool is_memblock_offlined(struct memory_block *mem)
684{
685 return mem->state == MEM_OFFLINE;
686}
687
688static struct attribute *memory_root_attrs[] = {
689#ifdef CONFIG_ARCH_MEMORY_PROBE
690 &dev_attr_probe.attr,
691#endif
692
693#ifdef CONFIG_MEMORY_FAILURE
694 &dev_attr_soft_offline_page.attr,
695 &dev_attr_hard_offline_page.attr,
696#endif
697
698 &dev_attr_block_size_bytes.attr,
699 NULL
700};
701
702static struct attribute_group memory_root_attr_group = {
703 .attrs = memory_root_attrs,
704};
705
706static const struct attribute_group *memory_root_attr_groups[] = {
707 &memory_root_attr_group,
708 NULL,
709};
710
711/*
712 * Initialize the sysfs support for memory devices...
713 */
714int __init memory_dev_init(void)
715{
716 unsigned int i;
717 int ret;
718 int err;
719 unsigned long block_sz;
720
721 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
722 if (ret)
723 goto out;
724
725 block_sz = get_memory_block_size();
726 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
727
728 /*
729 * Create entries for memory sections that were found
730 * during boot and have been initialized
731 */
732 mutex_lock(&mem_sysfs_mutex);
733 for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
734 err = add_memory_block(i);
735 if (!ret)
736 ret = err;
737 }
738 mutex_unlock(&mem_sysfs_mutex);
739
740out:
741 if (ret)
742 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
743 return ret;
744}