1 // SPDX-License-Identifier: GPL-2.0
3 * Memory subsystem support
5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
6 * Dave Hansen <haveblue@us.ibm.com>
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.
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>
22 #include <linux/stat.h>
23 #include <linux/slab.h>
24 #include <linux/xarray.h>
26 #include <linux/atomic.h>
27 #include <linux/uaccess.h>
29 #define MEMORY_CLASS_NAME "memory"
31 static 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",
38 int mhp_online_type_from_str(const char *str)
42 for (i = 0; i < ARRAY_SIZE(online_type_to_str); i++) {
43 if (sysfs_streq(str, online_type_to_str[i]))
49 #define to_memory_block(dev) container_of(dev, struct memory_block, dev)
51 static int sections_per_block;
53 static inline unsigned long memory_block_id(unsigned long section_nr)
55 return section_nr / sections_per_block;
58 static inline unsigned long pfn_to_block_id(unsigned long pfn)
60 return memory_block_id(pfn_to_section_nr(pfn));
63 static inline unsigned long phys_to_block_id(unsigned long phys)
65 return pfn_to_block_id(PFN_DOWN(phys));
68 static int memory_subsys_online(struct device *dev);
69 static int memory_subsys_offline(struct device *dev);
71 static 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,
79 * Memory blocks are cached in a local radix tree to avoid
80 * a costly linear search for the corresponding device on
83 static DEFINE_XARRAY(memory_blocks);
86 * Memory groups, indexed by memory group id (mgid).
88 static DEFINE_XARRAY_FLAGS(memory_groups, XA_FLAGS_ALLOC);
89 #define MEMORY_GROUP_MARK_DYNAMIC XA_MARK_1
91 static BLOCKING_NOTIFIER_HEAD(memory_chain);
93 int register_memory_notifier(struct notifier_block *nb)
95 return blocking_notifier_chain_register(&memory_chain, nb);
97 EXPORT_SYMBOL(register_memory_notifier);
99 void unregister_memory_notifier(struct notifier_block *nb)
101 blocking_notifier_chain_unregister(&memory_chain, nb);
103 EXPORT_SYMBOL(unregister_memory_notifier);
105 static void memory_block_release(struct device *dev)
107 struct memory_block *mem = to_memory_block(dev);
108 /* Verify that the altmap is freed */
109 WARN_ON(mem->altmap);
113 unsigned long __weak memory_block_size_bytes(void)
115 return MIN_MEMORY_BLOCK_SIZE;
117 EXPORT_SYMBOL_GPL(memory_block_size_bytes);
119 /* Show the memory block ID, relative to the memory block size */
120 static ssize_t phys_index_show(struct device *dev,
121 struct device_attribute *attr, char *buf)
123 struct memory_block *mem = to_memory_block(dev);
125 return sysfs_emit(buf, "%08lx\n", memory_block_id(mem->start_section_nr));
129 * Legacy interface that we cannot remove. Always indicate "removable"
130 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
132 static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
135 return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
139 * online, offline, going offline, etc.
141 static ssize_t state_show(struct device *dev, struct device_attribute *attr,
144 struct memory_block *mem = to_memory_block(dev);
148 * We can probably put these states in a nice little array
149 * so that they're not open-coded
151 switch (mem->state) {
158 case MEM_GOING_OFFLINE:
159 output = "going-offline";
163 return sysfs_emit(buf, "ERROR-UNKNOWN-%ld\n", mem->state);
166 return sysfs_emit(buf, "%s\n", output);
169 int memory_notify(unsigned long val, void *v)
171 return blocking_notifier_call_chain(&memory_chain, val, v);
174 #if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
175 static unsigned long memblk_nr_poison(struct memory_block *mem);
177 static inline unsigned long memblk_nr_poison(struct memory_block *mem)
184 * Must acquire mem_hotplug_lock in write mode.
186 static int memory_block_online(struct memory_block *mem)
188 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
189 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
190 unsigned long nr_vmemmap_pages = 0;
194 if (memblk_nr_poison(mem))
197 zone = zone_for_pfn_range(mem->online_type, mem->nid, mem->group,
198 start_pfn, nr_pages);
201 * Although vmemmap pages have a different lifecycle than the pages
202 * they describe (they remain until the memory is unplugged), doing
203 * their initialization and accounting at memory onlining/offlining
204 * stage helps to keep accounting easier to follow - e.g vmemmaps
205 * belong to the same zone as the memory they backed.
208 nr_vmemmap_pages = mem->altmap->free;
211 if (nr_vmemmap_pages) {
212 ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
217 ret = online_pages(start_pfn + nr_vmemmap_pages,
218 nr_pages - nr_vmemmap_pages, zone, mem->group);
220 if (nr_vmemmap_pages)
221 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
226 * Account once onlining succeeded. If the zone was unpopulated, it is
227 * now already properly populated.
229 if (nr_vmemmap_pages)
230 adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
240 * Must acquire mem_hotplug_lock in write mode.
242 static int memory_block_offline(struct memory_block *mem)
244 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
245 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
246 unsigned long nr_vmemmap_pages = 0;
253 * Unaccount before offlining, such that unpopulated zone and kthreads
254 * can properly be torn down in offline_pages().
257 nr_vmemmap_pages = mem->altmap->free;
260 if (nr_vmemmap_pages)
261 adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
264 ret = offline_pages(start_pfn + nr_vmemmap_pages,
265 nr_pages - nr_vmemmap_pages, mem->zone, mem->group);
267 /* offline_pages() failed. Account back. */
268 if (nr_vmemmap_pages)
269 adjust_present_page_count(pfn_to_page(start_pfn),
270 mem->group, nr_vmemmap_pages);
274 if (nr_vmemmap_pages)
275 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
284 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
285 * OK to have direct references to sparsemem variables in here.
288 memory_block_action(struct memory_block *mem, unsigned long action)
294 ret = memory_block_online(mem);
297 ret = memory_block_offline(mem);
300 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
301 "%ld\n", __func__, mem->start_section_nr, action, action);
308 static int memory_block_change_state(struct memory_block *mem,
309 unsigned long to_state, unsigned long from_state_req)
313 if (mem->state != from_state_req)
316 if (to_state == MEM_OFFLINE)
317 mem->state = MEM_GOING_OFFLINE;
319 ret = memory_block_action(mem, to_state);
320 mem->state = ret ? from_state_req : to_state;
325 /* The device lock serializes operations on memory_subsys_[online|offline] */
326 static int memory_subsys_online(struct device *dev)
328 struct memory_block *mem = to_memory_block(dev);
331 if (mem->state == MEM_ONLINE)
335 * When called via device_online() without configuring the online_type,
336 * we want to default to MMOP_ONLINE.
338 if (mem->online_type == MMOP_OFFLINE)
339 mem->online_type = MMOP_ONLINE;
341 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
342 mem->online_type = MMOP_OFFLINE;
347 static int memory_subsys_offline(struct device *dev)
349 struct memory_block *mem = to_memory_block(dev);
351 if (mem->state == MEM_OFFLINE)
354 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
357 static ssize_t state_store(struct device *dev, struct device_attribute *attr,
358 const char *buf, size_t count)
360 const int online_type = mhp_online_type_from_str(buf);
361 struct memory_block *mem = to_memory_block(dev);
367 ret = lock_device_hotplug_sysfs();
371 switch (online_type) {
372 case MMOP_ONLINE_KERNEL:
373 case MMOP_ONLINE_MOVABLE:
375 /* mem->online_type is protected by device_hotplug_lock */
376 mem->online_type = online_type;
377 ret = device_online(&mem->dev);
380 ret = device_offline(&mem->dev);
383 ret = -EINVAL; /* should never happen */
386 unlock_device_hotplug();
397 * Legacy interface that we cannot remove: s390x exposes the storage increment
398 * covered by a memory block, allowing for identifying which memory blocks
399 * comprise a storage increment. Since a memory block spans complete
400 * storage increments nowadays, this interface is basically unused. Other
401 * archs never exposed != 0.
403 static ssize_t phys_device_show(struct device *dev,
404 struct device_attribute *attr, char *buf)
406 struct memory_block *mem = to_memory_block(dev);
407 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
409 return sysfs_emit(buf, "%d\n",
410 arch_get_memory_phys_device(start_pfn));
413 #ifdef CONFIG_MEMORY_HOTREMOVE
414 static int print_allowed_zone(char *buf, int len, int nid,
415 struct memory_group *group,
416 unsigned long start_pfn, unsigned long nr_pages,
417 int online_type, struct zone *default_zone)
421 zone = zone_for_pfn_range(online_type, nid, group, start_pfn, nr_pages);
422 if (zone == default_zone)
425 return sysfs_emit_at(buf, len, " %s", zone->name);
428 static ssize_t valid_zones_show(struct device *dev,
429 struct device_attribute *attr, char *buf)
431 struct memory_block *mem = to_memory_block(dev);
432 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
433 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
434 struct memory_group *group = mem->group;
435 struct zone *default_zone;
440 * Check the existing zone. Make sure that we do that only on the
441 * online nodes otherwise the page_zone is not reliable
443 if (mem->state == MEM_ONLINE) {
445 * If !mem->zone, the memory block spans multiple zones and
446 * cannot get offlined.
448 default_zone = mem->zone;
450 return sysfs_emit(buf, "%s\n", "none");
451 len += sysfs_emit_at(buf, len, "%s", default_zone->name);
455 default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, group,
456 start_pfn, nr_pages);
458 len += sysfs_emit_at(buf, len, "%s", default_zone->name);
459 len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
460 MMOP_ONLINE_KERNEL, default_zone);
461 len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
462 MMOP_ONLINE_MOVABLE, default_zone);
464 len += sysfs_emit_at(buf, len, "\n");
467 static DEVICE_ATTR_RO(valid_zones);
470 static DEVICE_ATTR_RO(phys_index);
471 static DEVICE_ATTR_RW(state);
472 static DEVICE_ATTR_RO(phys_device);
473 static DEVICE_ATTR_RO(removable);
476 * Show the memory block size (shared by all memory blocks).
478 static ssize_t block_size_bytes_show(struct device *dev,
479 struct device_attribute *attr, char *buf)
481 return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
484 static DEVICE_ATTR_RO(block_size_bytes);
487 * Memory auto online policy.
490 static ssize_t auto_online_blocks_show(struct device *dev,
491 struct device_attribute *attr, char *buf)
493 return sysfs_emit(buf, "%s\n",
494 online_type_to_str[mhp_default_online_type]);
497 static ssize_t auto_online_blocks_store(struct device *dev,
498 struct device_attribute *attr,
499 const char *buf, size_t count)
501 const int online_type = mhp_online_type_from_str(buf);
506 mhp_default_online_type = online_type;
510 static DEVICE_ATTR_RW(auto_online_blocks);
512 #ifdef CONFIG_CRASH_HOTPLUG
513 #include <linux/kexec.h>
514 static ssize_t crash_hotplug_show(struct device *dev,
515 struct device_attribute *attr, char *buf)
517 return sysfs_emit(buf, "%d\n", crash_hotplug_memory_support());
519 static DEVICE_ATTR_RO(crash_hotplug);
523 * Some architectures will have custom drivers to do this, and
524 * will not need to do it from userspace. The fake hot-add code
525 * as well as ppc64 will do all of their discovery in userspace
526 * and will require this interface.
528 #ifdef CONFIG_ARCH_MEMORY_PROBE
529 static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
530 const char *buf, size_t count)
534 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
536 ret = kstrtoull(buf, 0, &phys_addr);
540 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
543 ret = lock_device_hotplug_sysfs();
547 nid = memory_add_physaddr_to_nid(phys_addr);
548 ret = __add_memory(nid, phys_addr,
549 MIN_MEMORY_BLOCK_SIZE * sections_per_block,
557 unlock_device_hotplug();
561 static DEVICE_ATTR_WO(probe);
564 #ifdef CONFIG_MEMORY_FAILURE
566 * Support for offlining pages of memory
569 /* Soft offline a page */
570 static ssize_t soft_offline_page_store(struct device *dev,
571 struct device_attribute *attr,
572 const char *buf, size_t count)
576 if (!capable(CAP_SYS_ADMIN))
578 if (kstrtoull(buf, 0, &pfn) < 0)
581 ret = soft_offline_page(pfn, 0);
582 return ret == 0 ? count : ret;
585 /* Forcibly offline a page, including killing processes. */
586 static ssize_t hard_offline_page_store(struct device *dev,
587 struct device_attribute *attr,
588 const char *buf, size_t count)
592 if (!capable(CAP_SYS_ADMIN))
594 if (kstrtoull(buf, 0, &pfn) < 0)
597 ret = memory_failure(pfn, MF_SW_SIMULATED);
598 if (ret == -EOPNOTSUPP)
600 return ret ? ret : count;
603 static DEVICE_ATTR_WO(soft_offline_page);
604 static DEVICE_ATTR_WO(hard_offline_page);
607 /* See phys_device_show(). */
608 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
614 * A reference for the returned memory block device is acquired.
616 * Called under device_hotplug_lock.
618 static struct memory_block *find_memory_block_by_id(unsigned long block_id)
620 struct memory_block *mem;
622 mem = xa_load(&memory_blocks, block_id);
624 get_device(&mem->dev);
629 * Called under device_hotplug_lock.
631 struct memory_block *find_memory_block(unsigned long section_nr)
633 unsigned long block_id = memory_block_id(section_nr);
635 return find_memory_block_by_id(block_id);
638 static struct attribute *memory_memblk_attrs[] = {
639 &dev_attr_phys_index.attr,
640 &dev_attr_state.attr,
641 &dev_attr_phys_device.attr,
642 &dev_attr_removable.attr,
643 #ifdef CONFIG_MEMORY_HOTREMOVE
644 &dev_attr_valid_zones.attr,
649 static const struct attribute_group memory_memblk_attr_group = {
650 .attrs = memory_memblk_attrs,
653 static const struct attribute_group *memory_memblk_attr_groups[] = {
654 &memory_memblk_attr_group,
658 static int __add_memory_block(struct memory_block *memory)
662 memory->dev.bus = &memory_subsys;
663 memory->dev.id = memory->start_section_nr / sections_per_block;
664 memory->dev.release = memory_block_release;
665 memory->dev.groups = memory_memblk_attr_groups;
666 memory->dev.offline = memory->state == MEM_OFFLINE;
668 ret = device_register(&memory->dev);
670 put_device(&memory->dev);
673 ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
676 device_unregister(&memory->dev);
681 static struct zone *early_node_zone_for_memory_block(struct memory_block *mem,
684 const unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
685 const unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
686 struct zone *zone, *matching_zone = NULL;
687 pg_data_t *pgdat = NODE_DATA(nid);
691 * This logic only works for early memory, when the applicable zones
692 * already span the memory block. We don't expect overlapping zones on
693 * a single node for early memory. So if we're told that some PFNs
694 * of a node fall into this memory block, we can assume that all node
695 * zones that intersect with the memory block are actually applicable.
696 * No need to look at the memmap.
698 for (i = 0; i < MAX_NR_ZONES; i++) {
699 zone = pgdat->node_zones + i;
700 if (!populated_zone(zone))
702 if (!zone_intersects(zone, start_pfn, nr_pages))
704 if (!matching_zone) {
705 matching_zone = zone;
708 /* Spans multiple zones ... */
709 matching_zone = NULL;
712 return matching_zone;
717 * memory_block_add_nid() - Indicate that system RAM falling into this memory
718 * block device (partially) belongs to the given node.
719 * @mem: The memory block device.
721 * @context: The memory initialization context.
723 * Indicate that system RAM falling into this memory block (partially) belongs
724 * to the given node. If the context indicates ("early") that we are adding the
725 * node during node device subsystem initialization, this will also properly
726 * set/adjust mem->zone based on the zone ranges of the given node.
728 void memory_block_add_nid(struct memory_block *mem, int nid,
729 enum meminit_context context)
731 if (context == MEMINIT_EARLY && mem->nid != nid) {
733 * For early memory we have to determine the zone when setting
734 * the node id and handle multiple nodes spanning a single
735 * memory block by indicate via zone == NULL that we're not
736 * dealing with a single zone. So if we're setting the node id
737 * the first time, determine if there is a single zone. If we're
738 * setting the node id a second time to a different node,
739 * invalidate the single detected zone.
741 if (mem->nid == NUMA_NO_NODE)
742 mem->zone = early_node_zone_for_memory_block(mem, nid);
748 * If this memory block spans multiple nodes, we only indicate
749 * the last processed node. If we span multiple nodes (not applicable
750 * to hotplugged memory), zone == NULL will prohibit memory offlining
751 * and consequently unplug.
757 static int add_memory_block(unsigned long block_id, unsigned long state,
758 struct vmem_altmap *altmap,
759 struct memory_group *group)
761 struct memory_block *mem;
764 mem = find_memory_block_by_id(block_id);
766 put_device(&mem->dev);
769 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
773 mem->start_section_nr = block_id * sections_per_block;
775 mem->nid = NUMA_NO_NODE;
776 mem->altmap = altmap;
777 INIT_LIST_HEAD(&mem->group_next);
780 if (state == MEM_ONLINE)
782 * MEM_ONLINE at this point implies early memory. With NUMA,
783 * we'll determine the zone when setting the node id via
784 * memory_block_add_nid(). Memory hotplug updated the zone
785 * manually when memory onlining/offlining succeeds.
787 mem->zone = early_node_zone_for_memory_block(mem, NUMA_NO_NODE);
788 #endif /* CONFIG_NUMA */
790 ret = __add_memory_block(mem);
796 list_add(&mem->group_next, &group->memory_blocks);
802 static int __init add_boot_memory_block(unsigned long base_section_nr)
804 int section_count = 0;
807 for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
809 if (present_section_nr(nr))
812 if (section_count == 0)
814 return add_memory_block(memory_block_id(base_section_nr),
815 MEM_ONLINE, NULL, NULL);
818 static int add_hotplug_memory_block(unsigned long block_id,
819 struct vmem_altmap *altmap,
820 struct memory_group *group)
822 return add_memory_block(block_id, MEM_OFFLINE, altmap, group);
825 static void remove_memory_block(struct memory_block *memory)
827 if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
830 WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
833 list_del(&memory->group_next);
834 memory->group = NULL;
837 /* drop the ref. we got via find_memory_block() */
838 put_device(&memory->dev);
839 device_unregister(&memory->dev);
843 * Create memory block devices for the given memory area. Start and size
844 * have to be aligned to memory block granularity. Memory block devices
845 * will be initialized as offline.
847 * Called under device_hotplug_lock.
849 int create_memory_block_devices(unsigned long start, unsigned long size,
850 struct vmem_altmap *altmap,
851 struct memory_group *group)
853 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
854 unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
855 struct memory_block *mem;
856 unsigned long block_id;
859 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
860 !IS_ALIGNED(size, memory_block_size_bytes())))
863 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
864 ret = add_hotplug_memory_block(block_id, altmap, group);
869 end_block_id = block_id;
870 for (block_id = start_block_id; block_id != end_block_id;
872 mem = find_memory_block_by_id(block_id);
873 if (WARN_ON_ONCE(!mem))
875 remove_memory_block(mem);
882 * Remove memory block devices for the given memory area. Start and size
883 * have to be aligned to memory block granularity. Memory block devices
884 * have to be offline.
886 * Called under device_hotplug_lock.
888 void remove_memory_block_devices(unsigned long start, unsigned long size)
890 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
891 const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
892 struct memory_block *mem;
893 unsigned long block_id;
895 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
896 !IS_ALIGNED(size, memory_block_size_bytes())))
899 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
900 mem = find_memory_block_by_id(block_id);
901 if (WARN_ON_ONCE(!mem))
903 num_poisoned_pages_sub(-1UL, memblk_nr_poison(mem));
904 unregister_memory_block_under_nodes(mem);
905 remove_memory_block(mem);
909 static struct attribute *memory_root_attrs[] = {
910 #ifdef CONFIG_ARCH_MEMORY_PROBE
911 &dev_attr_probe.attr,
914 #ifdef CONFIG_MEMORY_FAILURE
915 &dev_attr_soft_offline_page.attr,
916 &dev_attr_hard_offline_page.attr,
919 &dev_attr_block_size_bytes.attr,
920 &dev_attr_auto_online_blocks.attr,
921 #ifdef CONFIG_CRASH_HOTPLUG
922 &dev_attr_crash_hotplug.attr,
927 static const struct attribute_group memory_root_attr_group = {
928 .attrs = memory_root_attrs,
931 static const struct attribute_group *memory_root_attr_groups[] = {
932 &memory_root_attr_group,
937 * Initialize the sysfs support for memory devices. At the time this function
938 * is called, we cannot have concurrent creation/deletion of memory block
939 * devices, the device_hotplug_lock is not needed.
941 void __init memory_dev_init(void)
944 unsigned long block_sz, nr;
946 /* Validate the configured memory block size */
947 block_sz = memory_block_size_bytes();
948 if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
949 panic("Memory block size not suitable: 0x%lx\n", block_sz);
950 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
952 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
954 panic("%s() failed to register subsystem: %d\n", __func__, ret);
957 * Create entries for memory sections that were found
958 * during boot and have been initialized
960 for (nr = 0; nr <= __highest_present_section_nr;
961 nr += sections_per_block) {
962 ret = add_boot_memory_block(nr);
964 panic("%s() failed to add memory block: %d\n", __func__,
970 * walk_memory_blocks - walk through all present memory blocks overlapped
971 * by the range [start, start + size)
973 * @start: start address of the memory range
974 * @size: size of the memory range
975 * @arg: argument passed to func
976 * @func: callback for each memory section walked
978 * This function walks through all present memory blocks overlapped by the
979 * range [start, start + size), calling func on each memory block.
981 * In case func() returns an error, walking is aborted and the error is
984 * Called under device_hotplug_lock.
986 int walk_memory_blocks(unsigned long start, unsigned long size,
987 void *arg, walk_memory_blocks_func_t func)
989 const unsigned long start_block_id = phys_to_block_id(start);
990 const unsigned long end_block_id = phys_to_block_id(start + size - 1);
991 struct memory_block *mem;
992 unsigned long block_id;
998 for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
999 mem = find_memory_block_by_id(block_id);
1003 ret = func(mem, arg);
1004 put_device(&mem->dev);
1011 struct for_each_memory_block_cb_data {
1012 walk_memory_blocks_func_t func;
1016 static int for_each_memory_block_cb(struct device *dev, void *data)
1018 struct memory_block *mem = to_memory_block(dev);
1019 struct for_each_memory_block_cb_data *cb_data = data;
1021 return cb_data->func(mem, cb_data->arg);
1025 * for_each_memory_block - walk through all present memory blocks
1027 * @arg: argument passed to func
1028 * @func: callback for each memory block walked
1030 * This function walks through all present memory blocks, calling func on
1031 * each memory block.
1033 * In case func() returns an error, walking is aborted and the error is
1036 int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
1038 struct for_each_memory_block_cb_data cb_data = {
1043 return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
1044 for_each_memory_block_cb);
1048 * This is an internal helper to unify allocation and initialization of
1049 * memory groups. Note that the passed memory group will be copied to a
1050 * dynamically allocated memory group. After this call, the passed
1051 * memory group should no longer be used.
1053 static int memory_group_register(struct memory_group group)
1055 struct memory_group *new_group;
1059 if (!node_possible(group.nid))
1062 new_group = kzalloc(sizeof(group), GFP_KERNEL);
1066 INIT_LIST_HEAD(&new_group->memory_blocks);
1068 ret = xa_alloc(&memory_groups, &mgid, new_group, xa_limit_31b,
1073 } else if (group.is_dynamic) {
1074 xa_set_mark(&memory_groups, mgid, MEMORY_GROUP_MARK_DYNAMIC);
1080 * memory_group_register_static() - Register a static memory group.
1081 * @nid: The node id.
1082 * @max_pages: The maximum number of pages we'll have in this static memory
1085 * Register a new static memory group and return the memory group id.
1086 * All memory in the group belongs to a single unit, such as a DIMM. All
1087 * memory belonging to a static memory group is added in one go to be removed
1088 * in one go -- it's static.
1090 * Returns an error if out of memory, if the node id is invalid, if no new
1091 * memory groups can be registered, or if max_pages is invalid (0). Otherwise,
1092 * returns the new memory group id.
1094 int memory_group_register_static(int nid, unsigned long max_pages)
1096 struct memory_group group = {
1099 .max_pages = max_pages,
1105 return memory_group_register(group);
1107 EXPORT_SYMBOL_GPL(memory_group_register_static);
1110 * memory_group_register_dynamic() - Register a dynamic memory group.
1111 * @nid: The node id.
1112 * @unit_pages: Unit in pages in which is memory added/removed in this dynamic
1115 * Register a new dynamic memory group and return the memory group id.
1116 * Memory within a dynamic memory group is added/removed dynamically
1119 * Returns an error if out of memory, if the node id is invalid, if no new
1120 * memory groups can be registered, or if unit_pages is invalid (0, not a
1121 * power of two, smaller than a single memory block). Otherwise, returns the
1122 * new memory group id.
1124 int memory_group_register_dynamic(int nid, unsigned long unit_pages)
1126 struct memory_group group = {
1130 .unit_pages = unit_pages,
1134 if (!unit_pages || !is_power_of_2(unit_pages) ||
1135 unit_pages < PHYS_PFN(memory_block_size_bytes()))
1137 return memory_group_register(group);
1139 EXPORT_SYMBOL_GPL(memory_group_register_dynamic);
1142 * memory_group_unregister() - Unregister a memory group.
1143 * @mgid: the memory group id
1145 * Unregister a memory group. If any memory block still belongs to this
1146 * memory group, unregistering will fail.
1148 * Returns -EINVAL if the memory group id is invalid, returns -EBUSY if some
1149 * memory blocks still belong to this memory group and returns 0 if
1150 * unregistering succeeded.
1152 int memory_group_unregister(int mgid)
1154 struct memory_group *group;
1159 group = xa_load(&memory_groups, mgid);
1162 if (!list_empty(&group->memory_blocks))
1164 xa_erase(&memory_groups, mgid);
1168 EXPORT_SYMBOL_GPL(memory_group_unregister);
1171 * This is an internal helper only to be used in core memory hotplug code to
1172 * lookup a memory group. We don't care about locking, as we don't expect a
1173 * memory group to get unregistered while adding memory to it -- because
1174 * the group and the memory is managed by the same driver.
1176 struct memory_group *memory_group_find_by_id(int mgid)
1178 return xa_load(&memory_groups, mgid);
1182 * This is an internal helper only to be used in core memory hotplug code to
1183 * walk all dynamic memory groups excluding a given memory group, either
1184 * belonging to a specific node, or belonging to any node.
1186 int walk_dynamic_memory_groups(int nid, walk_memory_groups_func_t func,
1187 struct memory_group *excluded, void *arg)
1189 struct memory_group *group;
1190 unsigned long index;
1193 xa_for_each_marked(&memory_groups, index, group,
1194 MEMORY_GROUP_MARK_DYNAMIC) {
1195 if (group == excluded)
1198 if (nid != NUMA_NO_NODE && group->nid != nid)
1200 #endif /* CONFIG_NUMA */
1201 ret = func(group, arg);
1208 #if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
1209 void memblk_nr_poison_inc(unsigned long pfn)
1211 const unsigned long block_id = pfn_to_block_id(pfn);
1212 struct memory_block *mem = find_memory_block_by_id(block_id);
1215 atomic_long_inc(&mem->nr_hwpoison);
1218 void memblk_nr_poison_sub(unsigned long pfn, long i)
1220 const unsigned long block_id = pfn_to_block_id(pfn);
1221 struct memory_block *mem = find_memory_block_by_id(block_id);
1224 atomic_long_sub(i, &mem->nr_hwpoison);
1227 static unsigned long memblk_nr_poison(struct memory_block *mem)
1229 return atomic_long_read(&mem->nr_hwpoison);