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/mutex.h>
23 #include <linux/stat.h>
24 #include <linux/slab.h>
26 #include <linux/atomic.h>
27 #include <linux/uaccess.h>
29 static DEFINE_MUTEX(mem_sysfs_mutex);
31 #define MEMORY_CLASS_NAME "memory"
33 #define to_memory_block(dev) container_of(dev, struct memory_block, dev)
35 static int sections_per_block;
37 static inline unsigned long base_memory_block_id(unsigned long section_nr)
39 return section_nr / sections_per_block;
42 static inline unsigned long pfn_to_block_id(unsigned long pfn)
44 return base_memory_block_id(pfn_to_section_nr(pfn));
47 static inline unsigned long phys_to_block_id(unsigned long phys)
49 return pfn_to_block_id(PFN_DOWN(phys));
52 static int memory_subsys_online(struct device *dev);
53 static int memory_subsys_offline(struct device *dev);
55 static struct bus_type memory_subsys = {
56 .name = MEMORY_CLASS_NAME,
57 .dev_name = MEMORY_CLASS_NAME,
58 .online = memory_subsys_online,
59 .offline = memory_subsys_offline,
62 static BLOCKING_NOTIFIER_HEAD(memory_chain);
64 int register_memory_notifier(struct notifier_block *nb)
66 return blocking_notifier_chain_register(&memory_chain, nb);
68 EXPORT_SYMBOL(register_memory_notifier);
70 void unregister_memory_notifier(struct notifier_block *nb)
72 blocking_notifier_chain_unregister(&memory_chain, nb);
74 EXPORT_SYMBOL(unregister_memory_notifier);
76 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
78 int register_memory_isolate_notifier(struct notifier_block *nb)
80 return atomic_notifier_chain_register(&memory_isolate_chain, nb);
82 EXPORT_SYMBOL(register_memory_isolate_notifier);
84 void unregister_memory_isolate_notifier(struct notifier_block *nb)
86 atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
88 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
90 static void memory_block_release(struct device *dev)
92 struct memory_block *mem = to_memory_block(dev);
97 unsigned long __weak memory_block_size_bytes(void)
99 return MIN_MEMORY_BLOCK_SIZE;
101 EXPORT_SYMBOL_GPL(memory_block_size_bytes);
104 * Show the first physical section index (number) of this memory block.
106 static ssize_t phys_index_show(struct device *dev,
107 struct device_attribute *attr, char *buf)
109 struct memory_block *mem = to_memory_block(dev);
110 unsigned long phys_index;
112 phys_index = mem->start_section_nr / sections_per_block;
113 return sprintf(buf, "%08lx\n", phys_index);
117 * Show whether the memory block is likely to be offlineable (or is already
118 * offline). Once offline, the memory block could be removed. The return
119 * value does, however, not indicate that there is a way to remove the
122 static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
125 struct memory_block *mem = to_memory_block(dev);
129 if (mem->state != MEM_ONLINE)
132 for (i = 0; i < sections_per_block; i++) {
133 if (!present_section_nr(mem->start_section_nr + i))
135 pfn = section_nr_to_pfn(mem->start_section_nr + i);
136 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
140 return sprintf(buf, "%d\n", ret);
144 * online, offline, going offline, etc.
146 static ssize_t state_show(struct device *dev, struct device_attribute *attr,
149 struct memory_block *mem = to_memory_block(dev);
153 * We can probably put these states in a nice little array
154 * so that they're not open-coded
156 switch (mem->state) {
158 len = sprintf(buf, "online\n");
161 len = sprintf(buf, "offline\n");
163 case MEM_GOING_OFFLINE:
164 len = sprintf(buf, "going-offline\n");
167 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
176 int memory_notify(unsigned long val, void *v)
178 return blocking_notifier_call_chain(&memory_chain, val, v);
181 int memory_isolate_notify(unsigned long val, void *v)
183 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
187 * The probe routines leave the pages uninitialized, just as the bootmem code
188 * does. Make sure we do not access them, but instead use only information from
191 static bool pages_correctly_probed(unsigned long start_pfn)
193 unsigned long section_nr = pfn_to_section_nr(start_pfn);
194 unsigned long section_nr_end = section_nr + sections_per_block;
195 unsigned long pfn = start_pfn;
198 * memmap between sections is not contiguous except with
199 * SPARSEMEM_VMEMMAP. We lookup the page once per section
200 * and assume memmap is contiguous within each section
202 for (; section_nr < section_nr_end; section_nr++) {
203 if (WARN_ON_ONCE(!pfn_valid(pfn)))
206 if (!present_section_nr(section_nr)) {
207 pr_warn("section %ld pfn[%lx, %lx) not present\n",
208 section_nr, pfn, pfn + PAGES_PER_SECTION);
210 } else if (!valid_section_nr(section_nr)) {
211 pr_warn("section %ld pfn[%lx, %lx) no valid memmap\n",
212 section_nr, pfn, pfn + PAGES_PER_SECTION);
214 } else if (online_section_nr(section_nr)) {
215 pr_warn("section %ld pfn[%lx, %lx) is already online\n",
216 section_nr, pfn, pfn + PAGES_PER_SECTION);
219 pfn += PAGES_PER_SECTION;
226 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
227 * OK to have direct references to sparsemem variables in here.
230 memory_block_action(unsigned long start_section_nr, unsigned long action,
233 unsigned long start_pfn;
234 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
237 start_pfn = section_nr_to_pfn(start_section_nr);
241 if (!pages_correctly_probed(start_pfn))
244 ret = online_pages(start_pfn, nr_pages, online_type);
247 ret = offline_pages(start_pfn, nr_pages);
250 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
251 "%ld\n", __func__, start_section_nr, action, action);
258 static int memory_block_change_state(struct memory_block *mem,
259 unsigned long to_state, unsigned long from_state_req)
263 if (mem->state != from_state_req)
266 if (to_state == MEM_OFFLINE)
267 mem->state = MEM_GOING_OFFLINE;
269 ret = memory_block_action(mem->start_section_nr, to_state,
272 mem->state = ret ? from_state_req : to_state;
277 /* The device lock serializes operations on memory_subsys_[online|offline] */
278 static int memory_subsys_online(struct device *dev)
280 struct memory_block *mem = to_memory_block(dev);
283 if (mem->state == MEM_ONLINE)
287 * If we are called from state_store(), online_type will be
288 * set >= 0 Otherwise we were called from the device online
289 * attribute and need to set the online_type.
291 if (mem->online_type < 0)
292 mem->online_type = MMOP_ONLINE_KEEP;
294 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
296 /* clear online_type */
297 mem->online_type = -1;
302 static int memory_subsys_offline(struct device *dev)
304 struct memory_block *mem = to_memory_block(dev);
306 if (mem->state == MEM_OFFLINE)
309 /* Can't offline block with non-present sections */
310 if (mem->section_count != sections_per_block)
313 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
316 static ssize_t state_store(struct device *dev, struct device_attribute *attr,
317 const char *buf, size_t count)
319 struct memory_block *mem = to_memory_block(dev);
320 int ret, online_type;
322 ret = lock_device_hotplug_sysfs();
326 if (sysfs_streq(buf, "online_kernel"))
327 online_type = MMOP_ONLINE_KERNEL;
328 else if (sysfs_streq(buf, "online_movable"))
329 online_type = MMOP_ONLINE_MOVABLE;
330 else if (sysfs_streq(buf, "online"))
331 online_type = MMOP_ONLINE_KEEP;
332 else if (sysfs_streq(buf, "offline"))
333 online_type = MMOP_OFFLINE;
339 switch (online_type) {
340 case MMOP_ONLINE_KERNEL:
341 case MMOP_ONLINE_MOVABLE:
342 case MMOP_ONLINE_KEEP:
343 /* mem->online_type is protected by device_hotplug_lock */
344 mem->online_type = online_type;
345 ret = device_online(&mem->dev);
348 ret = device_offline(&mem->dev);
351 ret = -EINVAL; /* should never happen */
355 unlock_device_hotplug();
366 * phys_device is a bad name for this. What I really want
367 * is a way to differentiate between memory ranges that
368 * are part of physical devices that constitute
369 * a complete removable unit or fru.
370 * i.e. do these ranges belong to the same physical device,
371 * s.t. if I offline all of these sections I can then
372 * remove the physical device?
374 static ssize_t phys_device_show(struct device *dev,
375 struct device_attribute *attr, char *buf)
377 struct memory_block *mem = to_memory_block(dev);
378 return sprintf(buf, "%d\n", mem->phys_device);
381 #ifdef CONFIG_MEMORY_HOTREMOVE
382 static void print_allowed_zone(char *buf, int nid, unsigned long start_pfn,
383 unsigned long nr_pages, int online_type,
384 struct zone *default_zone)
388 zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
389 if (zone != default_zone) {
391 strcat(buf, zone->name);
395 static ssize_t valid_zones_show(struct device *dev,
396 struct device_attribute *attr, char *buf)
398 struct memory_block *mem = to_memory_block(dev);
399 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
400 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
401 unsigned long valid_start_pfn, valid_end_pfn;
402 struct zone *default_zone;
406 * Check the existing zone. Make sure that we do that only on the
407 * online nodes otherwise the page_zone is not reliable
409 if (mem->state == MEM_ONLINE) {
411 * The block contains more than one zone can not be offlined.
412 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
414 if (!test_pages_in_a_zone(start_pfn, start_pfn + nr_pages,
415 &valid_start_pfn, &valid_end_pfn))
416 return sprintf(buf, "none\n");
417 start_pfn = valid_start_pfn;
418 strcat(buf, page_zone(pfn_to_page(start_pfn))->name);
423 default_zone = zone_for_pfn_range(MMOP_ONLINE_KEEP, nid, start_pfn, nr_pages);
424 strcat(buf, default_zone->name);
426 print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_KERNEL,
428 print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_MOVABLE,
435 static DEVICE_ATTR_RO(valid_zones);
438 static DEVICE_ATTR_RO(phys_index);
439 static DEVICE_ATTR_RW(state);
440 static DEVICE_ATTR_RO(phys_device);
441 static DEVICE_ATTR_RO(removable);
444 * Show the memory block size (shared by all memory blocks).
446 static ssize_t block_size_bytes_show(struct device *dev,
447 struct device_attribute *attr, char *buf)
449 return sprintf(buf, "%lx\n", memory_block_size_bytes());
452 static DEVICE_ATTR_RO(block_size_bytes);
455 * Memory auto online policy.
458 static ssize_t auto_online_blocks_show(struct device *dev,
459 struct device_attribute *attr, char *buf)
461 if (memhp_auto_online)
462 return sprintf(buf, "online\n");
464 return sprintf(buf, "offline\n");
467 static ssize_t auto_online_blocks_store(struct device *dev,
468 struct device_attribute *attr,
469 const char *buf, size_t count)
471 if (sysfs_streq(buf, "online"))
472 memhp_auto_online = true;
473 else if (sysfs_streq(buf, "offline"))
474 memhp_auto_online = false;
481 static DEVICE_ATTR_RW(auto_online_blocks);
484 * Some architectures will have custom drivers to do this, and
485 * will not need to do it from userspace. The fake hot-add code
486 * as well as ppc64 will do all of their discovery in userspace
487 * and will require this interface.
489 #ifdef CONFIG_ARCH_MEMORY_PROBE
490 static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
491 const char *buf, size_t count)
495 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
497 ret = kstrtoull(buf, 0, &phys_addr);
501 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
504 ret = lock_device_hotplug_sysfs();
508 nid = memory_add_physaddr_to_nid(phys_addr);
509 ret = __add_memory(nid, phys_addr,
510 MIN_MEMORY_BLOCK_SIZE * sections_per_block);
517 unlock_device_hotplug();
521 static DEVICE_ATTR_WO(probe);
524 #ifdef CONFIG_MEMORY_FAILURE
526 * Support for offlining pages of memory
529 /* Soft offline a page */
530 static ssize_t soft_offline_page_store(struct device *dev,
531 struct device_attribute *attr,
532 const char *buf, size_t count)
536 if (!capable(CAP_SYS_ADMIN))
538 if (kstrtoull(buf, 0, &pfn) < 0)
543 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
544 if (!pfn_to_online_page(pfn))
546 ret = soft_offline_page(pfn_to_page(pfn), 0);
547 return ret == 0 ? count : ret;
550 /* Forcibly offline a page, including killing processes. */
551 static ssize_t hard_offline_page_store(struct device *dev,
552 struct device_attribute *attr,
553 const char *buf, size_t count)
557 if (!capable(CAP_SYS_ADMIN))
559 if (kstrtoull(buf, 0, &pfn) < 0)
562 ret = memory_failure(pfn, 0);
563 return ret ? ret : count;
566 static DEVICE_ATTR_WO(soft_offline_page);
567 static DEVICE_ATTR_WO(hard_offline_page);
571 * Note that phys_device is optional. It is here to allow for
572 * differentiation between which *physical* devices each
573 * section belongs to...
575 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
580 /* A reference for the returned memory block device is acquired. */
581 static struct memory_block *find_memory_block_by_id(unsigned long block_id)
585 dev = subsys_find_device_by_id(&memory_subsys, block_id, NULL);
586 return dev ? to_memory_block(dev) : NULL;
590 * For now, we have a linear search to go find the appropriate
591 * memory_block corresponding to a particular phys_index. If
592 * this gets to be a real problem, we can always use a radix
593 * tree or something here.
595 * This could be made generic for all device subsystems.
597 struct memory_block *find_memory_block(struct mem_section *section)
599 unsigned long block_id = base_memory_block_id(__section_nr(section));
601 return find_memory_block_by_id(block_id);
604 static struct attribute *memory_memblk_attrs[] = {
605 &dev_attr_phys_index.attr,
606 &dev_attr_state.attr,
607 &dev_attr_phys_device.attr,
608 &dev_attr_removable.attr,
609 #ifdef CONFIG_MEMORY_HOTREMOVE
610 &dev_attr_valid_zones.attr,
615 static struct attribute_group memory_memblk_attr_group = {
616 .attrs = memory_memblk_attrs,
619 static const struct attribute_group *memory_memblk_attr_groups[] = {
620 &memory_memblk_attr_group,
625 * register_memory - Setup a sysfs device for a memory block
628 int register_memory(struct memory_block *memory)
632 memory->dev.bus = &memory_subsys;
633 memory->dev.id = memory->start_section_nr / sections_per_block;
634 memory->dev.release = memory_block_release;
635 memory->dev.groups = memory_memblk_attr_groups;
636 memory->dev.offline = memory->state == MEM_OFFLINE;
638 ret = device_register(&memory->dev);
640 put_device(&memory->dev);
645 static int init_memory_block(struct memory_block **memory,
646 unsigned long block_id, unsigned long state)
648 struct memory_block *mem;
649 unsigned long start_pfn;
652 mem = find_memory_block_by_id(block_id);
654 put_device(&mem->dev);
657 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
661 mem->start_section_nr = block_id * sections_per_block;
663 start_pfn = section_nr_to_pfn(mem->start_section_nr);
664 mem->phys_device = arch_get_memory_phys_device(start_pfn);
665 mem->nid = NUMA_NO_NODE;
667 ret = register_memory(mem);
673 static int add_memory_block(unsigned long base_section_nr)
675 int ret, section_count = 0;
676 struct memory_block *mem;
679 for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
681 if (present_section_nr(nr))
684 if (section_count == 0)
686 ret = init_memory_block(&mem, base_memory_block_id(base_section_nr),
690 mem->section_count = section_count;
694 static void unregister_memory(struct memory_block *memory)
696 if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
699 /* drop the ref. we got via find_memory_block() */
700 put_device(&memory->dev);
701 device_unregister(&memory->dev);
705 * Create memory block devices for the given memory area. Start and size
706 * have to be aligned to memory block granularity. Memory block devices
707 * will be initialized as offline.
709 int create_memory_block_devices(unsigned long start, unsigned long size)
711 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
712 unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
713 struct memory_block *mem;
714 unsigned long block_id;
717 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
718 !IS_ALIGNED(size, memory_block_size_bytes())))
721 mutex_lock(&mem_sysfs_mutex);
722 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
723 ret = init_memory_block(&mem, block_id, MEM_OFFLINE);
726 mem->section_count = sections_per_block;
729 end_block_id = block_id;
730 for (block_id = start_block_id; block_id != end_block_id;
732 mem = find_memory_block_by_id(block_id);
733 mem->section_count = 0;
734 unregister_memory(mem);
737 mutex_unlock(&mem_sysfs_mutex);
742 * Remove memory block devices for the given memory area. Start and size
743 * have to be aligned to memory block granularity. Memory block devices
744 * have to be offline.
746 void remove_memory_block_devices(unsigned long start, unsigned long size)
748 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
749 const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
750 struct memory_block *mem;
751 unsigned long block_id;
753 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
754 !IS_ALIGNED(size, memory_block_size_bytes())))
757 mutex_lock(&mem_sysfs_mutex);
758 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
759 mem = find_memory_block_by_id(block_id);
760 if (WARN_ON_ONCE(!mem))
762 mem->section_count = 0;
763 unregister_memory_block_under_nodes(mem);
764 unregister_memory(mem);
766 mutex_unlock(&mem_sysfs_mutex);
769 /* return true if the memory block is offlined, otherwise, return false */
770 bool is_memblock_offlined(struct memory_block *mem)
772 return mem->state == MEM_OFFLINE;
775 static struct attribute *memory_root_attrs[] = {
776 #ifdef CONFIG_ARCH_MEMORY_PROBE
777 &dev_attr_probe.attr,
780 #ifdef CONFIG_MEMORY_FAILURE
781 &dev_attr_soft_offline_page.attr,
782 &dev_attr_hard_offline_page.attr,
785 &dev_attr_block_size_bytes.attr,
786 &dev_attr_auto_online_blocks.attr,
790 static struct attribute_group memory_root_attr_group = {
791 .attrs = memory_root_attrs,
794 static const struct attribute_group *memory_root_attr_groups[] = {
795 &memory_root_attr_group,
800 * Initialize the sysfs support for memory devices...
802 void __init memory_dev_init(void)
806 unsigned long block_sz, nr;
808 /* Validate the configured memory block size */
809 block_sz = memory_block_size_bytes();
810 if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
811 panic("Memory block size not suitable: 0x%lx\n", block_sz);
812 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
814 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
819 * Create entries for memory sections that were found
820 * during boot and have been initialized
822 mutex_lock(&mem_sysfs_mutex);
823 for (nr = 0; nr <= __highest_present_section_nr;
824 nr += sections_per_block) {
825 err = add_memory_block(nr);
829 mutex_unlock(&mem_sysfs_mutex);
833 panic("%s() failed: %d\n", __func__, ret);
837 * walk_memory_blocks - walk through all present memory blocks overlapped
838 * by the range [start, start + size)
840 * @start: start address of the memory range
841 * @size: size of the memory range
842 * @arg: argument passed to func
843 * @func: callback for each memory section walked
845 * This function walks through all present memory blocks overlapped by the
846 * range [start, start + size), calling func on each memory block.
848 * In case func() returns an error, walking is aborted and the error is
851 int walk_memory_blocks(unsigned long start, unsigned long size,
852 void *arg, walk_memory_blocks_func_t func)
854 const unsigned long start_block_id = phys_to_block_id(start);
855 const unsigned long end_block_id = phys_to_block_id(start + size - 1);
856 struct memory_block *mem;
857 unsigned long block_id;
863 for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
864 mem = find_memory_block_by_id(block_id);
868 ret = func(mem, arg);
869 put_device(&mem->dev);
876 struct for_each_memory_block_cb_data {
877 walk_memory_blocks_func_t func;
881 static int for_each_memory_block_cb(struct device *dev, void *data)
883 struct memory_block *mem = to_memory_block(dev);
884 struct for_each_memory_block_cb_data *cb_data = data;
886 return cb_data->func(mem, cb_data->arg);
890 * for_each_memory_block - walk through all present memory blocks
892 * @arg: argument passed to func
893 * @func: callback for each memory block walked
895 * This function walks through all present memory blocks, calling func on
898 * In case func() returns an error, walking is aborted and the error is
901 int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
903 struct for_each_memory_block_cb_data cb_data = {
908 return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
909 for_each_memory_block_cb);