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 int base_memory_block_id(int section_nr)
39 return section_nr / sections_per_block;
42 static inline int pfn_to_block_id(unsigned long pfn)
44 return base_memory_block_id(pfn_to_section_nr(pfn));
47 static int memory_subsys_online(struct device *dev);
48 static int memory_subsys_offline(struct device *dev);
50 static struct bus_type memory_subsys = {
51 .name = MEMORY_CLASS_NAME,
52 .dev_name = MEMORY_CLASS_NAME,
53 .online = memory_subsys_online,
54 .offline = memory_subsys_offline,
57 static BLOCKING_NOTIFIER_HEAD(memory_chain);
59 int register_memory_notifier(struct notifier_block *nb)
61 return blocking_notifier_chain_register(&memory_chain, nb);
63 EXPORT_SYMBOL(register_memory_notifier);
65 void unregister_memory_notifier(struct notifier_block *nb)
67 blocking_notifier_chain_unregister(&memory_chain, nb);
69 EXPORT_SYMBOL(unregister_memory_notifier);
71 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
73 int register_memory_isolate_notifier(struct notifier_block *nb)
75 return atomic_notifier_chain_register(&memory_isolate_chain, nb);
77 EXPORT_SYMBOL(register_memory_isolate_notifier);
79 void unregister_memory_isolate_notifier(struct notifier_block *nb)
81 atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
83 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
85 static void memory_block_release(struct device *dev)
87 struct memory_block *mem = to_memory_block(dev);
92 unsigned long __weak memory_block_size_bytes(void)
94 return MIN_MEMORY_BLOCK_SIZE;
97 static unsigned long get_memory_block_size(void)
99 unsigned long block_sz;
101 block_sz = memory_block_size_bytes();
103 /* Validate blk_sz is a power of 2 and not less than section size */
104 if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
106 block_sz = MIN_MEMORY_BLOCK_SIZE;
113 * use this as the physical section index that this memsection
117 static ssize_t show_mem_start_phys_index(struct device *dev,
118 struct device_attribute *attr, char *buf)
120 struct memory_block *mem = to_memory_block(dev);
121 unsigned long phys_index;
123 phys_index = mem->start_section_nr / sections_per_block;
124 return sprintf(buf, "%08lx\n", phys_index);
128 * Show whether the section of memory is likely to be hot-removable
130 static ssize_t show_mem_removable(struct device *dev,
131 struct device_attribute *attr, char *buf)
133 unsigned long i, pfn;
135 struct memory_block *mem = to_memory_block(dev);
137 if (mem->state != MEM_ONLINE)
140 for (i = 0; i < sections_per_block; i++) {
141 if (!present_section_nr(mem->start_section_nr + i))
143 pfn = section_nr_to_pfn(mem->start_section_nr + i);
144 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
148 return sprintf(buf, "%d\n", ret);
152 * online, offline, going offline, etc.
154 static ssize_t show_mem_state(struct device *dev,
155 struct device_attribute *attr, char *buf)
157 struct memory_block *mem = to_memory_block(dev);
161 * We can probably put these states in a nice little array
162 * so that they're not open-coded
164 switch (mem->state) {
166 len = sprintf(buf, "online\n");
169 len = sprintf(buf, "offline\n");
171 case MEM_GOING_OFFLINE:
172 len = sprintf(buf, "going-offline\n");
175 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
184 int memory_notify(unsigned long val, void *v)
186 return blocking_notifier_call_chain(&memory_chain, val, v);
189 int memory_isolate_notify(unsigned long val, void *v)
191 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
195 * The probe routines leave the pages uninitialized, just as the bootmem code
196 * does. Make sure we do not access them, but instead use only information from
199 static bool pages_correctly_probed(unsigned long start_pfn)
201 unsigned long section_nr = pfn_to_section_nr(start_pfn);
202 unsigned long section_nr_end = section_nr + sections_per_block;
203 unsigned long pfn = start_pfn;
206 * memmap between sections is not contiguous except with
207 * SPARSEMEM_VMEMMAP. We lookup the page once per section
208 * and assume memmap is contiguous within each section
210 for (; section_nr < section_nr_end; section_nr++) {
211 if (WARN_ON_ONCE(!pfn_valid(pfn)))
214 if (!present_section_nr(section_nr)) {
215 pr_warn("section %ld pfn[%lx, %lx) not present",
216 section_nr, pfn, pfn + PAGES_PER_SECTION);
218 } else if (!valid_section_nr(section_nr)) {
219 pr_warn("section %ld pfn[%lx, %lx) no valid memmap",
220 section_nr, pfn, pfn + PAGES_PER_SECTION);
222 } else if (online_section_nr(section_nr)) {
223 pr_warn("section %ld pfn[%lx, %lx) is already online",
224 section_nr, pfn, pfn + PAGES_PER_SECTION);
227 pfn += PAGES_PER_SECTION;
234 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
235 * OK to have direct references to sparsemem variables in here.
238 memory_block_action(unsigned long start_section_nr, unsigned long action,
241 unsigned long start_pfn;
242 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
245 start_pfn = section_nr_to_pfn(start_section_nr);
249 if (!pages_correctly_probed(start_pfn))
252 ret = online_pages(start_pfn, nr_pages, online_type);
255 ret = offline_pages(start_pfn, nr_pages);
258 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
259 "%ld\n", __func__, start_section_nr, action, action);
266 static int memory_block_change_state(struct memory_block *mem,
267 unsigned long to_state, unsigned long from_state_req)
271 if (mem->state != from_state_req)
274 if (to_state == MEM_OFFLINE)
275 mem->state = MEM_GOING_OFFLINE;
277 ret = memory_block_action(mem->start_section_nr, to_state,
280 mem->state = ret ? from_state_req : to_state;
285 /* The device lock serializes operations on memory_subsys_[online|offline] */
286 static int memory_subsys_online(struct device *dev)
288 struct memory_block *mem = to_memory_block(dev);
291 if (mem->state == MEM_ONLINE)
295 * If we are called from store_mem_state(), online_type will be
296 * set >= 0 Otherwise we were called from the device online
297 * attribute and need to set the online_type.
299 if (mem->online_type < 0)
300 mem->online_type = MMOP_ONLINE_KEEP;
302 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
304 /* clear online_type */
305 mem->online_type = -1;
310 static int memory_subsys_offline(struct device *dev)
312 struct memory_block *mem = to_memory_block(dev);
314 if (mem->state == MEM_OFFLINE)
317 /* Can't offline block with non-present sections */
318 if (mem->section_count != sections_per_block)
321 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
325 store_mem_state(struct device *dev,
326 struct device_attribute *attr, const char *buf, size_t count)
328 struct memory_block *mem = to_memory_block(dev);
329 int ret, online_type;
331 ret = lock_device_hotplug_sysfs();
335 if (sysfs_streq(buf, "online_kernel"))
336 online_type = MMOP_ONLINE_KERNEL;
337 else if (sysfs_streq(buf, "online_movable"))
338 online_type = MMOP_ONLINE_MOVABLE;
339 else if (sysfs_streq(buf, "online"))
340 online_type = MMOP_ONLINE_KEEP;
341 else if (sysfs_streq(buf, "offline"))
342 online_type = MMOP_OFFLINE;
348 switch (online_type) {
349 case MMOP_ONLINE_KERNEL:
350 case MMOP_ONLINE_MOVABLE:
351 case MMOP_ONLINE_KEEP:
352 /* mem->online_type is protected by device_hotplug_lock */
353 mem->online_type = online_type;
354 ret = device_online(&mem->dev);
357 ret = device_offline(&mem->dev);
360 ret = -EINVAL; /* should never happen */
364 unlock_device_hotplug();
375 * phys_device is a bad name for this. What I really want
376 * is a way to differentiate between memory ranges that
377 * are part of physical devices that constitute
378 * a complete removable unit or fru.
379 * i.e. do these ranges belong to the same physical device,
380 * s.t. if I offline all of these sections I can then
381 * remove the physical device?
383 static ssize_t show_phys_device(struct device *dev,
384 struct device_attribute *attr, char *buf)
386 struct memory_block *mem = to_memory_block(dev);
387 return sprintf(buf, "%d\n", mem->phys_device);
390 #ifdef CONFIG_MEMORY_HOTREMOVE
391 static void print_allowed_zone(char *buf, int nid, unsigned long start_pfn,
392 unsigned long nr_pages, int online_type,
393 struct zone *default_zone)
397 zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
398 if (zone != default_zone) {
400 strcat(buf, zone->name);
404 static ssize_t show_valid_zones(struct device *dev,
405 struct device_attribute *attr, char *buf)
407 struct memory_block *mem = to_memory_block(dev);
408 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
409 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
410 unsigned long valid_start_pfn, valid_end_pfn;
411 struct zone *default_zone;
415 * Check the existing zone. Make sure that we do that only on the
416 * online nodes otherwise the page_zone is not reliable
418 if (mem->state == MEM_ONLINE) {
420 * The block contains more than one zone can not be offlined.
421 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
423 if (!test_pages_in_a_zone(start_pfn, start_pfn + nr_pages,
424 &valid_start_pfn, &valid_end_pfn))
425 return sprintf(buf, "none\n");
426 start_pfn = valid_start_pfn;
427 strcat(buf, page_zone(pfn_to_page(start_pfn))->name);
432 default_zone = zone_for_pfn_range(MMOP_ONLINE_KEEP, nid, start_pfn, nr_pages);
433 strcat(buf, default_zone->name);
435 print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_KERNEL,
437 print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_MOVABLE,
444 static DEVICE_ATTR(valid_zones, 0444, show_valid_zones, NULL);
447 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
448 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
449 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
450 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
453 * Block size attribute stuff
456 print_block_size(struct device *dev, struct device_attribute *attr,
459 return sprintf(buf, "%lx\n", get_memory_block_size());
462 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
465 * Memory auto online policy.
469 show_auto_online_blocks(struct device *dev, struct device_attribute *attr,
472 if (memhp_auto_online)
473 return sprintf(buf, "online\n");
475 return sprintf(buf, "offline\n");
479 store_auto_online_blocks(struct device *dev, struct device_attribute *attr,
480 const char *buf, size_t count)
482 if (sysfs_streq(buf, "online"))
483 memhp_auto_online = true;
484 else if (sysfs_streq(buf, "offline"))
485 memhp_auto_online = false;
492 static DEVICE_ATTR(auto_online_blocks, 0644, show_auto_online_blocks,
493 store_auto_online_blocks);
496 * Some architectures will have custom drivers to do this, and
497 * will not need to do it from userspace. The fake hot-add code
498 * as well as ppc64 will do all of their discovery in userspace
499 * and will require this interface.
501 #ifdef CONFIG_ARCH_MEMORY_PROBE
503 memory_probe_store(struct device *dev, struct device_attribute *attr,
504 const char *buf, size_t count)
508 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
510 ret = kstrtoull(buf, 0, &phys_addr);
514 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
517 ret = lock_device_hotplug_sysfs();
521 nid = memory_add_physaddr_to_nid(phys_addr);
522 ret = __add_memory(nid, phys_addr,
523 MIN_MEMORY_BLOCK_SIZE * sections_per_block);
530 unlock_device_hotplug();
534 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
537 #ifdef CONFIG_MEMORY_FAILURE
539 * Support for offlining pages of memory
542 /* Soft offline a page */
544 store_soft_offline_page(struct device *dev,
545 struct device_attribute *attr,
546 const char *buf, size_t count)
550 if (!capable(CAP_SYS_ADMIN))
552 if (kstrtoull(buf, 0, &pfn) < 0)
557 /* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
558 if (!pfn_to_online_page(pfn))
560 ret = soft_offline_page(pfn_to_page(pfn), 0);
561 return ret == 0 ? count : ret;
564 /* Forcibly offline a page, including killing processes. */
566 store_hard_offline_page(struct device *dev,
567 struct device_attribute *attr,
568 const char *buf, size_t count)
572 if (!capable(CAP_SYS_ADMIN))
574 if (kstrtoull(buf, 0, &pfn) < 0)
577 ret = memory_failure(pfn, 0);
578 return ret ? ret : count;
581 static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
582 static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
586 * Note that phys_device is optional. It is here to allow for
587 * differentiation between which *physical* devices each
588 * section belongs to...
590 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
596 * A reference for the returned object is held and the reference for the
597 * hinted object is released.
599 static struct memory_block *find_memory_block_by_id(int block_id,
600 struct memory_block *hint)
602 struct device *hintdev = hint ? &hint->dev : NULL;
605 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
607 put_device(&hint->dev);
610 return to_memory_block(dev);
613 struct memory_block *find_memory_block_hinted(struct mem_section *section,
614 struct memory_block *hint)
616 int block_id = base_memory_block_id(__section_nr(section));
618 return find_memory_block_by_id(block_id, hint);
622 * For now, we have a linear search to go find the appropriate
623 * memory_block corresponding to a particular phys_index. If
624 * this gets to be a real problem, we can always use a radix
625 * tree or something here.
627 * This could be made generic for all device subsystems.
629 struct memory_block *find_memory_block(struct mem_section *section)
631 return find_memory_block_hinted(section, NULL);
634 static struct attribute *memory_memblk_attrs[] = {
635 &dev_attr_phys_index.attr,
636 &dev_attr_state.attr,
637 &dev_attr_phys_device.attr,
638 &dev_attr_removable.attr,
639 #ifdef CONFIG_MEMORY_HOTREMOVE
640 &dev_attr_valid_zones.attr,
645 static struct attribute_group memory_memblk_attr_group = {
646 .attrs = memory_memblk_attrs,
649 static const struct attribute_group *memory_memblk_attr_groups[] = {
650 &memory_memblk_attr_group,
655 * register_memory - Setup a sysfs device for a memory block
658 int register_memory(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);
675 static int init_memory_block(struct memory_block **memory, int block_id,
678 struct memory_block *mem;
679 unsigned long start_pfn;
682 mem = find_memory_block_by_id(block_id, NULL);
684 put_device(&mem->dev);
687 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
691 mem->start_section_nr = block_id * sections_per_block;
692 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
694 start_pfn = section_nr_to_pfn(mem->start_section_nr);
695 mem->phys_device = arch_get_memory_phys_device(start_pfn);
696 mem->nid = NUMA_NO_NODE;
698 ret = register_memory(mem);
704 static int add_memory_block(int base_section_nr)
706 struct memory_block *mem;
707 int i, ret, section_count = 0;
709 for (i = base_section_nr;
710 i < base_section_nr + sections_per_block;
712 if (present_section_nr(i))
715 if (section_count == 0)
717 ret = init_memory_block(&mem, base_memory_block_id(base_section_nr),
721 mem->section_count = section_count;
725 static void unregister_memory(struct memory_block *memory)
727 if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
730 /* drop the ref. we got via find_memory_block() */
731 put_device(&memory->dev);
732 device_unregister(&memory->dev);
736 * Create memory block devices for the given memory area. Start and size
737 * have to be aligned to memory block granularity. Memory block devices
738 * will be initialized as offline.
740 int create_memory_block_devices(unsigned long start, unsigned long size)
742 const int start_block_id = pfn_to_block_id(PFN_DOWN(start));
743 int end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
744 struct memory_block *mem;
745 unsigned long block_id;
748 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
749 !IS_ALIGNED(size, memory_block_size_bytes())))
752 mutex_lock(&mem_sysfs_mutex);
753 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
754 ret = init_memory_block(&mem, block_id, MEM_OFFLINE);
757 mem->section_count = sections_per_block;
760 end_block_id = block_id;
761 for (block_id = start_block_id; block_id != end_block_id;
763 mem = find_memory_block_by_id(block_id, NULL);
764 mem->section_count = 0;
765 unregister_memory(mem);
768 mutex_unlock(&mem_sysfs_mutex);
773 * Remove memory block devices for the given memory area. Start and size
774 * have to be aligned to memory block granularity. Memory block devices
775 * have to be offline.
777 void remove_memory_block_devices(unsigned long start, unsigned long size)
779 const int start_block_id = pfn_to_block_id(PFN_DOWN(start));
780 const int end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
781 struct memory_block *mem;
784 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
785 !IS_ALIGNED(size, memory_block_size_bytes())))
788 mutex_lock(&mem_sysfs_mutex);
789 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
790 mem = find_memory_block_by_id(block_id, NULL);
791 if (WARN_ON_ONCE(!mem))
793 mem->section_count = 0;
794 unregister_memory_block_under_nodes(mem);
795 unregister_memory(mem);
797 mutex_unlock(&mem_sysfs_mutex);
800 /* return true if the memory block is offlined, otherwise, return false */
801 bool is_memblock_offlined(struct memory_block *mem)
803 return mem->state == MEM_OFFLINE;
806 static struct attribute *memory_root_attrs[] = {
807 #ifdef CONFIG_ARCH_MEMORY_PROBE
808 &dev_attr_probe.attr,
811 #ifdef CONFIG_MEMORY_FAILURE
812 &dev_attr_soft_offline_page.attr,
813 &dev_attr_hard_offline_page.attr,
816 &dev_attr_block_size_bytes.attr,
817 &dev_attr_auto_online_blocks.attr,
821 static struct attribute_group memory_root_attr_group = {
822 .attrs = memory_root_attrs,
825 static const struct attribute_group *memory_root_attr_groups[] = {
826 &memory_root_attr_group,
831 * Initialize the sysfs support for memory devices...
833 int __init memory_dev_init(void)
838 unsigned long block_sz;
840 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
844 block_sz = get_memory_block_size();
845 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
848 * Create entries for memory sections that were found
849 * during boot and have been initialized
851 mutex_lock(&mem_sysfs_mutex);
852 for (i = 0; i <= __highest_present_section_nr;
853 i += sections_per_block) {
854 err = add_memory_block(i);
858 mutex_unlock(&mem_sysfs_mutex);
862 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
866 struct for_each_memory_block_cb_data {
867 walk_memory_blocks_func_t func;
871 static int for_each_memory_block_cb(struct device *dev, void *data)
873 struct memory_block *mem = to_memory_block(dev);
874 struct for_each_memory_block_cb_data *cb_data = data;
876 return cb_data->func(mem, cb_data->arg);
880 * for_each_memory_block - walk through all present memory blocks
882 * @arg: argument passed to func
883 * @func: callback for each memory block walked
885 * This function walks through all present memory blocks, calling func on
888 * In case func() returns an error, walking is aborted and the error is
891 int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
893 struct for_each_memory_block_cb_data cb_data = {
898 return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
899 for_each_memory_block_cb);