2 * Memory subsystem support
4 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
5 * Dave Hansen <haveblue@us.ibm.com>
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.
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/kobject.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 <asm/uaccess.h>
29 static DEFINE_MUTEX(mem_sysfs_mutex);
31 #define MEMORY_CLASS_NAME "memory"
33 static int sections_per_block;
35 static inline int base_memory_block_id(int section_nr)
37 return section_nr / sections_per_block;
40 static struct bus_type memory_subsys = {
41 .name = MEMORY_CLASS_NAME,
42 .dev_name = MEMORY_CLASS_NAME,
45 static BLOCKING_NOTIFIER_HEAD(memory_chain);
47 int register_memory_notifier(struct notifier_block *nb)
49 return blocking_notifier_chain_register(&memory_chain, nb);
51 EXPORT_SYMBOL(register_memory_notifier);
53 void unregister_memory_notifier(struct notifier_block *nb)
55 blocking_notifier_chain_unregister(&memory_chain, nb);
57 EXPORT_SYMBOL(unregister_memory_notifier);
59 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
61 int register_memory_isolate_notifier(struct notifier_block *nb)
63 return atomic_notifier_chain_register(&memory_isolate_chain, nb);
65 EXPORT_SYMBOL(register_memory_isolate_notifier);
67 void unregister_memory_isolate_notifier(struct notifier_block *nb)
69 atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
71 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
73 static void memory_block_release(struct device *dev)
75 struct memory_block *mem = container_of(dev, struct memory_block, dev);
81 * register_memory - Setup a sysfs device for a memory block
84 int register_memory(struct memory_block *memory)
88 memory->dev.bus = &memory_subsys;
89 memory->dev.id = memory->start_section_nr / sections_per_block;
90 memory->dev.release = memory_block_release;
92 error = device_register(&memory->dev);
96 unsigned long __weak memory_block_size_bytes(void)
98 return MIN_MEMORY_BLOCK_SIZE;
101 static unsigned long get_memory_block_size(void)
103 unsigned long block_sz;
105 block_sz = memory_block_size_bytes();
107 /* Validate blk_sz is a power of 2 and not less than section size */
108 if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
110 block_sz = MIN_MEMORY_BLOCK_SIZE;
117 * use this as the physical section index that this memsection
121 static ssize_t show_mem_start_phys_index(struct device *dev,
122 struct device_attribute *attr, char *buf)
124 struct memory_block *mem =
125 container_of(dev, struct memory_block, dev);
126 unsigned long phys_index;
128 phys_index = mem->start_section_nr / sections_per_block;
129 return sprintf(buf, "%08lx\n", phys_index);
132 static ssize_t show_mem_end_phys_index(struct device *dev,
133 struct device_attribute *attr, char *buf)
135 struct memory_block *mem =
136 container_of(dev, struct memory_block, dev);
137 unsigned long phys_index;
139 phys_index = mem->end_section_nr / sections_per_block;
140 return sprintf(buf, "%08lx\n", phys_index);
144 * Show whether the section of memory is likely to be hot-removable
146 static ssize_t show_mem_removable(struct device *dev,
147 struct device_attribute *attr, char *buf)
149 unsigned long i, pfn;
151 struct memory_block *mem =
152 container_of(dev, struct memory_block, dev);
154 for (i = 0; i < sections_per_block; i++) {
155 pfn = section_nr_to_pfn(mem->start_section_nr + i);
156 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
159 return sprintf(buf, "%d\n", ret);
163 * online, offline, going offline, etc.
165 static ssize_t show_mem_state(struct device *dev,
166 struct device_attribute *attr, char *buf)
168 struct memory_block *mem =
169 container_of(dev, struct memory_block, dev);
173 * We can probably put these states in a nice little array
174 * so that they're not open-coded
176 switch (mem->state) {
178 len = sprintf(buf, "online\n");
181 len = sprintf(buf, "offline\n");
183 case MEM_GOING_OFFLINE:
184 len = sprintf(buf, "going-offline\n");
187 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
196 int memory_notify(unsigned long val, void *v)
198 return blocking_notifier_call_chain(&memory_chain, val, v);
201 int memory_isolate_notify(unsigned long val, void *v)
203 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
207 * The probe routines leave the pages reserved, just as the bootmem code does.
208 * Make sure they're still that way.
210 static bool pages_correctly_reserved(unsigned long start_pfn,
211 unsigned long nr_pages)
215 unsigned long pfn = start_pfn;
218 * memmap between sections is not contiguous except with
219 * SPARSEMEM_VMEMMAP. We lookup the page once per section
220 * and assume memmap is contiguous within each section
222 for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
223 if (WARN_ON_ONCE(!pfn_valid(pfn)))
225 page = pfn_to_page(pfn);
227 for (j = 0; j < PAGES_PER_SECTION; j++) {
228 if (PageReserved(page + j))
231 printk(KERN_WARNING "section number %ld page number %d "
232 "not reserved, was it already online?\n",
233 pfn_to_section_nr(pfn), j);
243 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
244 * OK to have direct references to sparsemem variables in here.
247 memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
249 unsigned long start_pfn;
250 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
251 struct page *first_page;
254 first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
255 start_pfn = page_to_pfn(first_page);
259 if (!pages_correctly_reserved(start_pfn, nr_pages))
262 ret = online_pages(start_pfn, nr_pages, online_type);
265 ret = offline_pages(start_pfn, nr_pages);
268 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
269 "%ld\n", __func__, phys_index, action, action);
276 static int __memory_block_change_state(struct memory_block *mem,
277 unsigned long to_state, unsigned long from_state_req,
282 if (mem->state != from_state_req) {
287 if (to_state == MEM_OFFLINE)
288 mem->state = MEM_GOING_OFFLINE;
290 ret = memory_block_action(mem->start_section_nr, to_state, online_type);
293 mem->state = from_state_req;
297 mem->state = to_state;
298 switch (mem->state) {
300 kobject_uevent(&mem->dev.kobj, KOBJ_OFFLINE);
303 kobject_uevent(&mem->dev.kobj, KOBJ_ONLINE);
312 static int memory_block_change_state(struct memory_block *mem,
313 unsigned long to_state, unsigned long from_state_req,
318 mutex_lock(&mem->state_mutex);
319 ret = __memory_block_change_state(mem, to_state, from_state_req,
321 mutex_unlock(&mem->state_mutex);
326 store_mem_state(struct device *dev,
327 struct device_attribute *attr, const char *buf, size_t count)
329 struct memory_block *mem;
332 mem = container_of(dev, struct memory_block, dev);
334 if (!strncmp(buf, "online_kernel", min_t(int, count, 13)))
335 ret = memory_block_change_state(mem, MEM_ONLINE,
336 MEM_OFFLINE, ONLINE_KERNEL);
337 else if (!strncmp(buf, "online_movable", min_t(int, count, 14)))
338 ret = memory_block_change_state(mem, MEM_ONLINE,
339 MEM_OFFLINE, ONLINE_MOVABLE);
340 else if (!strncmp(buf, "online", min_t(int, count, 6)))
341 ret = memory_block_change_state(mem, MEM_ONLINE,
342 MEM_OFFLINE, ONLINE_KEEP);
343 else if(!strncmp(buf, "offline", min_t(int, count, 7)))
344 ret = memory_block_change_state(mem, MEM_OFFLINE,
353 * phys_device is a bad name for this. What I really want
354 * is a way to differentiate between memory ranges that
355 * are part of physical devices that constitute
356 * a complete removable unit or fru.
357 * i.e. do these ranges belong to the same physical device,
358 * s.t. if I offline all of these sections I can then
359 * remove the physical device?
361 static ssize_t show_phys_device(struct device *dev,
362 struct device_attribute *attr, char *buf)
364 struct memory_block *mem =
365 container_of(dev, struct memory_block, dev);
366 return sprintf(buf, "%d\n", mem->phys_device);
369 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
370 static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
371 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
372 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
373 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
375 #define mem_create_simple_file(mem, attr_name) \
376 device_create_file(&mem->dev, &dev_attr_##attr_name)
377 #define mem_remove_simple_file(mem, attr_name) \
378 device_remove_file(&mem->dev, &dev_attr_##attr_name)
381 * Block size attribute stuff
384 print_block_size(struct device *dev, struct device_attribute *attr,
387 return sprintf(buf, "%lx\n", get_memory_block_size());
390 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
392 static int block_size_init(void)
394 return device_create_file(memory_subsys.dev_root,
395 &dev_attr_block_size_bytes);
399 * Some architectures will have custom drivers to do this, and
400 * will not need to do it from userspace. The fake hot-add code
401 * as well as ppc64 will do all of their discovery in userspace
402 * and will require this interface.
404 #ifdef CONFIG_ARCH_MEMORY_PROBE
406 memory_probe_store(struct device *dev, struct device_attribute *attr,
407 const char *buf, size_t count)
412 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
414 phys_addr = simple_strtoull(buf, NULL, 0);
416 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
419 for (i = 0; i < sections_per_block; i++) {
420 nid = memory_add_physaddr_to_nid(phys_addr);
421 ret = add_memory(nid, phys_addr,
422 PAGES_PER_SECTION << PAGE_SHIFT);
426 phys_addr += MIN_MEMORY_BLOCK_SIZE;
433 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
435 static int memory_probe_init(void)
437 return device_create_file(memory_subsys.dev_root, &dev_attr_probe);
440 static inline int memory_probe_init(void)
446 #ifdef CONFIG_MEMORY_FAILURE
448 * Support for offlining pages of memory
451 /* Soft offline a page */
453 store_soft_offline_page(struct device *dev,
454 struct device_attribute *attr,
455 const char *buf, size_t count)
459 if (!capable(CAP_SYS_ADMIN))
461 if (strict_strtoull(buf, 0, &pfn) < 0)
466 ret = soft_offline_page(pfn_to_page(pfn), 0);
467 return ret == 0 ? count : ret;
470 /* Forcibly offline a page, including killing processes. */
472 store_hard_offline_page(struct device *dev,
473 struct device_attribute *attr,
474 const char *buf, size_t count)
478 if (!capable(CAP_SYS_ADMIN))
480 if (strict_strtoull(buf, 0, &pfn) < 0)
483 ret = memory_failure(pfn, 0, 0);
484 return ret ? ret : count;
487 static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
488 static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
490 static __init int memory_fail_init(void)
494 err = device_create_file(memory_subsys.dev_root,
495 &dev_attr_soft_offline_page);
497 err = device_create_file(memory_subsys.dev_root,
498 &dev_attr_hard_offline_page);
502 static inline int memory_fail_init(void)
509 * Note that phys_device is optional. It is here to allow for
510 * differentiation between which *physical* devices each
511 * section belongs to...
513 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
519 * A reference for the returned object is held and the reference for the
520 * hinted object is released.
522 struct memory_block *find_memory_block_hinted(struct mem_section *section,
523 struct memory_block *hint)
525 int block_id = base_memory_block_id(__section_nr(section));
526 struct device *hintdev = hint ? &hint->dev : NULL;
529 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
531 put_device(&hint->dev);
534 return container_of(dev, struct memory_block, dev);
538 * For now, we have a linear search to go find the appropriate
539 * memory_block corresponding to a particular phys_index. If
540 * this gets to be a real problem, we can always use a radix
541 * tree or something here.
543 * This could be made generic for all device subsystems.
545 struct memory_block *find_memory_block(struct mem_section *section)
547 return find_memory_block_hinted(section, NULL);
550 static int init_memory_block(struct memory_block **memory,
551 struct mem_section *section, unsigned long state)
553 struct memory_block *mem;
554 unsigned long start_pfn;
558 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
562 scn_nr = __section_nr(section);
563 mem->start_section_nr =
564 base_memory_block_id(scn_nr) * sections_per_block;
565 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
567 mem->section_count++;
568 mutex_init(&mem->state_mutex);
569 start_pfn = section_nr_to_pfn(mem->start_section_nr);
570 mem->phys_device = arch_get_memory_phys_device(start_pfn);
572 ret = register_memory(mem);
574 ret = mem_create_simple_file(mem, phys_index);
576 ret = mem_create_simple_file(mem, end_phys_index);
578 ret = mem_create_simple_file(mem, state);
580 ret = mem_create_simple_file(mem, phys_device);
582 ret = mem_create_simple_file(mem, removable);
588 static int add_memory_section(int nid, struct mem_section *section,
589 struct memory_block **mem_p,
590 unsigned long state, enum mem_add_context context)
592 struct memory_block *mem = NULL;
593 int scn_nr = __section_nr(section);
596 mutex_lock(&mem_sysfs_mutex);
598 if (context == BOOT) {
599 /* same memory block ? */
601 if (scn_nr >= (*mem_p)->start_section_nr &&
602 scn_nr <= (*mem_p)->end_section_nr) {
604 kobject_get(&mem->dev.kobj);
607 mem = find_memory_block(section);
610 mem->section_count++;
611 kobject_put(&mem->dev.kobj);
613 ret = init_memory_block(&mem, section, state);
614 /* store memory_block pointer for next loop */
615 if (!ret && context == BOOT)
621 if (context == HOTPLUG &&
622 mem->section_count == sections_per_block)
623 ret = register_mem_sect_under_node(mem, nid);
626 mutex_unlock(&mem_sysfs_mutex);
631 * need an interface for the VM to add new memory regions,
632 * but without onlining it.
634 int register_new_memory(int nid, struct mem_section *section)
636 return add_memory_section(nid, section, NULL, MEM_OFFLINE, HOTPLUG);
639 #ifdef CONFIG_MEMORY_HOTREMOVE
641 unregister_memory(struct memory_block *memory)
643 BUG_ON(memory->dev.bus != &memory_subsys);
645 /* drop the ref. we got in remove_memory_block() */
646 kobject_put(&memory->dev.kobj);
647 device_unregister(&memory->dev);
650 static int remove_memory_block(unsigned long node_id,
651 struct mem_section *section, int phys_device)
653 struct memory_block *mem;
655 mutex_lock(&mem_sysfs_mutex);
656 mem = find_memory_block(section);
657 unregister_mem_sect_under_nodes(mem, __section_nr(section));
659 mem->section_count--;
660 if (mem->section_count == 0) {
661 mem_remove_simple_file(mem, phys_index);
662 mem_remove_simple_file(mem, end_phys_index);
663 mem_remove_simple_file(mem, state);
664 mem_remove_simple_file(mem, phys_device);
665 mem_remove_simple_file(mem, removable);
666 unregister_memory(mem);
668 kobject_put(&mem->dev.kobj);
670 mutex_unlock(&mem_sysfs_mutex);
674 int unregister_memory_section(struct mem_section *section)
676 if (!present_section(section))
679 return remove_memory_block(0, section, 0);
681 #endif /* CONFIG_MEMORY_HOTREMOVE */
684 * offline one memory block. If the memory block has been offlined, do nothing.
686 int offline_memory_block(struct memory_block *mem)
690 mutex_lock(&mem->state_mutex);
691 if (mem->state != MEM_OFFLINE)
692 ret = __memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE, -1);
693 mutex_unlock(&mem->state_mutex);
698 /* return true if the memory block is offlined, otherwise, return false */
699 bool is_memblock_offlined(struct memory_block *mem)
701 return mem->state == MEM_OFFLINE;
705 * Initialize the sysfs support for memory devices...
707 int __init memory_dev_init(void)
712 unsigned long block_sz;
713 struct memory_block *mem = NULL;
715 ret = subsys_system_register(&memory_subsys, NULL);
719 block_sz = get_memory_block_size();
720 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
723 * Create entries for memory sections that were found
724 * during boot and have been initialized
726 for (i = 0; i < NR_MEM_SECTIONS; i++) {
727 if (!present_section_nr(i))
729 /* don't need to reuse memory_block if only one per block */
730 err = add_memory_section(0, __nr_to_section(i),
731 (sections_per_block == 1) ? NULL : &mem,
738 err = memory_probe_init();
741 err = memory_fail_init();
744 err = block_size_init();
749 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);