2 * drivers/base/memory.c - basic Memory class support
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/sysdev.h>
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/kobject.h>
21 #include <linux/memory_hotplug.h>
23 #include <linux/mutex.h>
24 #include <linux/stat.h>
25 #include <linux/slab.h>
27 #include <linux/atomic.h>
28 #include <asm/uaccess.h>
30 static DEFINE_MUTEX(mem_sysfs_mutex);
32 #define MEMORY_CLASS_NAME "memory"
34 static int sections_per_block;
36 static inline int base_memory_block_id(int section_nr)
38 return section_nr / sections_per_block;
41 static struct sysdev_class memory_sysdev_class = {
42 .name = MEMORY_CLASS_NAME,
45 static const char *memory_uevent_name(struct kset *kset, struct kobject *kobj)
47 return MEMORY_CLASS_NAME;
50 static int memory_uevent(struct kset *kset, struct kobject *obj,
51 struct kobj_uevent_env *env)
58 static const struct kset_uevent_ops memory_uevent_ops = {
59 .name = memory_uevent_name,
60 .uevent = memory_uevent,
63 static BLOCKING_NOTIFIER_HEAD(memory_chain);
65 int register_memory_notifier(struct notifier_block *nb)
67 return blocking_notifier_chain_register(&memory_chain, nb);
69 EXPORT_SYMBOL(register_memory_notifier);
71 void unregister_memory_notifier(struct notifier_block *nb)
73 blocking_notifier_chain_unregister(&memory_chain, nb);
75 EXPORT_SYMBOL(unregister_memory_notifier);
77 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
79 int register_memory_isolate_notifier(struct notifier_block *nb)
81 return atomic_notifier_chain_register(&memory_isolate_chain, nb);
83 EXPORT_SYMBOL(register_memory_isolate_notifier);
85 void unregister_memory_isolate_notifier(struct notifier_block *nb)
87 atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
89 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
92 * register_memory - Setup a sysfs device for a memory block
95 int register_memory(struct memory_block *memory)
99 memory->sysdev.cls = &memory_sysdev_class;
100 memory->sysdev.id = memory->start_section_nr / sections_per_block;
102 error = sysdev_register(&memory->sysdev);
107 unregister_memory(struct memory_block *memory)
109 BUG_ON(memory->sysdev.cls != &memory_sysdev_class);
111 /* drop the ref. we got in remove_memory_block() */
112 kobject_put(&memory->sysdev.kobj);
113 sysdev_unregister(&memory->sysdev);
116 unsigned long __weak memory_block_size_bytes(void)
118 return MIN_MEMORY_BLOCK_SIZE;
121 static unsigned long get_memory_block_size(void)
123 unsigned long block_sz;
125 block_sz = memory_block_size_bytes();
127 /* Validate blk_sz is a power of 2 and not less than section size */
128 if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
130 block_sz = MIN_MEMORY_BLOCK_SIZE;
137 * use this as the physical section index that this memsection
141 static ssize_t show_mem_start_phys_index(struct sys_device *dev,
142 struct sysdev_attribute *attr, char *buf)
144 struct memory_block *mem =
145 container_of(dev, struct memory_block, sysdev);
146 unsigned long phys_index;
148 phys_index = mem->start_section_nr / sections_per_block;
149 return sprintf(buf, "%08lx\n", phys_index);
152 static ssize_t show_mem_end_phys_index(struct sys_device *dev,
153 struct sysdev_attribute *attr, char *buf)
155 struct memory_block *mem =
156 container_of(dev, struct memory_block, sysdev);
157 unsigned long phys_index;
159 phys_index = mem->end_section_nr / sections_per_block;
160 return sprintf(buf, "%08lx\n", phys_index);
164 * Show whether the section of memory is likely to be hot-removable
166 static ssize_t show_mem_removable(struct sys_device *dev,
167 struct sysdev_attribute *attr, char *buf)
169 unsigned long i, pfn;
171 struct memory_block *mem =
172 container_of(dev, struct memory_block, sysdev);
174 for (i = 0; i < sections_per_block; i++) {
175 pfn = section_nr_to_pfn(mem->start_section_nr + i);
176 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
179 return sprintf(buf, "%d\n", ret);
183 * online, offline, going offline, etc.
185 static ssize_t show_mem_state(struct sys_device *dev,
186 struct sysdev_attribute *attr, char *buf)
188 struct memory_block *mem =
189 container_of(dev, struct memory_block, sysdev);
193 * We can probably put these states in a nice little array
194 * so that they're not open-coded
196 switch (mem->state) {
198 len = sprintf(buf, "online\n");
201 len = sprintf(buf, "offline\n");
203 case MEM_GOING_OFFLINE:
204 len = sprintf(buf, "going-offline\n");
207 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
216 int memory_notify(unsigned long val, void *v)
218 return blocking_notifier_call_chain(&memory_chain, val, v);
221 int memory_isolate_notify(unsigned long val, void *v)
223 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
227 * The probe routines leave the pages reserved, just as the bootmem code does.
228 * Make sure they're still that way.
230 static bool pages_correctly_reserved(unsigned long start_pfn,
231 unsigned long nr_pages)
235 unsigned long pfn = start_pfn;
238 * memmap between sections is not contiguous except with
239 * SPARSEMEM_VMEMMAP. We lookup the page once per section
240 * and assume memmap is contiguous within each section
242 for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
243 if (WARN_ON_ONCE(!pfn_valid(pfn)))
245 page = pfn_to_page(pfn);
247 for (j = 0; j < PAGES_PER_SECTION; j++) {
248 if (PageReserved(page + j))
251 printk(KERN_WARNING "section number %ld page number %d "
252 "not reserved, was it already online?\n",
253 pfn_to_section_nr(pfn), j);
263 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
264 * OK to have direct references to sparsemem variables in here.
267 memory_block_action(unsigned long phys_index, unsigned long action)
269 unsigned long start_pfn, start_paddr;
270 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
271 struct page *first_page;
274 first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
278 start_pfn = page_to_pfn(first_page);
280 if (!pages_correctly_reserved(start_pfn, nr_pages))
283 ret = online_pages(start_pfn, nr_pages);
286 start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
287 ret = remove_memory(start_paddr,
288 nr_pages << PAGE_SHIFT);
291 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
292 "%ld\n", __func__, phys_index, action, action);
299 static int memory_block_change_state(struct memory_block *mem,
300 unsigned long to_state, unsigned long from_state_req)
304 mutex_lock(&mem->state_mutex);
306 if (mem->state != from_state_req) {
311 if (to_state == MEM_OFFLINE)
312 mem->state = MEM_GOING_OFFLINE;
314 ret = memory_block_action(mem->start_section_nr, to_state);
317 mem->state = from_state_req;
319 mem->state = to_state;
322 mutex_unlock(&mem->state_mutex);
327 store_mem_state(struct sys_device *dev,
328 struct sysdev_attribute *attr, const char *buf, size_t count)
330 struct memory_block *mem;
333 mem = container_of(dev, struct memory_block, sysdev);
335 if (!strncmp(buf, "online", min((int)count, 6)))
336 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
337 else if(!strncmp(buf, "offline", min((int)count, 7)))
338 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
346 * phys_device is a bad name for this. What I really want
347 * is a way to differentiate between memory ranges that
348 * are part of physical devices that constitute
349 * a complete removable unit or fru.
350 * i.e. do these ranges belong to the same physical device,
351 * s.t. if I offline all of these sections I can then
352 * remove the physical device?
354 static ssize_t show_phys_device(struct sys_device *dev,
355 struct sysdev_attribute *attr, char *buf)
357 struct memory_block *mem =
358 container_of(dev, struct memory_block, sysdev);
359 return sprintf(buf, "%d\n", mem->phys_device);
362 static SYSDEV_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
363 static SYSDEV_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
364 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
365 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
366 static SYSDEV_ATTR(removable, 0444, show_mem_removable, NULL);
368 #define mem_create_simple_file(mem, attr_name) \
369 sysdev_create_file(&mem->sysdev, &attr_##attr_name)
370 #define mem_remove_simple_file(mem, attr_name) \
371 sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
374 * Block size attribute stuff
377 print_block_size(struct sysdev_class *class, struct sysdev_class_attribute *attr,
380 return sprintf(buf, "%lx\n", get_memory_block_size());
383 static SYSDEV_CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
385 static int block_size_init(void)
387 return sysfs_create_file(&memory_sysdev_class.kset.kobj,
388 &attr_block_size_bytes.attr);
392 * Some architectures will have custom drivers to do this, and
393 * will not need to do it from userspace. The fake hot-add code
394 * as well as ppc64 will do all of their discovery in userspace
395 * and will require this interface.
397 #ifdef CONFIG_ARCH_MEMORY_PROBE
399 memory_probe_store(struct class *class, struct class_attribute *attr,
400 const char *buf, size_t count)
405 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
407 phys_addr = simple_strtoull(buf, NULL, 0);
409 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
412 for (i = 0; i < sections_per_block; i++) {
413 nid = memory_add_physaddr_to_nid(phys_addr);
414 ret = add_memory(nid, phys_addr,
415 PAGES_PER_SECTION << PAGE_SHIFT);
419 phys_addr += MIN_MEMORY_BLOCK_SIZE;
426 static CLASS_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
428 static int memory_probe_init(void)
430 return sysfs_create_file(&memory_sysdev_class.kset.kobj,
431 &class_attr_probe.attr);
434 static inline int memory_probe_init(void)
440 #ifdef CONFIG_MEMORY_FAILURE
442 * Support for offlining pages of memory
445 /* Soft offline a page */
447 store_soft_offline_page(struct class *class,
448 struct class_attribute *attr,
449 const char *buf, size_t count)
453 if (!capable(CAP_SYS_ADMIN))
455 if (strict_strtoull(buf, 0, &pfn) < 0)
460 ret = soft_offline_page(pfn_to_page(pfn), 0);
461 return ret == 0 ? count : ret;
464 /* Forcibly offline a page, including killing processes. */
466 store_hard_offline_page(struct class *class,
467 struct class_attribute *attr,
468 const char *buf, size_t count)
472 if (!capable(CAP_SYS_ADMIN))
474 if (strict_strtoull(buf, 0, &pfn) < 0)
477 ret = __memory_failure(pfn, 0, 0);
478 return ret ? ret : count;
481 static CLASS_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
482 static CLASS_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
484 static __init int memory_fail_init(void)
488 err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
489 &class_attr_soft_offline_page.attr);
491 err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
492 &class_attr_hard_offline_page.attr);
496 static inline int memory_fail_init(void)
503 * Note that phys_device is optional. It is here to allow for
504 * differentiation between which *physical* devices each
505 * section belongs to...
507 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
512 struct memory_block *find_memory_block_hinted(struct mem_section *section,
513 struct memory_block *hint)
515 struct kobject *kobj;
516 struct sys_device *sysdev;
517 struct memory_block *mem;
518 char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
519 int block_id = base_memory_block_id(__section_nr(section));
521 kobj = hint ? &hint->sysdev.kobj : NULL;
524 * This only works because we know that section == sysdev->id
525 * slightly redundant with sysdev_register()
527 sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, block_id);
529 kobj = kset_find_obj_hinted(&memory_sysdev_class.kset, name, kobj);
533 sysdev = container_of(kobj, struct sys_device, kobj);
534 mem = container_of(sysdev, struct memory_block, sysdev);
540 * For now, we have a linear search to go find the appropriate
541 * memory_block corresponding to a particular phys_index. If
542 * this gets to be a real problem, we can always use a radix
543 * tree or something here.
545 * This could be made generic for all sysdev classes.
547 struct memory_block *find_memory_block(struct mem_section *section)
549 return find_memory_block_hinted(section, NULL);
552 static int init_memory_block(struct memory_block **memory,
553 struct mem_section *section, unsigned long state)
555 struct memory_block *mem;
556 unsigned long start_pfn;
560 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
564 scn_nr = __section_nr(section);
565 mem->start_section_nr =
566 base_memory_block_id(scn_nr) * sections_per_block;
567 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
569 mem->section_count++;
570 mutex_init(&mem->state_mutex);
571 start_pfn = section_nr_to_pfn(mem->start_section_nr);
572 mem->phys_device = arch_get_memory_phys_device(start_pfn);
574 ret = register_memory(mem);
576 ret = mem_create_simple_file(mem, phys_index);
578 ret = mem_create_simple_file(mem, end_phys_index);
580 ret = mem_create_simple_file(mem, state);
582 ret = mem_create_simple_file(mem, phys_device);
584 ret = mem_create_simple_file(mem, removable);
590 static int add_memory_section(int nid, struct mem_section *section,
591 unsigned long state, enum mem_add_context context)
593 struct memory_block *mem;
596 mutex_lock(&mem_sysfs_mutex);
598 mem = find_memory_block(section);
600 mem->section_count++;
601 kobject_put(&mem->sysdev.kobj);
603 ret = init_memory_block(&mem, section, state);
606 if (context == HOTPLUG &&
607 mem->section_count == sections_per_block)
608 ret = register_mem_sect_under_node(mem, nid);
611 mutex_unlock(&mem_sysfs_mutex);
615 int remove_memory_block(unsigned long node_id, struct mem_section *section,
618 struct memory_block *mem;
620 mutex_lock(&mem_sysfs_mutex);
621 mem = find_memory_block(section);
622 unregister_mem_sect_under_nodes(mem, __section_nr(section));
624 mem->section_count--;
625 if (mem->section_count == 0) {
626 mem_remove_simple_file(mem, phys_index);
627 mem_remove_simple_file(mem, end_phys_index);
628 mem_remove_simple_file(mem, state);
629 mem_remove_simple_file(mem, phys_device);
630 mem_remove_simple_file(mem, removable);
631 unregister_memory(mem);
634 kobject_put(&mem->sysdev.kobj);
636 mutex_unlock(&mem_sysfs_mutex);
641 * need an interface for the VM to add new memory regions,
642 * but without onlining it.
644 int register_new_memory(int nid, struct mem_section *section)
646 return add_memory_section(nid, section, MEM_OFFLINE, HOTPLUG);
649 int unregister_memory_section(struct mem_section *section)
651 if (!present_section(section))
654 return remove_memory_block(0, section, 0);
658 * Initialize the sysfs support for memory devices...
660 int __init memory_dev_init(void)
665 unsigned long block_sz;
667 memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
668 ret = sysdev_class_register(&memory_sysdev_class);
672 block_sz = get_memory_block_size();
673 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
676 * Create entries for memory sections that were found
677 * during boot and have been initialized
679 for (i = 0; i < NR_MEM_SECTIONS; i++) {
680 if (!present_section_nr(i))
682 err = add_memory_section(0, __nr_to_section(i), MEM_ONLINE,
688 err = memory_probe_init();
691 err = memory_fail_init();
694 err = block_size_init();
699 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);