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);
74 * register_memory - Setup a sysfs device for a memory block
77 int register_memory(struct memory_block *memory)
81 memory->dev.bus = &memory_subsys;
82 memory->dev.id = memory->start_section_nr / sections_per_block;
84 error = device_register(&memory->dev);
89 unregister_memory(struct memory_block *memory)
91 BUG_ON(memory->dev.bus != &memory_subsys);
93 /* drop the ref. we got in remove_memory_block() */
94 kobject_put(&memory->dev.kobj);
95 device_unregister(&memory->dev);
98 unsigned long __weak memory_block_size_bytes(void)
100 return MIN_MEMORY_BLOCK_SIZE;
103 static unsigned long get_memory_block_size(void)
105 unsigned long block_sz;
107 block_sz = memory_block_size_bytes();
109 /* Validate blk_sz is a power of 2 and not less than section size */
110 if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
112 block_sz = MIN_MEMORY_BLOCK_SIZE;
119 * use this as the physical section index that this memsection
123 static ssize_t show_mem_start_phys_index(struct device *dev,
124 struct device_attribute *attr, char *buf)
126 struct memory_block *mem =
127 container_of(dev, struct memory_block, dev);
128 unsigned long phys_index;
130 phys_index = mem->start_section_nr / sections_per_block;
131 return sprintf(buf, "%08lx\n", phys_index);
134 static ssize_t show_mem_end_phys_index(struct device *dev,
135 struct device_attribute *attr, char *buf)
137 struct memory_block *mem =
138 container_of(dev, struct memory_block, dev);
139 unsigned long phys_index;
141 phys_index = mem->end_section_nr / sections_per_block;
142 return sprintf(buf, "%08lx\n", phys_index);
146 * Show whether the section of memory is likely to be hot-removable
148 static ssize_t show_mem_removable(struct device *dev,
149 struct device_attribute *attr, char *buf)
151 unsigned long i, pfn;
153 struct memory_block *mem =
154 container_of(dev, struct memory_block, dev);
156 for (i = 0; i < sections_per_block; i++) {
157 if (!present_section_nr(mem->start_section_nr + i))
159 pfn = section_nr_to_pfn(mem->start_section_nr + i);
160 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
163 return sprintf(buf, "%d\n", ret);
167 * online, offline, going offline, etc.
169 static ssize_t show_mem_state(struct device *dev,
170 struct device_attribute *attr, char *buf)
172 struct memory_block *mem =
173 container_of(dev, struct memory_block, dev);
177 * We can probably put these states in a nice little array
178 * so that they're not open-coded
180 switch (mem->state) {
182 len = sprintf(buf, "online\n");
185 len = sprintf(buf, "offline\n");
187 case MEM_GOING_OFFLINE:
188 len = sprintf(buf, "going-offline\n");
191 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
200 int memory_notify(unsigned long val, void *v)
202 return blocking_notifier_call_chain(&memory_chain, val, v);
205 int memory_isolate_notify(unsigned long val, void *v)
207 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
211 * The probe routines leave the pages reserved, just as the bootmem code does.
212 * Make sure they're still that way.
214 static bool pages_correctly_reserved(unsigned long start_pfn,
215 unsigned long nr_pages)
219 unsigned long pfn = start_pfn;
222 * memmap between sections is not contiguous except with
223 * SPARSEMEM_VMEMMAP. We lookup the page once per section
224 * and assume memmap is contiguous within each section
226 for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
227 if (WARN_ON_ONCE(!pfn_valid(pfn)))
229 page = pfn_to_page(pfn);
231 for (j = 0; j < PAGES_PER_SECTION; j++) {
232 if (PageReserved(page + j))
235 printk(KERN_WARNING "section number %ld page number %d "
236 "not reserved, was it already online?\n",
237 pfn_to_section_nr(pfn), j);
247 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
248 * OK to have direct references to sparsemem variables in here.
251 memory_block_action(unsigned long phys_index, unsigned long action)
253 unsigned long start_pfn, start_paddr;
254 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
255 struct page *first_page;
258 first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
262 start_pfn = page_to_pfn(first_page);
264 if (!pages_correctly_reserved(start_pfn, nr_pages))
267 ret = online_pages(start_pfn, nr_pages);
270 start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
271 ret = remove_memory(start_paddr,
272 nr_pages << PAGE_SHIFT);
275 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
276 "%ld\n", __func__, phys_index, action, action);
283 static int memory_block_change_state(struct memory_block *mem,
284 unsigned long to_state, unsigned long from_state_req)
288 mutex_lock(&mem->state_mutex);
290 if (mem->state != from_state_req) {
295 if (to_state == MEM_OFFLINE)
296 mem->state = MEM_GOING_OFFLINE;
298 ret = memory_block_action(mem->start_section_nr, to_state);
301 mem->state = from_state_req;
303 mem->state = to_state;
306 mutex_unlock(&mem->state_mutex);
311 store_mem_state(struct device *dev,
312 struct device_attribute *attr, const char *buf, size_t count)
314 struct memory_block *mem;
317 mem = container_of(dev, struct memory_block, dev);
319 if (!strncmp(buf, "online", min((int)count, 6)))
320 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
321 else if(!strncmp(buf, "offline", min((int)count, 7)))
322 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
330 * phys_device is a bad name for this. What I really want
331 * is a way to differentiate between memory ranges that
332 * are part of physical devices that constitute
333 * a complete removable unit or fru.
334 * i.e. do these ranges belong to the same physical device,
335 * s.t. if I offline all of these sections I can then
336 * remove the physical device?
338 static ssize_t show_phys_device(struct device *dev,
339 struct device_attribute *attr, char *buf)
341 struct memory_block *mem =
342 container_of(dev, struct memory_block, dev);
343 return sprintf(buf, "%d\n", mem->phys_device);
346 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
347 static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
348 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
349 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
350 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
352 #define mem_create_simple_file(mem, attr_name) \
353 device_create_file(&mem->dev, &dev_attr_##attr_name)
354 #define mem_remove_simple_file(mem, attr_name) \
355 device_remove_file(&mem->dev, &dev_attr_##attr_name)
358 * Block size attribute stuff
361 print_block_size(struct device *dev, struct device_attribute *attr,
364 return sprintf(buf, "%lx\n", get_memory_block_size());
367 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
369 static int block_size_init(void)
371 return device_create_file(memory_subsys.dev_root,
372 &dev_attr_block_size_bytes);
376 * Some architectures will have custom drivers to do this, and
377 * will not need to do it from userspace. The fake hot-add code
378 * as well as ppc64 will do all of their discovery in userspace
379 * and will require this interface.
381 #ifdef CONFIG_ARCH_MEMORY_PROBE
383 memory_probe_store(struct device *dev, struct device_attribute *attr,
384 const char *buf, size_t count)
389 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
391 phys_addr = simple_strtoull(buf, NULL, 0);
393 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
396 for (i = 0; i < sections_per_block; i++) {
397 nid = memory_add_physaddr_to_nid(phys_addr);
398 ret = add_memory(nid, phys_addr,
399 PAGES_PER_SECTION << PAGE_SHIFT);
403 phys_addr += MIN_MEMORY_BLOCK_SIZE;
410 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
412 static int memory_probe_init(void)
414 return device_create_file(memory_subsys.dev_root, &dev_attr_probe);
417 static inline int memory_probe_init(void)
423 #ifdef CONFIG_MEMORY_FAILURE
425 * Support for offlining pages of memory
428 /* Soft offline a page */
430 store_soft_offline_page(struct device *dev,
431 struct device_attribute *attr,
432 const char *buf, size_t count)
436 if (!capable(CAP_SYS_ADMIN))
438 if (strict_strtoull(buf, 0, &pfn) < 0)
443 ret = soft_offline_page(pfn_to_page(pfn), 0);
444 return ret == 0 ? count : ret;
447 /* Forcibly offline a page, including killing processes. */
449 store_hard_offline_page(struct device *dev,
450 struct device_attribute *attr,
451 const char *buf, size_t count)
455 if (!capable(CAP_SYS_ADMIN))
457 if (strict_strtoull(buf, 0, &pfn) < 0)
460 ret = __memory_failure(pfn, 0, 0);
461 return ret ? ret : count;
464 static DEVICE_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
465 static DEVICE_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
467 static __init int memory_fail_init(void)
471 err = device_create_file(memory_subsys.dev_root,
472 &dev_attr_soft_offline_page);
474 err = device_create_file(memory_subsys.dev_root,
475 &dev_attr_hard_offline_page);
479 static inline int memory_fail_init(void)
486 * Note that phys_device is optional. It is here to allow for
487 * differentiation between which *physical* devices each
488 * section belongs to...
490 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
496 * A reference for the returned object is held and the reference for the
497 * hinted object is released.
499 struct memory_block *find_memory_block_hinted(struct mem_section *section,
500 struct memory_block *hint)
502 int block_id = base_memory_block_id(__section_nr(section));
503 struct device *hintdev = hint ? &hint->dev : NULL;
506 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
508 put_device(&hint->dev);
511 return container_of(dev, struct memory_block, dev);
515 * For now, we have a linear search to go find the appropriate
516 * memory_block corresponding to a particular phys_index. If
517 * this gets to be a real problem, we can always use a radix
518 * tree or something here.
520 * This could be made generic for all device subsystems.
522 struct memory_block *find_memory_block(struct mem_section *section)
524 return find_memory_block_hinted(section, NULL);
527 static int init_memory_block(struct memory_block **memory,
528 struct mem_section *section, unsigned long state)
530 struct memory_block *mem;
531 unsigned long start_pfn;
535 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
539 scn_nr = __section_nr(section);
540 mem->start_section_nr =
541 base_memory_block_id(scn_nr) * sections_per_block;
542 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
544 mem->section_count++;
545 mutex_init(&mem->state_mutex);
546 start_pfn = section_nr_to_pfn(mem->start_section_nr);
547 mem->phys_device = arch_get_memory_phys_device(start_pfn);
549 ret = register_memory(mem);
551 ret = mem_create_simple_file(mem, phys_index);
553 ret = mem_create_simple_file(mem, end_phys_index);
555 ret = mem_create_simple_file(mem, state);
557 ret = mem_create_simple_file(mem, phys_device);
559 ret = mem_create_simple_file(mem, removable);
565 static int add_memory_section(int nid, struct mem_section *section,
566 unsigned long state, enum mem_add_context context)
568 struct memory_block *mem;
571 mutex_lock(&mem_sysfs_mutex);
573 mem = find_memory_block(section);
575 mem->section_count++;
576 kobject_put(&mem->dev.kobj);
578 ret = init_memory_block(&mem, section, state);
581 if (context == HOTPLUG &&
582 mem->section_count == sections_per_block)
583 ret = register_mem_sect_under_node(mem, nid);
586 mutex_unlock(&mem_sysfs_mutex);
590 int remove_memory_block(unsigned long node_id, struct mem_section *section,
593 struct memory_block *mem;
595 mutex_lock(&mem_sysfs_mutex);
596 mem = find_memory_block(section);
597 unregister_mem_sect_under_nodes(mem, __section_nr(section));
599 mem->section_count--;
600 if (mem->section_count == 0) {
601 mem_remove_simple_file(mem, phys_index);
602 mem_remove_simple_file(mem, end_phys_index);
603 mem_remove_simple_file(mem, state);
604 mem_remove_simple_file(mem, phys_device);
605 mem_remove_simple_file(mem, removable);
606 unregister_memory(mem);
609 kobject_put(&mem->dev.kobj);
611 mutex_unlock(&mem_sysfs_mutex);
616 * need an interface for the VM to add new memory regions,
617 * but without onlining it.
619 int register_new_memory(int nid, struct mem_section *section)
621 return add_memory_section(nid, section, MEM_OFFLINE, HOTPLUG);
624 int unregister_memory_section(struct mem_section *section)
626 if (!present_section(section))
629 return remove_memory_block(0, section, 0);
633 * Initialize the sysfs support for memory devices...
635 int __init memory_dev_init(void)
640 unsigned long block_sz;
642 ret = subsys_system_register(&memory_subsys, NULL);
646 block_sz = get_memory_block_size();
647 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
650 * Create entries for memory sections that were found
651 * during boot and have been initialized
653 for (i = 0; i < NR_MEM_SECTIONS; i++) {
654 if (!present_section_nr(i))
656 err = add_memory_section(0, __nr_to_section(i), MEM_ONLINE,
662 err = memory_probe_init();
665 err = memory_fail_init();
668 err = block_size_init();
673 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);