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 pfn = section_nr_to_pfn(mem->start_section_nr + i);
158 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
161 return sprintf(buf, "%d\n", ret);
165 * online, offline, going offline, etc.
167 static ssize_t show_mem_state(struct device *dev,
168 struct device_attribute *attr, char *buf)
170 struct memory_block *mem =
171 container_of(dev, struct memory_block, dev);
175 * We can probably put these states in a nice little array
176 * so that they're not open-coded
178 switch (mem->state) {
180 len = sprintf(buf, "online\n");
183 len = sprintf(buf, "offline\n");
185 case MEM_GOING_OFFLINE:
186 len = sprintf(buf, "going-offline\n");
189 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
198 int memory_notify(unsigned long val, void *v)
200 return blocking_notifier_call_chain(&memory_chain, val, v);
203 int memory_isolate_notify(unsigned long val, void *v)
205 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
209 * The probe routines leave the pages reserved, just as the bootmem code does.
210 * Make sure they're still that way.
212 static bool pages_correctly_reserved(unsigned long start_pfn,
213 unsigned long nr_pages)
217 unsigned long pfn = start_pfn;
220 * memmap between sections is not contiguous except with
221 * SPARSEMEM_VMEMMAP. We lookup the page once per section
222 * and assume memmap is contiguous within each section
224 for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
225 if (WARN_ON_ONCE(!pfn_valid(pfn)))
227 page = pfn_to_page(pfn);
229 for (j = 0; j < PAGES_PER_SECTION; j++) {
230 if (PageReserved(page + j))
233 printk(KERN_WARNING "section number %ld page number %d "
234 "not reserved, was it already online?\n",
235 pfn_to_section_nr(pfn), j);
245 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
246 * OK to have direct references to sparsemem variables in here.
249 memory_block_action(unsigned long phys_index, unsigned long action)
251 unsigned long start_pfn, start_paddr;
252 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
253 struct page *first_page;
256 first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
260 start_pfn = page_to_pfn(first_page);
262 if (!pages_correctly_reserved(start_pfn, nr_pages))
265 ret = online_pages(start_pfn, nr_pages);
268 start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
269 ret = remove_memory(start_paddr,
270 nr_pages << PAGE_SHIFT);
273 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
274 "%ld\n", __func__, phys_index, action, action);
281 static int memory_block_change_state(struct memory_block *mem,
282 unsigned long to_state, unsigned long from_state_req)
286 mutex_lock(&mem->state_mutex);
288 if (mem->state != from_state_req) {
293 if (to_state == MEM_OFFLINE)
294 mem->state = MEM_GOING_OFFLINE;
296 ret = memory_block_action(mem->start_section_nr, to_state);
299 mem->state = from_state_req;
301 mem->state = to_state;
304 mutex_unlock(&mem->state_mutex);
309 store_mem_state(struct device *dev,
310 struct device_attribute *attr, const char *buf, size_t count)
312 struct memory_block *mem;
315 mem = container_of(dev, struct memory_block, dev);
317 if (!strncmp(buf, "online", min((int)count, 6)))
318 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
319 else if(!strncmp(buf, "offline", min((int)count, 7)))
320 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
328 * phys_device is a bad name for this. What I really want
329 * is a way to differentiate between memory ranges that
330 * are part of physical devices that constitute
331 * a complete removable unit or fru.
332 * i.e. do these ranges belong to the same physical device,
333 * s.t. if I offline all of these sections I can then
334 * remove the physical device?
336 static ssize_t show_phys_device(struct device *dev,
337 struct device_attribute *attr, char *buf)
339 struct memory_block *mem =
340 container_of(dev, struct memory_block, dev);
341 return sprintf(buf, "%d\n", mem->phys_device);
344 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
345 static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
346 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
347 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
348 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
350 #define mem_create_simple_file(mem, attr_name) \
351 device_create_file(&mem->dev, &dev_attr_##attr_name)
352 #define mem_remove_simple_file(mem, attr_name) \
353 device_remove_file(&mem->dev, &dev_attr_##attr_name)
356 * Block size attribute stuff
359 print_block_size(struct device *dev, struct device_attribute *attr,
362 return sprintf(buf, "%lx\n", get_memory_block_size());
365 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
367 static int block_size_init(void)
369 return device_create_file(memory_subsys.dev_root,
370 &dev_attr_block_size_bytes);
374 * Some architectures will have custom drivers to do this, and
375 * will not need to do it from userspace. The fake hot-add code
376 * as well as ppc64 will do all of their discovery in userspace
377 * and will require this interface.
379 #ifdef CONFIG_ARCH_MEMORY_PROBE
381 memory_probe_store(struct device *dev, struct device_attribute *attr,
382 const char *buf, size_t count)
387 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
389 phys_addr = simple_strtoull(buf, NULL, 0);
391 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
394 for (i = 0; i < sections_per_block; i++) {
395 nid = memory_add_physaddr_to_nid(phys_addr);
396 ret = add_memory(nid, phys_addr,
397 PAGES_PER_SECTION << PAGE_SHIFT);
401 phys_addr += MIN_MEMORY_BLOCK_SIZE;
408 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
410 static int memory_probe_init(void)
412 return device_create_file(memory_subsys.dev_root, &dev_attr_probe);
415 static inline int memory_probe_init(void)
421 #ifdef CONFIG_MEMORY_FAILURE
423 * Support for offlining pages of memory
426 /* Soft offline a page */
428 store_soft_offline_page(struct device *dev,
429 struct device_attribute *attr,
430 const char *buf, size_t count)
434 if (!capable(CAP_SYS_ADMIN))
436 if (strict_strtoull(buf, 0, &pfn) < 0)
441 ret = soft_offline_page(pfn_to_page(pfn), 0);
442 return ret == 0 ? count : ret;
445 /* Forcibly offline a page, including killing processes. */
447 store_hard_offline_page(struct device *dev,
448 struct device_attribute *attr,
449 const char *buf, size_t count)
453 if (!capable(CAP_SYS_ADMIN))
455 if (strict_strtoull(buf, 0, &pfn) < 0)
458 ret = __memory_failure(pfn, 0, 0);
459 return ret ? ret : count;
462 static DEVICE_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
463 static DEVICE_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
465 static __init int memory_fail_init(void)
469 err = device_create_file(memory_subsys.dev_root,
470 &dev_attr_soft_offline_page);
472 err = device_create_file(memory_subsys.dev_root,
473 &dev_attr_hard_offline_page);
477 static inline int memory_fail_init(void)
484 * Note that phys_device is optional. It is here to allow for
485 * differentiation between which *physical* devices each
486 * section belongs to...
488 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
494 * A reference for the returned object is held and the reference for the
495 * hinted object is released.
497 struct memory_block *find_memory_block_hinted(struct mem_section *section,
498 struct memory_block *hint)
500 int block_id = base_memory_block_id(__section_nr(section));
501 struct device *hintdev = hint ? &hint->dev : NULL;
504 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
506 put_device(&hint->dev);
509 return container_of(dev, struct memory_block, dev);
513 * For now, we have a linear search to go find the appropriate
514 * memory_block corresponding to a particular phys_index. If
515 * this gets to be a real problem, we can always use a radix
516 * tree or something here.
518 * This could be made generic for all device subsystems.
520 struct memory_block *find_memory_block(struct mem_section *section)
522 return find_memory_block_hinted(section, NULL);
525 static int init_memory_block(struct memory_block **memory,
526 struct mem_section *section, unsigned long state)
528 struct memory_block *mem;
529 unsigned long start_pfn;
533 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
537 scn_nr = __section_nr(section);
538 mem->start_section_nr =
539 base_memory_block_id(scn_nr) * sections_per_block;
540 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
542 mem->section_count++;
543 mutex_init(&mem->state_mutex);
544 start_pfn = section_nr_to_pfn(mem->start_section_nr);
545 mem->phys_device = arch_get_memory_phys_device(start_pfn);
547 ret = register_memory(mem);
549 ret = mem_create_simple_file(mem, phys_index);
551 ret = mem_create_simple_file(mem, end_phys_index);
553 ret = mem_create_simple_file(mem, state);
555 ret = mem_create_simple_file(mem, phys_device);
557 ret = mem_create_simple_file(mem, removable);
563 static int add_memory_section(int nid, struct mem_section *section,
564 unsigned long state, enum mem_add_context context)
566 struct memory_block *mem;
569 mutex_lock(&mem_sysfs_mutex);
571 mem = find_memory_block(section);
573 mem->section_count++;
574 kobject_put(&mem->dev.kobj);
576 ret = init_memory_block(&mem, section, state);
579 if (context == HOTPLUG &&
580 mem->section_count == sections_per_block)
581 ret = register_mem_sect_under_node(mem, nid);
584 mutex_unlock(&mem_sysfs_mutex);
588 int remove_memory_block(unsigned long node_id, struct mem_section *section,
591 struct memory_block *mem;
593 mutex_lock(&mem_sysfs_mutex);
594 mem = find_memory_block(section);
595 unregister_mem_sect_under_nodes(mem, __section_nr(section));
597 mem->section_count--;
598 if (mem->section_count == 0) {
599 mem_remove_simple_file(mem, phys_index);
600 mem_remove_simple_file(mem, end_phys_index);
601 mem_remove_simple_file(mem, state);
602 mem_remove_simple_file(mem, phys_device);
603 mem_remove_simple_file(mem, removable);
604 unregister_memory(mem);
607 kobject_put(&mem->dev.kobj);
609 mutex_unlock(&mem_sysfs_mutex);
614 * need an interface for the VM to add new memory regions,
615 * but without onlining it.
617 int register_new_memory(int nid, struct mem_section *section)
619 return add_memory_section(nid, section, MEM_OFFLINE, HOTPLUG);
622 int unregister_memory_section(struct mem_section *section)
624 if (!present_section(section))
627 return remove_memory_block(0, section, 0);
631 * Initialize the sysfs support for memory devices...
633 int __init memory_dev_init(void)
638 unsigned long block_sz;
640 ret = subsys_system_register(&memory_subsys, NULL);
644 block_sz = get_memory_block_size();
645 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
648 * Create entries for memory sections that were found
649 * during boot and have been initialized
651 for (i = 0; i < NR_MEM_SECTIONS; i++) {
652 if (!present_section_nr(i))
654 err = add_memory_section(0, __nr_to_section(i), MEM_ONLINE,
660 err = memory_probe_init();
663 err = memory_fail_init();
666 err = block_size_init();
671 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);