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/memory_hotplug.h>
21 #include <linux/mutex.h>
22 #include <linux/stat.h>
23 #include <linux/slab.h>
25 #include <linux/atomic.h>
26 #include <asm/uaccess.h>
28 static DEFINE_MUTEX(mem_sysfs_mutex);
30 #define MEMORY_CLASS_NAME "memory"
32 #define to_memory_block(dev) container_of(dev, struct memory_block, dev)
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 int memory_subsys_online(struct device *dev);
42 static int memory_subsys_offline(struct device *dev);
44 static struct bus_type memory_subsys = {
45 .name = MEMORY_CLASS_NAME,
46 .dev_name = MEMORY_CLASS_NAME,
47 .online = memory_subsys_online,
48 .offline = memory_subsys_offline,
51 static BLOCKING_NOTIFIER_HEAD(memory_chain);
53 int register_memory_notifier(struct notifier_block *nb)
55 return blocking_notifier_chain_register(&memory_chain, nb);
57 EXPORT_SYMBOL(register_memory_notifier);
59 void unregister_memory_notifier(struct notifier_block *nb)
61 blocking_notifier_chain_unregister(&memory_chain, nb);
63 EXPORT_SYMBOL(unregister_memory_notifier);
65 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
67 int register_memory_isolate_notifier(struct notifier_block *nb)
69 return atomic_notifier_chain_register(&memory_isolate_chain, nb);
71 EXPORT_SYMBOL(register_memory_isolate_notifier);
73 void unregister_memory_isolate_notifier(struct notifier_block *nb)
75 atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
77 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
79 static void memory_block_release(struct device *dev)
81 struct memory_block *mem = to_memory_block(dev);
86 unsigned long __weak memory_block_size_bytes(void)
88 return MIN_MEMORY_BLOCK_SIZE;
91 static unsigned long get_memory_block_size(void)
93 unsigned long block_sz;
95 block_sz = memory_block_size_bytes();
97 /* Validate blk_sz is a power of 2 and not less than section size */
98 if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
100 block_sz = MIN_MEMORY_BLOCK_SIZE;
107 * use this as the physical section index that this memsection
111 static ssize_t show_mem_start_phys_index(struct device *dev,
112 struct device_attribute *attr, char *buf)
114 struct memory_block *mem = to_memory_block(dev);
115 unsigned long phys_index;
117 phys_index = mem->start_section_nr / sections_per_block;
118 return sprintf(buf, "%08lx\n", phys_index);
122 * Show whether the section of memory is likely to be hot-removable
124 static ssize_t show_mem_removable(struct device *dev,
125 struct device_attribute *attr, char *buf)
127 unsigned long i, pfn;
129 struct memory_block *mem = to_memory_block(dev);
131 for (i = 0; i < sections_per_block; i++) {
132 if (!present_section_nr(mem->start_section_nr + i))
134 pfn = section_nr_to_pfn(mem->start_section_nr + i);
135 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
138 return sprintf(buf, "%d\n", ret);
142 * online, offline, going offline, etc.
144 static ssize_t show_mem_state(struct device *dev,
145 struct device_attribute *attr, char *buf)
147 struct memory_block *mem = to_memory_block(dev);
151 * We can probably put these states in a nice little array
152 * so that they're not open-coded
154 switch (mem->state) {
156 len = sprintf(buf, "online\n");
159 len = sprintf(buf, "offline\n");
161 case MEM_GOING_OFFLINE:
162 len = sprintf(buf, "going-offline\n");
165 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
174 int memory_notify(unsigned long val, void *v)
176 return blocking_notifier_call_chain(&memory_chain, val, v);
179 int memory_isolate_notify(unsigned long val, void *v)
181 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
185 * The probe routines leave the pages reserved, just as the bootmem code does.
186 * Make sure they're still that way.
188 static bool pages_correctly_reserved(unsigned long start_pfn)
192 unsigned long pfn = start_pfn;
195 * memmap between sections is not contiguous except with
196 * SPARSEMEM_VMEMMAP. We lookup the page once per section
197 * and assume memmap is contiguous within each section
199 for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
200 if (WARN_ON_ONCE(!pfn_valid(pfn)))
202 page = pfn_to_page(pfn);
204 for (j = 0; j < PAGES_PER_SECTION; j++) {
205 if (PageReserved(page + j))
208 printk(KERN_WARNING "section number %ld page number %d "
209 "not reserved, was it already online?\n",
210 pfn_to_section_nr(pfn), j);
220 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
221 * OK to have direct references to sparsemem variables in here.
224 memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
226 unsigned long start_pfn;
227 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
228 struct page *first_page;
231 first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
232 start_pfn = page_to_pfn(first_page);
236 if (!pages_correctly_reserved(start_pfn))
239 ret = online_pages(start_pfn, nr_pages, online_type);
242 ret = offline_pages(start_pfn, nr_pages);
245 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
246 "%ld\n", __func__, phys_index, action, action);
253 static int memory_block_change_state(struct memory_block *mem,
254 unsigned long to_state, unsigned long from_state_req)
258 if (mem->state != from_state_req)
261 if (to_state == MEM_OFFLINE)
262 mem->state = MEM_GOING_OFFLINE;
264 ret = memory_block_action(mem->start_section_nr, to_state,
267 mem->state = ret ? from_state_req : to_state;
272 /* The device lock serializes operations on memory_subsys_[online|offline] */
273 static int memory_subsys_online(struct device *dev)
275 struct memory_block *mem = to_memory_block(dev);
278 if (mem->state == MEM_ONLINE)
282 * If we are called from store_mem_state(), online_type will be
283 * set >= 0 Otherwise we were called from the device online
284 * attribute and need to set the online_type.
286 if (mem->online_type < 0)
287 mem->online_type = MMOP_ONLINE_KEEP;
289 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
291 /* clear online_type */
292 mem->online_type = -1;
297 static int memory_subsys_offline(struct device *dev)
299 struct memory_block *mem = to_memory_block(dev);
301 if (mem->state == MEM_OFFLINE)
304 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
308 store_mem_state(struct device *dev,
309 struct device_attribute *attr, const char *buf, size_t count)
311 struct memory_block *mem = to_memory_block(dev);
312 int ret, online_type;
314 ret = lock_device_hotplug_sysfs();
318 if (sysfs_streq(buf, "online_kernel"))
319 online_type = MMOP_ONLINE_KERNEL;
320 else if (sysfs_streq(buf, "online_movable"))
321 online_type = MMOP_ONLINE_MOVABLE;
322 else if (sysfs_streq(buf, "online"))
323 online_type = MMOP_ONLINE_KEEP;
324 else if (sysfs_streq(buf, "offline"))
325 online_type = MMOP_OFFLINE;
331 switch (online_type) {
332 case MMOP_ONLINE_KERNEL:
333 case MMOP_ONLINE_MOVABLE:
334 case MMOP_ONLINE_KEEP:
336 * mem->online_type is not protected so there can be a
337 * race here. However, when racing online, the first
338 * will succeed and the second will just return as the
339 * block will already be online. The online type
340 * could be either one, but that is expected.
342 mem->online_type = online_type;
343 ret = device_online(&mem->dev);
346 ret = device_offline(&mem->dev);
349 ret = -EINVAL; /* should never happen */
353 unlock_device_hotplug();
361 * phys_device is a bad name for this. What I really want
362 * is a way to differentiate between memory ranges that
363 * are part of physical devices that constitute
364 * a complete removable unit or fru.
365 * i.e. do these ranges belong to the same physical device,
366 * s.t. if I offline all of these sections I can then
367 * remove the physical device?
369 static ssize_t show_phys_device(struct device *dev,
370 struct device_attribute *attr, char *buf)
372 struct memory_block *mem = to_memory_block(dev);
373 return sprintf(buf, "%d\n", mem->phys_device);
376 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
377 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
378 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
379 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
382 * Block size attribute stuff
385 print_block_size(struct device *dev, struct device_attribute *attr,
388 return sprintf(buf, "%lx\n", get_memory_block_size());
391 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
394 * Some architectures will have custom drivers to do this, and
395 * will not need to do it from userspace. The fake hot-add code
396 * as well as ppc64 will do all of their discovery in userspace
397 * and will require this interface.
399 #ifdef CONFIG_ARCH_MEMORY_PROBE
401 memory_probe_store(struct device *dev, struct device_attribute *attr,
402 const char *buf, size_t count)
407 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
409 ret = kstrtoull(buf, 0, &phys_addr);
413 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
416 for (i = 0; i < sections_per_block; i++) {
417 nid = memory_add_physaddr_to_nid(phys_addr);
418 ret = add_memory(nid, phys_addr,
419 PAGES_PER_SECTION << PAGE_SHIFT);
423 phys_addr += MIN_MEMORY_BLOCK_SIZE;
431 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
434 #ifdef CONFIG_MEMORY_FAILURE
436 * Support for offlining pages of memory
439 /* Soft offline a page */
441 store_soft_offline_page(struct device *dev,
442 struct device_attribute *attr,
443 const char *buf, size_t count)
447 if (!capable(CAP_SYS_ADMIN))
449 if (kstrtoull(buf, 0, &pfn) < 0)
454 ret = soft_offline_page(pfn_to_page(pfn), 0);
455 return ret == 0 ? count : ret;
458 /* Forcibly offline a page, including killing processes. */
460 store_hard_offline_page(struct device *dev,
461 struct device_attribute *attr,
462 const char *buf, size_t count)
466 if (!capable(CAP_SYS_ADMIN))
468 if (kstrtoull(buf, 0, &pfn) < 0)
471 ret = memory_failure(pfn, 0, 0);
472 return ret ? ret : count;
475 static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
476 static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
480 * Note that phys_device is optional. It is here to allow for
481 * differentiation between which *physical* devices each
482 * section belongs to...
484 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
490 * A reference for the returned object is held and the reference for the
491 * hinted object is released.
493 struct memory_block *find_memory_block_hinted(struct mem_section *section,
494 struct memory_block *hint)
496 int block_id = base_memory_block_id(__section_nr(section));
497 struct device *hintdev = hint ? &hint->dev : NULL;
500 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
502 put_device(&hint->dev);
505 return to_memory_block(dev);
509 * For now, we have a linear search to go find the appropriate
510 * memory_block corresponding to a particular phys_index. If
511 * this gets to be a real problem, we can always use a radix
512 * tree or something here.
514 * This could be made generic for all device subsystems.
516 struct memory_block *find_memory_block(struct mem_section *section)
518 return find_memory_block_hinted(section, NULL);
521 static struct attribute *memory_memblk_attrs[] = {
522 &dev_attr_phys_index.attr,
523 &dev_attr_state.attr,
524 &dev_attr_phys_device.attr,
525 &dev_attr_removable.attr,
529 static struct attribute_group memory_memblk_attr_group = {
530 .attrs = memory_memblk_attrs,
533 static const struct attribute_group *memory_memblk_attr_groups[] = {
534 &memory_memblk_attr_group,
539 * register_memory - Setup a sysfs device for a memory block
542 int register_memory(struct memory_block *memory)
544 memory->dev.bus = &memory_subsys;
545 memory->dev.id = memory->start_section_nr / sections_per_block;
546 memory->dev.release = memory_block_release;
547 memory->dev.groups = memory_memblk_attr_groups;
548 memory->dev.offline = memory->state == MEM_OFFLINE;
550 return device_register(&memory->dev);
553 static int init_memory_block(struct memory_block **memory,
554 struct mem_section *section, unsigned long state)
556 struct memory_block *mem;
557 unsigned long start_pfn;
561 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
565 scn_nr = __section_nr(section);
566 mem->start_section_nr =
567 base_memory_block_id(scn_nr) * sections_per_block;
568 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
570 mem->section_count++;
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);
580 static int add_memory_block(int base_section_nr)
582 struct memory_block *mem;
583 int i, ret, section_count = 0, section_nr;
585 for (i = base_section_nr;
586 (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
588 if (!present_section_nr(i))
590 if (section_count == 0)
595 if (section_count == 0)
597 ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
600 mem->section_count = section_count;
606 * need an interface for the VM to add new memory regions,
607 * but without onlining it.
609 int register_new_memory(int nid, struct mem_section *section)
612 struct memory_block *mem;
614 mutex_lock(&mem_sysfs_mutex);
616 mem = find_memory_block(section);
618 mem->section_count++;
619 put_device(&mem->dev);
621 ret = init_memory_block(&mem, section, MEM_OFFLINE);
626 if (mem->section_count == sections_per_block)
627 ret = register_mem_sect_under_node(mem, nid);
629 mutex_unlock(&mem_sysfs_mutex);
633 #ifdef CONFIG_MEMORY_HOTREMOVE
635 unregister_memory(struct memory_block *memory)
637 BUG_ON(memory->dev.bus != &memory_subsys);
639 /* drop the ref. we got in remove_memory_block() */
640 put_device(&memory->dev);
641 device_unregister(&memory->dev);
644 static int remove_memory_block(unsigned long node_id,
645 struct mem_section *section, int phys_device)
647 struct memory_block *mem;
649 mutex_lock(&mem_sysfs_mutex);
650 mem = find_memory_block(section);
651 unregister_mem_sect_under_nodes(mem, __section_nr(section));
653 mem->section_count--;
654 if (mem->section_count == 0)
655 unregister_memory(mem);
657 put_device(&mem->dev);
659 mutex_unlock(&mem_sysfs_mutex);
663 int unregister_memory_section(struct mem_section *section)
665 if (!present_section(section))
668 return remove_memory_block(0, section, 0);
670 #endif /* CONFIG_MEMORY_HOTREMOVE */
672 /* return true if the memory block is offlined, otherwise, return false */
673 bool is_memblock_offlined(struct memory_block *mem)
675 return mem->state == MEM_OFFLINE;
678 static struct attribute *memory_root_attrs[] = {
679 #ifdef CONFIG_ARCH_MEMORY_PROBE
680 &dev_attr_probe.attr,
683 #ifdef CONFIG_MEMORY_FAILURE
684 &dev_attr_soft_offline_page.attr,
685 &dev_attr_hard_offline_page.attr,
688 &dev_attr_block_size_bytes.attr,
692 static struct attribute_group memory_root_attr_group = {
693 .attrs = memory_root_attrs,
696 static const struct attribute_group *memory_root_attr_groups[] = {
697 &memory_root_attr_group,
702 * Initialize the sysfs support for memory devices...
704 int __init memory_dev_init(void)
709 unsigned long block_sz;
711 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
715 block_sz = get_memory_block_size();
716 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
719 * Create entries for memory sections that were found
720 * during boot and have been initialized
722 mutex_lock(&mem_sysfs_mutex);
723 for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
724 err = add_memory_block(i);
728 mutex_unlock(&mem_sysfs_mutex);
732 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);