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);
121 static ssize_t show_mem_end_phys_index(struct device *dev,
122 struct device_attribute *attr, char *buf)
124 struct memory_block *mem = to_memory_block(dev);
125 unsigned long phys_index;
127 phys_index = mem->end_section_nr / sections_per_block;
128 return sprintf(buf, "%08lx\n", phys_index);
132 * Show whether the section of memory is likely to be hot-removable
134 static ssize_t show_mem_removable(struct device *dev,
135 struct device_attribute *attr, char *buf)
137 unsigned long i, pfn;
139 struct memory_block *mem = to_memory_block(dev);
141 for (i = 0; i < sections_per_block; i++) {
142 if (!present_section_nr(mem->start_section_nr + i))
144 pfn = section_nr_to_pfn(mem->start_section_nr + i);
145 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
148 return sprintf(buf, "%d\n", ret);
152 * online, offline, going offline, etc.
154 static ssize_t show_mem_state(struct device *dev,
155 struct device_attribute *attr, char *buf)
157 struct memory_block *mem = to_memory_block(dev);
161 * We can probably put these states in a nice little array
162 * so that they're not open-coded
164 switch (mem->state) {
166 len = sprintf(buf, "online\n");
169 len = sprintf(buf, "offline\n");
171 case MEM_GOING_OFFLINE:
172 len = sprintf(buf, "going-offline\n");
175 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
184 int memory_notify(unsigned long val, void *v)
186 return blocking_notifier_call_chain(&memory_chain, val, v);
189 int memory_isolate_notify(unsigned long val, void *v)
191 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
195 * The probe routines leave the pages reserved, just as the bootmem code does.
196 * Make sure they're still that way.
198 static bool pages_correctly_reserved(unsigned long start_pfn)
202 unsigned long pfn = start_pfn;
205 * memmap between sections is not contiguous except with
206 * SPARSEMEM_VMEMMAP. We lookup the page once per section
207 * and assume memmap is contiguous within each section
209 for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
210 if (WARN_ON_ONCE(!pfn_valid(pfn)))
212 page = pfn_to_page(pfn);
214 for (j = 0; j < PAGES_PER_SECTION; j++) {
215 if (PageReserved(page + j))
218 printk(KERN_WARNING "section number %ld page number %d "
219 "not reserved, was it already online?\n",
220 pfn_to_section_nr(pfn), j);
230 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
231 * OK to have direct references to sparsemem variables in here.
234 memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
236 unsigned long start_pfn;
237 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
238 struct page *first_page;
241 first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
242 start_pfn = page_to_pfn(first_page);
246 if (!pages_correctly_reserved(start_pfn))
249 ret = online_pages(start_pfn, nr_pages, online_type);
252 ret = offline_pages(start_pfn, nr_pages);
255 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
256 "%ld\n", __func__, phys_index, action, action);
263 static int memory_block_change_state(struct memory_block *mem,
264 unsigned long to_state, unsigned long from_state_req)
268 if (mem->state != from_state_req)
271 if (to_state == MEM_OFFLINE)
272 mem->state = MEM_GOING_OFFLINE;
274 ret = memory_block_action(mem->start_section_nr, to_state,
277 mem->state = ret ? from_state_req : to_state;
282 /* The device lock serializes operations on memory_subsys_[online|offline] */
283 static int memory_subsys_online(struct device *dev)
285 struct memory_block *mem = to_memory_block(dev);
288 if (mem->state == MEM_ONLINE)
292 * If we are called from store_mem_state(), online_type will be
293 * set >= 0 Otherwise we were called from the device online
294 * attribute and need to set the online_type.
296 if (mem->online_type < 0)
297 mem->online_type = ONLINE_KEEP;
299 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
301 /* clear online_type */
302 mem->online_type = -1;
307 static int memory_subsys_offline(struct device *dev)
309 struct memory_block *mem = to_memory_block(dev);
311 if (mem->state == MEM_OFFLINE)
314 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
318 store_mem_state(struct device *dev,
319 struct device_attribute *attr, const char *buf, size_t count)
321 struct memory_block *mem = to_memory_block(dev);
322 int ret, online_type;
324 ret = lock_device_hotplug_sysfs();
328 if (!strncmp(buf, "online_kernel", min_t(int, count, 13)))
329 online_type = ONLINE_KERNEL;
330 else if (!strncmp(buf, "online_movable", min_t(int, count, 14)))
331 online_type = ONLINE_MOVABLE;
332 else if (!strncmp(buf, "online", min_t(int, count, 6)))
333 online_type = ONLINE_KEEP;
334 else if (!strncmp(buf, "offline", min_t(int, count, 7)))
341 switch (online_type) {
346 * mem->online_type is not protected so there can be a
347 * race here. However, when racing online, the first
348 * will succeed and the second will just return as the
349 * block will already be online. The online type
350 * could be either one, but that is expected.
352 mem->online_type = online_type;
353 ret = device_online(&mem->dev);
356 ret = device_offline(&mem->dev);
359 ret = -EINVAL; /* should never happen */
363 unlock_device_hotplug();
371 * phys_device is a bad name for this. What I really want
372 * is a way to differentiate between memory ranges that
373 * are part of physical devices that constitute
374 * a complete removable unit or fru.
375 * i.e. do these ranges belong to the same physical device,
376 * s.t. if I offline all of these sections I can then
377 * remove the physical device?
379 static ssize_t show_phys_device(struct device *dev,
380 struct device_attribute *attr, char *buf)
382 struct memory_block *mem = to_memory_block(dev);
383 return sprintf(buf, "%d\n", mem->phys_device);
386 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
387 static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
388 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
389 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
390 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
393 * Block size attribute stuff
396 print_block_size(struct device *dev, struct device_attribute *attr,
399 return sprintf(buf, "%lx\n", get_memory_block_size());
402 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
405 * Some architectures will have custom drivers to do this, and
406 * will not need to do it from userspace. The fake hot-add code
407 * as well as ppc64 will do all of their discovery in userspace
408 * and will require this interface.
410 #ifdef CONFIG_ARCH_MEMORY_PROBE
412 memory_probe_store(struct device *dev, struct device_attribute *attr,
413 const char *buf, size_t count)
418 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
420 phys_addr = simple_strtoull(buf, NULL, 0);
422 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
425 for (i = 0; i < sections_per_block; i++) {
426 nid = memory_add_physaddr_to_nid(phys_addr);
427 ret = add_memory(nid, phys_addr,
428 PAGES_PER_SECTION << PAGE_SHIFT);
432 phys_addr += MIN_MEMORY_BLOCK_SIZE;
440 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
443 #ifdef CONFIG_MEMORY_FAILURE
445 * Support for offlining pages of memory
448 /* Soft offline a page */
450 store_soft_offline_page(struct device *dev,
451 struct device_attribute *attr,
452 const char *buf, size_t count)
456 if (!capable(CAP_SYS_ADMIN))
458 if (kstrtoull(buf, 0, &pfn) < 0)
463 ret = soft_offline_page(pfn_to_page(pfn), 0);
464 return ret == 0 ? count : ret;
467 /* Forcibly offline a page, including killing processes. */
469 store_hard_offline_page(struct device *dev,
470 struct device_attribute *attr,
471 const char *buf, size_t count)
475 if (!capable(CAP_SYS_ADMIN))
477 if (kstrtoull(buf, 0, &pfn) < 0)
480 ret = memory_failure(pfn, 0, 0);
481 return ret ? ret : count;
484 static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
485 static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
489 * Note that phys_device is optional. It is here to allow for
490 * differentiation between which *physical* devices each
491 * section belongs to...
493 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
499 * A reference for the returned object is held and the reference for the
500 * hinted object is released.
502 struct memory_block *find_memory_block_hinted(struct mem_section *section,
503 struct memory_block *hint)
505 int block_id = base_memory_block_id(__section_nr(section));
506 struct device *hintdev = hint ? &hint->dev : NULL;
509 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
511 put_device(&hint->dev);
514 return to_memory_block(dev);
518 * For now, we have a linear search to go find the appropriate
519 * memory_block corresponding to a particular phys_index. If
520 * this gets to be a real problem, we can always use a radix
521 * tree or something here.
523 * This could be made generic for all device subsystems.
525 struct memory_block *find_memory_block(struct mem_section *section)
527 return find_memory_block_hinted(section, NULL);
530 static struct attribute *memory_memblk_attrs[] = {
531 &dev_attr_phys_index.attr,
532 &dev_attr_end_phys_index.attr,
533 &dev_attr_state.attr,
534 &dev_attr_phys_device.attr,
535 &dev_attr_removable.attr,
539 static struct attribute_group memory_memblk_attr_group = {
540 .attrs = memory_memblk_attrs,
543 static const struct attribute_group *memory_memblk_attr_groups[] = {
544 &memory_memblk_attr_group,
549 * register_memory - Setup a sysfs device for a memory block
552 int register_memory(struct memory_block *memory)
554 memory->dev.bus = &memory_subsys;
555 memory->dev.id = memory->start_section_nr / sections_per_block;
556 memory->dev.release = memory_block_release;
557 memory->dev.groups = memory_memblk_attr_groups;
558 memory->dev.offline = memory->state == MEM_OFFLINE;
560 return device_register(&memory->dev);
563 static int init_memory_block(struct memory_block **memory,
564 struct mem_section *section, unsigned long state)
566 struct memory_block *mem;
567 unsigned long start_pfn;
571 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
575 scn_nr = __section_nr(section);
576 mem->start_section_nr =
577 base_memory_block_id(scn_nr) * sections_per_block;
578 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
580 mem->section_count++;
581 start_pfn = section_nr_to_pfn(mem->start_section_nr);
582 mem->phys_device = arch_get_memory_phys_device(start_pfn);
584 ret = register_memory(mem);
590 static int add_memory_block(int base_section_nr)
592 struct memory_block *mem;
593 int i, ret, section_count = 0, section_nr;
595 for (i = base_section_nr;
596 (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
598 if (!present_section_nr(i))
600 if (section_count == 0)
605 if (section_count == 0)
607 ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
610 mem->section_count = section_count;
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)
622 struct memory_block *mem;
624 mutex_lock(&mem_sysfs_mutex);
626 mem = find_memory_block(section);
628 mem->section_count++;
629 put_device(&mem->dev);
631 ret = init_memory_block(&mem, section, MEM_OFFLINE);
636 if (mem->section_count == sections_per_block)
637 ret = register_mem_sect_under_node(mem, nid);
639 mutex_unlock(&mem_sysfs_mutex);
643 #ifdef CONFIG_MEMORY_HOTREMOVE
645 unregister_memory(struct memory_block *memory)
647 BUG_ON(memory->dev.bus != &memory_subsys);
649 /* drop the ref. we got in remove_memory_block() */
650 put_device(&memory->dev);
651 device_unregister(&memory->dev);
654 static int remove_memory_block(unsigned long node_id,
655 struct mem_section *section, int phys_device)
657 struct memory_block *mem;
659 mutex_lock(&mem_sysfs_mutex);
660 mem = find_memory_block(section);
661 unregister_mem_sect_under_nodes(mem, __section_nr(section));
663 mem->section_count--;
664 if (mem->section_count == 0)
665 unregister_memory(mem);
667 put_device(&mem->dev);
669 mutex_unlock(&mem_sysfs_mutex);
673 int unregister_memory_section(struct mem_section *section)
675 if (!present_section(section))
678 return remove_memory_block(0, section, 0);
680 #endif /* CONFIG_MEMORY_HOTREMOVE */
682 /* return true if the memory block is offlined, otherwise, return false */
683 bool is_memblock_offlined(struct memory_block *mem)
685 return mem->state == MEM_OFFLINE;
688 static struct attribute *memory_root_attrs[] = {
689 #ifdef CONFIG_ARCH_MEMORY_PROBE
690 &dev_attr_probe.attr,
693 #ifdef CONFIG_MEMORY_FAILURE
694 &dev_attr_soft_offline_page.attr,
695 &dev_attr_hard_offline_page.attr,
698 &dev_attr_block_size_bytes.attr,
702 static struct attribute_group memory_root_attr_group = {
703 .attrs = memory_root_attrs,
706 static const struct attribute_group *memory_root_attr_groups[] = {
707 &memory_root_attr_group,
712 * Initialize the sysfs support for memory devices...
714 int __init memory_dev_init(void)
719 unsigned long block_sz;
721 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
725 block_sz = get_memory_block_size();
726 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
729 * Create entries for memory sections that were found
730 * during boot and have been initialized
732 mutex_lock(&mem_sysfs_mutex);
733 for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
734 err = add_memory_block(i);
738 mutex_unlock(&mem_sysfs_mutex);
742 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);