Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 1 | /* |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2 | * Memory subsystem support |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 3 | * |
| 4 | * Written by Matt Tolentino <matthew.e.tolentino@intel.com> |
| 5 | * Dave Hansen <haveblue@us.ibm.com> |
| 6 | * |
| 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. |
| 11 | */ |
| 12 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 13 | #include <linux/module.h> |
| 14 | #include <linux/init.h> |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 15 | #include <linux/topology.h> |
Randy.Dunlap | c59ede7 | 2006-01-11 12:17:46 -0800 | [diff] [blame] | 16 | #include <linux/capability.h> |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 17 | #include <linux/device.h> |
| 18 | #include <linux/memory.h> |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 19 | #include <linux/memory_hotplug.h> |
| 20 | #include <linux/mm.h> |
Daniel Walker | da19cbc | 2008-02-04 23:35:47 -0800 | [diff] [blame] | 21 | #include <linux/mutex.h> |
Shaohua Li | 9f1b16a | 2008-10-18 20:27:12 -0700 | [diff] [blame] | 22 | #include <linux/stat.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 23 | #include <linux/slab.h> |
Shaohua Li | 9f1b16a | 2008-10-18 20:27:12 -0700 | [diff] [blame] | 24 | |
Arun Sharma | 60063497 | 2011-07-26 16:09:06 -0700 | [diff] [blame] | 25 | #include <linux/atomic.h> |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 26 | #include <asm/uaccess.h> |
| 27 | |
Nathan Fontenot | 2938ffb | 2010-10-19 12:45:24 -0500 | [diff] [blame] | 28 | static DEFINE_MUTEX(mem_sysfs_mutex); |
| 29 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 30 | #define MEMORY_CLASS_NAME "memory" |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 31 | |
Gu Zheng | 7315f0c | 2013-08-28 14:38:27 +0800 | [diff] [blame] | 32 | #define to_memory_block(dev) container_of(dev, struct memory_block, dev) |
| 33 | |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 34 | static int sections_per_block; |
| 35 | |
| 36 | static inline int base_memory_block_id(int section_nr) |
| 37 | { |
| 38 | return section_nr / sections_per_block; |
| 39 | } |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 40 | |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 41 | static int memory_subsys_online(struct device *dev); |
| 42 | static int memory_subsys_offline(struct device *dev); |
| 43 | |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 44 | static struct bus_type memory_subsys = { |
Kay Sievers | af5ca3f | 2007-12-20 02:09:39 +0100 | [diff] [blame] | 45 | .name = MEMORY_CLASS_NAME, |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 46 | .dev_name = MEMORY_CLASS_NAME, |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 47 | .online = memory_subsys_online, |
| 48 | .offline = memory_subsys_offline, |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 49 | }; |
| 50 | |
Alan Stern | e041c68 | 2006-03-27 01:16:30 -0800 | [diff] [blame] | 51 | static BLOCKING_NOTIFIER_HEAD(memory_chain); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 52 | |
Andy Whitcroft | 98a38eb | 2006-01-06 00:10:35 -0800 | [diff] [blame] | 53 | int register_memory_notifier(struct notifier_block *nb) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 54 | { |
Ioana Ciornei | 2aeebca | 2015-03-08 12:48:35 +0200 | [diff] [blame] | 55 | return blocking_notifier_chain_register(&memory_chain, nb); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 56 | } |
Hannes Hering | 3c82c30 | 2008-05-07 14:43:01 +0200 | [diff] [blame] | 57 | EXPORT_SYMBOL(register_memory_notifier); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 58 | |
Andy Whitcroft | 98a38eb | 2006-01-06 00:10:35 -0800 | [diff] [blame] | 59 | void unregister_memory_notifier(struct notifier_block *nb) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 60 | { |
Ioana Ciornei | 2aeebca | 2015-03-08 12:48:35 +0200 | [diff] [blame] | 61 | blocking_notifier_chain_unregister(&memory_chain, nb); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 62 | } |
Hannes Hering | 3c82c30 | 2008-05-07 14:43:01 +0200 | [diff] [blame] | 63 | EXPORT_SYMBOL(unregister_memory_notifier); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 64 | |
Robert Jennings | 925cc71 | 2009-12-17 14:44:38 +0000 | [diff] [blame] | 65 | static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain); |
| 66 | |
| 67 | int register_memory_isolate_notifier(struct notifier_block *nb) |
| 68 | { |
| 69 | return atomic_notifier_chain_register(&memory_isolate_chain, nb); |
| 70 | } |
| 71 | EXPORT_SYMBOL(register_memory_isolate_notifier); |
| 72 | |
| 73 | void unregister_memory_isolate_notifier(struct notifier_block *nb) |
| 74 | { |
| 75 | atomic_notifier_chain_unregister(&memory_isolate_chain, nb); |
| 76 | } |
| 77 | EXPORT_SYMBOL(unregister_memory_isolate_notifier); |
| 78 | |
Yasuaki Ishimatsu | fa7194e | 2012-12-11 16:00:44 -0800 | [diff] [blame] | 79 | static void memory_block_release(struct device *dev) |
| 80 | { |
Gu Zheng | 7315f0c | 2013-08-28 14:38:27 +0800 | [diff] [blame] | 81 | struct memory_block *mem = to_memory_block(dev); |
Yasuaki Ishimatsu | fa7194e | 2012-12-11 16:00:44 -0800 | [diff] [blame] | 82 | |
| 83 | kfree(mem); |
| 84 | } |
| 85 | |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 86 | unsigned long __weak memory_block_size_bytes(void) |
| 87 | { |
| 88 | return MIN_MEMORY_BLOCK_SIZE; |
| 89 | } |
| 90 | |
| 91 | static unsigned long get_memory_block_size(void) |
| 92 | { |
| 93 | unsigned long block_sz; |
| 94 | |
| 95 | block_sz = memory_block_size_bytes(); |
| 96 | |
| 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)) { |
| 99 | WARN_ON(1); |
| 100 | block_sz = MIN_MEMORY_BLOCK_SIZE; |
| 101 | } |
| 102 | |
| 103 | return block_sz; |
| 104 | } |
| 105 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 106 | /* |
| 107 | * use this as the physical section index that this memsection |
| 108 | * uses. |
| 109 | */ |
| 110 | |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 111 | static ssize_t show_mem_start_phys_index(struct device *dev, |
| 112 | struct device_attribute *attr, char *buf) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 113 | { |
Gu Zheng | 7315f0c | 2013-08-28 14:38:27 +0800 | [diff] [blame] | 114 | struct memory_block *mem = to_memory_block(dev); |
Nathan Fontenot | d336016 | 2011-01-20 10:44:29 -0600 | [diff] [blame] | 115 | unsigned long phys_index; |
| 116 | |
| 117 | phys_index = mem->start_section_nr / sections_per_block; |
| 118 | return sprintf(buf, "%08lx\n", phys_index); |
| 119 | } |
| 120 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 121 | /* |
Badari Pulavarty | 5c755e9 | 2008-07-23 21:28:19 -0700 | [diff] [blame] | 122 | * Show whether the section of memory is likely to be hot-removable |
| 123 | */ |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 124 | static ssize_t show_mem_removable(struct device *dev, |
| 125 | struct device_attribute *attr, char *buf) |
Badari Pulavarty | 5c755e9 | 2008-07-23 21:28:19 -0700 | [diff] [blame] | 126 | { |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 127 | unsigned long i, pfn; |
| 128 | int ret = 1; |
Gu Zheng | 7315f0c | 2013-08-28 14:38:27 +0800 | [diff] [blame] | 129 | struct memory_block *mem = to_memory_block(dev); |
Badari Pulavarty | 5c755e9 | 2008-07-23 21:28:19 -0700 | [diff] [blame] | 130 | |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 131 | for (i = 0; i < sections_per_block; i++) { |
Russ Anderson | 21ea9f5 | 2013-08-28 16:35:18 -0700 | [diff] [blame] | 132 | if (!present_section_nr(mem->start_section_nr + i)) |
| 133 | continue; |
Nathan Fontenot | d336016 | 2011-01-20 10:44:29 -0600 | [diff] [blame] | 134 | pfn = section_nr_to_pfn(mem->start_section_nr + i); |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 135 | ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION); |
| 136 | } |
| 137 | |
Badari Pulavarty | 5c755e9 | 2008-07-23 21:28:19 -0700 | [diff] [blame] | 138 | return sprintf(buf, "%d\n", ret); |
| 139 | } |
| 140 | |
| 141 | /* |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 142 | * online, offline, going offline, etc. |
| 143 | */ |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 144 | static ssize_t show_mem_state(struct device *dev, |
| 145 | struct device_attribute *attr, char *buf) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 146 | { |
Gu Zheng | 7315f0c | 2013-08-28 14:38:27 +0800 | [diff] [blame] | 147 | struct memory_block *mem = to_memory_block(dev); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 148 | ssize_t len = 0; |
| 149 | |
| 150 | /* |
| 151 | * We can probably put these states in a nice little array |
| 152 | * so that they're not open-coded |
| 153 | */ |
| 154 | switch (mem->state) { |
Ioana Ciornei | 3d3af6a | 2015-03-08 12:29:04 +0200 | [diff] [blame] | 155 | case MEM_ONLINE: |
| 156 | len = sprintf(buf, "online\n"); |
| 157 | break; |
| 158 | case MEM_OFFLINE: |
| 159 | len = sprintf(buf, "offline\n"); |
| 160 | break; |
| 161 | case MEM_GOING_OFFLINE: |
| 162 | len = sprintf(buf, "going-offline\n"); |
| 163 | break; |
| 164 | default: |
| 165 | len = sprintf(buf, "ERROR-UNKNOWN-%ld\n", |
| 166 | mem->state); |
| 167 | WARN_ON(1); |
| 168 | break; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 169 | } |
| 170 | |
| 171 | return len; |
| 172 | } |
| 173 | |
Yasunori Goto | 7b78d33 | 2007-10-21 16:41:36 -0700 | [diff] [blame] | 174 | int memory_notify(unsigned long val, void *v) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 175 | { |
Alan Stern | e041c68 | 2006-03-27 01:16:30 -0800 | [diff] [blame] | 176 | return blocking_notifier_call_chain(&memory_chain, val, v); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 177 | } |
| 178 | |
Robert Jennings | 925cc71 | 2009-12-17 14:44:38 +0000 | [diff] [blame] | 179 | int memory_isolate_notify(unsigned long val, void *v) |
| 180 | { |
| 181 | return atomic_notifier_call_chain(&memory_isolate_chain, val, v); |
| 182 | } |
| 183 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 184 | /* |
Mel Gorman | 2bbcb878 | 2011-10-17 16:38:20 +0200 | [diff] [blame] | 185 | * The probe routines leave the pages reserved, just as the bootmem code does. |
| 186 | * Make sure they're still that way. |
| 187 | */ |
Tang Chen | 6056d61 | 2013-04-29 15:08:40 -0700 | [diff] [blame] | 188 | static bool pages_correctly_reserved(unsigned long start_pfn) |
Mel Gorman | 2bbcb878 | 2011-10-17 16:38:20 +0200 | [diff] [blame] | 189 | { |
| 190 | int i, j; |
| 191 | struct page *page; |
| 192 | unsigned long pfn = start_pfn; |
| 193 | |
| 194 | /* |
| 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 |
| 198 | */ |
| 199 | for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) { |
| 200 | if (WARN_ON_ONCE(!pfn_valid(pfn))) |
| 201 | return false; |
| 202 | page = pfn_to_page(pfn); |
| 203 | |
| 204 | for (j = 0; j < PAGES_PER_SECTION; j++) { |
| 205 | if (PageReserved(page + j)) |
| 206 | continue; |
| 207 | |
| 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); |
| 211 | |
| 212 | return false; |
| 213 | } |
| 214 | } |
| 215 | |
| 216 | return true; |
| 217 | } |
| 218 | |
| 219 | /* |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 220 | * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is |
| 221 | * OK to have direct references to sparsemem variables in here. |
David Rientjes | 30467e0 | 2015-04-14 15:45:11 -0700 | [diff] [blame] | 222 | * Must already be protected by mem_hotplug_begin(). |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 223 | */ |
| 224 | static int |
Lai Jiangshan | 511c2ab | 2012-12-11 16:03:16 -0800 | [diff] [blame] | 225 | memory_block_action(unsigned long phys_index, unsigned long action, int online_type) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 226 | { |
Wen Congyang | a16cee1 | 2012-10-08 16:33:58 -0700 | [diff] [blame] | 227 | unsigned long start_pfn; |
Anton Blanchard | 5409d2c | 2011-05-11 17:25:14 +1000 | [diff] [blame] | 228 | unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block; |
Greg Kroah-Hartman | de0ed36 | 2011-10-18 14:00:57 -0700 | [diff] [blame] | 229 | struct page *first_page; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 230 | int ret; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 231 | |
Sheng Yong | 19c07d5 | 2015-04-14 15:44:54 -0700 | [diff] [blame] | 232 | start_pfn = section_nr_to_pfn(phys_index); |
Zhang Zhen | 71fbd55 | 2014-12-12 16:55:33 -0800 | [diff] [blame] | 233 | first_page = pfn_to_page(start_pfn); |
Greg Kroah-Hartman | de0ed36 | 2011-10-18 14:00:57 -0700 | [diff] [blame] | 234 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 235 | switch (action) { |
Ioana Ciornei | 3d3af6a | 2015-03-08 12:29:04 +0200 | [diff] [blame] | 236 | case MEM_ONLINE: |
| 237 | if (!pages_correctly_reserved(start_pfn)) |
| 238 | return -EBUSY; |
Mel Gorman | 2bbcb878 | 2011-10-17 16:38:20 +0200 | [diff] [blame] | 239 | |
Ioana Ciornei | 3d3af6a | 2015-03-08 12:29:04 +0200 | [diff] [blame] | 240 | ret = online_pages(start_pfn, nr_pages, online_type); |
| 241 | break; |
| 242 | case MEM_OFFLINE: |
| 243 | ret = offline_pages(start_pfn, nr_pages); |
| 244 | break; |
| 245 | default: |
| 246 | WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: " |
| 247 | "%ld\n", __func__, phys_index, action, action); |
| 248 | ret = -EINVAL; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 249 | } |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 250 | |
| 251 | return ret; |
| 252 | } |
| 253 | |
Vitaly Kuznetsov | 31bc385 | 2016-03-15 14:56:48 -0700 | [diff] [blame] | 254 | int memory_block_change_state(struct memory_block *mem, |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 255 | unsigned long to_state, unsigned long from_state_req) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 256 | { |
Greg Kroah-Hartman | de0ed36 | 2011-10-18 14:00:57 -0700 | [diff] [blame] | 257 | int ret = 0; |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 258 | |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 259 | if (mem->state != from_state_req) |
| 260 | return -EINVAL; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 261 | |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 262 | if (to_state == MEM_OFFLINE) |
| 263 | mem->state = MEM_GOING_OFFLINE; |
| 264 | |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 265 | ret = memory_block_action(mem->start_section_nr, to_state, |
| 266 | mem->online_type); |
| 267 | |
Rafael J. Wysocki | b2c064b | 2013-05-23 10:38:55 +0200 | [diff] [blame] | 268 | mem->state = ret ? from_state_req : to_state; |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 269 | |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 270 | return ret; |
| 271 | } |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 272 | |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 273 | /* The device lock serializes operations on memory_subsys_[online|offline] */ |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 274 | static int memory_subsys_online(struct device *dev) |
| 275 | { |
Gu Zheng | 7315f0c | 2013-08-28 14:38:27 +0800 | [diff] [blame] | 276 | struct memory_block *mem = to_memory_block(dev); |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 277 | int ret; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 278 | |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 279 | if (mem->state == MEM_ONLINE) |
| 280 | return 0; |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 281 | |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 282 | /* |
| 283 | * If we are called from store_mem_state(), online_type will be |
| 284 | * set >= 0 Otherwise we were called from the device online |
| 285 | * attribute and need to set the online_type. |
| 286 | */ |
| 287 | if (mem->online_type < 0) |
Tang Chen | 4f7c6b4 | 2014-08-06 16:05:13 -0700 | [diff] [blame] | 288 | mem->online_type = MMOP_ONLINE_KEEP; |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 289 | |
David Rientjes | 30467e0 | 2015-04-14 15:45:11 -0700 | [diff] [blame] | 290 | /* Already under protection of mem_hotplug_begin() */ |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 291 | ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE); |
| 292 | |
| 293 | /* clear online_type */ |
| 294 | mem->online_type = -1; |
| 295 | |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 296 | return ret; |
| 297 | } |
| 298 | |
| 299 | static int memory_subsys_offline(struct device *dev) |
| 300 | { |
Gu Zheng | 7315f0c | 2013-08-28 14:38:27 +0800 | [diff] [blame] | 301 | struct memory_block *mem = to_memory_block(dev); |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 302 | |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 303 | if (mem->state == MEM_OFFLINE) |
| 304 | return 0; |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 305 | |
Seth Jennings | 26bbe7e | 2015-12-11 13:40:57 -0800 | [diff] [blame] | 306 | /* Can't offline block with non-present sections */ |
| 307 | if (mem->section_count != sections_per_block) |
| 308 | return -EINVAL; |
| 309 | |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 310 | return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE); |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 311 | } |
| 312 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 313 | static ssize_t |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 314 | store_mem_state(struct device *dev, |
| 315 | struct device_attribute *attr, const char *buf, size_t count) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 316 | { |
Gu Zheng | 7315f0c | 2013-08-28 14:38:27 +0800 | [diff] [blame] | 317 | struct memory_block *mem = to_memory_block(dev); |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 318 | int ret, online_type; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 319 | |
Rafael J. Wysocki | 5e33bc4 | 2013-08-28 21:41:01 +0200 | [diff] [blame] | 320 | ret = lock_device_hotplug_sysfs(); |
| 321 | if (ret) |
| 322 | return ret; |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 323 | |
Tang Chen | 1f6a6cc | 2014-08-06 16:05:11 -0700 | [diff] [blame] | 324 | if (sysfs_streq(buf, "online_kernel")) |
Tang Chen | 4f7c6b4 | 2014-08-06 16:05:13 -0700 | [diff] [blame] | 325 | online_type = MMOP_ONLINE_KERNEL; |
Tang Chen | 1f6a6cc | 2014-08-06 16:05:11 -0700 | [diff] [blame] | 326 | else if (sysfs_streq(buf, "online_movable")) |
Tang Chen | 4f7c6b4 | 2014-08-06 16:05:13 -0700 | [diff] [blame] | 327 | online_type = MMOP_ONLINE_MOVABLE; |
Tang Chen | 1f6a6cc | 2014-08-06 16:05:11 -0700 | [diff] [blame] | 328 | else if (sysfs_streq(buf, "online")) |
Tang Chen | 4f7c6b4 | 2014-08-06 16:05:13 -0700 | [diff] [blame] | 329 | online_type = MMOP_ONLINE_KEEP; |
Tang Chen | 1f6a6cc | 2014-08-06 16:05:11 -0700 | [diff] [blame] | 330 | else if (sysfs_streq(buf, "offline")) |
Tang Chen | 4f7c6b4 | 2014-08-06 16:05:13 -0700 | [diff] [blame] | 331 | online_type = MMOP_OFFLINE; |
Yasuaki Ishimatsu | a37f863 | 2013-10-11 15:36:25 +0900 | [diff] [blame] | 332 | else { |
| 333 | ret = -EINVAL; |
| 334 | goto err; |
| 335 | } |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 336 | |
David Rientjes | 30467e0 | 2015-04-14 15:45:11 -0700 | [diff] [blame] | 337 | /* |
| 338 | * Memory hotplug needs to hold mem_hotplug_begin() for probe to find |
| 339 | * the correct memory block to online before doing device_online(dev), |
| 340 | * which will take dev->mutex. Take the lock early to prevent an |
| 341 | * inversion, memory_subsys_online() callbacks will be implemented by |
| 342 | * assuming it's already protected. |
| 343 | */ |
| 344 | mem_hotplug_begin(); |
| 345 | |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 346 | switch (online_type) { |
Tang Chen | 4f7c6b4 | 2014-08-06 16:05:13 -0700 | [diff] [blame] | 347 | case MMOP_ONLINE_KERNEL: |
| 348 | case MMOP_ONLINE_MOVABLE: |
| 349 | case MMOP_ONLINE_KEEP: |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 350 | mem->online_type = online_type; |
| 351 | ret = device_online(&mem->dev); |
| 352 | break; |
Tang Chen | 4f7c6b4 | 2014-08-06 16:05:13 -0700 | [diff] [blame] | 353 | case MMOP_OFFLINE: |
Seth Jennings | fa2be40 | 2013-08-20 16:05:05 -0500 | [diff] [blame] | 354 | ret = device_offline(&mem->dev); |
| 355 | break; |
| 356 | default: |
| 357 | ret = -EINVAL; /* should never happen */ |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 358 | } |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 359 | |
David Rientjes | 30467e0 | 2015-04-14 15:45:11 -0700 | [diff] [blame] | 360 | mem_hotplug_done(); |
Yasuaki Ishimatsu | a37f863 | 2013-10-11 15:36:25 +0900 | [diff] [blame] | 361 | err: |
Rafael J. Wysocki | 4960e05 | 2013-05-08 14:18:37 +0200 | [diff] [blame] | 362 | unlock_device_hotplug(); |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 363 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 364 | if (ret) |
| 365 | return ret; |
| 366 | return count; |
| 367 | } |
| 368 | |
| 369 | /* |
| 370 | * phys_device is a bad name for this. What I really want |
| 371 | * is a way to differentiate between memory ranges that |
| 372 | * are part of physical devices that constitute |
| 373 | * a complete removable unit or fru. |
| 374 | * i.e. do these ranges belong to the same physical device, |
| 375 | * s.t. if I offline all of these sections I can then |
| 376 | * remove the physical device? |
| 377 | */ |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 378 | static ssize_t show_phys_device(struct device *dev, |
| 379 | struct device_attribute *attr, char *buf) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 380 | { |
Gu Zheng | 7315f0c | 2013-08-28 14:38:27 +0800 | [diff] [blame] | 381 | struct memory_block *mem = to_memory_block(dev); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 382 | return sprintf(buf, "%d\n", mem->phys_device); |
| 383 | } |
| 384 | |
Zhang Zhen | ed2f240 | 2014-10-09 15:26:31 -0700 | [diff] [blame] | 385 | #ifdef CONFIG_MEMORY_HOTREMOVE |
| 386 | static ssize_t show_valid_zones(struct device *dev, |
| 387 | struct device_attribute *attr, char *buf) |
| 388 | { |
| 389 | struct memory_block *mem = to_memory_block(dev); |
| 390 | unsigned long start_pfn, end_pfn; |
| 391 | unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block; |
| 392 | struct page *first_page; |
| 393 | struct zone *zone; |
| 394 | |
| 395 | start_pfn = section_nr_to_pfn(mem->start_section_nr); |
| 396 | end_pfn = start_pfn + nr_pages; |
| 397 | first_page = pfn_to_page(start_pfn); |
| 398 | |
| 399 | /* The block contains more than one zone can not be offlined. */ |
| 400 | if (!test_pages_in_a_zone(start_pfn, end_pfn)) |
| 401 | return sprintf(buf, "none\n"); |
| 402 | |
| 403 | zone = page_zone(first_page); |
| 404 | |
| 405 | if (zone_idx(zone) == ZONE_MOVABLE - 1) { |
| 406 | /*The mem block is the last memoryblock of this zone.*/ |
| 407 | if (end_pfn == zone_end_pfn(zone)) |
| 408 | return sprintf(buf, "%s %s\n", |
| 409 | zone->name, (zone + 1)->name); |
| 410 | } |
| 411 | |
| 412 | if (zone_idx(zone) == ZONE_MOVABLE) { |
| 413 | /*The mem block is the first memoryblock of ZONE_MOVABLE.*/ |
| 414 | if (start_pfn == zone->zone_start_pfn) |
| 415 | return sprintf(buf, "%s %s\n", |
| 416 | zone->name, (zone - 1)->name); |
| 417 | } |
| 418 | |
| 419 | return sprintf(buf, "%s\n", zone->name); |
| 420 | } |
| 421 | static DEVICE_ATTR(valid_zones, 0444, show_valid_zones, NULL); |
| 422 | #endif |
| 423 | |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 424 | static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL); |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 425 | static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state); |
| 426 | static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL); |
| 427 | static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 428 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 429 | /* |
| 430 | * Block size attribute stuff |
| 431 | */ |
| 432 | static ssize_t |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 433 | print_block_size(struct device *dev, struct device_attribute *attr, |
Andi Kleen | 8564a6c | 2010-01-05 12:48:06 +0100 | [diff] [blame] | 434 | char *buf) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 435 | { |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 436 | return sprintf(buf, "%lx\n", get_memory_block_size()); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 437 | } |
| 438 | |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 439 | static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 440 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 441 | /* |
Vitaly Kuznetsov | 31bc385 | 2016-03-15 14:56:48 -0700 | [diff] [blame] | 442 | * Memory auto online policy. |
| 443 | */ |
| 444 | |
| 445 | static ssize_t |
| 446 | show_auto_online_blocks(struct device *dev, struct device_attribute *attr, |
| 447 | char *buf) |
| 448 | { |
| 449 | if (memhp_auto_online) |
| 450 | return sprintf(buf, "online\n"); |
| 451 | else |
| 452 | return sprintf(buf, "offline\n"); |
| 453 | } |
| 454 | |
| 455 | static ssize_t |
| 456 | store_auto_online_blocks(struct device *dev, struct device_attribute *attr, |
| 457 | const char *buf, size_t count) |
| 458 | { |
| 459 | if (sysfs_streq(buf, "online")) |
| 460 | memhp_auto_online = true; |
| 461 | else if (sysfs_streq(buf, "offline")) |
| 462 | memhp_auto_online = false; |
| 463 | else |
| 464 | return -EINVAL; |
| 465 | |
| 466 | return count; |
| 467 | } |
| 468 | |
| 469 | static DEVICE_ATTR(auto_online_blocks, 0644, show_auto_online_blocks, |
| 470 | store_auto_online_blocks); |
| 471 | |
| 472 | /* |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 473 | * Some architectures will have custom drivers to do this, and |
| 474 | * will not need to do it from userspace. The fake hot-add code |
| 475 | * as well as ppc64 will do all of their discovery in userspace |
| 476 | * and will require this interface. |
| 477 | */ |
| 478 | #ifdef CONFIG_ARCH_MEMORY_PROBE |
| 479 | static ssize_t |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 480 | memory_probe_store(struct device *dev, struct device_attribute *attr, |
Andi Kleen | 28812fe | 2010-01-05 12:48:07 +0100 | [diff] [blame] | 481 | const char *buf, size_t count) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 482 | { |
| 483 | u64 phys_addr; |
John Allen | cb5490a | 2016-01-14 15:22:16 -0800 | [diff] [blame] | 484 | int nid, ret; |
Anton Blanchard | 61b94fe | 2011-09-15 06:26:15 +1000 | [diff] [blame] | 485 | unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 486 | |
Zhang Zhen | b69deb2 | 2014-08-06 16:06:06 -0700 | [diff] [blame] | 487 | ret = kstrtoull(buf, 0, &phys_addr); |
| 488 | if (ret) |
| 489 | return ret; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 490 | |
Anton Blanchard | 61b94fe | 2011-09-15 06:26:15 +1000 | [diff] [blame] | 491 | if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1)) |
| 492 | return -EINVAL; |
| 493 | |
John Allen | cb5490a | 2016-01-14 15:22:16 -0800 | [diff] [blame] | 494 | nid = memory_add_physaddr_to_nid(phys_addr); |
| 495 | ret = add_memory(nid, phys_addr, |
| 496 | MIN_MEMORY_BLOCK_SIZE * sections_per_block); |
Nathan Fontenot | 6add7cd | 2011-01-31 10:55:23 -0600 | [diff] [blame] | 497 | |
John Allen | cb5490a | 2016-01-14 15:22:16 -0800 | [diff] [blame] | 498 | if (ret) |
| 499 | goto out; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 500 | |
Nikanth Karthikesan | 9f0af69 | 2011-03-24 11:46:18 +0530 | [diff] [blame] | 501 | ret = count; |
| 502 | out: |
| 503 | return ret; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 504 | } |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 505 | |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 506 | static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 507 | #endif |
| 508 | |
Andi Kleen | facb601 | 2009-12-16 12:20:00 +0100 | [diff] [blame] | 509 | #ifdef CONFIG_MEMORY_FAILURE |
| 510 | /* |
| 511 | * Support for offlining pages of memory |
| 512 | */ |
| 513 | |
| 514 | /* Soft offline a page */ |
| 515 | static ssize_t |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 516 | store_soft_offline_page(struct device *dev, |
| 517 | struct device_attribute *attr, |
Andi Kleen | 28812fe | 2010-01-05 12:48:07 +0100 | [diff] [blame] | 518 | const char *buf, size_t count) |
Andi Kleen | facb601 | 2009-12-16 12:20:00 +0100 | [diff] [blame] | 519 | { |
| 520 | int ret; |
| 521 | u64 pfn; |
| 522 | if (!capable(CAP_SYS_ADMIN)) |
| 523 | return -EPERM; |
Jingoo Han | 34da5e6 | 2013-07-26 13:10:22 +0900 | [diff] [blame] | 524 | if (kstrtoull(buf, 0, &pfn) < 0) |
Andi Kleen | facb601 | 2009-12-16 12:20:00 +0100 | [diff] [blame] | 525 | return -EINVAL; |
| 526 | pfn >>= PAGE_SHIFT; |
| 527 | if (!pfn_valid(pfn)) |
| 528 | return -ENXIO; |
| 529 | ret = soft_offline_page(pfn_to_page(pfn), 0); |
| 530 | return ret == 0 ? count : ret; |
| 531 | } |
| 532 | |
| 533 | /* Forcibly offline a page, including killing processes. */ |
| 534 | static ssize_t |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 535 | store_hard_offline_page(struct device *dev, |
| 536 | struct device_attribute *attr, |
Andi Kleen | 28812fe | 2010-01-05 12:48:07 +0100 | [diff] [blame] | 537 | const char *buf, size_t count) |
Andi Kleen | facb601 | 2009-12-16 12:20:00 +0100 | [diff] [blame] | 538 | { |
| 539 | int ret; |
| 540 | u64 pfn; |
| 541 | if (!capable(CAP_SYS_ADMIN)) |
| 542 | return -EPERM; |
Jingoo Han | 34da5e6 | 2013-07-26 13:10:22 +0900 | [diff] [blame] | 543 | if (kstrtoull(buf, 0, &pfn) < 0) |
Andi Kleen | facb601 | 2009-12-16 12:20:00 +0100 | [diff] [blame] | 544 | return -EINVAL; |
| 545 | pfn >>= PAGE_SHIFT; |
Tony Luck | cd42f4a | 2011-12-15 10:48:12 -0800 | [diff] [blame] | 546 | ret = memory_failure(pfn, 0, 0); |
Andi Kleen | facb601 | 2009-12-16 12:20:00 +0100 | [diff] [blame] | 547 | return ret ? ret : count; |
| 548 | } |
| 549 | |
Felipe Balbi | 74fef7a | 2013-02-18 21:09:03 +0200 | [diff] [blame] | 550 | static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page); |
| 551 | static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page); |
Andi Kleen | facb601 | 2009-12-16 12:20:00 +0100 | [diff] [blame] | 552 | #endif |
| 553 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 554 | /* |
| 555 | * Note that phys_device is optional. It is here to allow for |
| 556 | * differentiation between which *physical* devices each |
| 557 | * section belongs to... |
| 558 | */ |
Heiko Carstens | bc32df0 | 2010-03-15 00:35:03 -0400 | [diff] [blame] | 559 | int __weak arch_get_memory_phys_device(unsigned long start_pfn) |
| 560 | { |
| 561 | return 0; |
| 562 | } |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 563 | |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 564 | /* |
| 565 | * A reference for the returned object is held and the reference for the |
| 566 | * hinted object is released. |
| 567 | */ |
Robin Holt | 9838303 | 2010-09-29 14:00:55 -0500 | [diff] [blame] | 568 | struct memory_block *find_memory_block_hinted(struct mem_section *section, |
| 569 | struct memory_block *hint) |
| 570 | { |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 571 | int block_id = base_memory_block_id(__section_nr(section)); |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 572 | struct device *hintdev = hint ? &hint->dev : NULL; |
| 573 | struct device *dev; |
Robin Holt | 9838303 | 2010-09-29 14:00:55 -0500 | [diff] [blame] | 574 | |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 575 | dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev); |
| 576 | if (hint) |
| 577 | put_device(&hint->dev); |
| 578 | if (!dev) |
Robin Holt | 9838303 | 2010-09-29 14:00:55 -0500 | [diff] [blame] | 579 | return NULL; |
Gu Zheng | 7315f0c | 2013-08-28 14:38:27 +0800 | [diff] [blame] | 580 | return to_memory_block(dev); |
Robin Holt | 9838303 | 2010-09-29 14:00:55 -0500 | [diff] [blame] | 581 | } |
| 582 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 583 | /* |
| 584 | * For now, we have a linear search to go find the appropriate |
| 585 | * memory_block corresponding to a particular phys_index. If |
| 586 | * this gets to be a real problem, we can always use a radix |
| 587 | * tree or something here. |
| 588 | * |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 589 | * This could be made generic for all device subsystems. |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 590 | */ |
Gary Hade | c04fc58 | 2009-01-06 14:39:14 -0800 | [diff] [blame] | 591 | struct memory_block *find_memory_block(struct mem_section *section) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 592 | { |
Robin Holt | 9838303 | 2010-09-29 14:00:55 -0500 | [diff] [blame] | 593 | return find_memory_block_hinted(section, NULL); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 594 | } |
| 595 | |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 596 | static struct attribute *memory_memblk_attrs[] = { |
| 597 | &dev_attr_phys_index.attr, |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 598 | &dev_attr_state.attr, |
| 599 | &dev_attr_phys_device.attr, |
| 600 | &dev_attr_removable.attr, |
Zhang Zhen | ed2f240 | 2014-10-09 15:26:31 -0700 | [diff] [blame] | 601 | #ifdef CONFIG_MEMORY_HOTREMOVE |
| 602 | &dev_attr_valid_zones.attr, |
| 603 | #endif |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 604 | NULL |
| 605 | }; |
| 606 | |
| 607 | static struct attribute_group memory_memblk_attr_group = { |
| 608 | .attrs = memory_memblk_attrs, |
| 609 | }; |
| 610 | |
| 611 | static const struct attribute_group *memory_memblk_attr_groups[] = { |
| 612 | &memory_memblk_attr_group, |
| 613 | NULL, |
| 614 | }; |
| 615 | |
| 616 | /* |
| 617 | * register_memory - Setup a sysfs device for a memory block |
| 618 | */ |
| 619 | static |
| 620 | int register_memory(struct memory_block *memory) |
| 621 | { |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 622 | memory->dev.bus = &memory_subsys; |
| 623 | memory->dev.id = memory->start_section_nr / sections_per_block; |
| 624 | memory->dev.release = memory_block_release; |
| 625 | memory->dev.groups = memory_memblk_attr_groups; |
Linus Torvalds | f991fae | 2013-07-03 14:35:40 -0700 | [diff] [blame] | 626 | memory->dev.offline = memory->state == MEM_OFFLINE; |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 627 | |
Seth Jennings | 879f1be | 2013-08-20 12:12:58 -0500 | [diff] [blame] | 628 | return device_register(&memory->dev); |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 629 | } |
| 630 | |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 631 | static int init_memory_block(struct memory_block **memory, |
| 632 | struct mem_section *section, unsigned long state) |
Nathan Fontenot | e4619c8 | 2010-10-19 12:44:20 -0500 | [diff] [blame] | 633 | { |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 634 | struct memory_block *mem; |
Nathan Fontenot | e4619c8 | 2010-10-19 12:44:20 -0500 | [diff] [blame] | 635 | unsigned long start_pfn; |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 636 | int scn_nr; |
Nathan Fontenot | e4619c8 | 2010-10-19 12:44:20 -0500 | [diff] [blame] | 637 | int ret = 0; |
| 638 | |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 639 | mem = kzalloc(sizeof(*mem), GFP_KERNEL); |
Nathan Fontenot | e4619c8 | 2010-10-19 12:44:20 -0500 | [diff] [blame] | 640 | if (!mem) |
| 641 | return -ENOMEM; |
| 642 | |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 643 | scn_nr = __section_nr(section); |
Nathan Fontenot | d336016 | 2011-01-20 10:44:29 -0600 | [diff] [blame] | 644 | mem->start_section_nr = |
| 645 | base_memory_block_id(scn_nr) * sections_per_block; |
| 646 | mem->end_section_nr = mem->start_section_nr + sections_per_block - 1; |
Nathan Fontenot | e4619c8 | 2010-10-19 12:44:20 -0500 | [diff] [blame] | 647 | mem->state = state; |
Nathan Fontenot | d336016 | 2011-01-20 10:44:29 -0600 | [diff] [blame] | 648 | start_pfn = section_nr_to_pfn(mem->start_section_nr); |
Nathan Fontenot | e4619c8 | 2010-10-19 12:44:20 -0500 | [diff] [blame] | 649 | mem->phys_device = arch_get_memory_phys_device(start_pfn); |
| 650 | |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 651 | ret = register_memory(mem); |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 652 | |
| 653 | *memory = mem; |
| 654 | return ret; |
| 655 | } |
| 656 | |
Seth Jennings | cb5e39b | 2013-08-20 12:13:03 -0500 | [diff] [blame] | 657 | static int add_memory_block(int base_section_nr) |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 658 | { |
Seth Jennings | cb5e39b | 2013-08-20 12:13:03 -0500 | [diff] [blame] | 659 | struct memory_block *mem; |
| 660 | int i, ret, section_count = 0, section_nr; |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 661 | |
Seth Jennings | cb5e39b | 2013-08-20 12:13:03 -0500 | [diff] [blame] | 662 | for (i = base_section_nr; |
| 663 | (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS; |
| 664 | i++) { |
| 665 | if (!present_section_nr(i)) |
| 666 | continue; |
| 667 | if (section_count == 0) |
| 668 | section_nr = i; |
| 669 | section_count++; |
Yinghai Lu | 321bf4e | 2012-01-30 13:57:12 -0800 | [diff] [blame] | 670 | } |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 671 | |
Seth Jennings | cb5e39b | 2013-08-20 12:13:03 -0500 | [diff] [blame] | 672 | if (section_count == 0) |
| 673 | return 0; |
| 674 | ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE); |
| 675 | if (ret) |
| 676 | return ret; |
| 677 | mem->section_count = section_count; |
| 678 | return 0; |
Nathan Fontenot | e4619c8 | 2010-10-19 12:44:20 -0500 | [diff] [blame] | 679 | } |
| 680 | |
Dan Williams | 260ae3f | 2016-01-15 16:56:17 -0800 | [diff] [blame] | 681 | static bool is_zone_device_section(struct mem_section *ms) |
| 682 | { |
| 683 | struct page *page; |
| 684 | |
| 685 | page = sparse_decode_mem_map(ms->section_mem_map, __section_nr(ms)); |
| 686 | return is_zone_device_page(page); |
| 687 | } |
Seth Jennings | cb5e39b | 2013-08-20 12:13:03 -0500 | [diff] [blame] | 688 | |
David Rientjes | 4edd7ce | 2013-04-29 15:08:22 -0700 | [diff] [blame] | 689 | /* |
| 690 | * need an interface for the VM to add new memory regions, |
| 691 | * but without onlining it. |
| 692 | */ |
| 693 | int register_new_memory(int nid, struct mem_section *section) |
| 694 | { |
Seth Jennings | d7f8053 | 2013-08-20 12:13:00 -0500 | [diff] [blame] | 695 | int ret = 0; |
| 696 | struct memory_block *mem; |
Seth Jennings | b1eaef3 | 2013-08-20 12:12:57 -0500 | [diff] [blame] | 697 | |
Dan Williams | 260ae3f | 2016-01-15 16:56:17 -0800 | [diff] [blame] | 698 | if (is_zone_device_section(section)) |
| 699 | return 0; |
| 700 | |
Seth Jennings | b1eaef3 | 2013-08-20 12:12:57 -0500 | [diff] [blame] | 701 | mutex_lock(&mem_sysfs_mutex); |
Seth Jennings | b1eaef3 | 2013-08-20 12:12:57 -0500 | [diff] [blame] | 702 | |
Seth Jennings | d7f8053 | 2013-08-20 12:13:00 -0500 | [diff] [blame] | 703 | mem = find_memory_block(section); |
| 704 | if (mem) { |
| 705 | mem->section_count++; |
| 706 | put_device(&mem->dev); |
| 707 | } else { |
| 708 | ret = init_memory_block(&mem, section, MEM_OFFLINE); |
| 709 | if (ret) |
| 710 | goto out; |
Seth Jennings | 56c6b5d | 2016-01-14 15:20:21 -0800 | [diff] [blame] | 711 | mem->section_count++; |
Seth Jennings | d7f8053 | 2013-08-20 12:13:00 -0500 | [diff] [blame] | 712 | } |
| 713 | |
| 714 | if (mem->section_count == sections_per_block) |
| 715 | ret = register_mem_sect_under_node(mem, nid); |
| 716 | out: |
| 717 | mutex_unlock(&mem_sysfs_mutex); |
Seth Jennings | b1eaef3 | 2013-08-20 12:12:57 -0500 | [diff] [blame] | 718 | return ret; |
David Rientjes | 4edd7ce | 2013-04-29 15:08:22 -0700 | [diff] [blame] | 719 | } |
| 720 | |
| 721 | #ifdef CONFIG_MEMORY_HOTREMOVE |
| 722 | static void |
| 723 | unregister_memory(struct memory_block *memory) |
| 724 | { |
| 725 | BUG_ON(memory->dev.bus != &memory_subsys); |
| 726 | |
| 727 | /* drop the ref. we got in remove_memory_block() */ |
Seth Jennings | df2b717 | 2013-08-20 12:12:59 -0500 | [diff] [blame] | 728 | put_device(&memory->dev); |
David Rientjes | 4edd7ce | 2013-04-29 15:08:22 -0700 | [diff] [blame] | 729 | device_unregister(&memory->dev); |
| 730 | } |
| 731 | |
Seth Jennings | cc292b0 | 2016-01-14 15:20:24 -0800 | [diff] [blame] | 732 | static int remove_memory_section(unsigned long node_id, |
David Rientjes | 4edd7ce | 2013-04-29 15:08:22 -0700 | [diff] [blame] | 733 | struct mem_section *section, int phys_device) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 734 | { |
| 735 | struct memory_block *mem; |
| 736 | |
Dan Williams | 260ae3f | 2016-01-15 16:56:17 -0800 | [diff] [blame] | 737 | if (is_zone_device_section(section)) |
| 738 | return 0; |
| 739 | |
Nathan Fontenot | 2938ffb | 2010-10-19 12:45:24 -0500 | [diff] [blame] | 740 | mutex_lock(&mem_sysfs_mutex); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 741 | mem = find_memory_block(section); |
Nathan Fontenot | d336016 | 2011-01-20 10:44:29 -0600 | [diff] [blame] | 742 | unregister_mem_sect_under_nodes(mem, __section_nr(section)); |
Nathan Fontenot | 0768121 | 2010-10-19 12:46:19 -0500 | [diff] [blame] | 743 | |
| 744 | mem->section_count--; |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 745 | if (mem->section_count == 0) |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 746 | unregister_memory(mem); |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 747 | else |
Seth Jennings | df2b717 | 2013-08-20 12:12:59 -0500 | [diff] [blame] | 748 | put_device(&mem->dev); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 749 | |
Nathan Fontenot | 2938ffb | 2010-10-19 12:45:24 -0500 | [diff] [blame] | 750 | mutex_unlock(&mem_sysfs_mutex); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 751 | return 0; |
| 752 | } |
| 753 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 754 | int unregister_memory_section(struct mem_section *section) |
| 755 | { |
Andy Whitcroft | 540557b | 2007-10-16 01:24:11 -0700 | [diff] [blame] | 756 | if (!present_section(section)) |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 757 | return -EINVAL; |
| 758 | |
Seth Jennings | cc292b0 | 2016-01-14 15:20:24 -0800 | [diff] [blame] | 759 | return remove_memory_section(0, section, 0); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 760 | } |
David Rientjes | 4edd7ce | 2013-04-29 15:08:22 -0700 | [diff] [blame] | 761 | #endif /* CONFIG_MEMORY_HOTREMOVE */ |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 762 | |
Yasuaki Ishimatsu | 6677e3e | 2013-02-22 16:32:52 -0800 | [diff] [blame] | 763 | /* return true if the memory block is offlined, otherwise, return false */ |
| 764 | bool is_memblock_offlined(struct memory_block *mem) |
| 765 | { |
| 766 | return mem->state == MEM_OFFLINE; |
| 767 | } |
| 768 | |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 769 | static struct attribute *memory_root_attrs[] = { |
| 770 | #ifdef CONFIG_ARCH_MEMORY_PROBE |
| 771 | &dev_attr_probe.attr, |
| 772 | #endif |
| 773 | |
| 774 | #ifdef CONFIG_MEMORY_FAILURE |
| 775 | &dev_attr_soft_offline_page.attr, |
| 776 | &dev_attr_hard_offline_page.attr, |
| 777 | #endif |
| 778 | |
| 779 | &dev_attr_block_size_bytes.attr, |
Vitaly Kuznetsov | 31bc385 | 2016-03-15 14:56:48 -0700 | [diff] [blame] | 780 | &dev_attr_auto_online_blocks.attr, |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 781 | NULL |
| 782 | }; |
| 783 | |
| 784 | static struct attribute_group memory_root_attr_group = { |
| 785 | .attrs = memory_root_attrs, |
| 786 | }; |
| 787 | |
| 788 | static const struct attribute_group *memory_root_attr_groups[] = { |
| 789 | &memory_root_attr_group, |
| 790 | NULL, |
| 791 | }; |
| 792 | |
Wen Congyang | e90bdb7 | 2012-10-08 16:34:01 -0700 | [diff] [blame] | 793 | /* |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 794 | * Initialize the sysfs support for memory devices... |
| 795 | */ |
| 796 | int __init memory_dev_init(void) |
| 797 | { |
| 798 | unsigned int i; |
| 799 | int ret; |
Andrew Morton | 28ec24e | 2006-12-06 20:37:29 -0800 | [diff] [blame] | 800 | int err; |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 801 | unsigned long block_sz; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 802 | |
Nathan Fontenot | 96b2c0f | 2013-06-04 14:42:28 -0500 | [diff] [blame] | 803 | ret = subsys_system_register(&memory_subsys, memory_root_attr_groups); |
Andrew Morton | 28ec24e | 2006-12-06 20:37:29 -0800 | [diff] [blame] | 804 | if (ret) |
| 805 | goto out; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 806 | |
Nathan Fontenot | 0c2c99b | 2011-01-20 10:43:34 -0600 | [diff] [blame] | 807 | block_sz = get_memory_block_size(); |
| 808 | sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE; |
| 809 | |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 810 | /* |
| 811 | * Create entries for memory sections that were found |
| 812 | * during boot and have been initialized |
| 813 | */ |
Seth Jennings | b1eaef3 | 2013-08-20 12:12:57 -0500 | [diff] [blame] | 814 | mutex_lock(&mem_sysfs_mutex); |
Seth Jennings | cb5e39b | 2013-08-20 12:13:03 -0500 | [diff] [blame] | 815 | for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) { |
| 816 | err = add_memory_block(i); |
Andrew Morton | 28ec24e | 2006-12-06 20:37:29 -0800 | [diff] [blame] | 817 | if (!ret) |
| 818 | ret = err; |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 819 | } |
Seth Jennings | b1eaef3 | 2013-08-20 12:12:57 -0500 | [diff] [blame] | 820 | mutex_unlock(&mem_sysfs_mutex); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 821 | |
Andrew Morton | 28ec24e | 2006-12-06 20:37:29 -0800 | [diff] [blame] | 822 | out: |
| 823 | if (ret) |
Harvey Harrison | 2b3a302 | 2008-03-04 16:41:05 -0800 | [diff] [blame] | 824 | printk(KERN_ERR "%s() failed: %d\n", __func__, ret); |
Dave Hansen | 3947be1 | 2005-10-29 18:16:54 -0700 | [diff] [blame] | 825 | return ret; |
| 826 | } |