blob: 2804aed3f416aea878efde04937937e238ff4e5e [file] [log] [blame]
Dave Hansen3947be12005-10-29 18:16:54 -07001/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002 * Memory subsystem support
Dave Hansen3947be12005-10-29 18:16:54 -07003 *
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 Hansen3947be12005-10-29 18:16:54 -070013#include <linux/module.h>
14#include <linux/init.h>
Dave Hansen3947be12005-10-29 18:16:54 -070015#include <linux/topology.h>
Randy.Dunlapc59ede72006-01-11 12:17:46 -080016#include <linux/capability.h>
Dave Hansen3947be12005-10-29 18:16:54 -070017#include <linux/device.h>
18#include <linux/memory.h>
Dave Hansen3947be12005-10-29 18:16:54 -070019#include <linux/memory_hotplug.h>
20#include <linux/mm.h>
Daniel Walkerda19cbc2008-02-04 23:35:47 -080021#include <linux/mutex.h>
Shaohua Li9f1b16a2008-10-18 20:27:12 -070022#include <linux/stat.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090023#include <linux/slab.h>
Shaohua Li9f1b16a2008-10-18 20:27:12 -070024
Arun Sharma600634972011-07-26 16:09:06 -070025#include <linux/atomic.h>
Dave Hansen3947be12005-10-29 18:16:54 -070026#include <asm/uaccess.h>
27
Nathan Fontenot2938ffb2010-10-19 12:45:24 -050028static DEFINE_MUTEX(mem_sysfs_mutex);
29
Dave Hansen3947be12005-10-29 18:16:54 -070030#define MEMORY_CLASS_NAME "memory"
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -060031
Gu Zheng7315f0c2013-08-28 14:38:27 +080032#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
33
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -060034static int sections_per_block;
35
36static inline int base_memory_block_id(int section_nr)
37{
38 return section_nr / sections_per_block;
39}
Dave Hansen3947be12005-10-29 18:16:54 -070040
Rafael J. Wysocki4960e052013-05-08 14:18:37 +020041static int memory_subsys_online(struct device *dev);
42static int memory_subsys_offline(struct device *dev);
43
Kay Sievers10fbcf42011-12-21 14:48:43 -080044static struct bus_type memory_subsys = {
Kay Sieversaf5ca3f2007-12-20 02:09:39 +010045 .name = MEMORY_CLASS_NAME,
Kay Sievers10fbcf42011-12-21 14:48:43 -080046 .dev_name = MEMORY_CLASS_NAME,
Rafael J. Wysocki4960e052013-05-08 14:18:37 +020047 .online = memory_subsys_online,
48 .offline = memory_subsys_offline,
Dave Hansen3947be12005-10-29 18:16:54 -070049};
50
Alan Sterne041c682006-03-27 01:16:30 -080051static BLOCKING_NOTIFIER_HEAD(memory_chain);
Dave Hansen3947be12005-10-29 18:16:54 -070052
Andy Whitcroft98a38eb2006-01-06 00:10:35 -080053int register_memory_notifier(struct notifier_block *nb)
Dave Hansen3947be12005-10-29 18:16:54 -070054{
Ioana Ciornei2aeebca2015-03-08 12:48:35 +020055 return blocking_notifier_chain_register(&memory_chain, nb);
Dave Hansen3947be12005-10-29 18:16:54 -070056}
Hannes Hering3c82c302008-05-07 14:43:01 +020057EXPORT_SYMBOL(register_memory_notifier);
Dave Hansen3947be12005-10-29 18:16:54 -070058
Andy Whitcroft98a38eb2006-01-06 00:10:35 -080059void unregister_memory_notifier(struct notifier_block *nb)
Dave Hansen3947be12005-10-29 18:16:54 -070060{
Ioana Ciornei2aeebca2015-03-08 12:48:35 +020061 blocking_notifier_chain_unregister(&memory_chain, nb);
Dave Hansen3947be12005-10-29 18:16:54 -070062}
Hannes Hering3c82c302008-05-07 14:43:01 +020063EXPORT_SYMBOL(unregister_memory_notifier);
Dave Hansen3947be12005-10-29 18:16:54 -070064
Robert Jennings925cc712009-12-17 14:44:38 +000065static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
66
67int register_memory_isolate_notifier(struct notifier_block *nb)
68{
69 return atomic_notifier_chain_register(&memory_isolate_chain, nb);
70}
71EXPORT_SYMBOL(register_memory_isolate_notifier);
72
73void unregister_memory_isolate_notifier(struct notifier_block *nb)
74{
75 atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
76}
77EXPORT_SYMBOL(unregister_memory_isolate_notifier);
78
Yasuaki Ishimatsufa7194e2012-12-11 16:00:44 -080079static void memory_block_release(struct device *dev)
80{
Gu Zheng7315f0c2013-08-28 14:38:27 +080081 struct memory_block *mem = to_memory_block(dev);
Yasuaki Ishimatsufa7194e2012-12-11 16:00:44 -080082
83 kfree(mem);
84}
85
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -060086unsigned long __weak memory_block_size_bytes(void)
87{
88 return MIN_MEMORY_BLOCK_SIZE;
89}
90
91static 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 Hansen3947be12005-10-29 18:16:54 -0700106/*
107 * use this as the physical section index that this memsection
108 * uses.
109 */
110
Kay Sievers10fbcf42011-12-21 14:48:43 -0800111static ssize_t show_mem_start_phys_index(struct device *dev,
112 struct device_attribute *attr, char *buf)
Dave Hansen3947be12005-10-29 18:16:54 -0700113{
Gu Zheng7315f0c2013-08-28 14:38:27 +0800114 struct memory_block *mem = to_memory_block(dev);
Nathan Fontenotd3360162011-01-20 10:44:29 -0600115 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 Hansen3947be12005-10-29 18:16:54 -0700121/*
Badari Pulavarty5c755e92008-07-23 21:28:19 -0700122 * Show whether the section of memory is likely to be hot-removable
123 */
Kay Sievers10fbcf42011-12-21 14:48:43 -0800124static ssize_t show_mem_removable(struct device *dev,
125 struct device_attribute *attr, char *buf)
Badari Pulavarty5c755e92008-07-23 21:28:19 -0700126{
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600127 unsigned long i, pfn;
128 int ret = 1;
Gu Zheng7315f0c2013-08-28 14:38:27 +0800129 struct memory_block *mem = to_memory_block(dev);
Badari Pulavarty5c755e92008-07-23 21:28:19 -0700130
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600131 for (i = 0; i < sections_per_block; i++) {
Russ Anderson21ea9f52013-08-28 16:35:18 -0700132 if (!present_section_nr(mem->start_section_nr + i))
133 continue;
Nathan Fontenotd3360162011-01-20 10:44:29 -0600134 pfn = section_nr_to_pfn(mem->start_section_nr + i);
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600135 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
136 }
137
Badari Pulavarty5c755e92008-07-23 21:28:19 -0700138 return sprintf(buf, "%d\n", ret);
139}
140
141/*
Dave Hansen3947be12005-10-29 18:16:54 -0700142 * online, offline, going offline, etc.
143 */
Kay Sievers10fbcf42011-12-21 14:48:43 -0800144static ssize_t show_mem_state(struct device *dev,
145 struct device_attribute *attr, char *buf)
Dave Hansen3947be12005-10-29 18:16:54 -0700146{
Gu Zheng7315f0c2013-08-28 14:38:27 +0800147 struct memory_block *mem = to_memory_block(dev);
Dave Hansen3947be12005-10-29 18:16:54 -0700148 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 Ciornei3d3af6a2015-03-08 12:29:04 +0200155 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 Hansen3947be12005-10-29 18:16:54 -0700169 }
170
171 return len;
172}
173
Yasunori Goto7b78d332007-10-21 16:41:36 -0700174int memory_notify(unsigned long val, void *v)
Dave Hansen3947be12005-10-29 18:16:54 -0700175{
Alan Sterne041c682006-03-27 01:16:30 -0800176 return blocking_notifier_call_chain(&memory_chain, val, v);
Dave Hansen3947be12005-10-29 18:16:54 -0700177}
178
Robert Jennings925cc712009-12-17 14:44:38 +0000179int memory_isolate_notify(unsigned long val, void *v)
180{
181 return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
182}
183
Dave Hansen3947be12005-10-29 18:16:54 -0700184/*
Mel Gorman2bbcb8782011-10-17 16:38:20 +0200185 * The probe routines leave the pages reserved, just as the bootmem code does.
186 * Make sure they're still that way.
187 */
Tang Chen6056d612013-04-29 15:08:40 -0700188static bool pages_correctly_reserved(unsigned long start_pfn)
Mel Gorman2bbcb8782011-10-17 16:38:20 +0200189{
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 Hansen3947be12005-10-29 18:16:54 -0700220 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
221 * OK to have direct references to sparsemem variables in here.
David Rientjes30467e02015-04-14 15:45:11 -0700222 * Must already be protected by mem_hotplug_begin().
Dave Hansen3947be12005-10-29 18:16:54 -0700223 */
224static int
Lai Jiangshan511c2ab2012-12-11 16:03:16 -0800225memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
Dave Hansen3947be12005-10-29 18:16:54 -0700226{
Wen Congyanga16cee12012-10-08 16:33:58 -0700227 unsigned long start_pfn;
Anton Blanchard5409d2c2011-05-11 17:25:14 +1000228 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
Greg Kroah-Hartmande0ed362011-10-18 14:00:57 -0700229 struct page *first_page;
Dave Hansen3947be12005-10-29 18:16:54 -0700230 int ret;
Dave Hansen3947be12005-10-29 18:16:54 -0700231
Sheng Yong19c07d52015-04-14 15:44:54 -0700232 start_pfn = section_nr_to_pfn(phys_index);
Zhang Zhen71fbd552014-12-12 16:55:33 -0800233 first_page = pfn_to_page(start_pfn);
Greg Kroah-Hartmande0ed362011-10-18 14:00:57 -0700234
Dave Hansen3947be12005-10-29 18:16:54 -0700235 switch (action) {
Ioana Ciornei3d3af6a2015-03-08 12:29:04 +0200236 case MEM_ONLINE:
237 if (!pages_correctly_reserved(start_pfn))
238 return -EBUSY;
Mel Gorman2bbcb8782011-10-17 16:38:20 +0200239
Ioana Ciornei3d3af6a2015-03-08 12:29:04 +0200240 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 Hansen3947be12005-10-29 18:16:54 -0700249 }
Dave Hansen3947be12005-10-29 18:16:54 -0700250
251 return ret;
252}
253
Seth Jenningsfa2be402013-08-20 16:05:05 -0500254static int memory_block_change_state(struct memory_block *mem,
255 unsigned long to_state, unsigned long from_state_req)
Dave Hansen3947be12005-10-29 18:16:54 -0700256{
Greg Kroah-Hartmande0ed362011-10-18 14:00:57 -0700257 int ret = 0;
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600258
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200259 if (mem->state != from_state_req)
260 return -EINVAL;
Dave Hansen3947be12005-10-29 18:16:54 -0700261
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600262 if (to_state == MEM_OFFLINE)
263 mem->state = MEM_GOING_OFFLINE;
264
Seth Jenningsfa2be402013-08-20 16:05:05 -0500265 ret = memory_block_action(mem->start_section_nr, to_state,
266 mem->online_type);
267
Rafael J. Wysockib2c064b2013-05-23 10:38:55 +0200268 mem->state = ret ? from_state_req : to_state;
Seth Jenningsfa2be402013-08-20 16:05:05 -0500269
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200270 return ret;
271}
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600272
Seth Jenningsfa2be402013-08-20 16:05:05 -0500273/* The device lock serializes operations on memory_subsys_[online|offline] */
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200274static int memory_subsys_online(struct device *dev)
275{
Gu Zheng7315f0c2013-08-28 14:38:27 +0800276 struct memory_block *mem = to_memory_block(dev);
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200277 int ret;
Dave Hansen3947be12005-10-29 18:16:54 -0700278
Seth Jenningsfa2be402013-08-20 16:05:05 -0500279 if (mem->state == MEM_ONLINE)
280 return 0;
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200281
Seth Jenningsfa2be402013-08-20 16:05:05 -0500282 /*
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 Chen4f7c6b42014-08-06 16:05:13 -0700288 mem->online_type = MMOP_ONLINE_KEEP;
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200289
David Rientjes30467e02015-04-14 15:45:11 -0700290 /* Already under protection of mem_hotplug_begin() */
Seth Jenningsfa2be402013-08-20 16:05:05 -0500291 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
292
293 /* clear online_type */
294 mem->online_type = -1;
295
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200296 return ret;
297}
298
299static int memory_subsys_offline(struct device *dev)
300{
Gu Zheng7315f0c2013-08-28 14:38:27 +0800301 struct memory_block *mem = to_memory_block(dev);
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200302
Seth Jenningsfa2be402013-08-20 16:05:05 -0500303 if (mem->state == MEM_OFFLINE)
304 return 0;
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200305
Seth Jenningsfa2be402013-08-20 16:05:05 -0500306 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200307}
308
Dave Hansen3947be12005-10-29 18:16:54 -0700309static ssize_t
Kay Sievers10fbcf42011-12-21 14:48:43 -0800310store_mem_state(struct device *dev,
311 struct device_attribute *attr, const char *buf, size_t count)
Dave Hansen3947be12005-10-29 18:16:54 -0700312{
Gu Zheng7315f0c2013-08-28 14:38:27 +0800313 struct memory_block *mem = to_memory_block(dev);
Seth Jenningsfa2be402013-08-20 16:05:05 -0500314 int ret, online_type;
Dave Hansen3947be12005-10-29 18:16:54 -0700315
Rafael J. Wysocki5e33bc42013-08-28 21:41:01 +0200316 ret = lock_device_hotplug_sysfs();
317 if (ret)
318 return ret;
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200319
Tang Chen1f6a6cc2014-08-06 16:05:11 -0700320 if (sysfs_streq(buf, "online_kernel"))
Tang Chen4f7c6b42014-08-06 16:05:13 -0700321 online_type = MMOP_ONLINE_KERNEL;
Tang Chen1f6a6cc2014-08-06 16:05:11 -0700322 else if (sysfs_streq(buf, "online_movable"))
Tang Chen4f7c6b42014-08-06 16:05:13 -0700323 online_type = MMOP_ONLINE_MOVABLE;
Tang Chen1f6a6cc2014-08-06 16:05:11 -0700324 else if (sysfs_streq(buf, "online"))
Tang Chen4f7c6b42014-08-06 16:05:13 -0700325 online_type = MMOP_ONLINE_KEEP;
Tang Chen1f6a6cc2014-08-06 16:05:11 -0700326 else if (sysfs_streq(buf, "offline"))
Tang Chen4f7c6b42014-08-06 16:05:13 -0700327 online_type = MMOP_OFFLINE;
Yasuaki Ishimatsua37f8632013-10-11 15:36:25 +0900328 else {
329 ret = -EINVAL;
330 goto err;
331 }
Seth Jenningsfa2be402013-08-20 16:05:05 -0500332
David Rientjes30467e02015-04-14 15:45:11 -0700333 /*
334 * Memory hotplug needs to hold mem_hotplug_begin() for probe to find
335 * the correct memory block to online before doing device_online(dev),
336 * which will take dev->mutex. Take the lock early to prevent an
337 * inversion, memory_subsys_online() callbacks will be implemented by
338 * assuming it's already protected.
339 */
340 mem_hotplug_begin();
341
Seth Jenningsfa2be402013-08-20 16:05:05 -0500342 switch (online_type) {
Tang Chen4f7c6b42014-08-06 16:05:13 -0700343 case MMOP_ONLINE_KERNEL:
344 case MMOP_ONLINE_MOVABLE:
345 case MMOP_ONLINE_KEEP:
Seth Jenningsfa2be402013-08-20 16:05:05 -0500346 mem->online_type = online_type;
347 ret = device_online(&mem->dev);
348 break;
Tang Chen4f7c6b42014-08-06 16:05:13 -0700349 case MMOP_OFFLINE:
Seth Jenningsfa2be402013-08-20 16:05:05 -0500350 ret = device_offline(&mem->dev);
351 break;
352 default:
353 ret = -EINVAL; /* should never happen */
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200354 }
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200355
David Rientjes30467e02015-04-14 15:45:11 -0700356 mem_hotplug_done();
Yasuaki Ishimatsua37f8632013-10-11 15:36:25 +0900357err:
Rafael J. Wysocki4960e052013-05-08 14:18:37 +0200358 unlock_device_hotplug();
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600359
Dave Hansen3947be12005-10-29 18:16:54 -0700360 if (ret)
361 return ret;
362 return count;
363}
364
365/*
366 * phys_device is a bad name for this. What I really want
367 * is a way to differentiate between memory ranges that
368 * are part of physical devices that constitute
369 * a complete removable unit or fru.
370 * i.e. do these ranges belong to the same physical device,
371 * s.t. if I offline all of these sections I can then
372 * remove the physical device?
373 */
Kay Sievers10fbcf42011-12-21 14:48:43 -0800374static ssize_t show_phys_device(struct device *dev,
375 struct device_attribute *attr, char *buf)
Dave Hansen3947be12005-10-29 18:16:54 -0700376{
Gu Zheng7315f0c2013-08-28 14:38:27 +0800377 struct memory_block *mem = to_memory_block(dev);
Dave Hansen3947be12005-10-29 18:16:54 -0700378 return sprintf(buf, "%d\n", mem->phys_device);
379}
380
Zhang Zhened2f2402014-10-09 15:26:31 -0700381#ifdef CONFIG_MEMORY_HOTREMOVE
382static ssize_t show_valid_zones(struct device *dev,
383 struct device_attribute *attr, char *buf)
384{
385 struct memory_block *mem = to_memory_block(dev);
386 unsigned long start_pfn, end_pfn;
387 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
388 struct page *first_page;
389 struct zone *zone;
390
391 start_pfn = section_nr_to_pfn(mem->start_section_nr);
392 end_pfn = start_pfn + nr_pages;
393 first_page = pfn_to_page(start_pfn);
394
395 /* The block contains more than one zone can not be offlined. */
396 if (!test_pages_in_a_zone(start_pfn, end_pfn))
397 return sprintf(buf, "none\n");
398
399 zone = page_zone(first_page);
400
401 if (zone_idx(zone) == ZONE_MOVABLE - 1) {
402 /*The mem block is the last memoryblock of this zone.*/
403 if (end_pfn == zone_end_pfn(zone))
404 return sprintf(buf, "%s %s\n",
405 zone->name, (zone + 1)->name);
406 }
407
408 if (zone_idx(zone) == ZONE_MOVABLE) {
409 /*The mem block is the first memoryblock of ZONE_MOVABLE.*/
410 if (start_pfn == zone->zone_start_pfn)
411 return sprintf(buf, "%s %s\n",
412 zone->name, (zone - 1)->name);
413 }
414
415 return sprintf(buf, "%s\n", zone->name);
416}
417static DEVICE_ATTR(valid_zones, 0444, show_valid_zones, NULL);
418#endif
419
Kay Sievers10fbcf42011-12-21 14:48:43 -0800420static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
Kay Sievers10fbcf42011-12-21 14:48:43 -0800421static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
422static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
423static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
Dave Hansen3947be12005-10-29 18:16:54 -0700424
Dave Hansen3947be12005-10-29 18:16:54 -0700425/*
426 * Block size attribute stuff
427 */
428static ssize_t
Kay Sievers10fbcf42011-12-21 14:48:43 -0800429print_block_size(struct device *dev, struct device_attribute *attr,
Andi Kleen8564a6c2010-01-05 12:48:06 +0100430 char *buf)
Dave Hansen3947be12005-10-29 18:16:54 -0700431{
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600432 return sprintf(buf, "%lx\n", get_memory_block_size());
Dave Hansen3947be12005-10-29 18:16:54 -0700433}
434
Kay Sievers10fbcf42011-12-21 14:48:43 -0800435static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
Dave Hansen3947be12005-10-29 18:16:54 -0700436
Dave Hansen3947be12005-10-29 18:16:54 -0700437/*
438 * Some architectures will have custom drivers to do this, and
439 * will not need to do it from userspace. The fake hot-add code
440 * as well as ppc64 will do all of their discovery in userspace
441 * and will require this interface.
442 */
443#ifdef CONFIG_ARCH_MEMORY_PROBE
444static ssize_t
Kay Sievers10fbcf42011-12-21 14:48:43 -0800445memory_probe_store(struct device *dev, struct device_attribute *attr,
Andi Kleen28812fe2010-01-05 12:48:07 +0100446 const char *buf, size_t count)
Dave Hansen3947be12005-10-29 18:16:54 -0700447{
448 u64 phys_addr;
Yasunori Gotobc02af92006-06-27 02:53:30 -0700449 int nid;
Nathan Fontenot6add7cd2011-01-31 10:55:23 -0600450 int i, ret;
Anton Blanchard61b94fe2011-09-15 06:26:15 +1000451 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
Dave Hansen3947be12005-10-29 18:16:54 -0700452
Zhang Zhenb69deb22014-08-06 16:06:06 -0700453 ret = kstrtoull(buf, 0, &phys_addr);
454 if (ret)
455 return ret;
Dave Hansen3947be12005-10-29 18:16:54 -0700456
Anton Blanchard61b94fe2011-09-15 06:26:15 +1000457 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
458 return -EINVAL;
459
Nathan Fontenot6add7cd2011-01-31 10:55:23 -0600460 for (i = 0; i < sections_per_block; i++) {
461 nid = memory_add_physaddr_to_nid(phys_addr);
462 ret = add_memory(nid, phys_addr,
463 PAGES_PER_SECTION << PAGE_SHIFT);
464 if (ret)
Nikanth Karthikesan9f0af692011-03-24 11:46:18 +0530465 goto out;
Nathan Fontenot6add7cd2011-01-31 10:55:23 -0600466
467 phys_addr += MIN_MEMORY_BLOCK_SIZE;
468 }
Dave Hansen3947be12005-10-29 18:16:54 -0700469
Nikanth Karthikesan9f0af692011-03-24 11:46:18 +0530470 ret = count;
471out:
472 return ret;
Dave Hansen3947be12005-10-29 18:16:54 -0700473}
Dave Hansen3947be12005-10-29 18:16:54 -0700474
Nathan Fontenot96b2c0f2013-06-04 14:42:28 -0500475static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
Dave Hansen3947be12005-10-29 18:16:54 -0700476#endif
477
Andi Kleenfacb6012009-12-16 12:20:00 +0100478#ifdef CONFIG_MEMORY_FAILURE
479/*
480 * Support for offlining pages of memory
481 */
482
483/* Soft offline a page */
484static ssize_t
Kay Sievers10fbcf42011-12-21 14:48:43 -0800485store_soft_offline_page(struct device *dev,
486 struct device_attribute *attr,
Andi Kleen28812fe2010-01-05 12:48:07 +0100487 const char *buf, size_t count)
Andi Kleenfacb6012009-12-16 12:20:00 +0100488{
489 int ret;
490 u64 pfn;
491 if (!capable(CAP_SYS_ADMIN))
492 return -EPERM;
Jingoo Han34da5e62013-07-26 13:10:22 +0900493 if (kstrtoull(buf, 0, &pfn) < 0)
Andi Kleenfacb6012009-12-16 12:20:00 +0100494 return -EINVAL;
495 pfn >>= PAGE_SHIFT;
496 if (!pfn_valid(pfn))
497 return -ENXIO;
498 ret = soft_offline_page(pfn_to_page(pfn), 0);
499 return ret == 0 ? count : ret;
500}
501
502/* Forcibly offline a page, including killing processes. */
503static ssize_t
Kay Sievers10fbcf42011-12-21 14:48:43 -0800504store_hard_offline_page(struct device *dev,
505 struct device_attribute *attr,
Andi Kleen28812fe2010-01-05 12:48:07 +0100506 const char *buf, size_t count)
Andi Kleenfacb6012009-12-16 12:20:00 +0100507{
508 int ret;
509 u64 pfn;
510 if (!capable(CAP_SYS_ADMIN))
511 return -EPERM;
Jingoo Han34da5e62013-07-26 13:10:22 +0900512 if (kstrtoull(buf, 0, &pfn) < 0)
Andi Kleenfacb6012009-12-16 12:20:00 +0100513 return -EINVAL;
514 pfn >>= PAGE_SHIFT;
Tony Luckcd42f4a2011-12-15 10:48:12 -0800515 ret = memory_failure(pfn, 0, 0);
Andi Kleenfacb6012009-12-16 12:20:00 +0100516 return ret ? ret : count;
517}
518
Felipe Balbi74fef7a2013-02-18 21:09:03 +0200519static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
520static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
Andi Kleenfacb6012009-12-16 12:20:00 +0100521#endif
522
Dave Hansen3947be12005-10-29 18:16:54 -0700523/*
524 * Note that phys_device is optional. It is here to allow for
525 * differentiation between which *physical* devices each
526 * section belongs to...
527 */
Heiko Carstensbc32df02010-03-15 00:35:03 -0400528int __weak arch_get_memory_phys_device(unsigned long start_pfn)
529{
530 return 0;
531}
Dave Hansen3947be12005-10-29 18:16:54 -0700532
Kay Sievers10fbcf42011-12-21 14:48:43 -0800533/*
534 * A reference for the returned object is held and the reference for the
535 * hinted object is released.
536 */
Robin Holt98383032010-09-29 14:00:55 -0500537struct memory_block *find_memory_block_hinted(struct mem_section *section,
538 struct memory_block *hint)
539{
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600540 int block_id = base_memory_block_id(__section_nr(section));
Kay Sievers10fbcf42011-12-21 14:48:43 -0800541 struct device *hintdev = hint ? &hint->dev : NULL;
542 struct device *dev;
Robin Holt98383032010-09-29 14:00:55 -0500543
Kay Sievers10fbcf42011-12-21 14:48:43 -0800544 dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
545 if (hint)
546 put_device(&hint->dev);
547 if (!dev)
Robin Holt98383032010-09-29 14:00:55 -0500548 return NULL;
Gu Zheng7315f0c2013-08-28 14:38:27 +0800549 return to_memory_block(dev);
Robin Holt98383032010-09-29 14:00:55 -0500550}
551
Dave Hansen3947be12005-10-29 18:16:54 -0700552/*
553 * For now, we have a linear search to go find the appropriate
554 * memory_block corresponding to a particular phys_index. If
555 * this gets to be a real problem, we can always use a radix
556 * tree or something here.
557 *
Kay Sievers10fbcf42011-12-21 14:48:43 -0800558 * This could be made generic for all device subsystems.
Dave Hansen3947be12005-10-29 18:16:54 -0700559 */
Gary Hadec04fc582009-01-06 14:39:14 -0800560struct memory_block *find_memory_block(struct mem_section *section)
Dave Hansen3947be12005-10-29 18:16:54 -0700561{
Robin Holt98383032010-09-29 14:00:55 -0500562 return find_memory_block_hinted(section, NULL);
Dave Hansen3947be12005-10-29 18:16:54 -0700563}
564
Nathan Fontenot96b2c0f2013-06-04 14:42:28 -0500565static struct attribute *memory_memblk_attrs[] = {
566 &dev_attr_phys_index.attr,
Nathan Fontenot96b2c0f2013-06-04 14:42:28 -0500567 &dev_attr_state.attr,
568 &dev_attr_phys_device.attr,
569 &dev_attr_removable.attr,
Zhang Zhened2f2402014-10-09 15:26:31 -0700570#ifdef CONFIG_MEMORY_HOTREMOVE
571 &dev_attr_valid_zones.attr,
572#endif
Nathan Fontenot96b2c0f2013-06-04 14:42:28 -0500573 NULL
574};
575
576static struct attribute_group memory_memblk_attr_group = {
577 .attrs = memory_memblk_attrs,
578};
579
580static const struct attribute_group *memory_memblk_attr_groups[] = {
581 &memory_memblk_attr_group,
582 NULL,
583};
584
585/*
586 * register_memory - Setup a sysfs device for a memory block
587 */
588static
589int register_memory(struct memory_block *memory)
590{
Nathan Fontenot96b2c0f2013-06-04 14:42:28 -0500591 memory->dev.bus = &memory_subsys;
592 memory->dev.id = memory->start_section_nr / sections_per_block;
593 memory->dev.release = memory_block_release;
594 memory->dev.groups = memory_memblk_attr_groups;
Linus Torvaldsf991fae2013-07-03 14:35:40 -0700595 memory->dev.offline = memory->state == MEM_OFFLINE;
Nathan Fontenot96b2c0f2013-06-04 14:42:28 -0500596
Seth Jennings879f1be2013-08-20 12:12:58 -0500597 return device_register(&memory->dev);
Nathan Fontenot96b2c0f2013-06-04 14:42:28 -0500598}
599
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600600static int init_memory_block(struct memory_block **memory,
601 struct mem_section *section, unsigned long state)
Nathan Fontenote4619c82010-10-19 12:44:20 -0500602{
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600603 struct memory_block *mem;
Nathan Fontenote4619c82010-10-19 12:44:20 -0500604 unsigned long start_pfn;
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600605 int scn_nr;
Nathan Fontenote4619c82010-10-19 12:44:20 -0500606 int ret = 0;
607
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600608 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
Nathan Fontenote4619c82010-10-19 12:44:20 -0500609 if (!mem)
610 return -ENOMEM;
611
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600612 scn_nr = __section_nr(section);
Nathan Fontenotd3360162011-01-20 10:44:29 -0600613 mem->start_section_nr =
614 base_memory_block_id(scn_nr) * sections_per_block;
615 mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
Nathan Fontenote4619c82010-10-19 12:44:20 -0500616 mem->state = state;
Nathan Fontenot07681212010-10-19 12:46:19 -0500617 mem->section_count++;
Nathan Fontenotd3360162011-01-20 10:44:29 -0600618 start_pfn = section_nr_to_pfn(mem->start_section_nr);
Nathan Fontenote4619c82010-10-19 12:44:20 -0500619 mem->phys_device = arch_get_memory_phys_device(start_pfn);
620
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600621 ret = register_memory(mem);
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600622
623 *memory = mem;
624 return ret;
625}
626
Seth Jenningscb5e39b2013-08-20 12:13:03 -0500627static int add_memory_block(int base_section_nr)
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600628{
Seth Jenningscb5e39b2013-08-20 12:13:03 -0500629 struct memory_block *mem;
630 int i, ret, section_count = 0, section_nr;
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600631
Seth Jenningscb5e39b2013-08-20 12:13:03 -0500632 for (i = base_section_nr;
633 (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
634 i++) {
635 if (!present_section_nr(i))
636 continue;
637 if (section_count == 0)
638 section_nr = i;
639 section_count++;
Yinghai Lu321bf4e2012-01-30 13:57:12 -0800640 }
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600641
Seth Jenningscb5e39b2013-08-20 12:13:03 -0500642 if (section_count == 0)
643 return 0;
644 ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
645 if (ret)
646 return ret;
647 mem->section_count = section_count;
648 return 0;
Nathan Fontenote4619c82010-10-19 12:44:20 -0500649}
650
Seth Jenningscb5e39b2013-08-20 12:13:03 -0500651
David Rientjes4edd7ce2013-04-29 15:08:22 -0700652/*
653 * need an interface for the VM to add new memory regions,
654 * but without onlining it.
655 */
656int register_new_memory(int nid, struct mem_section *section)
657{
Seth Jenningsd7f80532013-08-20 12:13:00 -0500658 int ret = 0;
659 struct memory_block *mem;
Seth Jenningsb1eaef32013-08-20 12:12:57 -0500660
661 mutex_lock(&mem_sysfs_mutex);
Seth Jenningsb1eaef32013-08-20 12:12:57 -0500662
Seth Jenningsd7f80532013-08-20 12:13:00 -0500663 mem = find_memory_block(section);
664 if (mem) {
665 mem->section_count++;
666 put_device(&mem->dev);
667 } else {
668 ret = init_memory_block(&mem, section, MEM_OFFLINE);
669 if (ret)
670 goto out;
671 }
672
673 if (mem->section_count == sections_per_block)
674 ret = register_mem_sect_under_node(mem, nid);
675out:
676 mutex_unlock(&mem_sysfs_mutex);
Seth Jenningsb1eaef32013-08-20 12:12:57 -0500677 return ret;
David Rientjes4edd7ce2013-04-29 15:08:22 -0700678}
679
680#ifdef CONFIG_MEMORY_HOTREMOVE
681static void
682unregister_memory(struct memory_block *memory)
683{
684 BUG_ON(memory->dev.bus != &memory_subsys);
685
686 /* drop the ref. we got in remove_memory_block() */
Seth Jenningsdf2b7172013-08-20 12:12:59 -0500687 put_device(&memory->dev);
David Rientjes4edd7ce2013-04-29 15:08:22 -0700688 device_unregister(&memory->dev);
689}
690
691static int remove_memory_block(unsigned long node_id,
692 struct mem_section *section, int phys_device)
Dave Hansen3947be12005-10-29 18:16:54 -0700693{
694 struct memory_block *mem;
695
Nathan Fontenot2938ffb2010-10-19 12:45:24 -0500696 mutex_lock(&mem_sysfs_mutex);
Dave Hansen3947be12005-10-29 18:16:54 -0700697 mem = find_memory_block(section);
Nathan Fontenotd3360162011-01-20 10:44:29 -0600698 unregister_mem_sect_under_nodes(mem, __section_nr(section));
Nathan Fontenot07681212010-10-19 12:46:19 -0500699
700 mem->section_count--;
Nathan Fontenot96b2c0f2013-06-04 14:42:28 -0500701 if (mem->section_count == 0)
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600702 unregister_memory(mem);
Nathan Fontenot96b2c0f2013-06-04 14:42:28 -0500703 else
Seth Jenningsdf2b7172013-08-20 12:12:59 -0500704 put_device(&mem->dev);
Dave Hansen3947be12005-10-29 18:16:54 -0700705
Nathan Fontenot2938ffb2010-10-19 12:45:24 -0500706 mutex_unlock(&mem_sysfs_mutex);
Dave Hansen3947be12005-10-29 18:16:54 -0700707 return 0;
708}
709
Dave Hansen3947be12005-10-29 18:16:54 -0700710int unregister_memory_section(struct mem_section *section)
711{
Andy Whitcroft540557b2007-10-16 01:24:11 -0700712 if (!present_section(section))
Dave Hansen3947be12005-10-29 18:16:54 -0700713 return -EINVAL;
714
715 return remove_memory_block(0, section, 0);
716}
David Rientjes4edd7ce2013-04-29 15:08:22 -0700717#endif /* CONFIG_MEMORY_HOTREMOVE */
Dave Hansen3947be12005-10-29 18:16:54 -0700718
Yasuaki Ishimatsu6677e3e2013-02-22 16:32:52 -0800719/* return true if the memory block is offlined, otherwise, return false */
720bool is_memblock_offlined(struct memory_block *mem)
721{
722 return mem->state == MEM_OFFLINE;
723}
724
Nathan Fontenot96b2c0f2013-06-04 14:42:28 -0500725static struct attribute *memory_root_attrs[] = {
726#ifdef CONFIG_ARCH_MEMORY_PROBE
727 &dev_attr_probe.attr,
728#endif
729
730#ifdef CONFIG_MEMORY_FAILURE
731 &dev_attr_soft_offline_page.attr,
732 &dev_attr_hard_offline_page.attr,
733#endif
734
735 &dev_attr_block_size_bytes.attr,
736 NULL
737};
738
739static struct attribute_group memory_root_attr_group = {
740 .attrs = memory_root_attrs,
741};
742
743static const struct attribute_group *memory_root_attr_groups[] = {
744 &memory_root_attr_group,
745 NULL,
746};
747
Wen Congyange90bdb72012-10-08 16:34:01 -0700748/*
Dave Hansen3947be12005-10-29 18:16:54 -0700749 * Initialize the sysfs support for memory devices...
750 */
751int __init memory_dev_init(void)
752{
753 unsigned int i;
754 int ret;
Andrew Morton28ec24e2006-12-06 20:37:29 -0800755 int err;
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600756 unsigned long block_sz;
Dave Hansen3947be12005-10-29 18:16:54 -0700757
Nathan Fontenot96b2c0f2013-06-04 14:42:28 -0500758 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
Andrew Morton28ec24e2006-12-06 20:37:29 -0800759 if (ret)
760 goto out;
Dave Hansen3947be12005-10-29 18:16:54 -0700761
Nathan Fontenot0c2c99b2011-01-20 10:43:34 -0600762 block_sz = get_memory_block_size();
763 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
764
Dave Hansen3947be12005-10-29 18:16:54 -0700765 /*
766 * Create entries for memory sections that were found
767 * during boot and have been initialized
768 */
Seth Jenningsb1eaef32013-08-20 12:12:57 -0500769 mutex_lock(&mem_sysfs_mutex);
Seth Jenningscb5e39b2013-08-20 12:13:03 -0500770 for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
771 err = add_memory_block(i);
Andrew Morton28ec24e2006-12-06 20:37:29 -0800772 if (!ret)
773 ret = err;
Dave Hansen3947be12005-10-29 18:16:54 -0700774 }
Seth Jenningsb1eaef32013-08-20 12:12:57 -0500775 mutex_unlock(&mem_sysfs_mutex);
Dave Hansen3947be12005-10-29 18:16:54 -0700776
Andrew Morton28ec24e2006-12-06 20:37:29 -0800777out:
778 if (ret)
Harvey Harrison2b3a3022008-03-04 16:41:05 -0800779 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
Dave Hansen3947be12005-10-29 18:16:54 -0700780 return ret;
781}