blob: 0d7cb955aa01db7017cc63b47bf1a6d55b5b880e
1 | ============== |
2 | Memory Hotplug |
3 | ============== |
4 | |
5 | Created: Jul 28 2007 |
6 | Add description of notifier of memory hotplug Oct 11 2007 |
7 | |
8 | This document is about memory hotplug including how-to-use and current status. |
9 | Because Memory Hotplug is still under development, contents of this text will |
10 | be changed often. |
11 | |
12 | 1. Introduction |
13 | 1.1 purpose of memory hotplug |
14 | 1.2. Phases of memory hotplug |
15 | 1.3. Unit of Memory online/offline operation |
16 | 2. Kernel Configuration |
17 | 3. sysfs files for memory hotplug |
18 | 4. Physical memory hot-add phase |
19 | 4.1 Hardware(Firmware) Support |
20 | 4.2 Notify memory hot-add event by hand |
21 | 5. Logical Memory hot-add phase |
22 | 5.1. State of memory |
23 | 5.2. How to online memory |
24 | 6. Logical memory remove |
25 | 6.1 Memory offline and ZONE_MOVABLE |
26 | 6.2. How to offline memory |
27 | 7. Physical memory remove |
28 | 8. Memory hotplug event notifier |
29 | 9. Future Work List |
30 | |
31 | Note(1): x86_64's has special implementation for memory hotplug. |
32 | This text does not describe it. |
33 | Note(2): This text assumes that sysfs is mounted at /sys. |
34 | |
35 | |
36 | --------------- |
37 | 1. Introduction |
38 | --------------- |
39 | |
40 | 1.1 purpose of memory hotplug |
41 | ------------ |
42 | Memory Hotplug allows users to increase/decrease the amount of memory. |
43 | Generally, there are two purposes. |
44 | |
45 | (A) For changing the amount of memory. |
46 | This is to allow a feature like capacity on demand. |
47 | (B) For installing/removing DIMMs or NUMA-nodes physically. |
48 | This is to exchange DIMMs/NUMA-nodes, reduce power consumption, etc. |
49 | |
50 | (A) is required by highly virtualized environments and (B) is required by |
51 | hardware which supports memory power management. |
52 | |
53 | Linux memory hotplug is designed for both purpose. |
54 | |
55 | |
56 | 1.2. Phases of memory hotplug |
57 | --------------- |
58 | There are 2 phases in Memory Hotplug. |
59 | 1) Physical Memory Hotplug phase |
60 | 2) Logical Memory Hotplug phase. |
61 | |
62 | The First phase is to communicate hardware/firmware and make/erase |
63 | environment for hotplugged memory. Basically, this phase is necessary |
64 | for the purpose (B), but this is good phase for communication between |
65 | highly virtualized environments too. |
66 | |
67 | When memory is hotplugged, the kernel recognizes new memory, makes new memory |
68 | management tables, and makes sysfs files for new memory's operation. |
69 | |
70 | If firmware supports notification of connection of new memory to OS, |
71 | this phase is triggered automatically. ACPI can notify this event. If not, |
72 | "probe" operation by system administration is used instead. |
73 | (see Section 4.). |
74 | |
75 | Logical Memory Hotplug phase is to change memory state into |
76 | available/unavailable for users. Amount of memory from user's view is |
77 | changed by this phase. The kernel makes all memory in it as free pages |
78 | when a memory range is available. |
79 | |
80 | In this document, this phase is described as online/offline. |
81 | |
82 | Logical Memory Hotplug phase is triggered by write of sysfs file by system |
83 | administrator. For the hot-add case, it must be executed after Physical Hotplug |
84 | phase by hand. |
85 | (However, if you writes udev's hotplug scripts for memory hotplug, these |
86 | phases can be execute in seamless way.) |
87 | |
88 | |
89 | 1.3. Unit of Memory online/offline operation |
90 | ------------ |
91 | Memory hotplug uses SPARSEMEM memory model which allows memory to be divided |
92 | into chunks of the same size. These chunks are called "sections". The size of |
93 | a memory section is architecture dependent. For example, power uses 16MiB, ia64 |
94 | uses 1GiB. |
95 | |
96 | Memory sections are combined into chunks referred to as "memory blocks". The |
97 | size of a memory block is architecture dependent and represents the logical |
98 | unit upon which memory online/offline operations are to be performed. The |
99 | default size of a memory block is the same as memory section size unless an |
100 | architecture specifies otherwise. (see Section 3.) |
101 | |
102 | To determine the size (in bytes) of a memory block please read this file: |
103 | |
104 | /sys/devices/system/memory/block_size_bytes |
105 | |
106 | |
107 | ----------------------- |
108 | 2. Kernel Configuration |
109 | ----------------------- |
110 | To use memory hotplug feature, kernel must be compiled with following |
111 | config options. |
112 | |
113 | - For all memory hotplug |
114 | Memory model -> Sparse Memory (CONFIG_SPARSEMEM) |
115 | Allow for memory hot-add (CONFIG_MEMORY_HOTPLUG) |
116 | |
117 | - To enable memory removal, the followings are also necessary |
118 | Allow for memory hot remove (CONFIG_MEMORY_HOTREMOVE) |
119 | Page Migration (CONFIG_MIGRATION) |
120 | |
121 | - For ACPI memory hotplug, the followings are also necessary |
122 | Memory hotplug (under ACPI Support menu) (CONFIG_ACPI_HOTPLUG_MEMORY) |
123 | This option can be kernel module. |
124 | |
125 | - As a related configuration, if your box has a feature of NUMA-node hotplug |
126 | via ACPI, then this option is necessary too. |
127 | ACPI0004,PNP0A05 and PNP0A06 Container Driver (under ACPI Support menu) |
128 | (CONFIG_ACPI_CONTAINER). |
129 | This option can be kernel module too. |
130 | |
131 | |
132 | -------------------------------- |
133 | 3 sysfs files for memory hotplug |
134 | -------------------------------- |
135 | All memory blocks have their device information in sysfs. Each memory block |
136 | is described under /sys/devices/system/memory as |
137 | |
138 | /sys/devices/system/memory/memoryXXX |
139 | (XXX is the memory block id.) |
140 | |
141 | For the memory block covered by the sysfs directory. It is expected that all |
142 | memory sections in this range are present and no memory holes exist in the |
143 | range. Currently there is no way to determine if there is a memory hole, but |
144 | the existence of one should not affect the hotplug capabilities of the memory |
145 | block. |
146 | |
147 | For example, assume 1GiB memory block size. A device for a memory starting at |
148 | 0x100000000 is /sys/device/system/memory/memory4 |
149 | (0x100000000 / 1Gib = 4) |
150 | This device covers address range [0x100000000 ... 0x140000000) |
151 | |
152 | Under each memory block, you can see 5 files: |
153 | |
154 | /sys/devices/system/memory/memoryXXX/phys_index |
155 | /sys/devices/system/memory/memoryXXX/phys_device |
156 | /sys/devices/system/memory/memoryXXX/state |
157 | /sys/devices/system/memory/memoryXXX/removable |
158 | /sys/devices/system/memory/memoryXXX/valid_zones |
159 | |
160 | 'phys_index' : read-only and contains memory block id, same as XXX. |
161 | 'state' : read-write |
162 | at read: contains online/offline state of memory. |
163 | at write: user can specify "online_kernel", |
164 | "online_movable", "online", "offline" command |
165 | which will be performed on all sections in the block. |
166 | 'phys_device' : read-only: designed to show the name of physical memory |
167 | device. This is not well implemented now. |
168 | 'removable' : read-only: contains an integer value indicating |
169 | whether the memory block is removable or not |
170 | removable. A value of 1 indicates that the memory |
171 | block is removable and a value of 0 indicates that |
172 | it is not removable. A memory block is removable only if |
173 | every section in the block is removable. |
174 | 'valid_zones' : read-only: designed to show which zones this memory block |
175 | can be onlined to. |
176 | The first column shows it's default zone. |
177 | "memory6/valid_zones: Normal Movable" shows this memoryblock |
178 | can be onlined to ZONE_NORMAL by default and to ZONE_MOVABLE |
179 | by online_movable. |
180 | "memory7/valid_zones: Movable Normal" shows this memoryblock |
181 | can be onlined to ZONE_MOVABLE by default and to ZONE_NORMAL |
182 | by online_kernel. |
183 | |
184 | NOTE: |
185 | These directories/files appear after physical memory hotplug phase. |
186 | |
187 | If CONFIG_NUMA is enabled the memoryXXX/ directories can also be accessed |
188 | via symbolic links located in the /sys/devices/system/node/node* directories. |
189 | |
190 | For example: |
191 | /sys/devices/system/node/node0/memory9 -> ../../memory/memory9 |
192 | |
193 | A backlink will also be created: |
194 | /sys/devices/system/memory/memory9/node0 -> ../../node/node0 |
195 | |
196 | |
197 | -------------------------------- |
198 | 4. Physical memory hot-add phase |
199 | -------------------------------- |
200 | |
201 | 4.1 Hardware(Firmware) Support |
202 | ------------ |
203 | On x86_64/ia64 platform, memory hotplug by ACPI is supported. |
204 | |
205 | In general, the firmware (ACPI) which supports memory hotplug defines |
206 | memory class object of _HID "PNP0C80". When a notify is asserted to PNP0C80, |
207 | Linux's ACPI handler does hot-add memory to the system and calls a hotplug udev |
208 | script. This will be done automatically. |
209 | |
210 | But scripts for memory hotplug are not contained in generic udev package(now). |
211 | You may have to write it by yourself or online/offline memory by hand. |
212 | Please see "How to online memory", "How to offline memory" in this text. |
213 | |
214 | If firmware supports NUMA-node hotplug, and defines an object _HID "ACPI0004", |
215 | "PNP0A05", or "PNP0A06", notification is asserted to it, and ACPI handler |
216 | calls hotplug code for all of objects which are defined in it. |
217 | If memory device is found, memory hotplug code will be called. |
218 | |
219 | |
220 | 4.2 Notify memory hot-add event by hand |
221 | ------------ |
222 | On some architectures, the firmware may not notify the kernel of a memory |
223 | hotplug event. Therefore, the memory "probe" interface is supported to |
224 | explicitly notify the kernel. This interface depends on |
225 | CONFIG_ARCH_MEMORY_PROBE and can be configured on powerpc, sh, and x86 |
226 | if hotplug is supported, although for x86 this should be handled by ACPI |
227 | notification. |
228 | |
229 | Probe interface is located at |
230 | /sys/devices/system/memory/probe |
231 | |
232 | You can tell the physical address of new memory to the kernel by |
233 | |
234 | % echo start_address_of_new_memory > /sys/devices/system/memory/probe |
235 | |
236 | Then, [start_address_of_new_memory, start_address_of_new_memory + |
237 | memory_block_size] memory range is hot-added. In this case, hotplug script is |
238 | not called (in current implementation). You'll have to online memory by |
239 | yourself. Please see "How to online memory" in this text. |
240 | |
241 | |
242 | ------------------------------ |
243 | 5. Logical Memory hot-add phase |
244 | ------------------------------ |
245 | |
246 | 5.1. State of memory |
247 | ------------ |
248 | To see (online/offline) state of a memory block, read 'state' file. |
249 | |
250 | % cat /sys/device/system/memory/memoryXXX/state |
251 | |
252 | |
253 | If the memory block is online, you'll read "online". |
254 | If the memory block is offline, you'll read "offline". |
255 | |
256 | |
257 | 5.2. How to online memory |
258 | ------------ |
259 | When the memory is hot-added, the kernel decides whether or not to "online" |
260 | it according to the policy which can be read from "auto_online_blocks" file: |
261 | |
262 | % cat /sys/devices/system/memory/auto_online_blocks |
263 | |
264 | The default depends on the CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE kernel config |
265 | option. If it is disabled the default is "offline" which means the newly added |
266 | memory is not in a ready-to-use state and you have to "online" the newly added |
267 | memory blocks manually. Automatic onlining can be requested by writing "online" |
268 | to "auto_online_blocks" file: |
269 | |
270 | % echo online > /sys/devices/system/memory/auto_online_blocks |
271 | |
272 | This sets a global policy and impacts all memory blocks that will subsequently |
273 | be hotplugged. Currently offline blocks keep their state. It is possible, under |
274 | certain circumstances, that some memory blocks will be added but will fail to |
275 | online. User space tools can check their "state" files |
276 | (/sys/devices/system/memory/memoryXXX/state) and try to online them manually. |
277 | |
278 | If the automatic onlining wasn't requested, failed, or some memory block was |
279 | offlined it is possible to change the individual block's state by writing to the |
280 | "state" file: |
281 | |
282 | % echo online > /sys/devices/system/memory/memoryXXX/state |
283 | |
284 | This onlining will not change the ZONE type of the target memory block, |
285 | If the memory block is in ZONE_NORMAL, you can change it to ZONE_MOVABLE: |
286 | |
287 | % echo online_movable > /sys/devices/system/memory/memoryXXX/state |
288 | (NOTE: current limit: this memory block must be adjacent to ZONE_MOVABLE) |
289 | |
290 | And if the memory block is in ZONE_MOVABLE, you can change it to ZONE_NORMAL: |
291 | |
292 | % echo online_kernel > /sys/devices/system/memory/memoryXXX/state |
293 | (NOTE: current limit: this memory block must be adjacent to ZONE_NORMAL) |
294 | |
295 | After this, memory block XXX's state will be 'online' and the amount of |
296 | available memory will be increased. |
297 | |
298 | Currently, newly added memory is added as ZONE_NORMAL (for powerpc, ZONE_DMA). |
299 | This may be changed in future. |
300 | |
301 | |
302 | |
303 | ------------------------ |
304 | 6. Logical memory remove |
305 | ------------------------ |
306 | |
307 | 6.1 Memory offline and ZONE_MOVABLE |
308 | ------------ |
309 | Memory offlining is more complicated than memory online. Because memory offline |
310 | has to make the whole memory block be unused, memory offline can fail if |
311 | the memory block includes memory which cannot be freed. |
312 | |
313 | In general, memory offline can use 2 techniques. |
314 | |
315 | (1) reclaim and free all memory in the memory block. |
316 | (2) migrate all pages in the memory block. |
317 | |
318 | In the current implementation, Linux's memory offline uses method (2), freeing |
319 | all pages in the memory block by page migration. But not all pages are |
320 | migratable. Under current Linux, migratable pages are anonymous pages and |
321 | page caches. For offlining a memory block by migration, the kernel has to |
322 | guarantee that the memory block contains only migratable pages. |
323 | |
324 | Now, a boot option for making a memory block which consists of migratable pages |
325 | is supported. By specifying "kernelcore=" or "movablecore=" boot option, you can |
326 | create ZONE_MOVABLE...a zone which is just used for movable pages. |
327 | (See also Documentation/kernel-parameters.txt) |
328 | |
329 | Assume the system has "TOTAL" amount of memory at boot time, this boot option |
330 | creates ZONE_MOVABLE as following. |
331 | |
332 | 1) When kernelcore=YYYY boot option is used, |
333 | Size of memory not for movable pages (not for offline) is YYYY. |
334 | Size of memory for movable pages (for offline) is TOTAL-YYYY. |
335 | |
336 | 2) When movablecore=ZZZZ boot option is used, |
337 | Size of memory not for movable pages (not for offline) is TOTAL - ZZZZ. |
338 | Size of memory for movable pages (for offline) is ZZZZ. |
339 | |
340 | |
341 | Note: Unfortunately, there is no information to show which memory block belongs |
342 | to ZONE_MOVABLE. This is TBD. |
343 | |
344 | |
345 | 6.2. How to offline memory |
346 | ------------ |
347 | You can offline a memory block by using the same sysfs interface that was used |
348 | in memory onlining. |
349 | |
350 | % echo offline > /sys/devices/system/memory/memoryXXX/state |
351 | |
352 | If offline succeeds, the state of the memory block is changed to be "offline". |
353 | If it fails, some error core (like -EBUSY) will be returned by the kernel. |
354 | Even if a memory block does not belong to ZONE_MOVABLE, you can try to offline |
355 | it. If it doesn't contain 'unmovable' memory, you'll get success. |
356 | |
357 | A memory block under ZONE_MOVABLE is considered to be able to be offlined |
358 | easily. But under some busy state, it may return -EBUSY. Even if a memory |
359 | block cannot be offlined due to -EBUSY, you can retry offlining it and may be |
360 | able to offline it (or not). (For example, a page is referred to by some kernel |
361 | internal call and released soon.) |
362 | |
363 | Consideration: |
364 | Memory hotplug's design direction is to make the possibility of memory offlining |
365 | higher and to guarantee unplugging memory under any situation. But it needs |
366 | more work. Returning -EBUSY under some situation may be good because the user |
367 | can decide to retry more or not by himself. Currently, memory offlining code |
368 | does some amount of retry with 120 seconds timeout. |
369 | |
370 | ------------------------- |
371 | 7. Physical memory remove |
372 | ------------------------- |
373 | Need more implementation yet.... |
374 | - Notification completion of remove works by OS to firmware. |
375 | - Guard from remove if not yet. |
376 | |
377 | -------------------------------- |
378 | 8. Memory hotplug event notifier |
379 | -------------------------------- |
380 | Hotplugging events are sent to a notification queue. |
381 | |
382 | There are six types of notification defined in include/linux/memory.h: |
383 | |
384 | MEM_GOING_ONLINE |
385 | Generated before new memory becomes available in order to be able to |
386 | prepare subsystems to handle memory. The page allocator is still unable |
387 | to allocate from the new memory. |
388 | |
389 | MEM_CANCEL_ONLINE |
390 | Generated if MEMORY_GOING_ONLINE fails. |
391 | |
392 | MEM_ONLINE |
393 | Generated when memory has successfully brought online. The callback may |
394 | allocate pages from the new memory. |
395 | |
396 | MEM_GOING_OFFLINE |
397 | Generated to begin the process of offlining memory. Allocations are no |
398 | longer possible from the memory but some of the memory to be offlined |
399 | is still in use. The callback can be used to free memory known to a |
400 | subsystem from the indicated memory block. |
401 | |
402 | MEM_CANCEL_OFFLINE |
403 | Generated if MEMORY_GOING_OFFLINE fails. Memory is available again from |
404 | the memory block that we attempted to offline. |
405 | |
406 | MEM_OFFLINE |
407 | Generated after offlining memory is complete. |
408 | |
409 | A callback routine can be registered by calling |
410 | |
411 | hotplug_memory_notifier(callback_func, priority) |
412 | |
413 | Callback functions with higher values of priority are called before callback |
414 | functions with lower values. |
415 | |
416 | A callback function must have the following prototype: |
417 | |
418 | int callback_func( |
419 | struct notifier_block *self, unsigned long action, void *arg); |
420 | |
421 | The first argument of the callback function (self) is a pointer to the block |
422 | of the notifier chain that points to the callback function itself. |
423 | The second argument (action) is one of the event types described above. |
424 | The third argument (arg) passes a pointer of struct memory_notify. |
425 | |
426 | struct memory_notify { |
427 | unsigned long start_pfn; |
428 | unsigned long nr_pages; |
429 | int status_change_nid_normal; |
430 | int status_change_nid_high; |
431 | int status_change_nid; |
432 | } |
433 | |
434 | start_pfn is start_pfn of online/offline memory. |
435 | nr_pages is # of pages of online/offline memory. |
436 | status_change_nid_normal is set node id when N_NORMAL_MEMORY of nodemask |
437 | is (will be) set/clear, if this is -1, then nodemask status is not changed. |
438 | status_change_nid_high is set node id when N_HIGH_MEMORY of nodemask |
439 | is (will be) set/clear, if this is -1, then nodemask status is not changed. |
440 | status_change_nid is set node id when N_MEMORY of nodemask is (will be) |
441 | set/clear. It means a new(memoryless) node gets new memory by online and a |
442 | node loses all memory. If this is -1, then nodemask status is not changed. |
443 | If status_changed_nid* >= 0, callback should create/discard structures for the |
444 | node if necessary. |
445 | |
446 | The callback routine shall return one of the values |
447 | NOTIFY_DONE, NOTIFY_OK, NOTIFY_BAD, NOTIFY_STOP |
448 | defined in include/linux/notifier.h |
449 | |
450 | NOTIFY_DONE and NOTIFY_OK have no effect on the further processing. |
451 | |
452 | NOTIFY_BAD is used as response to the MEM_GOING_ONLINE, MEM_GOING_OFFLINE, |
453 | MEM_ONLINE, or MEM_OFFLINE action to cancel hotplugging. It stops |
454 | further processing of the notification queue. |
455 | |
456 | NOTIFY_STOP stops further processing of the notification queue. |
457 | |
458 | -------------- |
459 | 9. Future Work |
460 | -------------- |
461 | - allowing memory hot-add to ZONE_MOVABLE. maybe we need some switch like |
462 | sysctl or new control file. |
463 | - showing memory block and physical device relationship. |
464 | - test and make it better memory offlining. |
465 | - support HugeTLB page migration and offlining. |
466 | - memmap removing at memory offline. |
467 | - physical remove memory. |
468 | |
469 |