cpu-hotplug.txt 16.8 KB
Newer Older
1 2 3 4
		CPU hotplug Support in Linux(tm) Kernel

		CPU Hotplug Core:
			Rusty Russell <rusty@rustcorp.com.au>
6 7
			Srivatsa Vaddagiri <vatsa@in.ibm.com>
			Zwane Mwaikambo <zwanem@gmail.com>
9 10 11 12 13
			Nathan Lynch <nathanl@austin.ibm.com>
			Joel Schopp <jschopp@austin.ibm.com>
			Ashok Raj <ashok.raj@intel.com>
14 15
			Heiko Carstens <heiko.carstens@de.ibm.com>
16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48

Authors: Ashok Raj <ashok.raj@intel.com>
Lots of feedback: Nathan Lynch <nathanl@austin.ibm.com>,
	     Joel Schopp <jschopp@austin.ibm.com>


Modern advances in system architectures have introduced advanced error
reporting and correction capabilities in processors. CPU architectures permit
partitioning support, where compute resources of a single CPU could be made
available to virtual machine environments. There are couple OEMS that
support NUMA hardware which are hot pluggable as well, where physical
node insertion and removal require support for CPU hotplug.

Such advances require CPUs available to a kernel to be removed either for
provisioning reasons, or for RAS purposes to keep an offending CPU off
system execution path. Hence the need for CPU hotplug support in the
Linux kernel.

A more novel use of CPU-hotplug support is its use today in suspend
resume support for SMP. Dual-core and HT support makes even
a laptop run SMP kernels which didn't support these methods. SMP support
for suspend/resume is a work in progress.

General Stuff about CPU Hotplug

Command Line Switches
maxcpus=n    Restrict boot time cpus to n. Say if you have 4 cpus, using
             maxcpus=2 will only boot 2. You can choose to bring the
             other cpus later online, read FAQ's for more info.

additional_cpus=n (*)	Use this to limit hotpluggable cpus. This option sets
  			cpu_possible_mask = cpu_present_mask + additional_cpus

52 53 54 55 56 57
cede_offline={"off","on"}  Use this option to disable/enable putting offlined
		            processors to an extended H_CEDE state on
			    supported pseries platforms.
			    If nothing is specified,
			    cede_offline is set to "on".

(*) Option valid only for following architectures
- ia64

61 62 63 64 65 66
ia64 uses the number of disabled local apics in ACPI tables MADT to
determine the number of potentially hot-pluggable cpus. The implementation
should only rely on this to count the # of cpus, but *MUST* not rely
on the apicid values in those tables for disabled apics. In the event
BIOS doesn't mark such hot-pluggable cpus as disabled entries, one could
use this parameter "additional_cpus=x" to represent those cpus in the

possible_cpus=n		[s390,x86_64] use this to set hotpluggable cpus.
			This option sets possible_cpus bits in
			cpu_possible_mask. Thus keeping the numbers of bits set
72 73
			constant even if the machine gets rebooted.

74 75 76 77 78
CPU maps and such
[More on cpumaps and primitive to manipulate, please check
include/linux/cpumask.h that has more descriptive text.]

cpu_possible_mask: Bitmap of possible CPUs that can ever be available in the
80 81 82 83 84 85 86
system. This is used to allocate some boot time memory for per_cpu variables
that aren't designed to grow/shrink as CPUs are made available or removed.
Once set during boot time discovery phase, the map is static, i.e no bits
are added or removed anytime.  Trimming it accurately for your system needs
upfront can save some boot time memory. See below for how we use heuristics
in x86_64 case to keep this under check.

cpu_online_mask: Bitmap of all CPUs currently online. Its set in __cpu_up()
88 89 90 91 92
after a cpu is available for kernel scheduling and ready to receive
interrupts from devices. Its cleared when a cpu is brought down using
__cpu_disable(), before which all OS services including interrupts are
migrated to another target CPU.

cpu_present_mask: Bitmap of CPUs currently present in the system. Not all
94 95 96 97 98 99 100 101
of them may be online. When physical hotplug is processed by the relevant
subsystem (e.g ACPI) can change and new bit either be added or removed
from the map depending on the event is hot-add/hot-remove. There are currently
no locking rules as of now. Typical usage is to init topology during boot,
at which time hotplug is disabled.

You really dont need to manipulate any of the system cpu maps. They should
be read-only for most use. When setting up per-cpu resources almost always use
cpu_possible_mask/for_each_possible_cpu() to iterate.
103 104 105

Never use anything other than cpumask_t to represent bitmap of CPUs.

	#include <linux/cpumask.h>

108 109 110
	for_each_possible_cpu     - Iterate over cpu_possible_mask
	for_each_online_cpu       - Iterate over cpu_online_mask
	for_each_present_cpu      - Iterate over cpu_present_mask
	for_each_cpu(x,mask)      - Iterate over some random collection of cpu mask.

	#include <linux/cpu.h>
	get_online_cpus() and put_online_cpus():

The above calls are used to inhibit cpu hotplug operations. While the
cpu_hotplug.refcount is non zero, the cpu_online_mask will not change.
118 119 120
If you merely need to avoid cpus going away, you could also use
preempt_disable() and preempt_enable() for those sections.
Just remember the critical section cannot call any
121 122 123 124 125
function that can sleep or schedule this process away. The preempt_disable()
will work as long as stop_machine_run() is used to take a cpu down.

CPU Hotplug - Frequently Asked Questions.

Q: How to enable my kernel to support CPU hotplug?
127 128 129 130
A: When doing make defconfig, Enable CPU hotplug support

   "Processor type and Features" -> Support for Hotpluggable CPUs

Make sure that you have CONFIG_SMP turned on as well.
132 133 134 135 136 137 138 139 140 141 142 143 144 145 146

You would need to enable CONFIG_HOTPLUG_CPU for SMP suspend/resume support
as well.

Q: What architectures support CPU hotplug?
A: As of 2.6.14, the following architectures support CPU hotplug.

i386 (Intel), ppc, ppc64, parisc, s390, ia64 and x86_64

Q: How to test if hotplug is supported on the newly built kernel?
A: You should now notice an entry in sysfs.

Check if sysfs is mounted, using the "mount" command. You should notice
an entry as shown below in the output.

147 148 149
	none on /sys type sysfs (rw)

If this is not mounted, do the following.

	#mkdir /sys
	#mount -t sysfs sys /sys

Now you should see entries for all present cpu, the following is an example
157 158
in a 8-way system.

159 160 161 162 163 164 165 166 167 168 169 170 171 172
	#ls -l
	total 0
	drwxr-xr-x  10 root root 0 Sep 19 07:44 .
	drwxr-xr-x  13 root root 0 Sep 19 07:45 ..
	drwxr-xr-x   3 root root 0 Sep 19 07:44 cpu0
	drwxr-xr-x   3 root root 0 Sep 19 07:44 cpu1
	drwxr-xr-x   3 root root 0 Sep 19 07:44 cpu2
	drwxr-xr-x   3 root root 0 Sep 19 07:44 cpu3
	drwxr-xr-x   3 root root 0 Sep 19 07:44 cpu4
	drwxr-xr-x   3 root root 0 Sep 19 07:44 cpu5
	drwxr-xr-x   3 root root 0 Sep 19 07:44 cpu6
	drwxr-xr-x   3 root root 0 Sep 19 07:48 cpu7
173 174 175 176 177 178

Under each directory you would find an "online" file which is the control
file to logically online/offline a processor.

Q: Does hot-add/hot-remove refer to physical add/remove of cpus?
A: The usage of hot-add/remove may not be very consistently used in the code.
CONFIG_HOTPLUG_CPU enables logical online/offline capability in the kernel.
180 181 182 183 184 185 186
To support physical addition/removal, one would need some BIOS hooks and
the platform should have something like an attention button in PCI hotplug.
CONFIG_ACPI_HOTPLUG_CPU enables ACPI support for physical add/remove of CPUs.

Q: How do i logically offline a CPU?
A: Do the following.

	#echo 0 > /sys/devices/system/cpu/cpuX/online

Once the logical offline is successful, check

	#cat /proc/interrupts

You should now not see the CPU that you removed. Also online file will report
194 195
the state as 0 when a cpu if offline and 1 when its online.

196 197
	#To display the current cpu state.
	#cat /sys/devices/system/cpu/cpuX/online

Q: Why can't i remove CPU0 on some systems?
200 201 202 203 204 205 206 207 208 209
A: Some architectures may have some special dependency on a certain CPU.

For e.g in IA64 platforms we have ability to sent platform interrupts to the
OS. a.k.a Corrected Platform Error Interrupts (CPEI). In current ACPI
specifications, we didn't have a way to change the target CPU. Hence if the
current ACPI version doesn't support such re-direction, we disable that CPU
by making it not-removable.

In such cases you will also notice that the online file is missing under cpu0.

210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233
Q: Is CPU0 removable on X86?
A: Yes. If kernel is compiled with CONFIG_BOOTPARAM_HOTPLUG_CPU0=y, CPU0 is
removable by default. Otherwise, CPU0 is also removable by kernel option

But some features depend on CPU0. Two known dependencies are:

1. Resume from hibernate/suspend depends on CPU0. Hibernate/suspend will fail if
CPU0 is offline and you need to online CPU0 before hibernate/suspend can
2. PIC interrupts also depend on CPU0. CPU0 can't be removed if a PIC interrupt
is detected.

It's said poweroff/reboot may depend on CPU0 on some machines although I haven't
seen any poweroff/reboot failure so far after CPU0 is offline on a few tested

Please let me know if you know or see any other dependencies of CPU0.

If the dependencies are under your control, you can turn on CPU0 hotplug feature
either by CONFIG_BOOTPARAM_HOTPLUG_CPU0 or by kernel parameter cpu0_hotplug.

--Fenghua Yu <fenghua.yu@intel.com>

234 235 236 237 238 239 240 241 242 243 244 245 246
Q: How do i find out if a particular CPU is not removable?
A: Depending on the implementation, some architectures may show this by the
absence of the "online" file. This is done if it can be determined ahead of
time that this CPU cannot be removed.

In some situations, this can be a run time check, i.e if you try to remove the
last CPU, this will not be permitted. You can find such failures by
investigating the return value of the "echo" command.

Q: What happens when a CPU is being logically offlined?
A: The following happen, listed in no particular order :-)

- A notification is sent to in-kernel registered modules by sending an event
247 248 249
  CPU_DOWN_PREPARE or CPU_DOWN_PREPARE_FROZEN, depending on whether or not the
  CPU is being offlined while tasks are frozen due to a suspend operation in
250 251 252
- All processes are migrated away from this outgoing CPU to new CPUs.
  The new CPU is chosen from each process' current cpuset, which may be
  a subset of all online CPUs.
253 254 255 256 257
- All interrupts targeted to this CPU is migrated to a new CPU
- timers/bottom half/task lets are also migrated to a new CPU
- Once all services are migrated, kernel calls an arch specific routine
  __cpu_disable() to perform arch specific cleanup.
- Once this is successful, an event for successful cleanup is sent by an event
258 259
  CPU_DEAD (or CPU_DEAD_FROZEN if tasks are frozen due to a suspend while the
  CPU is being offlined).
260 261 262 263 264 265 266 267 268

  "It is expected that each service cleans up when the CPU_DOWN_PREPARE
  notifier is called, when CPU_DEAD is called its expected there is nothing
  running on behalf of this CPU that was offlined"

Q: If i have some kernel code that needs to be aware of CPU arrival and
   departure, how to i arrange for proper notification?
A: This is what you would need in your kernel code to receive notifications.

	#include <linux/cpu.h>
270 271
	static int foobar_cpu_callback(struct notifier_block *nfb,
				       unsigned long action, void *hcpu)
272 273 274 275 276
		unsigned int cpu = (unsigned long)hcpu;

		switch (action) {
		case CPU_ONLINE:
278 279 280
		case CPU_DEAD:
282 283 284 285 286 287
		return NOTIFY_OK;

	static struct notifier_block foobar_cpu_notifier =
289 290 291 292
	   .notifier_call = foobar_cpu_callback,

293 294 295 296
You need to call register_cpu_notifier() from your init function.
Init functions could be of two types:
1. early init (init function called when only the boot processor is online).
2. late init (init function called _after_ all the CPUs are online).

For the first case, you should add the following to your init function
299 300 301


302 303 304 305
For the second case, you should add the following to your init function


306 307 308 309 310 311 312 313 314
You can fail PREPARE notifiers if something doesn't work to prepare resources.
This will stop the activity and send a following CANCELED event back.

CPU_DEAD should not be failed, its just a goodness indication, but bad
things will happen if a notifier in path sent a BAD notify code.

Q: I don't see my action being called for all CPUs already up and running?
A: Yes, CPU notifiers are called only when new CPUs are on-lined or offlined.
   If you need to perform some action for each cpu already in the system, then
   do this:

	for_each_online_cpu(i) {
		foobar_cpu_callback(&foobar_cpu_notifier, CPU_UP_PREPARE, i);
319 320
		foobar_cpu_callback(&foobar_cpu_notifier, CPU_ONLINE, i);

322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365
   However, if you want to register a hotplug callback, as well as perform
   some initialization for CPUs that are already online, then do this:

   Version 1: (Correct)


		for_each_online_cpu(i) {
					    CPU_UP_PREPARE, i);
					    CPU_ONLINE, i);

	/* Note the use of the double underscored version of the API */


   Note that the following code is *NOT* the right way to achieve this,
   because it is prone to an ABBA deadlock between the cpu_add_remove_lock
   and the cpu_hotplug.lock.

   Version 2: (Wrong!)


		for_each_online_cpu(i) {
					    CPU_UP_PREPARE, i);
					    CPU_ONLINE, i);



    So always use the first version shown above when you want to register
    callbacks as well as initialize the already online CPUs.

366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386
Q: If i would like to develop cpu hotplug support for a new architecture,
   what do i need at a minimum?
A: The following are what is required for CPU hotplug infrastructure to work

    - Make sure you have an entry in Kconfig to enable CONFIG_HOTPLUG_CPU
    - __cpu_up()        - Arch interface to bring up a CPU
    - __cpu_disable()   - Arch interface to shutdown a CPU, no more interrupts
                          can be handled by the kernel after the routine
                          returns. Including local APIC timers etc are
     - __cpu_die()      - This actually supposed to ensure death of the CPU.
                          Actually look at some example code in other arch
                          that implement CPU hotplug. The processor is taken
                          down from the idle() loop for that specific
                          architecture. __cpu_die() typically waits for some
                          per_cpu state to be set, to ensure the processor
                          dead routine is called to be sure positively.

Q: I need to ensure that a particular cpu is not removed when there is some
   work specific to this cpu is in progress.
387 388 389
A: There are two ways.  If your code can be run in interrupt context, use
   smp_call_function_single(), otherwise use work_on_cpu().  Note that
   work_on_cpu() is slow, and can fail due to out of memory:

391 392
	int my_func_on_cpu(int cpu)
393 394 395 396 397 398 399 400 401 402 403 404 405 406
		int err;
		if (!cpu_online(cpu))
			err = -EINVAL;
			err = work_on_cpu(cpu, __my_func_on_cpu, NULL);
			smp_call_function_single(cpu, __my_func_on_cpu, &err,
		return err;
407 408 409 410 411 412 413 414 415 416 417

Q: How do we determine how many CPUs are available for hotplug.
A: There is no clear spec defined way from ACPI that can give us that
   information today. Based on some input from Natalie of Unisys,
   that the ACPI MADT (Multiple APIC Description Tables) marks those possible
   CPUs in a system with disabled status.

   Andi implemented some simple heuristics that count the number of disabled
   CPUs in MADT as hotpluggable CPUS.  In the case there are no disabled CPUS
   we assume 1/2 the number of CPUs currently present can be hotplugged.

418 419 420 421
   Caveat: ACPI MADT can only provide 256 entries in systems with only ACPI 2.0c
   or earlier ACPI version supported, because the apicid field in MADT is only
   8 bits. From ACPI 3.0, this limitation was removed since the apicid field
   was extended to 32 bits with x2APIC introduced.
422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452

User Space Notification

Hotplug support for devices is common in Linux today. Its being used today to
support automatic configuration of network, usb and pci devices. A hotplug
event can be used to invoke an agent script to perform the configuration task.

You can add /etc/hotplug/cpu.agent to handle hotplug notification user space

	# $Id: cpu.agent
	# Kernel hotplug params include:
	#ACTION=%s [online or offline]
	cd /etc/hotplug
	. ./hotplug.functions

	case $ACTION in
			echo `date` ":cpu.agent" add cpu >> /tmp/hotplug.txt
			echo `date` ":cpu.agent" remove cpu >>/tmp/hotplug.txt
			debug_mesg CPU $ACTION event not supported
        exit 1