linux/Documentation/cpu-hotplug.txt

408 lines
16 KiB
Plaintext
Raw Normal View History

CPU hotplug Support in Linux(tm) Kernel
Maintainers:
CPU Hotplug Core:
Rusty Russell <rusty@rustcorp.com.au>
Srivatsa Vaddagiri <vatsa@in.ibm.com>
i386:
Zwane Mwaikambo <zwanem@gmail.com>
ppc64:
Nathan Lynch <nathanl@austin.ibm.com>
Joel Schopp <jschopp@austin.ibm.com>
ia64/x86_64:
Ashok Raj <ashok.raj@intel.com>
s390:
Heiko Carstens <heiko.carstens@de.ibm.com>
Authors: Ashok Raj <ashok.raj@intel.com>
Lots of feedback: Nathan Lynch <nathanl@austin.ibm.com>,
Joel Schopp <jschopp@austin.ibm.com>
Introduction
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
powerpc/pseries: Add hooks to put the CPU into an appropriate offline state When a CPU is offlined on POWER currently, we call rtas_stop_self() and hand the CPU back to the resource pool. This path is used for DLPAR which will cause a change in the LPAR configuration which will be visible outside. This patch changes the default state a CPU is put into when it is offlined. On platforms which support ceding the processor to the hypervisor with latency hint specifier value, during a cpu offline operation, instead of calling rtas_stop_self(), we cede the vCPU to the hypervisor while passing a latency hint specifier value. The Hypervisor can use this hint to provide better energy savings. Also, during the offline operation, the control of the vCPU remains with the LPAR as oppposed to returning it to the resource pool. The patch achieves this by creating an infrastructure to set the preferred_offline_state() which can be either - CPU_STATE_OFFLINE: which is the current behaviour of calling rtas_stop_self() - CPU_STATE_INACTIVE: which cedes the vCPU to the hypervisor with the latency hint specifier. The codepath which wants to perform a DLPAR operation can set the preferred_offline_state() of a CPU to CPU_STATE_OFFLINE before invoking cpu_down(). The patch also provides a boot-time command line argument to disable/enable CPU_STATE_INACTIVE. Signed-off-by: Gautham R Shenoy <ego@in.ibm.com> Signed-off-by: Nathan Fontenot <nfont@austin.ibm.com> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2009-10-30 03:22:53 +08:00
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
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
cpu_possible_mask.
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
constant even if the machine gets rebooted.
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
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()
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
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.
Never use anything other than cpumask_t to represent bitmap of CPUs.
#include <linux/cpumask.h>
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_mask(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.
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
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?
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.
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.
....
none on /sys type sysfs (rw)
....
If this is not mounted, do the following.
#mkdir /sysfs
#mount -t sysfs sys /sys
Now you should see entries for all present cpu, the following is an example
in a 8-way system.
#pwd
#/sys/devices/system/cpu
#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
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.
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
the state as 0 when a cpu if offline and 1 when its online.
#To display the current cpu state.
#cat /sys/devices/system/cpu/cpuX/online
Q: Why can't i remove CPU0 on some systems?
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.
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
cpu0_hotplug.
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
continue.
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
machines.
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>
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
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
progress
hotplug cpu: migrate a task within its cpuset When a cpu is disabled, move_task_off_dead_cpu() is called for tasks that have been running on that cpu. Currently, such a task is migrated: 1) to any cpu on the same node as the disabled cpu, which is both online and among that task's cpus_allowed 2) to any cpu which is both online and among that task's cpus_allowed It is typical of a multithreaded application running on a large NUMA system to have its tasks confined to a cpuset so as to cluster them near the memory that they share. Furthermore, it is typical to explicitly place such a task on a specific cpu in that cpuset. And in that case the task's cpus_allowed includes only a single cpu. This patch would insert a preference to migrate such a task to some cpu within its cpuset (and set its cpus_allowed to its entire cpuset). With this patch, migrate the task to: 1) to any cpu on the same node as the disabled cpu, which is both online and among that task's cpus_allowed 2) to any online cpu within the task's cpuset 3) to any cpu which is both online and among that task's cpus_allowed In order to do this, move_task_off_dead_cpu() must make a call to cpuset_cpus_allowed_locked(), a new subset of cpuset_cpus_allowed(), that will not block. (name change - per Oleg's suggestion) Calls are made to cpuset_lock() and cpuset_unlock() in migration_call() to set the cpuset mutex during the whole migrate_live_tasks() and migrate_dead_tasks() procedure. [akpm@linux-foundation.org: build fix] [pj@sgi.com: Fix indentation and spacing] Signed-off-by: Cliff Wickman <cpw@sgi.com> Cc: Oleg Nesterov <oleg@tv-sign.ru> Cc: Christoph Lameter <clameter@sgi.com> Cc: Paul Jackson <pj@sgi.com> Cc: Ingo Molnar <mingo@elte.hu> Signed-off-by: Paul Jackson <pj@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-19 14:40:46 +08:00
- 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.
- 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
CPU_DEAD (or CPU_DEAD_FROZEN if tasks are frozen due to a suspend while the
CPU is being offlined).
"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>
static int foobar_cpu_callback(struct notifier_block *nfb,
unsigned long action, void *hcpu)
{
unsigned int cpu = (unsigned long)hcpu;
switch (action) {
case CPU_ONLINE:
case CPU_ONLINE_FROZEN:
foobar_online_action(cpu);
break;
case CPU_DEAD:
case CPU_DEAD_FROZEN:
foobar_dead_action(cpu);
break;
}
return NOTIFY_OK;
}
static struct notifier_block foobar_cpu_notifer =
{
.notifier_call = foobar_cpu_callback,
};
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
register_cpu_notifier(&foobar_cpu_notifier);
For the second case, you should add the following to your init function
register_hotcpu_notifier(&foobar_cpu_notifier);
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
for_each_online_cpu(i) {
foobar_cpu_callback(&foobar_cpu_notifier, CPU_UP_PREPARE, i);
foobar_cpu_callback(&foobar_cpu_notifier, CPU_ONLINE, i);
}
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
correctly.
- 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
shutdown.
- __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.
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:
int my_func_on_cpu(int cpu)
{
int err;
get_online_cpus();
if (!cpu_online(cpu))
err = -EINVAL;
else
#if NEEDS_BLOCKING
err = work_on_cpu(cpu, __my_func_on_cpu, NULL);
#else
smp_call_function_single(cpu, __my_func_on_cpu, &err,
true);
#endif
put_online_cpus();
return err;
}
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.
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.
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
scripts.
#!/bin/bash
# $Id: cpu.agent
# Kernel hotplug params include:
#ACTION=%s [online or offline]
#DEVPATH=%s
#
cd /etc/hotplug
. ./hotplug.functions
case $ACTION in
online)
echo `date` ":cpu.agent" add cpu >> /tmp/hotplug.txt
;;
offline)
echo `date` ":cpu.agent" remove cpu >>/tmp/hotplug.txt
;;
*)
debug_mesg CPU $ACTION event not supported
exit 1
;;
esac