linux-sg2042/Documentation/power/suspend-and-cpuhotplug.txt

276 lines
12 KiB
Plaintext

Interaction of Suspend code (S3) with the CPU hotplug infrastructure
(C) 2011 Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
I. How does the regular CPU hotplug code differ from how the Suspend-to-RAM
infrastructure uses it internally? And where do they share common code?
Well, a picture is worth a thousand words... So ASCII art follows :-)
[This depicts the current design in the kernel, and focusses only on the
interactions involving the freezer and CPU hotplug and also tries to explain
the locking involved. It outlines the notifications involved as well.
But please note that here, only the call paths are illustrated, with the aim
of describing where they take different paths and where they share code.
What happens when regular CPU hotplug and Suspend-to-RAM race with each other
is not depicted here.]
On a high level, the suspend-resume cycle goes like this:
|Freeze| -> |Disable nonboot| -> |Do suspend| -> |Enable nonboot| -> |Thaw |
|tasks | | cpus | | | | cpus | |tasks|
More details follow:
Suspend call path
-----------------
Write 'mem' to
/sys/power/state
sysfs file
|
v
Acquire pm_mutex lock
|
v
Send PM_SUSPEND_PREPARE
notifications
|
v
Freeze tasks
|
|
v
disable_nonboot_cpus()
/* start */
|
v
Acquire cpu_add_remove_lock
|
v
Iterate over CURRENTLY
online CPUs
|
|
| ----------
v | L
======> _cpu_down() |
| [This takes cpuhotplug.lock |
Common | before taking down the CPU |
code | and releases it when done] | O
| While it is at it, notifications |
| are sent when notable events occur, |
======> by running all registered callbacks. |
| | O
| |
| |
v |
Note down these cpus in | P
frozen_cpus mask ----------
|
v
Disable regular cpu hotplug
by setting cpu_hotplug_disabled=1
|
v
Release cpu_add_remove_lock
|
v
/* disable_nonboot_cpus() complete */
|
v
Do suspend
Resuming back is likewise, with the counterparts being (in the order of
execution during resume):
* enable_nonboot_cpus() which involves:
| Acquire cpu_add_remove_lock
| Reset cpu_hotplug_disabled to 0, thereby enabling regular cpu hotplug
| Call _cpu_up() [for all those cpus in the frozen_cpus mask, in a loop]
| Release cpu_add_remove_lock
v
* thaw tasks
* send PM_POST_SUSPEND notifications
* Release pm_mutex lock.
It is to be noted here that the pm_mutex lock is acquired at the very
beginning, when we are just starting out to suspend, and then released only
after the entire cycle is complete (i.e., suspend + resume).
Regular CPU hotplug call path
-----------------------------
Write 0 (or 1) to
/sys/devices/system/cpu/cpu*/online
sysfs file
|
|
v
cpu_down()
|
v
Acquire cpu_add_remove_lock
|
v
If cpu_hotplug_disabled is 1
return gracefully
|
|
v
======> _cpu_down()
| [This takes cpuhotplug.lock
Common | before taking down the CPU
code | and releases it when done]
| While it is at it, notifications
| are sent when notable events occur,
======> by running all registered callbacks.
|
|
v
Release cpu_add_remove_lock
[That's it!, for
regular CPU hotplug]
So, as can be seen from the two diagrams (the parts marked as "Common code"),
regular CPU hotplug and the suspend code path converge at the _cpu_down() and
_cpu_up() functions. They differ in the arguments passed to these functions,
in that during regular CPU hotplug, 0 is passed for the 'tasks_frozen'
argument. But during suspend, since the tasks are already frozen by the time
the non-boot CPUs are offlined or onlined, the _cpu_*() functions are called
with the 'tasks_frozen' argument set to 1.
[See below for some known issues regarding this.]
Important files and functions/entry points:
------------------------------------------
kernel/power/process.c : freeze_processes(), thaw_processes()
kernel/power/suspend.c : suspend_prepare(), suspend_enter(), suspend_finish()
kernel/cpu.c: cpu_[up|down](), _cpu_[up|down](), [disable|enable]_nonboot_cpus()
II. What are the issues involved in CPU hotplug?
-------------------------------------------
There are some interesting situations involving CPU hotplug and microcode
update on the CPUs, as discussed below:
[Please bear in mind that the kernel requests the microcode images from
userspace, using the request_firmware() function defined in
drivers/base/firmware_class.c]
a. When all the CPUs are identical:
This is the most common situation and it is quite straightforward: we want
to apply the same microcode revision to each of the CPUs.
To give an example of x86, the collect_cpu_info() function defined in
arch/x86/kernel/microcode_core.c helps in discovering the type of the CPU
and thereby in applying the correct microcode revision to it.
But note that the kernel does not maintain a common microcode image for the
all CPUs, in order to handle case 'b' described below.
b. When some of the CPUs are different than the rest:
In this case since we probably need to apply different microcode revisions
to different CPUs, the kernel maintains a copy of the correct microcode
image for each CPU (after appropriate CPU type/model discovery using
functions such as collect_cpu_info()).
c. When a CPU is physically hot-unplugged and a new (and possibly different
type of) CPU is hot-plugged into the system:
In the current design of the kernel, whenever a CPU is taken offline during
a regular CPU hotplug operation, upon receiving the CPU_DEAD notification
(which is sent by the CPU hotplug code), the microcode update driver's
callback for that event reacts by freeing the kernel's copy of the
microcode image for that CPU.
Hence, when a new CPU is brought online, since the kernel finds that it
doesn't have the microcode image, it does the CPU type/model discovery
afresh and then requests the userspace for the appropriate microcode image
for that CPU, which is subsequently applied.
For example, in x86, the mc_cpu_callback() function (which is the microcode
update driver's callback registered for CPU hotplug events) calls
microcode_update_cpu() which would call microcode_init_cpu() in this case,
instead of microcode_resume_cpu() when it finds that the kernel doesn't
have a valid microcode image. This ensures that the CPU type/model
discovery is performed and the right microcode is applied to the CPU after
getting it from userspace.
d. Handling microcode update during suspend/hibernate:
Strictly speaking, during a CPU hotplug operation which does not involve
physically removing or inserting CPUs, the CPUs are not actually powered
off during a CPU offline. They are just put to the lowest C-states possible.
Hence, in such a case, it is not really necessary to re-apply microcode
when the CPUs are brought back online, since they wouldn't have lost the
image during the CPU offline operation.
This is the usual scenario encountered during a resume after a suspend.
However, in the case of hibernation, since all the CPUs are completely
powered off, during restore it becomes necessary to apply the microcode
images to all the CPUs.
[Note that we don't expect someone to physically pull out nodes and insert
nodes with a different type of CPUs in-between a suspend-resume or a
hibernate/restore cycle.]
In the current design of the kernel however, during a CPU offline operation
as part of the suspend/hibernate cycle (the CPU_DEAD_FROZEN notification),
the existing copy of microcode image in the kernel is not freed up.
And during the CPU online operations (during resume/restore), since the
kernel finds that it already has copies of the microcode images for all the
CPUs, it just applies them to the CPUs, avoiding any re-discovery of CPU
type/model and the need for validating whether the microcode revisions are
right for the CPUs or not (due to the above assumption that physical CPU
hotplug will not be done in-between suspend/resume or hibernate/restore
cycles).
III. Are there any known problems when regular CPU hotplug and suspend race
with each other?
Yes, they are listed below:
1. When invoking regular CPU hotplug, the 'tasks_frozen' argument passed to
the _cpu_down() and _cpu_up() functions is *always* 0.
This might not reflect the true current state of the system, since the
tasks could have been frozen by an out-of-band event such as a suspend
operation in progress. Hence, it will lead to wrong notifications being
sent during the cpu online/offline events (eg, CPU_ONLINE notification
instead of CPU_ONLINE_FROZEN) which in turn will lead to execution of
inappropriate code by the callbacks registered for such CPU hotplug events.
2. If a regular CPU hotplug stress test happens to race with the freezer due
to a suspend operation in progress at the same time, then we could hit the
situation described below:
* A regular cpu online operation continues its journey from userspace
into the kernel, since the freezing has not yet begun.
* Then freezer gets to work and freezes userspace.
* If cpu online has not yet completed the microcode update stuff by now,
it will now start waiting on the frozen userspace in the
TASK_UNINTERRUPTIBLE state, in order to get the microcode image.
* Now the freezer continues and tries to freeze the remaining tasks. But
due to this wait mentioned above, the freezer won't be able to freeze
the cpu online hotplug task and hence freezing of tasks fails.
As a result of this task freezing failure, the suspend operation gets
aborted.