Power Management for USB

                 Alan Stern <stern@rowland.harvard.edu>

                            November 10, 2009



        What is Power Management?
        -------------------------

Power Management (PM) is the practice of saving energy by suspending
parts of a computer system when they aren't being used.  While a
component is "suspended" it is in a nonfunctional low-power state; it
might even be turned off completely.  A suspended component can be
"resumed" (returned to a functional full-power state) when the kernel
needs to use it.  (There also are forms of PM in which components are
placed in a less functional but still usable state instead of being
suspended; an example would be reducing the CPU's clock rate.  This
document will not discuss those other forms.)

When the parts being suspended include the CPU and most of the rest of
the system, we speak of it as a "system suspend".  When a particular
device is turned off while the system as a whole remains running, we
call it a "dynamic suspend" (also known as a "runtime suspend" or
"selective suspend").  This document concentrates mostly on how
dynamic PM is implemented in the USB subsystem, although system PM is
covered to some extent (see Documentation/power/*.txt for more
information about system PM).

Note: Dynamic PM support for USB is present only if the kernel was
built with CONFIG_USB_SUSPEND enabled.  System PM support is present
only if the kernel was built with CONFIG_SUSPEND or CONFIG_HIBERNATION
enabled.


        What is Remote Wakeup?
        ----------------------

When a device has been suspended, it generally doesn't resume until
the computer tells it to.  Likewise, if the entire computer has been
suspended, it generally doesn't resume until the user tells it to, say
by pressing a power button or opening the cover.

However some devices have the capability of resuming by themselves, or
asking the kernel to resume them, or even telling the entire computer
to resume.  This capability goes by several names such as "Wake On
LAN"; we will refer to it generically as "remote wakeup".  When a
device is enabled for remote wakeup and it is suspended, it may resume
itself (or send a request to be resumed) in response to some external
event.  Examples include a suspended keyboard resuming when a key is
pressed, or a suspended USB hub resuming when a device is plugged in.


        When is a USB device idle?
        --------------------------

A device is idle whenever the kernel thinks it's not busy doing
anything important and thus is a candidate for being suspended.  The
exact definition depends on the device's driver; drivers are allowed
to declare that a device isn't idle even when there's no actual
communication taking place.  (For example, a hub isn't considered idle
unless all the devices plugged into that hub are already suspended.)
In addition, a device isn't considered idle so long as a program keeps
its usbfs file open, whether or not any I/O is going on.

If a USB device has no driver, its usbfs file isn't open, and it isn't
being accessed through sysfs, then it definitely is idle.


        Forms of dynamic PM
        -------------------

Dynamic suspends occur when the kernel decides to suspend an idle
device.  This is called "autosuspend" for short.  In general, a device
won't be autosuspended unless it has been idle for some minimum period
of time, the so-called idle-delay time.

Of course, nothing the kernel does on its own initiative should
prevent the computer or its devices from working properly.  If a
device has been autosuspended and a program tries to use it, the
kernel will automatically resume the device (autoresume).  For the
same reason, an autosuspended device will usually have remote wakeup
enabled, if the device supports remote wakeup.

It is worth mentioning that many USB drivers don't support
autosuspend.  In fact, at the time of this writing (Linux 2.6.23) the
only drivers which do support it are the hub driver, kaweth, asix,
usblp, usblcd, and usb-skeleton (which doesn't count).  If a
non-supporting driver is bound to a device, the device won't be
autosuspended.  In effect, the kernel pretends the device is never
idle.

We can categorize power management events in two broad classes:
external and internal.  External events are those triggered by some
agent outside the USB stack: system suspend/resume (triggered by
userspace), manual dynamic resume (also triggered by userspace), and
remote wakeup (triggered by the device).  Internal events are those
triggered within the USB stack: autosuspend and autoresume.  Note that
all dynamic suspend events are internal; external agents are not
allowed to issue dynamic suspends.


        The user interface for dynamic PM
        ---------------------------------

The user interface for controlling dynamic PM is located in the power/
subdirectory of each USB device's sysfs directory, that is, in
/sys/bus/usb/devices/.../power/ where "..." is the device's ID.  The
relevant attribute files are: wakeup, level, and autosuspend.

        power/wakeup

                This file is empty if the device does not support
                remote wakeup.  Otherwise the file contains either the
                word "enabled" or the word "disabled", and you can
                write those words to the file.  The setting determines
                whether or not remote wakeup will be enabled when the
                device is next suspended.  (If the setting is changed
                while the device is suspended, the change won't take
                effect until the following suspend.)

        power/level

                This file contains one of two words: "on" or "auto".
                You can write those words to the file to change the
                device's setting.

                "on" means that the device should be resumed and
                autosuspend is not allowed.  (Of course, system
                suspends are still allowed.)

                "auto" is the normal state in which the kernel is
                allowed to autosuspend and autoresume the device.

                (In kernels up to 2.6.32, you could also specify
                "suspend", meaning that the device should remain
                suspended and autoresume was not allowed.  This
                setting is no longer supported.)

        power/autosuspend

                This file contains an integer value, which is the
                number of seconds the device should remain idle before
                the kernel will autosuspend it (the idle-delay time).
                The default is 2.  0 means to autosuspend as soon as
                the device becomes idle, and negative values mean
                never to autosuspend.  You can write a number to the
                file to change the autosuspend idle-delay time.

Writing "-1" to power/autosuspend and writing "on" to power/level do
essentially the same thing -- they both prevent the device from being
autosuspended.  Yes, this is a redundancy in the API.

(In 2.6.21 writing "0" to power/autosuspend would prevent the device
from being autosuspended; the behavior was changed in 2.6.22.  The
power/autosuspend attribute did not exist prior to 2.6.21, and the
power/level attribute did not exist prior to 2.6.22.)


        Changing the default idle-delay time
        ------------------------------------

The default autosuspend idle-delay time is controlled by a module
parameter in usbcore.  You can specify the value when usbcore is
loaded.  For example, to set it to 5 seconds instead of 2 you would
do:

        modprobe usbcore autosuspend=5

Equivalently, you could add to /etc/modprobe.conf a line saying:

        options usbcore autosuspend=5

Some distributions load the usbcore module very early during the boot
process, by means of a program or script running from an initramfs
image.  To alter the parameter value you would have to rebuild that
image.

If usbcore is compiled into the kernel rather than built as a loadable
module, you can add

        usbcore.autosuspend=5

to the kernel's boot command line.

Finally, the parameter value can be changed while the system is
running.  If you do:

        echo 5 >/sys/module/usbcore/parameters/autosuspend

then each new USB device will have its autosuspend idle-delay
initialized to 5.  (The idle-delay values for already existing devices
will not be affected.)

Setting the initial default idle-delay to -1 will prevent any
autosuspend of any USB device.  This is a simple alternative to
disabling CONFIG_USB_SUSPEND and rebuilding the kernel, and it has the
added benefit of allowing you to enable autosuspend for selected
devices.


        Warnings
        --------

The USB specification states that all USB devices must support power
management.  Nevertheless, the sad fact is that many devices do not
support it very well.  You can suspend them all right, but when you
try to resume them they disconnect themselves from the USB bus or
they stop working entirely.  This seems to be especially prevalent
among printers and scanners, but plenty of other types of device have
the same deficiency.

For this reason, by default the kernel disables autosuspend (the
power/level attribute is initialized to "on") for all devices other
than hubs.  Hubs, at least, appear to be reasonably well-behaved in
this regard.

(In 2.6.21 and 2.6.22 this wasn't the case.  Autosuspend was enabled
by default for almost all USB devices.  A number of people experienced
problems as a result.)

This means that non-hub devices won't be autosuspended unless the user
or a program explicitly enables it.  As of this writing there aren't
any widespread programs which will do this; we hope that in the near
future device managers such as HAL will take on this added
responsibility.  In the meantime you can always carry out the
necessary operations by hand or add them to a udev script.  You can
also change the idle-delay time; 2 seconds is not the best choice for
every device.

Sometimes it turns out that even when a device does work okay with
autosuspend there are still problems.  For example, there are
experimental patches adding autosuspend support to the usbhid driver,
which manages keyboards and mice, among other things.  Tests with a
number of keyboards showed that typing on a suspended keyboard, while
causing the keyboard to do a remote wakeup all right, would
nonetheless frequently result in lost keystrokes.  Tests with mice
showed that some of them would issue a remote-wakeup request in
response to button presses but not to motion, and some in response to
neither.

The kernel will not prevent you from enabling autosuspend on devices
that can't handle it.  It is even possible in theory to damage a
device by suspending it at the wrong time -- for example, suspending a
USB hard disk might cause it to spin down without parking the heads.
(Highly unlikely, but possible.)  Take care.


        The driver interface for Power Management
        -----------------------------------------

The requirements for a USB driver to support external power management
are pretty modest; the driver need only define

        .suspend
        .resume
        .reset_resume

methods in its usb_driver structure, and the reset_resume method is
optional.  The methods' jobs are quite simple:

        The suspend method is called to warn the driver that the
        device is going to be suspended.  If the driver returns a
        negative error code, the suspend will be aborted.  Normally
        the driver will return 0, in which case it must cancel all
        outstanding URBs (usb_kill_urb()) and not submit any more.

        The resume method is called to tell the driver that the
        device has been resumed and the driver can return to normal
        operation.  URBs may once more be submitted.

        The reset_resume method is called to tell the driver that
        the device has been resumed and it also has been reset.
        The driver should redo any necessary device initialization,
        since the device has probably lost most or all of its state
        (although the interfaces will be in the same altsettings as
        before the suspend).

If the device is disconnected or powered down while it is suspended,
the disconnect method will be called instead of the resume or
reset_resume method.  This is also quite likely to happen when
waking up from hibernation, as many systems do not maintain suspend
current to the USB host controllers during hibernation.  (It's
possible to work around the hibernation-forces-disconnect problem by
using the USB Persist facility.)

The reset_resume method is used by the USB Persist facility (see
Documentation/usb/persist.txt) and it can also be used under certain
circumstances when CONFIG_USB_PERSIST is not enabled.  Currently, if a
device is reset during a resume and the driver does not have a
reset_resume method, the driver won't receive any notification about
the resume.  Later kernels will call the driver's disconnect method;
2.6.23 doesn't do this.

USB drivers are bound to interfaces, so their suspend and resume
methods get called when the interfaces are suspended or resumed.  In
principle one might want to suspend some interfaces on a device (i.e.,
force the drivers for those interface to stop all activity) without
suspending the other interfaces.  The USB core doesn't allow this; all
interfaces are suspended when the device itself is suspended and all
interfaces are resumed when the device is resumed.  It isn't possible
to suspend or resume some but not all of a device's interfaces.  The
closest you can come is to unbind the interfaces' drivers.


        The driver interface for autosuspend and autoresume
        ---------------------------------------------------

To support autosuspend and autoresume, a driver should implement all
three of the methods listed above.  In addition, a driver indicates
that it supports autosuspend by setting the .supports_autosuspend flag
in its usb_driver structure.  It is then responsible for informing the
USB core whenever one of its interfaces becomes busy or idle.  The
driver does so by calling these six functions:

        int  usb_autopm_get_interface(struct usb_interface *intf);
        void usb_autopm_put_interface(struct usb_interface *intf);
        int  usb_autopm_get_interface_async(struct usb_interface *intf);
        void usb_autopm_put_interface_async(struct usb_interface *intf);
        void usb_autopm_get_interface_no_resume(struct usb_interface *intf);
        void usb_autopm_put_interface_no_suspend(struct usb_interface *intf);

The functions work by maintaining a counter in the usb_interface
structure.  When intf->pm_usage_count is > 0 then the interface is
deemed to be busy, and the kernel will not autosuspend the interface's
device.  When intf->pm_usage_count is <= 0 then the interface is
considered to be idle, and the kernel may autosuspend the device.

(There is a similar pm_usage_count field in struct usb_device,
associated with the device itself rather than any of its interfaces.
This field is used only by the USB core.)

Drivers must not modify intf->pm_usage_count directly; its value
should be changed only be using the functions listed above.  Drivers
are responsible for insuring that the overall change to pm_usage_count
during their lifetime balances out to 0 (it may be necessary for the
disconnect method to call usb_autopm_put_interface() one or more times
to fulfill this requirement).  The first two routines use the PM mutex
in struct usb_device for mutual exclusion; drivers using the async
routines are responsible for their own synchronization and mutual
exclusion.

        usb_autopm_get_interface() increments pm_usage_count and
        attempts an autoresume if the new value is > 0 and the
        device is suspended.

        usb_autopm_put_interface() decrements pm_usage_count and
        attempts an autosuspend if the new value is <= 0 and the
        device isn't suspended.

        usb_autopm_get_interface_async() and
        usb_autopm_put_interface_async() do almost the same things as
        their non-async counterparts.  The differences are: they do
        not acquire the PM mutex, and they use a workqueue to do their
        jobs.  As a result they can be called in an atomic context,
        such as an URB's completion handler, but when they return the
        device will not generally not yet be in the desired state.

        usb_autopm_get_interface_no_resume() and
        usb_autopm_put_interface_no_suspend() merely increment or
        decrement the pm_usage_count value; they do not attempt to
        carry out an autoresume or an autosuspend.  Hence they can be
        called in an atomic context.

The conventional usage pattern is that a driver calls
usb_autopm_get_interface() in its open routine and
usb_autopm_put_interface() in its close or release routine.  But
other patterns are possible.

The autosuspend attempts mentioned above will often fail for one
reason or another.  For example, the power/level attribute might be
set to "on", or another interface in the same device might not be
idle.  This is perfectly normal.  If the reason for failure was that
the device hasn't been idle for long enough, a delayed workqueue
routine is automatically set up to carry out the operation when the
autosuspend idle-delay has expired.

Autoresume attempts also can fail, although failure would mean that
the device is no longer present or operating properly.  Unlike
autosuspend, there's no delay for an autoresume.


        Other parts of the driver interface
        -----------------------------------

Sometimes a driver needs to make sure that remote wakeup is enabled
during autosuspend.  For example, there's not much point
autosuspending a keyboard if the user can't cause the keyboard to do a
remote wakeup by typing on it.  If the driver sets
intf->needs_remote_wakeup to 1, the kernel won't autosuspend the
device if remote wakeup isn't available or has been disabled through
the power/wakeup attribute.  (If the device is already autosuspended,
though, setting this flag won't cause the kernel to autoresume it.
Normally a driver would set this flag in its probe method, at which
time the device is guaranteed not to be autosuspended.)

The synchronous usb_autopm_* routines have to run in a sleepable
process context; they must not be called from an interrupt handler or
while holding a spinlock.  In fact, the entire autosuspend mechanism
is not well geared toward interrupt-driven operation.  However there
is one thing a driver can do in an interrupt handler:

        usb_mark_last_busy(struct usb_device *udev);

This sets udev->last_busy to the current time.  udev->last_busy is the
field used for idle-delay calculations; updating it will cause any
pending autosuspend to be moved back.  The usb_autopm_* routines will
also set the last_busy field to the current time.

Calling urb_mark_last_busy() from within an URB completion handler is
subject to races: The kernel may have just finished deciding the
device has been idle for long enough but not yet gotten around to
calling the driver's suspend method.  The driver would have to be
responsible for synchronizing its suspend method with its URB
completion handler and causing the autosuspend to fail with -EBUSY if
an URB had completed too recently.

External suspend calls should never be allowed to fail in this way,
only autosuspend calls.  The driver can tell them apart by checking
the PM_EVENT_AUTO bit in the message.event argument to the suspend
method; this bit will be set for internal PM events (autosuspend) and
clear for external PM events.

Many of the ingredients in the autosuspend framework are oriented
towards interfaces: The usb_interface structure contains the
pm_usage_cnt field, and the usb_autopm_* routines take an interface
pointer as their argument.  But somewhat confusingly, a few of the
pieces (i.e., usb_mark_last_busy()) use the usb_device structure
instead.  Drivers need to keep this straight; they can call
interface_to_usbdev() to find the device structure for a given
interface.


        Locking requirements
        --------------------

All three suspend/resume methods are always called while holding the
usb_device's PM mutex.  For external events -- but not necessarily for
autosuspend or autoresume -- the device semaphore (udev->dev.sem) will
also be held.  This implies that external suspend/resume events are
mutually exclusive with calls to probe, disconnect, pre_reset, and
post_reset; the USB core guarantees that this is true of internal
suspend/resume events as well.

If a driver wants to block all suspend/resume calls during some
critical section, it can simply acquire udev->pm_mutex. Note that
calls to resume may be triggered indirectly. Block IO due to memory
allocations can make the vm subsystem resume a device. Thus while
holding this lock you must not allocate memory with GFP_KERNEL or
GFP_NOFS.

Alternatively, if the critical section might call some of the
usb_autopm_* routines, the driver can avoid deadlock by doing:

        down(&udev->dev.sem);
        rc = usb_autopm_get_interface(intf);

and at the end of the critical section:

        if (!rc)
                usb_autopm_put_interface(intf);
        up(&udev->dev.sem);

Holding the device semaphore will block all external PM calls, and the
usb_autopm_get_interface() will prevent any internal PM calls, even if
it fails.  (Exercise: Why?)

The rules for locking order are:

        Never acquire any device semaphore while holding any PM mutex.

        Never acquire udev->pm_mutex while holding the PM mutex for
        a device that isn't a descendant of udev.

In other words, PM mutexes should only be acquired going up the device
tree, and they should be acquired only after locking all the device
semaphores you need to hold.  These rules don't matter to drivers very
much; they usually affect just the USB core.

Still, drivers do need to be careful.  For example, many drivers use a
private mutex to synchronize their normal I/O activities with their
disconnect method.  Now if the driver supports autosuspend then it
must call usb_autopm_put_interface() from somewhere -- maybe from its
close method.  It should make the call while holding the private mutex,
since a driver shouldn't call any of the usb_autopm_* functions for an
interface from which it has been unbound.

But the usb_autpm_* routines always acquire the device's PM mutex, and
consequently the locking order has to be: private mutex first, PM
mutex second.  Since the suspend method is always called with the PM
mutex held, it mustn't try to acquire the private mutex.  It has to
synchronize with the driver's I/O activities in some other way.


        Interaction between dynamic PM and system PM
        --------------------------------------------

Dynamic power management and system power management can interact in
a couple of ways.

Firstly, a device may already be manually suspended or autosuspended
when a system suspend occurs.  Since system suspends are supposed to
be as transparent as possible, the device should remain suspended
following the system resume.  The 2.6.23 kernel obeys this principle
for manually suspended devices but not for autosuspended devices; they
do get resumed when the system wakes up.  (Presumably they will be
autosuspended again after their idle-delay time expires.)  In later
kernels this behavior will be fixed.

(There is an exception.  If a device would undergo a reset-resume
instead of a normal resume, and the device is enabled for remote
wakeup, then the reset-resume takes place even if the device was
already suspended when the system suspend began.  The justification is
that a reset-resume is a kind of remote-wakeup event.  Or to put it
another way, a device which needs a reset won't be able to generate
normal remote-wakeup signals, so it ought to be resumed immediately.)

Secondly, a dynamic power-management event may occur as a system
suspend is underway.  The window for this is short, since system
suspends don't take long (a few seconds usually), but it can happen.
For example, a suspended device may send a remote-wakeup signal while
the system is suspending.  The remote wakeup may succeed, which would
cause the system suspend to abort.  If the remote wakeup doesn't
succeed, it may still remain active and thus cause the system to
resume as soon as the system suspend is complete.  Or the remote
wakeup may fail and get lost.  Which outcome occurs depends on timing
and on the hardware and firmware design.