lkml.org 
[lkml]   [2010]   [Jun]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [linux-pm] suspend blockers & Android integration
On Mon, 7 Jun 2010, Arve Hjønnevåg wrote:

> The patch that modifies evdev (posted in this patchset) uses an ioctl
> to enable the suspend blocker. Not all input devices are used for
> wakeup events and those don't need to block suspend.

But you do have a 1-1 correspondence, right? That is, the input
devices that are used for wakeup events are exactly the ones that block
suspend?


> If you read an event that occurred after you blocked the task
> freezing, then tasks will never get frozen again (until more events
> occur). I think my original description was less confusing, but it
> seems you got completely distracted by my use of block and unblock
> suspend when referring to the user space api.

I still find your wording a little confusing. Task freezing can be
prevented (a more accurate term than "blocked") by two kinds of things:
a suspend blocker or an "active" wakeup source. I'm not sure which
kind you mean here.

> It has an indirect connection. You report a wakeup event when it
> occurs, but clear it when user space calls an api before reading the
> event. So:

Yes, that's right.

> Wakeup event occurs, and the driver:
> - report wakeup event type A
> - queue event for delivery to user-space

That's not really two distinct steps. Queuing the event for delivery
to userspace involves waking up any tasks that are waiting to read the
device file; that action (calling wake_up_all() or whatever the driver
does) is how the event gets reported.

> User space wakes up:
> - Calls api to block task freezing for event type A

Again, that's a confusing way of putting it. The API you're referring
to is simply the function that activates a suspend blocker. It does
prevent task freezing, but you shouldn't say it prevents freezing for
event type A. More like the other way around: In addition to
preventing freezing, the function tells the power manager that event
type A should no longer be considered active. Thus, in a sense it
_stops_ event type A from preventing freezing.

> Another wakeup event occurs, and the driver:
> - report wakeup event type A
> - queue event for delivery to user-space

Same as above.

> User space continues:
> - Read events
> - Wait for more events
>
> Result: Task are not frozen again.

Because the suspend blocker was never deactivated. The same thing
happens with wakelocks: If a task activates a wakelock and never
deactivates it, the system won't go into opportunistic suspend again.

Here's how my scheme is meant to work:

Wakeup event for input device A occurs.

A's driver adds an entry to the input device queue and
(if the queue was empty) does wake_up_all() on the device
file's wait_queue.

The PM process returns from poll() and sees that device
file A is now readable, so it adds A to its list of active
sources and unfreezes userspace.

Some other process sees that device file A is now readable,
so it activates a suspend blocker and reads events from A.

When the PM process receives the request to activate the
suspend blocker, it removes A from its list of active
sources. But it doesn't freeze userspace yet, because now
a suspend blocker is active.

The other process consumes events from A and does other
stuff. Maybe more input data arrives while this is happening
and the process reads it. Eventually the process decides to
deactivate the suspend blocker, perhaps when no more data
is available from the device file, perhaps not.

When the PM process receives the request to deactivate the
suspend blocker, it sees that now there are no active
sources and no active suspend blockers. Therefore it
freezes userspace and does a big poll() on all possible
sources. (If there are still events on the input device
queue, the poll() returns immediately.)

Rinse and repeat.

I don't see any dangerous races there. The scheme can be made a little
more efficient by having the PM process do another poll() (with 0
timeout) just before freezing userspace; if the result indicates that a
source is active then the freezing and unfreezing can be skipped.

The big assumption here is that a user process never consumes wakeup
events without first activating a suspend blocker. This seems like a
reasonable assumption, but we can work around it if necessary.

> >> It seems you would need a way to pass the wakeup source id to use from
> >> user space to the driver and for this to work
> >
> > No, nothing needs to be passed from userspace to the kernel.  However
> > the source ID (or a set of source IDs) does need to be passed to the
> > power manager process, probably when the suspend blocker is created.
> >
>
> Then the source id need to be passed from the kernel to user-space.

A source ID is a file descriptor. File descriptors are passed from the
kernel to userspace whenever a file is opened; I can't deny it. And
they are passed back to the kernel as part of the read() and poll()
system calls. Is that what you mean?

> No, that is not the unclear part. What is unclear to me is where the
> source IDs come from. Are they static and hardcoded in the driver and
> user-space, or are they passed between the driver and user-space
> client?

They are not static; they are file descriptors. I guess this should
have been made more clear originally, but this is still pretty new to
me too.

> I don't understand how you are planning to ensure that the driver and
> user-space code that consumes the real event use the same source id.

How can it be otherwise? The userspace code consumes the event by
reading from the device file. In order to do so, it has to use the
same file descriptor it received when it opened the device file
originally.

> The biggest problem I have with it though is that you have created a
> new race condition between reporting that a wakeup event has occurred
> and processing of the real event.

There is no race. The driver reports an event has occurred by making
the data available to be read from the device file, and the event is
processed by reading it from the device file (or at least, that's the
first step in processing the event).


There's one other thing worth mentioning. All along I've been talking
about a power manager process that coordinates all these activities.
In theory there's no reason that process couldn't be implemented as a
kernel thread. This would improve efficiency by reducing the number of
context switches, and it would change IPC calls into plain system
calls.

If you did implement it that way, it could be done as a standalone
kernel module, totally noninvasive. It would not need to be part of
the vanilla kernel and nobody would object to it.

Alan Stern

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2010-06-08 16:53    [from the cache]
©2003-2011 Jasper Spaans