PM: Opportunistic suspend support.
in [Kernel]
|
Prev: direct-io: do not merge logically non-contiguous requests
Next: [PATCH] linux/elfcore.h: hide kernel functions
From: Arve Hjønnevåg on 25 May 2010 20:00 2010/5/25 Dmitry Torokhov <dmitry.torokhov(a)gmail.com>: > On Tue, May 25, 2010 at 04:13:35PM -0700, Arve Hj�nnev�g wrote: >> 2010/5/25 Dmitry Torokhov <dmitry.torokhov(a)gmail.com>: >> > On Tue, May 25, 2010 at 03:37:48PM -0700, Arve Hj�nnev�g wrote: >> >> 2010/5/25 Dmitry Torokhov <dmitry.torokhov(a)gmail.com>: >> >> > On Tue, May 25, 2010 at 03:23:23PM -0700, Arve Hj�nnev�g wrote: >> >> >> On Tue, May 25, 2010 at 11:47 AM, Dmitry Torokhov >> >> >> <dmitry.torokhov(a)gmail.com> wrote: >> >> >> > On Tue, May 25, 2010 at 02:35:17PM -0400, Alan Stern wrote: >> >> >> >> On Tue, 25 May 2010, Dmitry Torokhov wrote: >> >> >> >> >> >> >> >> > > Here's the scenario: >> >> >> >> > > >> >> >> >> > > The system is awake, and the user presses a key. The keyboard driver >> >> >> >> > > processes the keystroke and puts it in an input queue. �A user process >> >> >> >> > > reads it from the event queue, thereby emptying the queue. >> >> >> >> > > >> >> >> >> > > At that moment, the system decides to go into opportunistic suspend. >> >> >> >> > > Since the input queue is empty, there's nothing to stop it. �As the >> >> >> >> > > first step, userspace is frozen -- before the process has a chance to >> >> >> >> > > do anything with the keystroke it just read. �As a result, the system >> >> >> >> > > stays asleep until something else wakes it up, even though the >> >> >> >> > > keystroke was important and should have prevented it from sleeping. >> >> >> >> > > >> >> >> >> > > Suspend blockers protect against this scenario. �Here's how: >> >> >> >> > > >> >> >> >> > > The user process doesn't read the input queue directly; instead it >> >> >> >> > > does a select or poll. �When it sees there is data in the queue, it >> >> >> >> > > first acquires a suspend blocker and then reads the data. >> >> >> >> > > >> >> >> >> > > Now the system _can't_ go into opportunistic suspend, because a suspend >> >> >> >> > > blocker is active. �The user process can do whatever it wants with the >> >> >> >> > > keystroke. �When it is finished, it releases the suspend blocker and >> >> >> >> > > loops back to the select/poll call. >> >> >> >> > > >> >> >> >> > >> >> >> >> > What you describe can be done in userspace though, via a "suspend manager" >> >> >> >> > process. Tasks reading input events will post "busy" events to stop the >> >> >> >> > manager process from sending system into suspend. But this can be confined to >> >> >> >> > Android userspace, leaving the kernel as is (well, kernel needs to be modified >> >> >> >> > to not go into suspend with full queues, but that is using existing kernel >> >> >> >> > APIs). >> >> >> >> >> >> >> >> I think that could be made to work. �And it might remove the need for >> >> >> >> the userspace suspend-blocker API, which would be an advantage. �It >> >> >> >> could even remove the need for the opportunistic-suspend workqueue -- >> >> >> >> opportunistic suspends would be initiated by the "suspend manager" >> >> >> >> process instead of by the kernel. >> >> >> >> >> >> >> >> However you still have the issue of modifying the kernel drivers to >> >> >> >> disallow opportunistic suspend if their queues are non-empty. �Doing >> >> >> >> that is more or less equivalent to implementing kernel-level suspend >> >> >> >> blockers. �(The suspend blocker approach is slightly more efficient, >> >> >> >> because it will prevent a suspend from starting if a queue is >> >> >> >> non-empty, instead of allowing the suspend to start and then aborting >> >> >> >> it partway through.) >> >> >> >> >> >> >> >> Maybe I'm missing something here... �No doubt someone will point it out >> >> >> >> if I am. >> >> >> >> >> >> >> > >> >> >> > Well, from my perspective that would limit changes to the evdev driver >> >> >> > (well, limited input core plumbing will be needed) but that is using the >> >> >> > current PM infrastructure. The HW driver changes will be limited to what >> >> >> > you described "type 2" in your other e-mail. >> >> >> > >> >> >> > Also, not suspending while events are in progress) is probably >> >> >> > beneficial for platforms other than Android as well. So unless I am >> >> >> > missing something this sounds like a win. >> >> >> > >> >> >> >> >> >> How would this limit the changes you need in the evdev driver? It need >> >> >> to block suspend when there are unprocessed events in some queues. >> >> >> Suspend blockers gives you an api to do this, without it, you check >> >> >> the queues in your suspend hook and abort suspend if they are not >> >> >> empty. Without suspend blockers you have no api to signal that it is >> >> >> OK to suspend again, so you are forcing the thread that tried to >> >> >> suspend to poll until you stop aborting suspend. >> >> > >> >> > No, you do not need to poll. You just set a timeout (short or long, >> >> > depending on your needs) and if no userspace task blocked suspend >> >> > durng that time you attempt to initiate suspend from your manager >> >> > process. If it succeeds - good, if not that means that more events came >> >> > your way and you have to do it later. >> >> > >> >> >> >> How is that not polling? If the user is holding down a key, the keypad >> >> driver has to block suspend, and user space will try to suspend again >> >> and again and again... >> >> >> > >> > If your userpsace is that stupid - sure. However, you can: >> > >> > 1. Notify the suspend manager process that he rest of your userspace is >> > busy handling keystrokes so that it does not try to suspend while there >> > are events pending. >> >> You are missing the point. There are no event pending. The kernel >> reported the key down event, it was handled, but the keypad driver is >> still scanning to see if the user presses another key, > > Employ reasonable timeout. Timeout for what? Stop trying to suspend altogether, stop scanning for key changes, or a more "reasonable" poll interval? > >> or releases the >> currently held key. >> > > Userspace consumer should wait for the key release and retract "busy" > once event is received and handled. > No it should not. User-space does not know if the key is coming from a keypad driver that needs to actively scan the matrix while keys are pressed. >> > >> > 2. Wait a tiny bit after last application notified you that it finished >> > processing events. >> > >> > So basically the difference is that with in-kernel suspend blockers, >> > there is a tiny window where we haven't started the suspend yet but are >> > about to the driver has a chance to prevent entire system from starting >> > sleep. >> >> No, the difference is that if a driver needs to prevent suspend for an >> extended period of time, you don't have user space continuously >> polling to see if it can suspend. > > Why would a driver, on its own, prevent suspend for extended periods of > time? I think that the decision should originate from userspace, kernel > is here just to serve the requests. > A driver prevents suspend if suspend would prevent it from working. For instance, the gpio keypad matrix code prevents suspend when a key is help down, since it has to activly scan the keypad for changes. Only no-keys-pressed versus one-or-more-keys-pressed can be detected with an interrupt. >> >> > >> > Without the blocker we may start suspending and will stop midcycle. We >> > may be even better off in the end since we could leave some devices >> > still powered down after aborting system-wide suspend. >> > >> >> That does not sound right. > > Why doesn't it? If a device implements runtime PM it may chose remain in > powered-down mode even if system is awake. > If the device implements runtime PM it should already be powered-down if it is not in use. -- Arve Hj�nnev�g -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Arve Hjønnevåg on 25 May 2010 20:10 2010/5/25 Kevin Hilman <khilman(a)deeprootsystems.com>: > Arve Hj�nnev�g <arve(a)android.com> writes: > >> 2010/5/25 Dmitry Torokhov <dmitry.torokhov(a)gmail.com>: >>> On Tue, May 25, 2010 at 03:23:23PM -0700, Arve Hj�nnev�g wrote: >>>> On Tue, May 25, 2010 at 11:47 AM, Dmitry Torokhov >>>> <dmitry.torokhov(a)gmail.com> wrote: >>>> > On Tue, May 25, 2010 at 02:35:17PM -0400, Alan Stern wrote: >>>> >> On Tue, 25 May 2010, Dmitry Torokhov wrote: >>>> >> >>>> >> > > Here's the scenario: >>>> >> > > >>>> >> > > The system is awake, and the user presses a key. The keyboard driver >>>> >> > > processes the keystroke and puts it in an input queue. �A user process >>>> >> > > reads it from the event queue, thereby emptying the queue. >>>> >> > > >>>> >> > > At that moment, the system decides to go into opportunistic suspend. >>>> >> > > Since the input queue is empty, there's nothing to stop it. �As the >>>> >> > > first step, userspace is frozen -- before the process has a chance to >>>> >> > > do anything with the keystroke it just read. �As a result, the system >>>> >> > > stays asleep until something else wakes it up, even though the >>>> >> > > keystroke was important and should have prevented it from sleeping. >>>> >> > > >>>> >> > > Suspend blockers protect against this scenario. �Here's how: >>>> >> > > >>>> >> > > The user process doesn't read the input queue directly; instead it >>>> >> > > does a select or poll. �When it sees there is data in the queue, it >>>> >> > > first acquires a suspend blocker and then reads the data. >>>> >> > > >>>> >> > > Now the system _can't_ go into opportunistic suspend, because a suspend >>>> >> > > blocker is active. �The user process can do whatever it wants with the >>>> >> > > keystroke. �When it is finished, it releases the suspend blocker and >>>> >> > > loops back to the select/poll call. >>>> >> > > >>>> >> > >>>> >> > What you describe can be done in userspace though, via a "suspend manager" >>>> >> > process. Tasks reading input events will post "busy" events to stop the >>>> >> > manager process from sending system into suspend. But this can be confined to >>>> >> > Android userspace, leaving the kernel as is (well, kernel needs to be modified >>>> >> > to not go into suspend with full queues, but that is using existing kernel >>>> >> > APIs). >>>> >> >>>> >> I think that could be made to work. �And it might remove the need for >>>> >> the userspace suspend-blocker API, which would be an advantage. �It >>>> >> could even remove the need for the opportunistic-suspend workqueue -- >>>> >> opportunistic suspends would be initiated by the "suspend manager" >>>> >> process instead of by the kernel. >>>> >> >>>> >> However you still have the issue of modifying the kernel drivers to >>>> >> disallow opportunistic suspend if their queues are non-empty. �Doing >>>> >> that is more or less equivalent to implementing kernel-level suspend >>>> >> blockers. �(The suspend blocker approach is slightly more efficient, >>>> >> because it will prevent a suspend from starting if a queue is >>>> >> non-empty, instead of allowing the suspend to start and then aborting >>>> >> it partway through.) >>>> >> >>>> >> Maybe I'm missing something here... �No doubt someone will point it out >>>> >> if I am. >>>> >> >>>> > >>>> > Well, from my perspective that would limit changes to the evdev driver >>>> > (well, limited input core plumbing will be needed) but that is using the >>>> > current PM infrastructure. The HW driver changes will be limited to what >>>> > you described "type 2" in your other e-mail. >>>> > >>>> > Also, not suspending while events are in progress) is probably >>>> > beneficial for platforms other than Android as well. So unless I am >>>> > missing something this sounds like a win. >>>> > >>>> >>>> How would this limit the changes you need in the evdev driver? It need >>>> to block suspend when there are unprocessed events in some queues. >>>> Suspend blockers gives you an api to do this, without it, you check >>>> the queues in your suspend hook and abort suspend if they are not >>>> empty. Without suspend blockers you have no api to signal that it is >>>> OK to suspend again, so you are forcing the thread that tried to >>>> suspend to poll until you stop aborting suspend. >>> >>> No, you do not need to poll. You just set a timeout (short or long, >>> depending on your needs) and if no userspace task blocked suspend >>> durng that time you attempt to initiate suspend from your manager >>> process. If it succeeds - good, if not that means that more events came >>> your way and you have to do it later. >>> >> >> How is that not polling? If the user is holding down a key, the keypad >> driver has to block suspend, and user space will try to suspend again >> and again and again... > > Then the userspace suspend manager should be a little more clever > and should not blindly retry continuously. > > It should be more like a governor which makes some simple decisions > based on previous events, simple heuristics, uses timeouts etc., > So instead of the kernel suspending as soon as the last driver stops blocking suspend, you want to add heuristics in user-space to guess when suspend will succeed. This would, in my opinion, be a much worse solution. -- Arve Hj�nnev�g -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Dmitry Torokhov on 25 May 2010 20:30 On Tue, May 25, 2010 at 04:54:34PM -0700, Arve Hj�nnev�g wrote: > 2010/5/25 Dmitry Torokhov <dmitry.torokhov(a)gmail.com>: > > On Tue, May 25, 2010 at 04:13:35PM -0700, Arve Hj�nnev�g wrote: > >> > >> You are missing the point. There are no event pending. The kernel > >> reported the key down event, it was handled, but the keypad driver is > >> still scanning to see if the user presses another key, > > > > Employ reasonable timeout. > > Timeout for what? Stop trying to suspend altogether, stop scanning for > key changes, or a more "reasonable" poll interval? Stop trying to suspend for a fraction of a second to see if user wants to press another key. > > > > >> or releases the > >> currently held key. > >> > > > > Userspace consumer should wait for the key release and retract "busy" > > once event is received and handled. > > > > No it should not. User-space does not know if the key is coming from a > keypad driver that needs to actively scan the matrix while keys are > pressed. OTOH nor does kernel driver know whether system suspend should be blocked while it is scanning. What should happen in I press KEY_SUSPEND? Do we always want to wait till user releases it? > > >> > > >> > 2. Wait a tiny bit after last application notified you that it finished > >> > processing events. > >> > > >> > So basically the difference is that with in-kernel suspend blockers, > >> > there is a tiny window where we haven't started the suspend yet but are > >> > about to the driver has a chance to prevent entire system from starting > >> > sleep. > >> > >> No, the difference is that if a driver needs to prevent suspend for an > >> extended period of time, you don't have user space continuously > >> polling to see if it can suspend. > > > > Why would a driver, on its own, prevent suspend for extended periods of > > time? I think that the decision should originate from userspace, kernel > > is here just to serve the requests. > > > > A driver prevents suspend if suspend would prevent it from working. > For instance, the gpio keypad matrix code prevents suspend when a key > is help down, since it has to activly scan the keypad for changes. > Only no-keys-pressed versus one-or-more-keys-pressed can be detected > with an interrupt. > > >> > >> > > >> > Without the blocker we may start suspending and will stop midcycle. We > >> > may be even better off in the end since we could leave some devices > >> > still powered down after aborting system-wide suspend. > >> > > >> > >> That does not sound right. > > > > Why doesn't it? If a device implements runtime PM it may chose remain in > > powered-down mode even if system is awake. > > > > If the device implements runtime PM it should already be powered-down > if it is not in use. It could have been in use but userspace-initiated suspend accelerated it entering low power state. -- Dmitry -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Arve Hjønnevåg on 25 May 2010 20:40 2010/5/25 Dmitry Torokhov <dmitry.torokhov(a)gmail.com>: > On Tue, May 25, 2010 at 04:54:34PM -0700, Arve Hj�nnev�g wrote: >> 2010/5/25 Dmitry Torokhov <dmitry.torokhov(a)gmail.com>: >> > On Tue, May 25, 2010 at 04:13:35PM -0700, Arve Hj�nnev�g wrote: >> >> >> >> You are missing the point. There are no event pending. The kernel >> >> reported the key down event, it was handled, but the keypad driver is >> >> still scanning to see if the user presses another key, >> > >> > Employ reasonable timeout. >> >> Timeout for what? Stop trying to suspend altogether, stop scanning for >> key changes, or a more "reasonable" poll interval? > > Stop trying to suspend for a fraction of a second to see if user wants > to press another key. > In other words, try to suspend several times a second until it succeeds. >> >> > >> >> or releases the >> >> currently held key. >> >> >> > >> > Userspace consumer should wait for the key release and retract "busy" >> > once event is received and handled. >> > >> >> No it should not. User-space does not know if the key is coming from a >> keypad driver that needs to actively scan the matrix while keys are >> pressed. > > OTOH nor does kernel driver know whether system suspend should be > blocked while it is scanning. Yes it does, it cannot deliver wakeup keys if it suspends. > What should happen in I press KEY_SUSPEND? > Do we always want to wait till user releases it? If KEY_SUSPEND is part of the keypad matrix, then yes. > >> >> >> > >> >> > 2. Wait a tiny bit after last application notified you that it finished >> >> > processing events. >> >> > >> >> > So basically the difference is that with in-kernel suspend blockers, >> >> > there is a tiny window where we haven't started the suspend yet but are >> >> > about to the driver has a chance to prevent entire system from starting >> >> > sleep. >> >> >> >> No, the difference is that if a driver needs to prevent suspend for an >> >> extended period of time, you don't have user space continuously >> >> polling to see if it can suspend. >> > >> > Why would a driver, on its own, prevent suspend for extended periods of >> > time? I think that the decision should originate from userspace, kernel >> > is here just to serve the requests. >> > >> >> A driver prevents suspend if suspend would prevent it from working. >> For instance, the gpio keypad matrix code prevents suspend when a key >> is help down, since it has to activly scan the keypad for changes. >> Only no-keys-pressed versus one-or-more-keys-pressed can be detected >> with an interrupt. >> >> >> >> >> > >> >> > Without the blocker we may start suspending and will stop midcycle. We >> >> > may be even better off in the end since we could leave some devices >> >> > still powered down after aborting system-wide suspend. >> >> > >> >> >> >> That does not sound right. >> > >> > Why doesn't it? If a device implements runtime PM it may chose remain in >> > powered-down mode even if system is awake. >> > >> >> If the device implements runtime PM it should already be powered-down >> if it is not in use. > > It could have been in use but userspace-initiated suspend accelerated it > entering low power state. > So you are suggesting we should repeatedly try to suspend the whole system to accelerate runtime PM powering down devices? Why not power the device down as soon as it is not in use? -- Arve Hj�nnev�g -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Peter Zijlstra on 26 May 2010 04:50
On Tue, 2010-05-25 at 15:33 -0700, Arve Hjønnevåg wrote: > The biggest problem here is not that it is hard to change our > user-space, but that the proposed change is inferior to what we have > now. It forces us to poll until all drivers stop aborting suspend. On > one hand we have people telling us that all code that polls is broken > and must be fixed (instead of suspending to limit the damage), and on > the other hand we have people suggesting we implement opportunistic > suspend by polling from user-space until suspend succeeds. No it does _not_. You're really not getting that Dmitry is proposing. So your proposal is that when we wake userspace, it opens /dev/suspend_blocker _before_ it consumes whatever event, consumes the event, deals with the event, then closes the suspend_blocker. Then the kernel, upon reaching a 0 suspend_blocker count, will try to suspend again. What Dmitry proposes is that, the app _before_ it consumes the event, pokes at this suspend manager, it increases a blocker count, then consumes the event (the kernel will _not_ auto-suspend), handles it and then again pokes the suspend manager, this time decreasing the blocker count. The suspend manager will, upon reaching a 0 block count, suspend the machine. If that fails, it means there's something to do, an app will inc, work, dec its count, and it will try again once it reaches 0 again. There is no polling what-so-ever in this model. The only thing is that the kernel will not try to auto-suspend and there is no user-space suspend blocker API. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ |