On Fri, 11 Nov 2005, David Brownell wrote: > On Tuesday 08 November 2005 11:43 am, Alan Stern wrote: > > A question has come up recently in private conversations concerning remote > > wakeup requests arriving while a system sleep transition is in progress. > > How should they be handled? > > > > There are at least three possibilities: > > > > (a) Carry out the wakeup request (i.e., resume the requesting device) > > and then abort the system sleep when the device's parent is told > > to suspend and it sees that its children aren't all suspended. > > This probably makes the most sense to me ... Me too (as must have been obvious from the following comments). > > (b) Carry out the wakeup request, then sometime later try to recover > > by re-suspending the device that was just resumed, and continue > > the sleep transition. > > > > (c) Ignore the wakeup request and proceed with the sleep transition. > > > > (a) is very simple and doesn't require any extra code, since drivers > > should already be checking to make sure that all the children are > > suspended before suspending their devices. I didn't mention this before, but it's possible that some drivers don't bother to make this check. We may want to do the checking in the PM core instead -- whenever a resume request occurs while a sleep transition is in progress, abort the sleep. > > When the sleep transition is > > aborted, we can just return -ERESTARTSYS and force the user to try again > > -- exactly the same as would happen with any other interrupted system > > call. > > That is, it requires faults on suspend to be handled reasonably. > Not a new requirement ... but quite possibly one that's not always > being satisfied. And certainly one that's awkward to test. Here's an easy way to test it. In an appropriate driver's suspend routine, add: printk(KERN_INFO "Delaying suspend -- do your stuff!\n"); msleep(2000); Then when that message pops up, you have two seconds to cause a wakeup event. > > (b) is not a good solution, IMO. It involves a pretty hairy recovery > > scheme that's not well defined. It runs the risk of leaving some devices > > unsuspended while suspending their parents, or leaving data structures out > > of sync with the actual hardware state. Furthermore I don't see any > > advantage in making the kernel do the re-suspend over making userspace > > restart the entire sleep transition, as in (a). > > Although I'm not sure I'd see the point to having userspace try to > suspend again ... user said to wake up, after all!! Right. Although it may have been a coincidence. Or something as silly as a release of the Enter key aborting a suspend which got started when that key was pressed. > In general, (b) sounds complex and error prone. Two things that > are IMO worth avoiding in the kernel. Agreed. > > (c) seems like a reasonable approach. It does, however, have a couple of > > problems. First, drivers can't (currently) tell the difference between an > > overall system sleep and a device-specific runtime suspend. Since we > > certainly _don't_ want to ignore device-wakeup requests during a runtime > > suspend, there's no way for a driver to ignore them during a system sleep > > transition. > > > > The second thing to consider about (c) is that ignoring a wakeup request > > might not make it go away. It will depend on the exact nature of the > > hardware and the type of request of course, but there's a good chance that > > the request will persist until the system finally goes to sleep. At that > > point the request will immediately wake the system right back up. The > > overall effect is the same as with (a) -- the system is awake at the end > > -- but you've wasted a lot of time and effort to get there. > > Both are reasons I'd prefer (a). > > > > Now it's true that sometimes there will be no choice. Once tasks are > > frozen, there's no way to resume a device since doing so requires a > > process context. Then we'd be forced to adopt (c). > > On the other hand, a clean failure of the suspend would restart those > tasks nicely... I don't see that as an issue. Neither do I. Note that if tasks are frozen, the driver doesn't have any reason to abort the suspend. As far as the driver can tell everything proceeds normally: It gets a resume request and it tells the appropriate process to carry out the resume. It's not the driver's fault that the process is frozen and so can't do the resume until the system awakens. > > However there are > > situations where tasks are _not_ frozen during a sleep transition. > > Suspend-To-RAM on PPC, for example. What should happen in such cases? > > Use (a). ;) My thought exactly. :-) But I'd still like to hear what other people on the list think. Maybe there's a fourth alternative... Alan Stern