Hi Kevin, On Mon, Nov 05, 2012 at 23:33:07, Kevin Hilman wrote: > "Bedia, Vaibhav" <vaibhav.bedia@xxxxxx> writes: > > > On Sat, Nov 03, 2012 at 18:34:30, Kevin Hilman wrote: > > [...] > >> >> > >> >> Doesn't this also mean that you won't get timer wakeups > >> >> in idle? Or are you keeping the domain where the clockevent is > >> >> on during idle? > >> >> > >> > > >> > The lowest idle state that we are targeting will have MPU powered > >> > off with external memory in self-refresh mode. Peripheral domain > >> > with the clockevent will be kept on. > >> > >> Is this a limitation of the hardware? or the software? > >> > > > > Well, making the lowest idle state same as the suspend state will > > require us to involve WKUP_M3 in the idle path and wakeup sources get > > limited to the IPs in the WKUP domain alone. There's no IO daisy > > chaining in AM33XX so that's one big difference compared to OMAP. The > > other potential problem is that the IPC mechanism that we have uses > > interrupts. > > It can still interrupt the M3, it's only the interrupt back to the MPU > that is the issue, right? That being said, there's no reason it > couldn't use polling in the idle path, right? > Yes we could use polling but I think we have a bigger problem in the chip architecture. > > Assuming that the lowest idle state, say Cx, is the same as the > > suspend state, we'll need to communicate with the WKUP_M3 using > > interrupts once we decide to enter Cx. I am not sure if we can do > > something in the cpuidle implementation to work around the "interrupt > > for idle" problem. > > > > We could probably not wait for an ACK when we want to enter Cx, > > why not? > > Are the response times from the M3 really up to 500ms (guessing based on > the timeout you used in the suspend path.) That seems rather unlikely. > No 500ms is too high. Actual delays would be much lower, I need to check with the design team on the expected number. > Hmm, but as I think about it. Why does the MPU need to wait for an ACK > at all? Why not just send the cmd and WFI? > I have myself being going back and forth on this. There are lot of things that we do in software, DDR being one of them. We can't do some of the DDR related stuff unless memory enter self-refresh AND EMIF gets disabled. Doing so essentially means that the drivers have entered sort of suspend state. Given this h/w limitation I don't see how we could handle without impacting a running system. > > but the problem of limited wakeup sources remains. If we let the > > various drivers block the entry to Cx, since almost all the IPs are in > > the peripheral domain a system which uses anything other than UART and > > Timer in WKUP domain will probably never be able enter Cx. > > Even so, I think the system needs to be designed to hit the same power > states in idle and suspend. Then, the states can be restricted based > wakeup capabilities as you described. This would be easy to do in the > runtime PM implementation for this device. > > IMO, assuming that idle will not be useful from the begining is leading > down the path to poor design choices that will be much more difficult to > fixup down the road in order to add idle support later. We need to > design both idle and suspend at the same time. > Getting PER to transition on a running system is something I can't figure out. Maybe MPU OFF is the lowest we can go. > Also, don't forget about GPIO0. Systems could easily be built such that > peripherals which want to wakeup but don't have native wakeup > capabilities could use a GPIO in bank 0 to wake the system. > > Similarily, I2C0 is in WKUP, and brought out to capes, so some simple > designs with with I2C devices on a cape might be perfectly capable of > hitting deep power states in idle. > Ok this is interesting. AFAIK I2C wakeup requires the device to be operating in slave mode. If so, is this something that's already supported on OMAP? Regards, Vaibhav -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
