> On Nov 12, 2015, at 12:49 AM, Michael Zimmermann <[email protected]> > wrote: > > Stall was just an example, I can also use DEBUG. My timer interval is > 100ms(I converted it from the 100ns unit). > > What I meant with "thread mode code" is the "normal" non IRQ context code > running on the CPU. That means that there are actually two contexts in > EDK2, the exception(i.e. IRQ's like the Timer/Watchdog) and the normal mode > all code is running in. >
OK thanks that helps. The timer ISR runs in interrupt context, and from an ARM point of view EFI is running the ARM "Thread mode". The ISR context is an implementation detail and not really defined by the specification. I'd also point out the Events dispatch independent of the interrupt context. gBS->RestoreTpl() can cause events to dispatch. For example a lot of the EFI Protocol services have locks that raise TPL to prevent recursion. When these locks are released and the TPL is restored events are dispatched. So just calling EFI services can cause events to run. So conceptually you can ignore the interrupt context, as that is used to implement the timer tick. Every thing else is an event that runs in the main EFI context. https://github.com/tianocore/edk2/blob/master/MdeModulePkg/Core/Dxe/Event/Tpl.c#L126 Thanks, Andrew Fish > On Thu, Nov 12, 2015 at 9:27 AM, Andrew Fish <[email protected]> wrote: > >> >>> On Nov 11, 2015, at 11:51 PM, Michael Zimmermann < >> [email protected]> wrote: >>> >>> I've started investigating in the timer event problem and I think I have >>> some weird problem with my platform drivers(I hope, so it's not a EDK2 >> bug). >>> >>> If I create a timer that runs every 100ms which does nothing but a >>> stall(1), the thread mode code stops after some random time(usually in >> edk2 >>> shell so I guess it's a race condition which needs some cpu load). >>> >> >> Michael, >> >> Watch out as Stall() is Microseconds, and SetTimer() is 100ns I've seen >> bugs like that before in code. >> >>> When threadmode code is stopped the timer continues getting called and >> even >>> if I stop the timer afterwards(with CloseEvent) it keeps being stopped. >>> >>> Is there a way to get the threadmode context from inside a timer >> callback? >>> This way I could read the PC to check what's going on. >>> >> >> There are no threads. EFI is an event model. If your code is running you >> are blocking every one else's forward progress. Only code running at a >> higher TPL can preempt. So when your event is running it is blocking the >> main flow and any event trying to run at <= TPL of your event from making >> forward progress. gBS->Stall() does not yield, it is no different than >> running code. >> >> So there is only one context, you can print it out any time you want. >> >> Thanks, >> >> Andrew Fish >> >> PS When folks yell at us for not having threads in EFI we point them at: >> https://web.stanford.edu/~ouster/cgi-bin/papers/threads.pdf >> >> >>> Michael >>> >>> On Wed, Nov 11, 2015 at 8:10 PM, Kinney, Michael D < >>> [email protected]> wrote: >>> >>>> Michael, >>>> >>>> A periodic event timer at 30 times a second should not cause pauses >>>> forever, unless the action you are performing in the event notification >>>> function takes more than 1/30 of a second to complete. You should be >> able >>>> to just add a periodic event handler that does nothing, so you can >> measure >>>> what the overhead is. >>>> >>>> Another option is to use the performance counter in the TimerLib each >> time >>>> a Blt() is called (GetPerformanceCounterProperties() and >>>> GetPerformanceCounter()). When Blt() is called frequently, the amount >> of >>>> time since last vsync will have elapsed, and you can go the vsync action >>>> within the Blt() call. If you also set a one shot timer event, so if >> the >>>> last call to Blt() did not do a vsync and there are no more Blt() calls, >>>> 1/30th of a second later, the vsync action can be done. Every time >> Blt() >>>> is called, the one shot timer can be re-armed. This way, the one shot >>>> timer event is not actually executed very often. >>>> >>>> Mike >>>> >>>>> -----Original Message----- >>>>> From: edk2-devel [mailto:[email protected]] On Behalf Of >>>> Michael Zimmermann >>>>> Sent: Wednesday, November 11, 2015 12:32 AM >>>>> To: [email protected] >>>>> Subject: [edk2] EFI GOP with manual vsync trigger >>>>> >>>>> Hi, >>>>> >>>>> my Graphics HW uses a manual vsync trigger. That means that after >> drawing >>>>> to the framebuffer I need to manually trigger vsync(you can compare it >> to >>>>> switching between double buffers). >>>>> >>>>> The problem is that UEFI's GraphicsOutputProtocol(GOP) doesn't take >> care >>>> of >>>>> HW that needs a flush. >>>>> While issuing the trigger after every Blt Operation works, this >> obviously >>>>> causes extremely slow rendering for applications like the Shell which >> call >>>>> Blt very often(like for every character). >>>>> >>>>> Also I can't use a timer to set the trigger(like 30times a second) >> because >>>>> it takes too much time and the Timer Interrupt ends up consuming too >> much >>>>> time and the "normal" code gets paused forever. >>>>> >>>>> Do you have any other ideas how to handle this? >>>>> >>>>> Thx >>>>> Michael >>>>> _______________________________________________ >>>>> edk2-devel mailing list >>>>> [email protected] >>>>> https://lists.01.org/mailman/listinfo/edk2-devel >>>> >>> _______________________________________________ >>> edk2-devel mailing list >>> [email protected] >>> https://lists.01.org/mailman/listinfo/edk2-devel >> >> > _______________________________________________ > edk2-devel mailing list > [email protected] > https://lists.01.org/mailman/listinfo/edk2-devel _______________________________________________ edk2-devel mailing list [email protected] https://lists.01.org/mailman/listinfo/edk2-devel
