On 07/10/2013 09:18 AM, Winkler, Tomas wrote: >>> suspend complete. >>> >>> I can start bi-sect of this problem on intel-display scope if you >>> would like me to. Please let me know if the bisect scope should be larger. >>> >>> -- Shuah >> >> I got finally an older system where this reproduces consistently, I'm trying to >> root cause that now. >> As soon I have something to test I will send it out. >> > > I'm not sure why but I'm not able to get the interrupt from the device until all the devices has resumed after that the mei reset is successful and device work as expected. > I need to fix the state machine so the reset won't go crazy on resetting actually I did that and it reduces resets into two but I'm still trying to figure out what's going on. > The regression coming from the fact that I've linearize the reset flow, it waits for interrupt to arrive, before that the flow gave up and has restarted directly from the interrupt thread. > > > Thanks > Tomas > > Tomas, This might be a shot in the dark. Commit mei: me: clear interrupts on the resume path (42f132febff3b7b42c6c9dbfc151f29233be3132), clears pending interrupts in mei_me_pci_resume(). Your comment about not seeing interrupt until mei reset is successful got me wondering, if clearing pending interrupts is causing this? I don't see mei_me_pci_suspend() calling mei_disable_interrupts() and pci_disable_msi(). I don't see a call to mei_enable_interrupts() from mei_me_pci_resume(). I don't think mei_enable_interrupts() is used anywhere. pci_dev_msi_enabled() enables the interrupts in mei_me_pci_resume() looks like. However, I did notice one thing, if pci_dev_msi_enabled() returns false, request_threaded_irq() is called with IRQF_SHARED. Again might just be fine, it leads me to the next question. mei_me_pci_suspend() has code that runs after disabling interrupts. Does this need to be split into suspend() and suspend_noirq() ops, since the IRQ could be shared? It is a possibility if pci_enable_msi() takes longer in resume path and pci_dev_msi_enabled() returns false, then IRQF_SHARED is requested. Don't know if this helpful, but I thought I would ask just in case, it helps you think of something you didn't before. Please let me know if you need help gathering data from my system. -- Shuah Shuah Khan, Linux Kernel Developer - Open Source Group Samsung Research America (Silicon Valley) shuah.kh at samsung.com | (970) 672-0658
