On 01.09.21 12:03, Frieder Schrempf wrote: > Hi Lucas, > > On 09.08.21 13:50, Frieder Schrempf wrote: >> On 09.08.21 13:01, Lucas Stach wrote: >>> Hi Frieder, >>> >>> Am Donnerstag, dem 05.08.2021 um 20:56 +0200 schrieb Frieder Schrempf: >>>> On 05.08.21 12:18, Frieder Schrempf wrote: >>>>> On 21.07.21 22:46, Lucas Stach wrote: >>>>>> Hi all, >>>>>> >>>>>> second revision of the GPC improvements and BLK_CTRL driver to make use >>>>>> of all the power-domains on the i.MX8MM. I'm not going to repeat the full >>>>>> blurb from the v1 cover letter here, but if you are not familiar with >>>>>> i.MX8MM power domains, it may be worth a read. >>>>>> >>>>>> This 2nd revision fixes the DT bindings to be valid yaml, some small >>>>>> failure path issues and most importantly the interaction with system >>>>>> suspend/resume. With the previous version some of the power domains >>>>>> would not come up correctly after a suspend/resume cycle. >>>>>> >>>>>> Updated testing git trees here, disclaimer still applies: >>>>>> https://eur04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgit.pengutronix.de%2Fcgit%2Flst%2Flinux%2Flog%2F%3Fh%3Dimx8m-power-domains&data=04%7C01%7Cfrieder.schrempf%40kontron.de%7Cbf3a4cacd1e047be747b08d96d39e713%7C8c9d3c973fd941c8a2b1646f3942daf1%7C0%7C0%7C637660917728575954%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=0A7jRJH16d3T1S868RHg57csVuDUtgB3lNl2A3QZdus%3D&reserved=0 >>>>>> https://eur04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgit.pengutronix.de%2Fcgit%2Flst%2Flinux%2Flog%2F%3Fh%3Dimx8m-power-domains-testing&data=04%7C01%7Cfrieder.schrempf%40kontron.de%7Cbf3a4cacd1e047be747b08d96d39e713%7C8c9d3c973fd941c8a2b1646f3942daf1%7C0%7C0%7C637660917728575954%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=gC5jcC0w3VP4HiJYQMBWD%2FHQzU2rr7KjtGG82Snh4X0%3D&reserved=0 >>>>> >>>>> I finally did some tests on my side using USB, GPU and DSI (no PCIe, VPU, CSI so far) and the results are promising. Thanks for the effort! >>>>> >>>>> I will try to run some more automated suspend/resume and reboot test cycles over the weekend and report the results here afterwards. >>>>> >>>> >>>> Unfortunately I got some results sooner than I had hoped. I set up a simple loop to suspend/resume every few seconds and on the first run it took around 2-3 hours for the device to lock up on resume. On the second run it took less than half an hour. I had glmark2-es2-drm running in the background, but it looks like it crashed at some point before the lockup occurred. >>>> >>>> Of course this could also be unrelated and caused by some peripheral driver or something but the first suspicion is definitely the power domains. >>>> >>>> If you have any suggestions for which debug options to enable or where to add some printks, please let me know. If I do another run I would like to make sure that the resulting logs are helpful for debugging. >>>> >>>> And I would appreciate if someone else could try to reproduce this problem on his/her side. I use this simple script for testing: >>>> >>>> #!/bin/sh >>>> >>>> glmark2-es2-drm & >>>> >>>> while true; >>>> do >>>> echo +10 > /sys/class/rtc/rtc0/wakealarm >>>> echo mem > /sys/power/state >>>> sleep 5 >>>> done; >>> >>> Hm, that's unfortunate. >>> >>> I'm back from a two week vacation, but it looks like I won't have much >>> time available to look into this issue soon. It would be very helpful >>> if you could try to pinpoint the hang a bit more. If you can reproduce >>> the hang with no_console_suspend you might be able to extract a bit >>> more info in which stage the hang happens (suspend, resume, TF-A, etc.) >>> If the hang is in the kernel you might be able to add some prints to >>> the suspend/resume paths to be able to track down the exact point of >>> the hang. >>> >>> I'm happy to look into the issue once it's better known where to look, >>> but I fear that I won't have time to do the above investigation myself >>> short term. Frieder, is this something you could help with over the >>> next few days? >> >> I will see if I can find some time to track down the issue at least a little bit more. But I imagine it could get quite tedious if it takes up to several hours to reproduce the issue and I don't have much time to spare. >> >> @Peng, @Adam and everyone else: Any chance you could setup a similar test and try to reproduce this? >> >> On the other hand reboot cycle testing didn't show any lockup problems over more than 24 hours, so it seems like the issue is limited to resume. > > I ran a few more suspend/resume cycles and watched the log. The first > 2.5 hours nothing noteworthy happened, except that glmark2 crashed again > at some point. Facepalm! Of course glmark2 didn't crash, it just doesn't loop endlessly as I expected it to do, which totally makes sense for a benchmark. Using --run-forever should do the trick.
