18:19 <Myy> Meow

18:21 <Myy> So this morning, I thought I had the VPU completely working again, however turns out that it was a false positive. Still, today I've been able to repair the IOMMU DMA ScatterGather lists mapping unmapping side and I'm now able to, at least, invoke mpv with MPP without having the VPU driver crashing horribly.

18:22 <Myy> I don't know if MPV or MPP is not passing the format correctly, or if the movie I tried to play wasn't supported by the VPU so now I'm trying various MP4/MKV files.

18:22 <Myy> Meanwhile, I'll still push my modifications on the driver and generate a big patch containing all the modifications done on the kernel side.

18:25 <Myy> Turns out that using the IOMMU_DMA API was a good idea. However casting the iommu_domain->iova_cookie to a iova_domain structure was not. When using that API, that kind of cast missed the Iova domain address by 4 bytes, causing a lot of issues due to not reading exactly at the right address.

18:27 <Myy> So, long story short : I still cannot check the "acceleration" part (The VPU driver informs MPV that it does not support the format) but at least it doesn't crash horribly when mapping DMA buffers.

18:28 <phh> vpu driver = mpp?

18:39 <Myy> Well, the VPU driver used through the MPP library informs MPV that the format was not recognized

18:40 <Myy> The VPU code I'm using is this one : https://github.com/Miouyouyou/rockchip-vcodec/tree/retry

18:41 <Myy> And these are the kernel modifications I applied to test this code : https://gist.github.com/Miouyouyou/945b5e1e9816afa24822aadeef486952

18:43 <Myy> Basically, the modifications are : export the iommu_dma_init function to initialize the cache used by the IOVA functions (else you get a NPE) and add two signatures for iommu_dma_(un)map_sg that accepts a domain as an argument directly, instead of trying to get it through iommu_get_domain_for_dev

18:43 <Myy> The reason being iommu_get_domain_for_dev returns NULL everytime :C

18:44 <phh> uh, sounds like iommu is not properly declared in dts?

18:45 <Myy> Ah... You might need...

18:45 <Myy> Wait a minute

18:46 <Myy> This one : https://github.com/Miouyouyou/RockMyy/blob/master/patches/kernel/v4.13/DTS/0009-Define-VPU-services-in-the-Rockchip-3288-DTS-files.patch

18:47 <phh> I'm speaking for iommu_get_domain_for_dev returning null

18:47 <Myy> Ah, it seems that it tries to get the iommu_group attached to the device structure

18:47 <Myy> And then ask the iommu group to give the domain attached

18:47 <Myy> However, sometimes the device does NOT have a iommu group for some reasons and then it fails miserably :C

18:48 <Myy> I tried a few blind combinations of iommu_attach and iommu_group_add but that only led to freezes so I'll have to poke some #armlinux guys for more infos

18:50 <Myy> So, yeah, I'll try to grab the hevc from that Google drive link again and see if I can decode it with MPV<->MPP<->VPU-Driver

18:50 <phh> but, iommu attachment should be done based on dts anyway, not by vpu driver

18:51 <Myy> That's entirely possible, some parts of the IOMMU driver are still witchcraft to me. If you have a good idea, I'm all ears. I've setup a small driver skeleton to test various things with IOMMU and the VPU so I could give a try again.

18:52 <Myy> https://github.com/Miouyouyou/Nyanbada

18:58 <phh> do you have a37b19a384914c60b7e1264a6c21e7bf96b637e8 ?

18:59 <Myy> You mean : https://github.com/torvalds/linux/commit/a37b19a384914c60b7e1264a6c21e7bf96b637e8 ?

19:00 <phh> yes

19:00 <phh> I'm not too sure if it will help, I can't tell if it makes probe_defer change behaviour or not

19:02 <Myy> I'll check that. Meanwhile playing some samples from stdint seem to work fine...

19:03 <Myy> However I'll need something longer than 7 seconds

19:05 <Myy> I'll try with the 140 MB WebM file. If that works, that means that the VPU works with 4.13 kernels. I'll then have to remove the ugly hacks and everything should be fine.

19:07 <Myy> Looks like it's working... Yay...

19:08 <Myy> I'll provide a build of the patched kernel

19:08 <Myy> So that anyone can test this

19:09 <Myy> I'll even do this right now

19:09 <phh> Nice!

19:16 <Myy> With the OOT module also added : https://github.com/Miouyouyou/RockMyy-Build/tree/VPU_Work_tests

19:17 <Myy> Without the OOT module added : https://github.com/Miouyouyou/RockMyy-Build/tree/109e377bbb8a7f9132ab5448aaeb611c000e6e4b

19:17 <Myy> I'll tag the one with the module for easy download

19:19 <Myy> https://github.com/Miouyouyou/RockMyy-Build/releases/tag/v4.13-VPU-Test

19:19 <Myy> Meanwhile, I'll update the RockMyy script and patches repo to include my (ugly) kernel patches so that anyone can modify it and recompile the OOT driver

19:26 <Myy> So, I'm testing a recompilation of my kernel with the new patches. If everything goes fine, I'm pushing it too and everything should be ready for testing.

20:00 <Myy> The patches and compilation script are now available here : https://github.com/Miouyouyou/RockMyy-Build/tree/VPU_Work_tests