#asahi-dev · 2023-08-17 08:18
jannaudcpext "works" again on M2. I saw a picture on the usb-c/dp display although not for long#asahi-dev · 2023-08-16 22:52
jannauat least I'm now in half-working state. dcp sees the monitor but displayport link training still fails#asahi-dev · 2023-08-16 22:46
jannauthe gcc exception appears to be something in the dcp module. issue vanished with clang compiled modules. dart/io-pgtbl is built-in#asahi-dev · 2023-08-16 22:44
jannauTIL: it's possible to load modules compiled with clang into a linux image compiled by gcc#asahi-dev · 2023-08-16 22:42
jannaueiln: looks green, err good. so 1/2 to 2/3 done#asahi-dev · 2023-08-16 19:55
jannausigh, this dcp/afk exception with gcc is annoying. I have in apple-dart's domain_flush_tlb() debug code to read the crashing pte. on AP side everything is fine after the allocation and before sending the command, dcp still crashes#asahi-dev · 2023-08-16 08:29
jannauthe syncing should be unrelated. that is not the first table entry written#asahi-dev · 2023-08-16 07:37
jannauthe level 1 page table syncing for locked darts was my first suspect but adding barriers there makes no difference#asahi-dev · 2023-08-16 07:31
jannaubut where? I don't see how it could be afk. after dma_alloc_coherent() everything should be fine there. That leaves dart/io-pgtbl but why would that specifically break on afk/epic commands#asahi-dev · 2023-08-16 06:59
jannaureproduced on asahi alarm. compiled with gcc breaks in the same way. HOW?!#asahi-dev · 2023-08-16 06:44
jannaumain non-source difference is that on fedora the kernel is compiled with gcc instead of llvm#asahi-dev · 2023-08-16 06:42
jannauold state on asahi alarm still has working dptxport communication and I can't spot an relevant difference #asahi-dev · 2023-08-16 06:06
jannauat least dptxport communication worked before and I do not see how my changes could be related#asahi-dev · 2023-08-16 06:01
jannausetting apple,dma-range to vm-base,vm-size from the ADT makes no difference. it should be impossible since dcpext must have read it from the AFK ringbuffer which is allocatedd earlier with an higher dva#asahi-dev · 2023-08-15 20:29
jannauthe tracer can read the command as well and just the dart is invalidated just before the command is logged#asahi-dev · 2023-08-15 20:27
jannauthe the two latter things where issues it should be issues for iomfb as well but that works on dcp and dcpext#asahi-dev · 2023-08-15 20:25
jannaulinux allocates iova top-down instead bottom-up but within vm-base,vm-size#asahi-dev · 2023-08-15 20:23
jannauour dart-dcpext PTEs have the non-cacheable bit set#asahi-dev · 2023-08-15 20:23
jannauknown differences: macos driver sends 2 REPORTs before but did that already in 13.3#asahi-dev · 2023-08-15 20:20
jannau13.5 fw on t8112, as far as I can tell no major differences to macos#asahi-dev · 2023-08-15 20:18
jannaufirst after boot. yes, the faulting dva is the expected one from linux and is mapped in dart-dcpext#asahi-dev · 2023-08-15 20:10
jannauany idea why dcpext would fault on reading the first afk/epic cmd txbuf although it is mapped and I can read the from the hypervisor#asahi-dev · 2023-08-13 19:47
jannaubut no need to remove your local fix#asahi-dev · 2023-08-13 19:47
jannaunst: https://github.com/AsahiLinux/linux/pull/174#asahi-dev · 2023-08-13 19:32
jannauhttps://github.com/AsahiLinux/linux/pull/174#asahi-dev · 2023-08-13 17:52
jannaunst: yes, that's the fix#asahi-dev · 2023-08-13 17:40
jannauthat results in a phandle property in the piodma node which blows up of_translate_dma_regon() due to the #{address,size}-cells proterties in the dcp node#asahi-dev · 2023-08-13 17:38
jannauChaosPrincess: you're seeing it with asahi-wip? I think it is only reproducible there because the bits/000-devicetree rebase reintroduced 'apple,piodma-mapper = <&disp0_piodma>;' errorneously in the dcp node#asahi-dev · 2023-08-13 17:33
jannauChaosPrincess: https://paste.bingner.com/paste/opncz fixes it#asahi-dev · 2023-08-13 17:12
jannauyes, with that change it blows up for me as well#asahi-dev · 2023-08-13 16:57
jannauyour dtb excerpts looked as expected but in mine the piodma <-> reverved memory node linking is clearly broken#asahi-dev · 2023-08-13 16:53
jannauor introduce the issue for me#asahi-dev · 2023-08-13 16:52
jannauChaosPrincess: I think https://paste.bingner.com/paste/j7jfo should fix the issue#asahi-dev · 2023-08-13 16:48
jannauChaosPrincess: I think I see the issue in my devicetree#asahi-dev · 2023-08-13 16:39
jannauChaosPrincess: scripts/faddr2line vmlinux $FRAME#asahi-dev · 2023-08-13 16:38
jannauthis is a different issue#asahi-dev · 2023-08-13 16:38
jannaunst: there is no piodma child node? your boot.bin then doesn't have the dtb from your kernel. update-m1n1 after an kernel update takes care of this in the standard setup #asahi-dev · 2023-08-13 16:14
jannaucan you try to get which line is 'of_iommu_get_resv_regions+0xc8/0x1e0'?#asahi-dev · 2023-08-13 16:08
jannauon j314c I think 13.3. on other devices 13.5 (m1 ultra). I don't see how the macos version could be related to this though. this looks linux issue#asahi-dev · 2023-08-13 16:04
jannaunst, ChaosPrincess: are you using a self-compiled kernel? if yes can you try with my config or compare for suspicious differences? https://www.jannau.net/asahi/6.4.0-asahi-6.config#asahi-dev · 2023-08-13 16:01
jannauit's a little strange that none of the current fedora users complained so far. #asahi-dev · 2023-08-13 15:59
jannaunst: I'm not sure. I would have said there are none but obviously that's not true#asahi-dev · 2023-08-13 12:19
jannauI have access to a razer core x (eGPU) case so I can test in principle any pcie3 card. I think that one is alpine ridge based#asahi-dev · 2023-08-13 12:17
jannauthat worked sometimes, usually needed a few connect/deconnect cycles#asahi-dev · 2023-08-13 12:15
jannaumarcan: atcphy-20230219#asahi-dev · 2023-08-12 15:21
jannausigh, why apple?#asahi-dev · 2023-08-12 15:12
jannauI'll review the code and see if I can get it merged#asahi-dev · 2023-08-12 15:12
jannauI think other coprocessors should be able to use that PR other coprocessors with little code changes in m1n1#asahi-dev · 2023-08-12 15:10
jannauin https://github.com/AsahiLinux/m1n1/pull/286 the dcp support is refactored to be used by sio#asahi-dev · 2023-08-12 15:09
jannaueiln: you mean the coprocessor's segments? we need to handle that in m1n1 via reserved-memory as currently done for dcp