#asahi-dev · 2022-04-20 14:41
marcanhttps://github.com/AsahiLinux/linux/blob/asahi/arch/arm64/boot/dts/apple/t8103.dtsi#L441 https://github.com/AsahiLinux/linux/blob/asahi/arch/arm64/boot/dts/apple/t8103-pmgr.dtsi#asahi-dev · 2022-04-20 14:40
marcan(the ADT is a complete mess)#asahi-dev · 2022-04-20 14:40
marcanapple love to sprinkle references to pmgr and other things in random places#asahi-dev · 2022-04-20 14:40
marcanthe error log stuff is noise#asahi-dev · 2022-04-20 14:39
marcanthat's pmgr#asahi-dev · 2022-04-20 02:10
marcanyeah that's been there forever#asahi-dev · 2022-04-19 15:38
marcanheh, missed a thing, force-pushed the fix to properly fix it now#asahi-dev · 2022-04-19 15:37
marcan(it wasn't properly running the tick on secondaries)#asahi-dev · 2022-04-19 15:36
marcanhopefully I didn't slow it down again :)#asahi-dev · 2022-04-19 15:36
marcanjannau: whoops, forgot a thing, pushed a fix#asahi-dev · 2022-04-19 15:32
marcaneverything should be faster in general, I think macos is just heavier so becomes even worse when everything is fighting#asahi-dev · 2022-04-19 15:23
marcanbut yeah#asahi-dev · 2022-04-19 15:23
marcannot sure exactly how that uncached bit interacts#asahi-dev · 2022-04-19 15:23
marcanthe issue I fixed was just the MMU mappings#asahi-dev · 2022-04-19 15:22
marcanthough the uncached bit seems unrelated here?#asahi-dev · 2022-04-19 15:22
marcanyeah#asahi-dev · 2022-04-19 15:21
marcan(only 1000 and from the same CPU in the easy case)#asahi-dev · 2022-04-19 15:21
marcanit no longer tries to take a single lock 20000 times per second :p#asahi-dev · 2022-04-19 15:21
marcanjannau: hopefully HV should scale to t6002 core counts better now#asahi-dev · 2022-04-19 15:21
marcanor rather, not#asahi-dev · 2022-04-19 15:19
marcanright..#asahi-dev · 2022-04-19 14:44
marcanthis won't work (right now) if we have to allocate our own page tables, of course#asahi-dev · 2022-04-19 14:44
marcanthere's a flag in pmap-ranges that I'm guessing tells us that; alternatively we could use the real-time DART property and the l2 ranges thing#asahi-dev · 2022-04-19 14:43
marcanjannau: fixed the macos thing. It's what I said; we need to map the pagetables as UC, since DART page table walks count as part of the realtime path, same as the framebuffer itself, and cannot hit the CPU cache.#asahi-dev · 2022-04-19 13:09
marcanso we probably need to map it uncached#asahi-dev · 2022-04-19 13:09
marcanbut we write it, right?#asahi-dev · 2022-04-19 13:07
marcanor maybe they are supposed to be RT, and having them in CPU cache also is a no-no?#asahi-dev · 2022-04-19 13:06
marcanwe might be mapping the *page tables* themselves as RT#asahi-dev · 2022-04-19 13:06
marcanI suspect that's wrong since DART_PTE_VALID sets both bits#asahi-dev · 2022-04-19 13:06
marcan dart->l1[ttbr][idx] = offset | DART_PTE_VALID;#asahi-dev · 2022-04-19 13:05
marcandid we set the uncached mode thing in the DART FB mapping?#asahi-dev · 2022-04-19 13:03
marcanyeah, need to look into that#asahi-dev · 2022-04-19 13:02
marcanlet me know if you need more; it's not 100% automated#asahi-dev · 2022-04-19 13:02
marcanjannau: https://mrcn.st/p/36QzY3JG partial dump#asahi-dev · 2022-04-19 12:44
marcanlet me take a look at that DCP thing#asahi-dev · 2022-04-19 12:44
marcanjannau: pushed your t6k2 stuff to main, with some adjustments#asahi-dev · 2022-04-19 11:16
marcanthough that seems c++-specific#asahi-dev · 2022-04-19 11:15
marcansuspect: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90691#asahi-dev · 2022-04-19 11:11
marcanI can't repro on godbolt on 9.4 either#asahi-dev · 2022-04-19 11:03
marcan11.2.1 throws no warnings for me#asahi-dev · 2022-04-19 10:22
marcansven: can we convert the tunables to some common format in m1n1?#asahi-dev · 2022-04-18 16:00
marcanmillenialhacker: sounds like you're going to be having a lot of fun :p#asahi-dev · 2022-04-18 15:59
marcanlol consistency#asahi-dev · 2022-04-18 15:59
marcannote BLK vs BULK#asahi-dev · 2022-04-18 15:59
marcanthis has LLT/BULK/RT apparently#asahi-dev · 2022-04-18 15:58
marcanUSB has LLT/BLK#asahi-dev · 2022-04-18 15:58
marcanlook like one specifically for realtime stuff which kinda makes sense, and no idea why they have 2 more#asahi-dev · 2022-04-18 15:57
marcanand some other SMMU nonsense#asahi-dev · 2022-04-18 15:57
marcanso then 3 DARTs for ISP#asahi-dev · 2022-04-18 15:57
marcanyeah, makes sense, it's before the DART in disp0 and I remember that