#asahi-dev · 2022-09-26 13:01
marcansounds like they just... carve out the entire GPU AS from userspace. lovely.#asahi-dev · 2022-09-26 13:00
marcanI think lina was wondering how they stopped the CPU tlbi from shooting down random userspace TLBs when shooting down GPU TLBs?#asahi-dev · 2022-09-26 13:00
marcan#define MACH_VM_MAX_GPU_CARVEOUT_ADDRESS_RAW 0x0000007000000000ULL#asahi-dev · 2022-09-26 13:00
marcan#define MACH_VM_MIN_GPU_CARVEOUT_ADDRESS_RAW 0x0000001000000000ULL#asahi-dev · 2022-09-26 13:00
marcanoh god#asahi-dev · 2022-09-26 12:57
marcan(and then fix all the fallout from it)#asahi-dev · 2022-09-26 12:57
marcanso I guess that's what Linux would have to do to make page size selectable *at boot time*#asahi-dev · 2022-09-26 12:57
marcan#define PAGE_MASK vm_page_mask#asahi-dev · 2022-09-26 12:57
marcan#define PAGE_SIZE vm_page_size#asahi-dev · 2022-09-26 12:57
marcan#define PAGE_SHIFT vm_page_shift#asahi-dev · 2022-09-26 12:57
marcanthey do have this though, but this is system-global#asahi-dev · 2022-09-26 12:55
marcanmaybe the rest of the kernel just doesn't assume PAGE_SIZE is real that much?#asahi-dev · 2022-09-26 12:54
marcansven: I gave the xnu thing a skim and I don't see any magic, other than that they made the entire VM subsystem for ARM support mixed page sizes.#asahi-dev · 2022-09-26 12:38
marcanI know they just have separate page ops for the actual VM management, but I mean the interaction with the rest of the kernel#asahi-dev · 2022-09-26 12:38
marcanI should look into how they pulled off mixed page size processes in xnu, I'm still surprised that actually works#asahi-dev · 2022-09-26 12:36
marcanlooks like it's 4K only, so yeah#asahi-dev · 2022-09-26 12:33
marcanI don't think porting an old version of NT that didn't even have 64-bit support to ARM64 is going to be exactly "fun"#asahi-dev · 2022-09-26 12:30
marcanyeah, heh#asahi-dev · 2022-09-26 12:30
marcanthat requires microsoft cooperation to implement#asahi-dev · 2022-09-26 12:30
marcanwindows cannot run on M1 bare metal because the core kernel does not support AIC/DART#asahi-dev · 2022-09-26 12:30
marcanit's not initialization, it has to be fully virtualized#asahi-dev · 2022-09-26 12:29
marcanbare metal is not without microsoft support, no#asahi-dev · 2022-09-26 12:29
marcanbasically virtualize enough for the core kernel to boot (mostly vGIC) and write drivers for the rest#asahi-dev · 2022-09-26 12:28
marcanit is, under a hypervisor with vGIC support#asahi-dev · 2022-09-26 12:23
marcanyeah, we should fix that#asahi-dev · 2022-09-26 12:22
marcanokay#asahi-dev · 2022-09-26 12:22
marcanthat gives you 32 bits of ias#asahi-dev · 2022-09-26 12:22
marcanohhh so *that* is why there were 4 TTBRs#asahi-dev · 2022-09-26 12:22
marcanone TTBR is enough for ias <= 36#asahi-dev · 2022-09-26 12:22
marcan_jannau_: did we even ever support multiple TTBRs? Right now we set ias to 32, which will always result in one TTBR#asahi-dev · 2022-09-26 12:18
marcan_jannau_: I think setting ias properly should take care of that#asahi-dev · 2022-09-26 12:15
marcan drivers/iommu/apple-dart.c | 559 ++++++++++++++++++++++++++++---------------------------------------------------------------------------------------------------------------------#asahi-dev · 2022-09-26 12:15
marcan$ git diff --stat bits/020-t6000-dart..24d6bbee0c252f0b0bfb7c855074c81da708a128#asahi-dev · 2022-09-26 12:14
marcanwere there? I don't recall any#asahi-dev · 2022-09-26 12:12
marcanI should submit t8110 soon then, hopefully that one won't ruffle any feathers since it's just driver stuff#asahi-dev · 2022-09-26 12:11
marcan\o/#asahi-dev · 2022-09-23 05:53
marcanI've added branch protection rules to asahi-soc/* on the linux repo so only I can push, since this is now significant attack surface (anything pushed there will automatically wind up on a *lot* of developers' machines)#asahi-dev · 2022-09-23 05:52
marcanlooks like we're in linux-next now!#asahi-dev · 2022-09-22 19:36
marcanI did work out that dcp.system.system.setProperty thing from firmware only#asahi-dev · 2022-09-22 19:36
marcanbut you can usually find thingsd#asahi-dev · 2022-09-22 19:36
marcanC++ so lots of vtables I'm too lazy to trace#asahi-dev · 2022-09-22 19:35
marcanthe firmware is, uh, variable#asahi-dev · 2022-09-22 19:35
marcansven: the problem with those error codes is a lot of them are highly overloaded, so it may or may not help find the place in the firmware, but you can certainly try#asahi-dev · 2022-09-22 14:57
marcanhave at it :)#asahi-dev · 2022-09-22 14:57
marcanpackage pushed, smoke tested on t6000, seems to work#asahi-dev · 2022-09-22 14:32
marcanasahi-6.0-rc6-1 tagged and pushed, that's my prior smoketest plus the reset fix change (assuming it works, about to find out with a proper kernel package now)#asahi-dev · 2022-09-22 14:29
marcanemail sent :)#asahi-dev · 2022-09-22 14:22
marcan(well, I already had one, I'll just fast forward it so it's more obvious it's all merged already)#asahi-dev · 2022-09-22 14:21
marcanah, right, I'll make a (trivial) fixes branch too and mention it then#asahi-dev · 2022-09-22 14:18
marcancool, makes sense!