#
✕ nick:marcan
#asahi · 2021-11-16 16:30
marcan width: 640
#asahi · 2021-11-16 16:27
marcanalso I still find it hilarious that without a screen it sets up a frambuffer with the portrait screen dimensions of the iPhone 4 or something
#asahi · 2021-11-16 16:26
marcanah, then it *is* connected to a "screen" :p
#asahi · 2021-11-16 16:26
marcan(dunno why, retries I guess)
#asahi · 2021-11-16 16:26
marcan< sven> (i think. my m1 isn't connected to any screen) <- you should fix that, it makes boot a good 10 seconds slower or so
#asahi · 2021-11-16 16:24
marcanm1n1 does the init that iBoot doesn't and bridges to the devicetree world; u-boot handles I/O and provides EFI services; grub is provided by distros and is what they know how to integrate and manage.
#asahi · 2021-11-16 16:23
marcanmort_: each boot stage does something discretely useful. there is nothing wrong with having things split up into stages.
#asahi · 2021-11-16 16:23
marcansven: yeah, I agree
#asahi · 2021-11-16 16:23
marcanfwiw I already have proper devicetrees for all the t6k models in my branch (at least for upstream stuff, plus the things I'm merging in)
#asahi · 2021-11-16 12:21
marcankettenis: yeah, working on that, just got sidetracked by MCC today
#asahi · 2021-11-16 12:14
marcanyes
#asahi · 2021-11-16 12:14
marcan(which is why adding them is a good idea before the drivers are ready)
#asahi · 2021-11-16 12:14
marcanyes, everything not used
#asahi · 2021-11-16 12:12
marcanmaybe things will be magically even faster now :)
#asahi · 2021-11-16 12:12
marcanpushed some MCC magic (in a new mcc.c) to handle the t8103/t6000 MCCs properly and ostensibly do the magic cache enable dance
#asahi · 2021-11-16 11:29
marcanproblem is if I can't repro I can't fix it :/
#asahi · 2021-11-16 11:11
marcan(I talked about this in -dev yesterday)
#asahi · 2021-11-16 11:11
marcansince we'll need something like that for sleep mode
#asahi · 2021-11-16 11:10
marcanwhich is also why we need to invent a new wheel so m1n1 can provide PSCI from the same EL2 as the kernel
#asahi · 2021-11-16 11:10
marcanno
#asahi · 2021-11-16 11:10
marcanthe hpm0 thing?
#asahi · 2021-11-16 11:01
marcan-s
#asahi · 2021-11-16 11:01
marcanthis isn't necessary in a "normal" OS which only maps things on demand, but in m1n1 we want to map all physical RAM by defaults to simplify experiments
#asahi · 2021-11-16 11:01
marcanwe need to unmap them because otherwise the CPU might speculatively read from them, and cause a SError
#asahi · 2021-11-16 11:00
marcanit's used for coprocessor carveouts instead
#asahi · 2021-11-16 11:00
marcanthere is no actual EL3/trustzone on these chips
#asahi · 2021-11-16 11:00
marcanmostly NVMe and SEP related
#asahi · 2021-11-16 11:00
marcanit just does that for some RAM regions
#asahi · 2021-11-16 10:56
marcanit's just the carveouts
#asahi · 2021-11-16 10:56
marcanyes, the only TrustZone feature Apple still use on paper
#asahi · 2021-11-16 10:53
marcan(and the m1n1 code that checks the TZ regs is not correct on t6000 and works by accident, heh)
#asahi · 2021-11-16 10:53
marcanalso AMCC got interesting in t6000... t8103 had 1 AMCC with 4 planes and 8 channels; t6000 seems to have 4 AMCCs with 4 planes and ? channels, and t6001 has 8 AMCCs.
#asahi · 2021-11-16 09:54
marcan(the AMCC cache, not the CPU cache)
#asahi · 2021-11-16 09:53
marcanalso it seems we might not be using the cache at all in m1n1/linux yet? there is explicit enable code...
#asahi · 2021-11-16 09:49
marcanbut LPDDR with 16-bit channels... what are you going to do, make them 24 bits?
#asahi · 2021-11-16 09:49
marcanDDR5 has 32-bit channels that become 40-bit channels with ECC (2x the ECC overhead of prior standards)
#asahi · 2021-11-16 09:49
marcanthat's a bit of a problem
#asahi · 2021-11-16 09:48
marcanthe thing is I'm not even sure how you'd do ECC on a system with 16-bit channels
#asahi · 2021-11-16 09:48
marcanhm
#asahi · 2021-11-16 09:48
marcanbank/way/entry would suggest cache
#asahi · 2021-11-16 09:48
marcanah, might be cache ECC
#asahi · 2021-11-16 09:46
marcanso on M1/t8103, there's an mcc_ecc=1 bootarg, but indeed the operation it does is a no-op: the register already has that value
#asahi · 2021-11-16 09:46
marcanyes
#asahi · 2021-11-16 09:45
marcanJTL: all of them
#asahi · 2021-11-16 09:28
marcanI wonder how we could test this...
#asahi · 2021-11-16 09:24
marcanthe T810x driver has a function to enable ECC, but the T6000 one doesn't; I think these days iBoot would enable it during mem init
#asahi · 2021-11-16 09:23
marcanit seems m1n1 already is, if I'm interpreting this right
#asahi · 2021-11-16 09:21
marcan(we don't have a driver for this yet of course, but it's the same reporting stuff that was barfing on T6000 when I mapped the framebuffer as cached and DISP0 realtime fetches were hitting the CPU cache)
#asahi · 2021-11-16 09:21
marcan"AMCC PLANE%d %s error: INTSTS 0x%016llx ECCONEBITERRLOG(Bank/Way/Entry/BitOff) %#x(%#x/%#x/%#x/%#x) ERRBITCNT %#x" @%s:%d
#asahi · 2021-11-16 09:21
marcan"AMCC PLANE%d %s error: INTSTS 0x%016llx ECCMULTIBITSERRLOG(Bank/Way/Entry) %#x(%#x/%#x/%#x)" @%s:%d

plus ›