#
✕ nick:marcan
#asahi-dev · 2021-11-03 10:50
marcanI strongly suspect the second block is where the second die's IRQs will show up on the rumored 2-die systems
#asahi-dev · 2021-11-03 10:50
marcanonly one of which is in use right now
#asahi-dev · 2021-11-03 10:50
marcanbut the current implementation has 1806 implemented, times 2 blocks
#asahi-dev · 2021-11-03 10:49
marcanwell, AICv2 has up to 4096 per block
#asahi-dev · 2021-11-03 10:48
marcanup to 1024 I think? 700-odd in practice IIRC
#asahi-dev · 2021-11-03 10:43
marcantree it is then, and then I'll just split the hwirq numbers into 16/16 bits and avoid the division
#asahi-dev · 2021-11-03 10:41
marcan(also comments are wrong etc)
#asahi-dev · 2021-11-03 10:40
marcanwhole thing needs to be redone and reworded
#asahi-dev · 2021-11-03 10:40
marcansven: thanks, and yeah, the commits are BS right now :-)
#asahi-dev · 2021-11-03 10:40
marcanhowever, with ~3000 hwirqs at this point even without holes (but with 2 groups), maybe that makes sense already?
#asahi-dev · 2021-11-03 10:40
marcanmaz: I was wondering about the division in the mask/unmask functions; alternatively I could make hwirq sparse and use bitfields, but at that point I really need to move to irq_domain_create_tree
#asahi-dev · 2021-11-03 10:39
marcanyeah
#asahi-dev · 2021-11-03 10:37
marcanhow did that ever work with 52-bit addrs?
#asahi-dev · 2021-11-03 10:37
marcanit's clearly a bug
#asahi-dev · 2021-11-03 10:37
marcansven: yeah
#asahi-dev · 2021-11-03 10:37
marcanyeah, I'll try to break it up into the compatible/fastipi handling first, and then add t6k / groups
#asahi-dev · 2021-11-03 10:27
marcan(plus the next commit in t6000/bringup-work for the DART fix)
#asahi-dev · 2021-11-03 10:27
marcanmaz: fwiw here's the initial bringup from last night: https://github.com/AsahiLinux/linux/commit/e4432729d38916392479252260c8434870a50922
#asahi-dev · 2021-11-03 10:26
marcaner, minus the second one
#asahi-dev · 2021-11-03 10:25
marcanMPIDRs are 00000[01] 000001 01010[0-3] 0102[0-3]
#asahi-dev · 2021-11-03 10:24
marcanmaz: yes
#asahi-dev · 2021-11-03 09:18
marcanbut a lot are copies/symlinks
#asahi-dev · 2021-11-03 09:18
marcan286 files is for all possible board/antenna/vendor/etc combinations
#asahi-dev · 2021-11-03 09:18
marcanj_ey: yes, once I commit it properly
#asahi-dev · 2021-11-03 08:29
marcanalso threw some shade at stan :p
#asahi-dev · 2021-11-03 08:28
marcanreplied
#asahi-dev · 2021-11-03 07:44
marcanI guess I need to review that one :p
#asahi-dev · 2021-11-03 07:44
marcansigh...
#asahi-dev · 2021-11-02 16:24
marcansgtm, I'll try that
#asahi-dev · 2021-11-02 16:17
marcanso IRQs are now identified as die, hwidx
#asahi-dev · 2021-11-02 16:17
marcanso that you can handle all the IRQs from one AIC, but you get one sub-block per die in it
#asahi-dev · 2021-11-02 16:16
marcanpresumably they have some kind of arrangement where AIC on one die forwards to AIC on the other or so
#asahi-dev · 2021-11-02 16:16
marcanthe irq *source*
#asahi-dev · 2021-11-02 16:14
marcan(cc robher)
#asahi-dev · 2021-11-02 16:14
marcanthoughts?
#asahi-dev · 2021-11-02 16:14
marcanwe could encode the die||irq into nr, or we could start accepting a 4-cell form
#asahi-dev · 2021-11-02 16:13
marcanfor M1 we used <0 nr flags> for IRQs and <1 nr flags> for FIQs
#asahi-dev · 2021-11-02 16:13
marcanhowever, now we have another dimension for the phandle args
#asahi-dev · 2021-11-02 16:13
marcanmaz: so this looks like it supports multi-chip systems (at least it has two interrupt blocks and I can figure out how the second one works even though it's idle)
#asahi-dev · 2021-11-02 16:02
marcanI'm not going to spend much time on it, but I also won't gratuitously break code that might work there
#asahi-dev · 2021-11-02 15:59
marcanyeah, I'd do that (and just trigger on the t8103 compatible, so the legacy IPIs can stay for retro chips)
#asahi-dev · 2021-11-02 15:58
marcanAIC itself is similar enough; I just need to indirect register offsets since they all moved, and have some conditionals in init paths
#asahi-dev · 2021-11-02 15:57
marcanthe vIPI stuff can be shared
#asahi-dev · 2021-11-02 15:56
marcanmaz: any preference as to how to organize it? stick it in the AIC driver, or try to split it into a separate irqchip?
#asahi-dev · 2021-11-02 15:56
marcanAIC just lost its legacy IPIs
#asahi-dev · 2021-11-02 15:56
marcanmaz: yes, they work the same way as on M1, that hasn't changed
#asahi-dev · 2021-11-02 15:31
marcan(there is a 4-bit per-irq config field that could be a routing config, but only 0 works for me)
#asahi-dev · 2021-11-02 15:31
marcan(I also can't for the life of me figure out IRQ affinity... it seems it just tries cores in order with a delay, with a global bit to select whether to start with pcores or ecores...)
#asahi-dev · 2021-11-02 15:30
marcanmaz: so it looks like AIC IPIs are gone in aic2, do you have any particular preference for implementing fast (sysreg/fiq) IPIs?
#asahi-dev · 2021-11-02 11:46
marcancurrently fighting AICv2

plus ›