#asahi-dev · 2025-04-04 09:08
mazrobher: any clue?#asahi-dev · 2025-04-04 09:08
mazI guess this should be part of describing the slots past the switch, but how to express that numbering is a mystery. wonder if there's prior art around that.#asahi-dev · 2025-04-04 09:03
mazhuh, what a mess.#asahi-dev · 2025-04-04 09:03
jannaunot sure how to model the reset and power enable gpios of devices behind the switch in the devicetree#asahi-dev · 2025-04-04 09:01
jannaumaz: mem assigment breaks down with hotplug but if all devices are powered up before booting it mostly works after putting "bus-range = <0 8>;" in the device tree#asahi-dev · 2025-04-04 08:58
jannauI haven't checked beyond quickly testing that old radeon hd4350#asahi-dev · 2025-04-04 08:57
chaos_princessdid they fix device mappings on pcie in that box?#asahi-dev · 2025-04-04 08:57
chaos_princessjannau: oh, wow, i salute you and your wallet for that sacrifice. #asahi-dev · 2025-04-04 08:56
jannauyes, switchtec. it's a microchip PM40100#asahi-dev · 2025-04-04 08:55
mazdo we have a native driver for the switch in this box? or relying on the generic Linux code?#asahi-dev · 2025-04-04 08:54
mazI usually have to play all sort of command-line tricks to deal with PCIe switches...#asahi-dev · 2025-04-04 08:52
mazyeah. it would take my M1 studio to die on me to be able to justify it. and even then, a 128 core Altra could make more sense, given what I do with it.#asahi-dev · 2025-04-04 08:52
jannaupcie setup is not fully working. the switch management endpoint is not working and the devices behind the switch only come up with PCIe gen1#asahi-dev · 2025-04-04 08:51
jannauI got it refurbished for a little bit over half price. still expensive though#asahi-dev · 2025-04-04 08:30
mazjannau: huh, right. oh well, less code anyway, and additional functionality. too bad I can't afford that box! :)#asahi-dev · 2025-04-04 08:28
jannaumaz: it has 2 root ports which was incompatible with the bus_notifier approach#asahi-dev · 2025-04-04 08:27
mazjannau: what was preventing this from working before my changes? I didn't really expect any functional difference with them.#asahi-dev · 2025-04-04 08:02
svenany of the deeper wfi states being broken would be very bad :(#asahi-dev · 2025-04-04 07:32
chadmednickchan: oh thats an apple register#asahi-dev · 2025-04-04 07:17
nickchanthere are some alternatives that have not been tried however like turning off cpu power domains in pmgr#asahi-dev · 2025-04-04 07:15
nickchanspace heater is unfun#asahi-dev · 2025-04-04 07:14
nickchanthough I can see that will make people much more willing to play cat and mouse#asahi-dev · 2025-04-04 07:14
nickchanchadmed: why would hiding cyc_ovrd defeat the purpose?#asahi-dev · 2025-04-04 07:12
chadmeddunno if arm would be happy about that#asahi-dev · 2025-04-04 07:12
chadmedif apple hid it in gl2 that would kinda defeat the purpose of doing this to make raw boot objects only see the in-spec architecture #asahi-dev · 2025-04-04 07:10
chaos_princessWould be interesting to see if cyc_ovrd is inaccessible just in el2#asahi-dev · 2025-04-04 07:10
nickchanand that means on m4 by default wfi is already the shallow wfi#asahi-dev · 2025-04-04 07:09
nickchanand yuka's fork forced the a7-a11 code path on wfi which is a shallow wfi#asahi-dev · 2025-04-04 07:09
nickchanchaos_princess: CYC_OVRD is supposedly completely inaccessible so there is no wfi configuration to speak of#asahi-dev · 2025-04-04 07:08
chaos_princessIs there any m4 fork that does chicken bits and wfi configuration correctly#asahi-dev · 2025-04-04 07:07
chadmedthats fucking stupid and surely not intended#asahi-dev · 2025-04-04 07:07
nickchanit does look like it assuming the m4 forks is not making a horribly incorrect assumption#asahi-dev · 2025-04-04 07:06
chadmedso if sptm is torn down due to a raw boot object we just lose deep wfi?#asahi-dev · 2025-04-04 07:06
nickchanchadmed: i think power management things and cpu start stop things like that are handled by sptm because of state loss#asahi-dev · 2025-04-04 07:05
chadmedtheres no way thats being trapped to sptm or anything right#asahi-dev · 2025-04-04 07:04
chadmedyeah id rather not go back to space heater laptops#asahi-dev · 2025-04-04 07:03
nickchanand we use that and not having that would make for a really bad experience#asahi-dev · 2025-04-04 07:03
nickchandeep wfi#asahi-dev · 2025-04-04 07:02
nickchanfrom some of the m1n1 m4 support code i think wfi is also broken somehow#asahi-dev · 2025-04-04 07:02
nickchanchadmed: another thing#asahi-dev · 2025-04-04 07:00
chadmedthis doesnt affect M3, only M4 (and all new socs going forward presumably) right? #asahi-dev · 2025-04-04 06:58
chadmedwel... i guess it's good that iboot is not encrypted now :/ #asahi-dev · 2025-04-04 06:56
chadmedwhos going to be doing x86 emu on an iphone anyway :p #asahi-dev · 2025-04-04 06:56
nickchannot like that matters in practice already confirmed it does not exist on a11#asahi-dev · 2025-04-04 06:55
nickchanand after looking into xnu sources it seems that a12 and possibly a13 have TSO too, but not ACTLR, it's HID9[16]#asahi-dev · 2025-04-04 06:55
chadmedyeah which is why i think arm probably had some words to apple and they picked the nuclear option. which is just kind of really dickish of them#asahi-dev · 2025-04-04 06:55
chadmedthen of course there was the whole apple-specific TSO discussion that kinda went nowhere which probably didnt help in that regard especially since FEAT_LRCPC was already a thing#asahi-dev · 2025-04-04 06:53
nickchanah, *that* got posted onto the mailing list#asahi-dev · 2025-04-04 06:51
nickchani think there's also a thing where some performance counters have some edge case regarding amx#asahi-dev · 2025-04-04 06:49
chadmedaaaages ago some folks were asking around about adding AMX support to arm64 linux. we always maintained that we would never _ever_ support anything even remotely like that but i think upstream caught wind of it