#asahi-dev · 2023-11-21 10:51
nullmaz: they're requested with IRQF_NO_THREAD, which should take effect regardless?#asahi-dev · 2023-01-30 17:26
nullcongrats#asahi-dev · 2023-01-30 17:26
nullhahaha#asahi-dev · 2023-01-30 17:10
nullI won't bet on the absolute time, but I will bet at least one RCU stall in the L2 guest#asahi-dev · 2022-12-15 18:55
nullmaz: I forget, did you end up with my T1000 in the end? IIRC I still have the rails in the cupboard under my stairs#asahi-dev · 2022-12-15 11:38
nullNah, I'm actually on an M1 macbook pro today; on the ergodox I'm a bit better#asahi-dev · 2022-12-15 11:38
nullI must really learn to type better#asahi-dev · 2022-12-15 11:37
nullI do appreciate that argument, but on the flip side there are a bunch of such things that over vendors do that could arguably be similar, and the calulus for upstream isn't just "enablnig this one thing for this one vendor's HW"#asahi-dev · 2022-12-15 11:34
nullNot blaming you, I just want ot be clear here so it's not a surprise on the list#asahi-dev · 2022-12-15 11:33
nullI don't think it's leaving users screwed as such, and while I appreciate people want every ounce of performance they can get, I think this is setting ourselves up for a much larger set of problems going forwards#asahi-dev · 2022-12-15 11:32
nullI appreciate that argument, but that's "twice as fast, on specific hardware, using an IMP-DEF feature', and crap everywhere else#asahi-dev · 2022-12-15 11:30
null... and the architecture is clearly going in a different direction, with special instructions rather than a mode#asahi-dev · 2022-12-15 11:30
nullI'm very much not keen on relaxing the policy, because it implies a very long tail of consequences, and we have no guarantee tthat a vendor won't completely change the thing in arbitrary ways in a future release, so we can't really commit to an ABI#asahi-dev · 2022-12-15 11:29
nullWIth neted virt it's not ;)#asahi-dev · 2022-12-15 11:28
null(and likewise, lower ELs should not assume the presence of IMP-DEF stuff unless explicilty advertised (e.g. in DT for PMUs and such)#asahi-dev · 2022-12-15 11:28
null... and the policy in general is simply "Do not expose IMP-DEF stuff to lower ELs"#asahi-dev · 2022-12-15 11:27
nullmarcan: sure, I'm just trying to say that in general it doesn't work, and to make it work requires a lot more finesse than may be initially obvious, since you need to reliably inform both ends, etc#asahi-dev · 2022-12-15 11:27
nullor any hypervisor, since when running at EL1 we don't know the hypervisor#asahi-dev · 2022-12-15 11:26
nullyes#asahi-dev · 2022-12-15 11:26
nullLikewise, so does advertising it#asahi-dev · 2022-12-15 11:26
nullEnabling the IMP-DEF bit *within* the kernel for this goes against the very strong upstream policy that we do not do so, doesn't work in VMs and so on (because it'll trap to the hypervisor)#asahi-dev · 2022-10-24 16:51
nullOh, now I re-read the above, to clarify: the set I listed was a "please include" rather than "please limit to"#asahi-dev · 2022-10-24 16:47
nullCool; have a good evening!#asahi-dev · 2022-10-24 16:46
nullYes#asahi-dev · 2022-10-24 16:46
nullI reckon just linux-arm-kernel would be good, with Ard Biesheuvel, Will Deacon, Catalin Marinas, and myselfg#asahi-dev · 2022-10-24 16:45
nullOh yes, I forgot which timezone you were in!#asahi-dev · 2022-10-24 16:44
nullI'm afraid I need to disappear shortly, but I can reply to things tomorrow and/or follow up on the list#asahi-dev · 2022-10-24 16:44
nullCool :)#asahi-dev · 2022-10-24 16:43
nullI guess my major point of contention here is "pretending this is PSCI"; if this is a "PSCI-like, but separate" UEFI runtime service, I'm not opposoed to that. That way we can manage the distinct parts (e.g. all the conduit subtleties above) separately and more manageably#asahi-dev · 2022-10-24 16:42
nullMaybe, but then you have to always have that mapped, and for things like control flow integrity we don't really want to have FW-provided pages mapped executable all the time#asahi-dev · 2022-10-24 16:40
nullFor the "one true way", I understand and agree with the principle that PSCI *should* be the one way, and we don't want any more implementations than strictly necessary. If we need a generic PSCI analogue for systems without EL2 I'm ok with that, and would prefer that we don't articifically try to force that into PSCI, but others seem to have different views there#asahi-dev · 2022-10-24 16:38
nullFor the *conduit*, HVC/SMC vs UEFI: HVC/SMC are easy to do in any kernel context, but UEFI means we have to mess with TTBRs, and use a lot more infrastrucure, which ends up interacting with things like lockdep/rcu/context-tracking more painfully, and will need a lot of thought. That's definitely an ML disucssion #asahi-dev · 2022-10-24 16:36
nullmhm#asahi-dev · 2022-10-24 16:35
nullThis is a bit difficult to write-up since there's a number of overlapping concerns, and I need to think for a bit to get the concepts in order#asahi-dev · 2022-10-24 16:33
nullIf that has to go thourhg UEFI runtime services then my main concern would be that it is plumbed separately from PSCI, even if it's PSCI-inspired#asahi-dev · 2022-10-24 16:32
nullI think wahatever we do we'll end up reinventing PSCI; I just don't want us to be under the illusion that what we're making is PSCI, since there are big differences in a number of areas#asahi-dev · 2022-10-24 16:31
nullMhmm, that is a fair point#asahi-dev · 2022-10-24 16:29
nullA first approximation would be spin-table with a cpu-return-address or similar to do hotplug, specified a bit like the ePAPR spin-table (which arm64's spin-table is a not-quite-clone of), but I'll admit I haven't thought much about idle rather than hotplug#asahi-dev · 2022-10-24 16:28
null... so I might just have missed that#asahi-dev · 2022-10-24 16:28
nullmarcan: I don't recall the discussions for the "reinventing UEFI" part, but I think there's been a lot of point-to-point conversions between people and not necessarily a single joined up story#asahi-dev · 2022-10-24 11:36
null... in case that's of any help#asahi-dev · 2022-10-24 11:36
nullto remove upstream branches I have an alias: prune-all = "!git remote prune $(git remote)"#asahi-dev · 2022-05-17 13:30
nullSOrry, that was meant to be `/window 9` but I hit the wrong key#asahi-dev · 2022-05-17 13:30
nullwinter: 9