#asahi-dev · 2022-09-05 17:12
marcan*ouch*, but nice work on the rework!#asahi-dev · 2022-09-05 17:12
marcanmy UPSes are 300VA / 600VA IIRC, this is small scale stuff :)#asahi-dev · 2022-09-05 17:11
marcanrmk: oof, make sure you don't accidentally overload it :D#asahi-dev · 2022-09-05 17:11
marcantomorrow I'll try to go through all the kernel merging stuff, especially audio, and push a 6.0 rebase to -dev#asahi-dev · 2022-09-05 17:10
marcananyway, I should sleep :)#asahi-dev · 2022-09-05 17:10
marcan(but hey)#asahi-dev · 2022-09-05 17:10
marcan(mind you, this is somewhat academic, since... I've had zero power outages and two split-second brownouts in 8 years in japan)#asahi-dev · 2022-09-05 17:09
marcan(intel machines don't help :D)#asahi-dev · 2022-09-05 17:08
marcanright now my UPS cries every now and then due to overload, it's serving both my studio *and* server closet#asahi-dev · 2022-09-05 17:08
marcanand I have a new UPS upstairs on the Threadripper temporarily, that needs to go in my studio (and the studio UPS in the server closet, probably - I think once it's all Macs that will have the lower power consumption ;))#asahi-dev · 2022-09-05 17:08
marcanreminds me I still need to redo my entire server setup... now that the Mac Studio has usable USB, I really can start the process#asahi-dev · 2022-09-05 17:07
marcanniiiice!#asahi-dev · 2022-09-05 17:06
marcanbut really, the sooner we get through all these growing pains the better, so thank you again for upstreaming this#asahi-dev · 2022-09-05 17:06
marcanand for alt distros, well, they know they're still in playing with fire territory right? ;)#asahi-dev · 2022-09-05 17:05
marcanso nobody will break#asahi-dev · 2022-09-05 17:05
marcanfor our reference distro, we sync kernel and DTs 1:1 on update#asahi-dev · 2022-09-05 17:05
marcanthat's fine, and I might drop the init in asahi anyway#asahi-dev · 2022-09-05 17:04
marcanyup#asahi-dev · 2022-09-05 17:03
marcan(and I don't want to come across as pushing unreviewed DTs on robher because OpenBSD picked them up either :) )#asahi-dev · 2022-09-05 17:03
marcanas long as it would work out for existing drivers, I don't mind keeping around an ugly DT for a while if it means the end resultis better#asahi-dev · 2022-09-05 17:03
marcanwe can also e.g. duplicate node names if we end up changing them#asahi-dev · 2022-09-05 17:00
marcanpersonally, I don't want to lock us into bad bindings due to openbsd being faster than us, but I'd rather avoid flag-day hell for kettenis, so ideally feel free to make any changes where a single (non-schema-compliant is fine) DT can work both the old and the new way, and if needed we can go that route#asahi-dev · 2022-09-05 16:59
marcanalso, I can keep transitional DTs in our tree#asahi-dev · 2022-09-05 16:59
marcanyeah; adding compatibles would be backwards-compatible though, so that much is no issue if we keep the same node names#asahi-dev · 2022-09-05 16:58
marcanI should get some sleep, it's 2AM here :)#asahi-dev · 2022-09-05 16:58
marcan(as I said I'm new to mfd so feel free to screw around with how that all hooks together at that level, if it makes things better)#asahi-dev · 2022-09-05 16:57
marcanre DT, fine with compatibles, not terribly clear on how that works myself, so if you want to try it, go for it. it could be useful in the future to have per-SoC compatibles anyway so we can quirk things when inevitably my idea of "SMC can just autoprobe funniness" breaks down...#asahi-dev · 2022-09-05 16:56
marcanrmk: I'll punt on the IRQ support comments since you're dropping that for now (and it could use more testing), so I think I covered everything else but DT#asahi-dev · 2022-09-05 16:56
marcanheh#asahi-dev · 2022-09-05 16:35
marcanwhen the whole cycle is fractions of a millisecond and the benchmark is naive and does not queue anything, all that time wasted is FPS you lose#asahi-dev · 2022-09-05 16:34
marcanespecially in the fully synchronous case, which I think glxgears is#asahi-dev · 2022-09-05 16:34
marcanbut AGX, yeah, that needs a doorbell that goes via mailbox for every command submission, unless the driver does some kind of batching/elision, and given how fast the hardware is apparently... I'm expecting it to matter.#asahi-dev · 2022-09-05 16:33
marcan(at least not until you're an Apple Watch or something)#asahi-dev · 2022-09-05 16:32
marcanDCP will never be a real bottleneck, modesetting/page flipping doesn't care about a few extra context switches#asahi-dev · 2022-09-05 16:31
marcanlet's just wait for Lina ;)#asahi-dev · 2022-09-05 16:31
marcanbut right now we don't have to care, nothing so far cares about RTKit performance#asahi-dev · 2022-09-05 16:30
marcanthat said, Apple also have some really funny overengineered stuff in their IOKit drivers, so I wouldn't be surprised if we get better FPS than them at some microbenchmarks in the end#asahi-dev · 2022-09-05 16:29
marcan(and that FIFO isn't even used right now, because mailbox)#asahi-dev · 2022-09-05 16:28
marcanand then we put a pile of nonsense in the final bottleneck for what should be a single MMIO write if there is FIFO space available? lol.#asahi-dev · 2022-09-05 16:28
marcanyes, for a GPU which seems to be explicitly designed with multiple queues just to avoid CPU lock contention#asahi-dev · 2022-09-05 16:27
marcan"increases glxgears FPS by 30%", and out the window it goes#asahi-dev · 2022-09-05 16:27
marcanyup, I'm waiting for that#asahi-dev · 2022-09-05 16:22
marcanso chances are the atomic stuff will only ever be used for the final shutdown/reboot messages, and honestly that's never going to hit any race conditions#asahi-dev · 2022-09-05 16:22
marcanIIRC I started running into that during an attempt to make the SMC IRQ stuff work in atomic context, before I discovered the trick to writing non-atomic irqchips#asahi-dev · 2022-09-05 16:21
marcanbut if you just want to make SMC work, the first two patches are probably all you need, even if there is lingering brokenness in the driver#asahi-dev · 2022-09-05 16:21
marcanand I'm this close to ripping out that entire driver and embedding the mailbox support into rtkit, since it's a massive impedance mismatch and has only caused us pain (abstractions/subsystems are only useful when they actually increase the useful combinatronics - this doesn't, it just causes pain between two sides that will only ever talk to each other)#asahi-dev · 2022-09-05 16:20
marcanthat last patch in particular, I swear the original code is backwards but I just find the design of that thing so confusing#asahi-dev · 2022-09-05 16:19
marcanso you might want to review the commits yourself and see if you agree#asahi-dev · 2022-09-05 16:19
marcanthe first 2 patches, but then I discovered some horribleness (that entire subsystem is a pet peeve of mine...)#asahi-dev · 2022-09-05 16:19
marcansee bits/010-mailbox