#asahi-dev · 2022-12-15 11:27
mazChaosPrincess: that's owing to its SPARC heritage. nobody forced them to build shit HW... ;-)#asahi-dev · 2022-12-15 11:26
mazthe distinction is that Linux itself doesn't need to know. the whole VM is TSO, end of story.#asahi-dev · 2022-12-15 11:25
mazmarcan: I wonder if we can somehow make the TSO stuff a VM property instead of a per-process thing.#asahi-dev · 2022-12-15 11:23
mazthe direction of travel so far is the exact opposite.#asahi-dev · 2022-12-15 11:23
mazChaosPrincess: you obviously have better information on the architecture than I do...#asahi-dev · 2022-12-15 11:20
maz(see, we're getting there! ;-)#asahi-dev · 2022-12-15 11:20
maz... and will be used by everyone because it will hide bugs.#asahi-dev · 2022-12-15 11:19
mazthe main issue is that SW that fails without TSO and works with it will simply always enable it instead of debugging the stuff.#asahi-dev · 2022-12-15 11:18
mazmarcan: gut feeling: at best, you'll get ignored. at worse, this will go down into a flame war around an effective fork of the whole userspace ABI... and all the variants in between.#asahi-dev · 2022-12-15 11:09
maz(understatement of 2022)#asahi-dev · 2022-12-09 17:47
mazkettenis: don't be sorry. I understand the complexity of this stuff, and I'm not moaning at anyone.#asahi-dev · 2022-12-09 15:15
mazmarcan: life is too short! I even depend on some ancient u-boot that still had the USB-A ports working...#asahi-dev · 2022-12-09 14:30
mazone day I'll need to update my mini... it is still running the hack I put there 18 months ago, with 11.x stubs.#asahi-dev · 2022-12-05 10:35
mazmarcan: I'm not too worried about the single-IPI behaviour. if you trace them, you'll see that multi-target IPIs are super rare.#asahi-dev · 2022-12-05 10:34
mazmarcan: yeah, it only took 15 revisions to get there...#asahi-dev · 2022-12-05 10:33
mazmarcan: thanks for having had a look at the IPI stuff.#asahi-dev · 2022-11-30 11:01
mazkettenis: no worries. for the few (and hopefully rare) cases where DT becomes a burden, we should consider alternatives...#asahi-dev · 2022-11-29 18:29
mazthis is mostly a copy-paste for the AIC driver, and I hope to try and move the driver over to that...#asahi-dev · 2022-11-29 18:28
mazmarcan: btw, re: irq and memory ordering. feel free to eyeball 20221129142449.886518-4-apatel@ventanamicro.com if you get a chance.#asahi-dev · 2022-11-29 16:07
mazI find the whole thing more and more depressing.#asahi-dev · 2022-11-29 14:03
mazprobably less eloquently ;-)#asahi-dev · 2022-11-29 14:03
mazI said what I had to say as well.#asahi-dev · 2022-11-29 12:24
mazyeah. classic example of better being the ennemy of good...#asahi-dev · 2022-11-29 11:49
mazand I thought the dust was settling on the cpufreq binding... oh well.#asahi-dev · 2022-11-24 11:51
mazjust to get this 20% frequency bump, you need to perform some significant uarch changes to meet the timing constraints.#asahi-dev · 2022-11-24 11:49
mazthat's sort of in line with my 15% perf difference.#asahi-dev · 2022-11-24 11:40
mazwow, that's a huge bump.#asahi-dev · 2022-11-24 11:35
mazthat is*#asahi-dev · 2022-11-24 11:35
mazmarcan: do you know whether that frequency, uarch or both?#asahi-dev · 2022-11-23 18:00
mazamarioguy: bah, there wasn't much overhead from the traps the first place (interrupts are a pretty rare event anyway)#asahi-dev · 2022-11-23 16:50
mazhuh, the E-cores are noticibly faster on M2 than on M1. about 15% quicker for a KVM guest pinned to the small cluster.#asahi-dev · 2022-11-23 15:57
mazYup, SError gets nicely delivered to the guest instead of landing on the host. progress!#asahi-dev · 2022-11-23 15:52
mazah, looks like M2 has the vgic SError bug fixed... need some more checking, but looking better so far.#asahi-dev · 2022-11-23 09:11
mazmarcan: solved my X issue with 6.1 on m2: the FB_EFI was getting in the way. with 6.0, simpledrm is able to take over from EFI, but not with 6.1, apparently. just disabled the option, and all is good.#asahi-dev · 2022-11-22 14:50
mazalso, none of that is kernel code. this looks like userspace tooling only.#asahi-dev · 2022-11-22 13:17
mazthe good news is that this pushes the release date to the year 4097, give or take.#asahi-dev · 2022-11-22 09:38
mazI'll diff the X logs when I get a chance.#asahi-dev · 2022-11-22 09:37
mazI saw that there is an APPLEDRM config, but didn't investigate that.#asahi-dev · 2022-11-22 09:36
mazyeah, it's not a fresh install. everything is working well with 6.0, and I compiled my own 6.1 kernel. all OK except for X (VTs are OK).#asahi-dev · 2022-11-22 09:35
mazah, talking about graphics. has there been a significant change in the asahi kernel between 6.0 and 6.1 that makes X trip over on M2?#asahi-dev · 2022-11-22 09:34
mazEmacs, man, Emacs!#asahi-dev · 2022-11-22 09:34
mazand I'm glad, less to do on my side! :D#asahi-dev · 2022-11-22 09:33
maz(this is starting to look a lot like $work, and I'm going to look away for a bit... ;-)#asahi-dev · 2022-11-22 09:32
mazsomething like that.#asahi-dev · 2022-11-22 09:31
mazunless you start doing carve-outs for the buffers to appear into.#asahi-dev · 2022-11-22 09:31
mazit means that the IPA space changes from under the guest's feet...#asahi-dev · 2022-11-22 09:31
mazhuh. that's pretty terrible.#asahi-dev · 2022-11-22 09:30
mazyeah, you'd need some scatter-gather mechanism for the GPU to linearize buffers.#asahi-dev · 2022-11-22 09:28
mazas I said, if this is just about a kick all the way down to userspace, and for userspace to peek at a shared buffer, there is 0 effort on the hypervisor side, which suits me.#asahi-dev · 2022-11-22 09:26
mazif, on the other hand, this is blindly forwarded to userspace, then I have no objection.