#asahi-dev · 2026-08-24 10:37
chadmedthank you for volunteering to add a binding for wait_for_completion_timeout :P#asahi-dev · 2026-08-24 10:36
tryhard[m]chadmed: seems like there is no rust method for this#asahi-dev · 2026-08-24 10:36
svenI wonder why #asahi-dev · 2026-08-24 10:36
svenjannau: :(#asahi-dev · 2026-08-24 10:35
chadmedor whatever the error condition is for failing to set the report state#asahi-dev · 2026-08-24 10:34
tryhard[m]luckily there is a rust api#asahi-dev · 2026-08-24 10:34
chadmedyou should also use the timeout variant here imo. if the thing is not ready in a short amout of time just bail and -EPROBE_DEFER or something and stop blocking a thread#asahi-dev · 2026-08-24 10:34
tryhard[m]very nice#asahi-dev · 2026-08-24 10:33
chaos_princessthis is the kernel's condition variable-adjacent thing#asahi-dev · 2026-08-24 10:32
chadmedoh god damn it#asahi-dev · 2026-08-24 10:32
chadmedhttps://elixir.bootlin.com/linux/v7.2/source/kernel/sched/completion.c#L141#asahi-dev · 2026-08-24 10:32
chaos_princesshttps://elixir.bootlin.com/linux/v7.2/source/kernel/sched/completion.c#L141#asahi-dev · 2026-08-24 10:31
tryhard[m]this loop runs only at the very beginning for a short time and afterwards there is only one iteration#asahi-dev · 2026-08-24 10:31
jannausven: the crazy tbt setup broke (different selection of DUTs though). only 2 host on DFPs are detected without apparent cause (without debug logs). 3rd host is running macos#asahi-dev · 2026-08-24 10:31
tryhard[m]chaos_princess: wdym with use a completion for the bridge ready check? i found something like thread::yield_now() we can maybe use but that's std#asahi-dev · 2026-08-24 10:29
chadmedah right okay#asahi-dev · 2026-08-24 10:28
jannauchadmed: I did not. I merged it because v7.2 did rename drm_atomic_state to drm_atomic_commit and I wanted to avoid causing conflicts. it's impossible for anyone to use that "accidentally"#asahi-dev · 2026-08-24 10:24
chadmedi may have time to revisit them shortly#asahi-dev · 2026-08-24 10:23
chadmedi guess it doesnt really matter since no one will be using it though#asahi-dev · 2026-08-24 10:23
chadmedjannau: did you end up looking at avservice and the surface clear iova faults? i wanted to avoid merging 14.7 support until those were fixed#asahi-dev · 2026-08-24 08:54
chaos_princesswrote some comments#asahi-dev · 2026-08-24 08:41
tryhard[m]driver: https://github.com/AsahiLinux/linux/pull/576#asahi-dev · 2026-08-24 08:41
tryhard[m]pmpv1 tracer: https://github.com/AsahiLinux/m1n1/pull/670#asahi-dev · 2026-08-24 08:28
jannauI seem to have compared the wrong kernel builds. remap helps in all cases but the effect seems to be negligible when using DMA-FQ and dropping the tlb flush after mapping#asahi-dev · 2026-08-24 08:21
svenoh, nice#asahi-dev · 2026-08-24 08:17
jannauso it has to be smarter than re-routing it to all remapped streams#asahi-dev · 2026-08-24 08:16
jannauremap helps if I keep the flush tlb in apple_dart_iotlb_sync_map. 1.86 Gbits/sec -> 10.7 Gbits/sec#asahi-dev · 2026-08-24 08:07
svenpity#asahi-dev · 2026-08-24 08:07
svenso remap internally then probably just replays any invalidation across all SID separately#asahi-dev · 2026-08-24 08:06
jannaumost if not all of the improvement comes from DMA-FQ / map_pages non-invalidation#asahi-dev · 2026-08-24 07:58
svenah right, i wanted to check where xnu invalidates because likely we don't need that invalidate after map_pages ever#asahi-dev · 2026-08-24 07:57
jannauacceptable '14.7 / 22.6 Gbits/sec' now but I did multiple changes: DMA-FQ and no invalidate after map_pages. I'll test the impact of remap now#asahi-dev · 2026-08-24 07:53
svenlet's hope it's actually faster then :D#asahi-dev · 2026-08-24 07:53
svennice!#asahi-dev · 2026-08-24 07:52
jannaujust done that. apple_dart_hw_enable_translation() only sets TTBR for the first sid and all other enabled sids just remap to the first. invalidate_tlb then breaks after invalidating the first sid#asahi-dev · 2026-08-24 07:42
svenmight be worth a try, i still think we could add that remap to just apple-dart.c without requiring any device tree changes#asahi-dev · 2026-08-24 07:21
jannaunot yet on m3*#asahi-dev · 2026-08-24 07:05
svenregarding those dart streams, did you check if using remap makes it any better?#asahi-dev · 2026-08-24 06:26
svenack#asahi-dev · 2026-08-23 22:41
jannausven: please pick https://github.com/AsahiLinux/linux/commit/71d72e6f41342d0afb45a530e333c8653ccf0d7e#asahi-dev · 2026-08-23 22:41
jannausmall t6020 acio dts fix added. tested that as well and the linux receive rate there is 6 time faster#asahi-dev · 2026-08-23 22:03
jannaupushed bits/* branches and asahi-wip-7.2. gpu driver is disabled. lightly tested (just on M3)#asahi-dev · 2026-08-23 21:59
jannaut8122/m3 thunderbolt works, but invalidating all those DARTs streams might be bad for performance. iperf3 results: 1.1 / 16.3 Gbits/sec#asahi-dev · 2026-08-23 19:53
jannauack#asahi-dev · 2026-08-23 19:53
chaos_princessthere is more, just ignore that version#asahi-dev · 2026-08-23 19:52
jannauI wasn't reviewing just applying ontop of sven tbt changes and noticed the single issue in the rejects#asahi-dev · 2026-08-23 19:50
chaos_princesswill send new version soon#asahi-dev · 2026-08-23 19:50
chaos_princessjannau: don't bother reviewing the current ace3 submit, i royally screwed up the rebase on that, and already noticed it#asahi-dev · 2026-08-23 18:21
svenSame old rule for anyone who asks or complains that PCIe or DP doesn’t work yet once it’s merged: I’ll make sure to double the eta for that again!#asahi-dev · 2026-08-23 18:18
jannauwill be a little annoying due to the m3 tipd/atc changes