#
✕ nick:marcan
#asahi-dev · 2022-10-23 12:36
marcanyeah lol
#asahi-dev · 2022-10-23 12:35
marcan(since mbox messages for AGX are just doorbells)
#asahi-dev · 2022-10-23 12:34
marcanthe shared mailbox code can keep a ring log of the last 16 submissions, and then a special TX function could poll the pending ones (reading the rptr from the hardware) and elide the send if there is already an identical message queued
#asahi-dev · 2022-10-23 12:33
marcanalso for lina, I think she can make the hardware tx queue unable to ever block, since she has 12 pipes I think? and 12 < 16
#asahi-dev · 2022-10-23 12:32
marcanyeah
#asahi-dev · 2022-10-23 12:31
marcanif you come up with something to migrate RTKit to an internal implementation in soc/apple I will merge it and push for it upstream
#asahi-dev · 2022-10-23 12:30
marcankill it with fire please.
#asahi-dev · 2022-10-23 12:30
marcanlovely.
#asahi-dev · 2022-10-23 12:29
marcanthis is a giant mess
#asahi-dev · 2022-10-23 12:29
marcancan we just yeet this thing already?
#asahi-dev · 2022-10-23 12:29
marcanso... that's a global define?
#asahi-dev · 2022-10-23 12:29
marcansven: lina's GPU stuff is synchronous right now, so you can't run into it unless you have MBOX_TX_QUEUE_LEN/2 apps actively rendering simultaneously I guess
#asahi-dev · 2022-10-22 22:59
marcankettenis: ack
#asahi-dev · 2022-10-22 16:58
marcanjannau: try tasksetting the shell to another cluster then checking again
#asahi-dev · 2022-10-22 08:36
marcankettenis: did you get a chance to find out how BSD would deal with the DT change?
#asahi-dev · 2022-10-22 08:35
marcanany oddities with cpufreq? if not, I want to send that out too
#asahi-dev · 2022-10-22 08:35
marcanlet's get that merged next week then
#asahi-dev · 2022-10-22 08:28
marcanjannau: that works with the m1n1 merge right?
#asahi-dev · 2022-10-20 16:54
marcanso the page table can first be built on an unreferenced, in-memory L1 and then swapped over to the real preallocated L1
#asahi-dev · 2022-10-20 16:53
marcanI think we'd need a new op that atomically transfers over the L1 or so
#asahi-dev · 2022-10-20 16:22
marcanit's *probably* safe given the TLBs but still...
#asahi-dev · 2022-10-20 16:21
marcanoof... okay, I can buy that interim solution but it really needs to be fixed at some point
#asahi-dev · 2022-10-20 16:12
marcanhow is that handled right now?
#asahi-dev · 2022-10-20 16:12
marcanYeah...
#asahi-dev · 2022-10-20 16:04
marcanneed to make sure it's solid on all the machines first (at least t8103/t600x ones)
#asahi-dev · 2022-10-20 16:04
marcanlet's get DCP merged soon, though I won't turn it on in the release Asahi kernel for a while (more reason for that linux-asahi-dev thing)
#asahi-dev · 2022-10-20 16:04
marcanokay, okay :p
#asahi-dev · 2022-10-20 13:33
marcanI don't know if regular scanout with full DCP will do fun PSR type tricks, but it will definitely stop scanout and shut down the display when told to turn off the output
#asahi-dev · 2022-10-20 13:33
marcannot the ASC, the display scanout
#asahi-dev · 2022-10-20 13:33
marcanyou mean the FB?
#asahi-dev · 2022-10-20 13:32
marcanj`ey: DCP ASC polling what?
#asahi-dev · 2022-10-20 13:32
marcanand they've had a bunch of bugs around this
#asahi-dev · 2022-10-20 13:32
marcanmacos... is too big and complex and has too much magic
#asahi-dev · 2022-10-20 13:32
marcanat least with linux s2idle you can know no random software is going to trigger wakeups (mostly because we don't support it :p)
#asahi-dev · 2022-10-20 13:32
marcanfor random reasons
#asahi-dev · 2022-10-20 13:32
marcanI've seen it last a week and die in a day
#asahi-dev · 2022-10-20 13:31
marcanmacos drain is very low when closed except when it isn't
#asahi-dev · 2022-10-20 13:30
marcanand CPU deep idle just because we're pretty sure that static power draw isn't negligible
#asahi-dev · 2022-10-20 13:29
marcanDCP because with a dumb framebuffer they *have* to poll memory so they can't do any PSR/power saving tricks (and it's also uncompressed, so memory bandwidth), so it *has* to be eating significant power relative to everything else idle, plus the panel itself being powered and refreshing
#asahi-dev · 2022-10-20 13:28
marcanI think the big battery saving wins are going to be DCP and CPU deep idle
#asahi-dev · 2022-10-20 13:28
marcan(to rtkit-helper really)
#asahi-dev · 2022-10-20 13:28
marcananyway, adding suspend to MTP is trivial
#asahi-dev · 2022-10-20 13:27
marcanwhat _jannau_ reports is more worrying, and makes me more inclined to believe the MTP theory since that's M2-exclusive
#asahi-dev · 2022-10-20 13:26
marcan< j`ey> that doesnt sound great(?) <- half battery in 7 hours matches the ~14h idle runtime so it sounds completely expected (no major savings)
#asahi-dev · 2022-10-20 13:22
marcanand then maybe MTP does something stupid like busyloop waiting for space
#asahi-dev · 2022-10-20 13:22
marcanbut I wonder if specifically touchpad events might quickly consume all FIFO space
#asahi-dev · 2022-10-20 13:21
marcanwhich we already know
#asahi-dev · 2022-10-20 13:21
marcan_jannau_: wild guess: something with MTP? the rtkit helper stuff doesn't do suspend yet IIRC, but the dockchannel IRQs would be masked by default in s2idle. that would result in the FIFO filling up with HID events
#asahi-dev · 2022-10-20 13:18
marcanhm? overnight would've probably been macOS, but the jam session thing was asahi
#asahi-dev · 2022-10-20 13:17
marcanmacsmc-hwmon is not a thing in official kernels yet

plus ›