#asahi-gpu · 2023-07-23 14:43
alyssaMesa doesn't currently use it so IDK if that kernel path was ever exercised#asahi-gpu · 2023-07-23 14:42
alyssalina: I see the uapi lets you submit multiple commands in one ioctl .. is this tested?#asahi-gpu · 2023-07-23 10:48
rosefromthedeadnot sure how conscious i was when i wrote it#asahi-gpu · 2023-07-23 10:48
rosefromthedeadit's also possible that my code is just mixing up values from different batches or something#asahi-gpu · 2023-07-23 10:47
rosefromthedeadonly on startup of tinywl for some reason, -3150us, -94us, and everything positive from there#asahi-gpu · 2023-07-23 10:46
marcanhow negative are you getting?#asahi-gpu · 2023-07-23 10:44
marcanrosefromthedead: I wonder if there's just a slight offset with the GPU timer for some reason. I don't really know what the internal architecture is like. I know there's one master AIC timer which I think everything derives from, but not how all the downstream timers are synced from that#asahi-gpu · 2023-07-23 10:42
marcanno ECV then, so no second offset#asahi-gpu · 2023-07-23 10:42
rosefromthedeadm1 mini#asahi-gpu · 2023-07-23 10:41
marcanrosefromthedead: what machine?#asahi-gpu · 2023-07-23 10:38
sventriggering that requires both a very fast role switch and a very fast cpu that likes to speculate a lot#asahi-gpu · 2023-07-23 10:37
sventhere was this one subtle role switch bug where they read the new role outside of the mutex or something#asahi-gpu · 2023-07-23 10:37
rosefromthedeadi'll check for an offset#asahi-gpu · 2023-07-23 10:37
rosefromthedeadlol#asahi-gpu · 2023-07-23 10:37
rosefromthedeadhmm, still getting some negative durations, but i think less now?#asahi-gpu · 2023-07-23 10:37
svensame thing inside the dwc3 driver#asahi-gpu · 2023-07-23 10:36
marcanno other ARM64 CPU managed to hit it apparently#asahi-gpu · 2023-07-23 10:36
marcanfound a years-old Linux core workqueue code bug#asahi-gpu · 2023-07-23 10:36
rosefromthedeadblimey#asahi-gpu · 2023-07-23 10:36
marcanI've seen these damn things speculate through an IRQ enough to break some code missing a barrier, it's nuts#asahi-gpu · 2023-07-23 10:35
marcanyeah, otherwise it could in principle speculate through the actual work submission to the GPU and all that#asahi-gpu · 2023-07-23 10:33
rosefromthedeadwill try with an extra isb, i already have one before the read but ig it would make sense to need ones either side?#asahi-gpu · 2023-07-23 10:33
rosefromthedeadmarcan: not on purpose, just a normal asahi install#asahi-gpu · 2023-07-23 02:47
marcanalso I assume you're not running this on the hypervisor, right? that *will* screw with timekeeping on purpose by default#asahi-gpu · 2023-07-23 02:46
marcanalso there could be an offset on the virtual counter for some reason#asahi-gpu · 2023-07-23 02:46
marcanrosefromthedead: you probably need an isb, those counters aren't synchronized#asahi-gpu · 2023-07-22 20:54
alyssa\shrug/#asahi-gpu · 2023-07-22 20:53
alyssarosefromthedead: the important thing is that it works on !ARM64 for FEX reasons#asahi-gpu · 2023-07-22 20:16
rosefromthedeadi wonder whether we could (/should) #ifdef the timestamp ioctl behind !ARM64 and use inline asm otherwise#asahi-gpu · 2023-07-22 20:15
rosefromthedeadgonna write that off as "wow, these machines really *are* quick!" for now#asahi-gpu · 2023-07-22 20:13
rosefromthedeadbut only slightly#asahi-gpu · 2023-07-22 20:13
rosefromthedeadhrm, i'm reading timestamps from CNTVCT_EL0 before submitting jobs and the finish timestamp on some of the jobs is ending up earlier than when they were submitted lol#asahi-gpu · 2023-07-22 18:22
linaI mean it could be mesa too, I'm just not sure it's a kernel issue at this point ^^#asahi-gpu · 2023-07-22 18:01
dottedmagIsn't it fun that there are so many userspace bugs uncovered by M1 that _any_ bug might actually be in the application code?#asahi-gpu · 2023-07-22 17:42
linaI think that might be a userspace bug then...#asahi-gpu · 2023-07-22 17:36
jannauunfortunately many missing symbols but nothing suspicious. killed cleanly and no further kernel messages#asahi-gpu · 2023-07-22 17:03
linaYou could try gdb and seeing where all the threads are waiting#asahi-gpu · 2023-07-22 17:02
linaAnd it went away after an update...#asahi-gpu · 2023-07-22 17:02
linaI saw a different thing earlier with an older version of firefox where right clicking on a link on the GitHub page would hang it for a while, until something timed out and killed it, but it wasn't the kernel side getting wedged like before I think#asahi-gpu · 2023-07-22 17:02
linaif firefox dies cleanly when you kill it and no kernel threads get stuck I don't think it's *that* problem#asahi-gpu · 2023-07-22 17:01
linajannau: see if the kernel reports any blocked tasks, and same after you kill firefox... but I don't really know at this point ;;#asahi-gpu · 2023-07-22 14:28
jannauis there anything useful I can do while the firefox is still "running"?#asahi-gpu · 2023-07-22 14:27
jannau61 times "Blocking due to fragment queue full", 4 times with vertex in dmesg#asahi-gpu · 2023-07-22 14:26
jannausigh, I think I have for the first time the firefox hang. unfortunately on asahi-6.4-3 / asahi-20230706#asahi-gpu · 2023-07-22 13:18
dottedmaglina: If this turns out to be on a hot path: http://paste.debian.net/plain/1286632#asahi-gpu · 2023-07-22 12:54
linaAnd of course the moment there is any overlap this whole mechanism gets disabled, it's strictly for the clean atlas use case#asahi-gpu · 2023-07-22 12:53
linaYeah, I think I can probably just say we track up to 128 regions and then bail or something like that.#asahi-gpu · 2023-07-22 12:52
dottedmag(maybe even SIMD it?)#asahi-gpu · 2023-07-22 12:52
dottedmaglina: 64 boxes with 32-bit coordinates is 1024 bytes, so 8 cache lines. Peanuts, for linear scanning.#asahi-gpu · 2023-07-22 12:50
linaI think I'm just going to try a region list thing...