00:04 airlied: Company: I'd probably go with lavapipe, but the api overheads is likely dwarfed by sw rendering backends anyways
00:11 alyssa: anyone object to adding a nir_if::cold flag?
00:11 alyssa: to model `if (unlikely(...)) {}`
00:12 alyssa: my current use case is tweaking the nir_opt_preamble cost model for my cold emulation paths
00:12 alyssa: but one can imagine more use cases
00:12 alyssa: or perhaps it could be a nir_selection_control
00:26 Company: airlied: I was mostly thinking about feature completeness - because first people advertised against lavapipe as it was still incomplete, then llvmpipe was missing damage regions, so no idea what the current state is
00:26 airlied: Company: no idea, guess if one works better pick it for now
00:27 Company: seems llvmpipe doesn't advertise dmabufs but lavapipe does
00:27 Company: so I guess lavapipe wins
01:00 permudaviews: actually this yesterdays idea only complicated things, as it's cryptoidea of indexes, better demo realization is to implement compressed ir + intrinsic for a compilation which designs new encoder to arbitrary precision (the other way around compared to yesterdays). and for linking and runtime there is also - intrinsic. so the content held inside the hash would normally go to multiple
01:00 permudaviews: computer words without compression, however decoding at that level is never done as it makes zero sense without elimination values which shrink things back unless theres bug to decode groups of magic values. But those magic values are complex to produce as 1024*1024*3 roughly is decimal range digit values for one alu bank. I hate to be a killer, but my tech is too strong for bunch of
01:00 permudaviews: jewish abusers active on this channel, to stalk and terror arian people everwyhere again and again, and claim that gas chamber isn't their place to be at. It is a decent primarly jewish and other semites bullshit show here on irc, that no normal real person can keep up with. So i part for good from such environment.
01:03 DemiMarie: Company: at least Qt and Iced have their own dedicated software renderers
02:02 cmarcelo: alyssa: maybe NIR selection control if that becomes a bitmask?
02:02 alyssa: cmarcelo: yeah, maybe?
02:04 cmarcelo: alyssa: semi-related (OpExpectKHR) ~> http://htmlpreview.github.io/?https://github.com/KhronosGroup/SPIRV-Registry/blob/main/extensions/KHR/SPV_KHR_expect_assume.html if we have a flag in nir selection control, we can identify identify expect(false) etc.
02:05 cmarcelo: I mean, expect(my condition, false)
02:05 alyssa: yeah
02:05 cmarcelo: which is what unlikely() in highlevel languages would produce
02:07 cmarcelo: right now we just ignore/pass-through. but if we encoded the expect somehow, have a nir "expect" thing that wraps a value and a constant, we could have a pass to "use it or drop it".
02:08 alyssa: sounds good yeah
02:08 cmarcelo: another place we would like to use this is range analysis. however, we don't know how common is to generators produce this.
02:08 alyssa: even in backend it can be useful
02:08 cmarcelo: back to original: I can't dig now why selection control is not a flag set instead of an enum, but seems a reasonable move.
02:08 alyssa: on most ISAs, it should be possible to sink a cold block to get it out of the icache
02:09 cmarcelo:*nods*
02:09 alyssa: (I don't know of any prior art for this on AGX but literally when has that ever stopped me lmao)
02:11 Company: DemiMarie: I hope it's worth the effort when they could just use llvmpipe instead
02:11 DemiMarie: Company: because llvmpipe is much slower, so it is worth it for them
02:12 DemiMarie: Qt in particular is used on embedded systems that might well have a tiny CPU and no GPU at all
02:15 Company: that's a limited featureset I guess
02:18 DemiMarie: Also there are surprisingly many cases (not just VMs) where GPU drivers are broken
02:21 Company: the fix in those cases is to fix the drivers, not add workarounds everywhere
02:21 DemiMarie: I will let the developers of those projects make the decision.
02:59 marc2377: It's so cool to see so many names here I'm familiar with from years of browsing linux kernel sources, patches, packages and such
07:06 tzimmermann: jfalempe, looks like the vblank support for matrox is not so reliable
07:14 mlankhorst: https://patchwork.freedesktop.org/project/intel-xe/series/ 504s?
07:21 tzimmermann: bbrezillon, AFAICT you get your patches merged into drm-misc-fixes, right?
07:27 bbrezillon: tzimmermann: yep, mlankhorst did the backmerge
07:44 mripard: jfalempe: https://lore.kernel.org/all/20241001210403.43535-3-ziyao@disroot.org/
09:40 Lynne: its nice to see radv generating exactly the same shader binary for different glslang optimizations
09:49 forwardleap: Your terrorists do not qualify to the top million programmers in the world. Yet you have enough arrogancy to commit tyranny. Go grab a burger DemiMarie try not to practice your spew publicly, there is many burgers waiting for you jewish pigonoid. https://people.engr.tamu.edu/djimenez/ut/utsa/cs3343/lecture25.html network is full of those implementations, too many to get a quick
09:49 forwardleap: enthusiasting kickoff, hashed alus are generated with such code modified a little. You have had every lecture in united states education system, it's likely you did understood nothing about them, so go on and fart and code the world to dissapearance. I do not even care, i fire bullets at you if you come again my way, and this time in cambodia i take a knife and scalpel and stitch strings
09:49 forwardleap: with me, if you humiliate and stalk and terror me there again, i and we chop your dicks off, and put it to your own mouth. Mark my words. I might need to go again, cause indrek bullshit man wasted all the money borrowed, so that hotel belongs to my dad and me now. But the parasites like finnish abuse whores have to be kicked out from that country before as i do not negoatiate with those
09:49 forwardleap: on terms.
10:10 zamundaaa[m]: <DemiMarie> "Company: because llvmpipe is..." <- Actually if you have multiple cores and only simple shaders, llvmpipe is a lot faster than QPainter, because QPainter is single threaded
10:12 zamundaaa[m]: It's one of the reasons why we try to use llvmpipe in KWin nowadays, rather than going for QPainter rendering when no hardware accelerated OpenGL contexts are available (or usable)
10:20 Company: zamundaaa[m]: I would expect kwin to benefit most from llvmpipe for its usecase, which is mainly large blits
10:20 Company: zamundaaa[m]: though I'd also expect that you can easily multithread that in QPainter
10:22 zamundaaa[m]: Company: it should indeed not be that difficult, but it's not effort anyone has been willing to put in
10:22 Company: what I did in GTK is optimize our fallback format conversion blit and mipmap functions to spawn threads and then have each thread convert a row and that was a huge speedup, even in the case where it's basically just memcpy
10:23 zamundaaa[m]: Largely because of reason number two for choosing llvmpipe: it supports all the OpenGL effects, like blur
10:23 Company: I did that mainly to enable Loupe to load large images (30k x 50k)
10:24 Company: it's downscaled to 600x1k or so because it's scaled to fit
10:24 Company: so I can use mipmap level 5, and if I only upload that, it's a lot faster
10:25 Company: but then I need to compute mipmap level 5 on the CPU by averaging 32x32 blocks of pixels
10:25 Company: and that was a massive speedup, I got it to go as fast as my RAM can transmit the data
10:27 Company: which is something like 50GB/s - but a single thread could only do like 5GB/s or so
10:27 Company: at which point I also used it for memcpys
10:27 Company: to copy stuff from shm buffers into Vulkan/GL buffers for example
10:29 Company: actually, that is something compositors could benefit from - though I guess you use glTexImage() for wl_shm buffers and then it's Mesa's job?
10:31 MrCooper: compositors could offload that to a separate thread though, glTexImage can't read data after it returns
10:32 MrCooper: zamundaaa[m]: coincidentally, mutter/gnome-shell disables effects if it detects software rendering, because they tend to be slow even with llvmpipe
10:33 jadahl: animations are disabled, at least
10:33 zamundaaa[m]: We do disable some by default as well (namely blur), but it sucks if you need to take screenshots
10:33 Company: the problem with llvmpipe is that it can be really slow and it can be really fast
10:34 Company: my benchmark runs with 200fps on my desktop and with <10fps on my rpi
10:34 zamundaaa[m]: As people love to use VMs to review new releases, being able to have pretty screenshots is pretty important to us
10:34 Company: because my desktop has 16 cores and llvmpipe benefits from that a lot
10:35 Company: and people tend to not give VMs all of their cores, so you can quickly run into perf issues I guess
10:35 jadahl: zamundaaa[m]: in gnome we have runtime knobs people can poke at to get back things we turn off when on llvmpipe
10:36 jadahl: to get a more "real" behaviour
10:36 Company: might be worth having some benchmark on first start to see how fast llvmpipe goes and toggle based on that
10:37 zamundaaa[m]: Yeah you can just go into effects settings and turn blur back on in Plasma as well
10:37 Company: or to throw up some "your computer seems slows, do you want to disable stuff?" warning
10:37 jadahl: zamundaaa[m]: I guess that is good enough, if you know about it :)
10:38 Company: so people don't think Gnome/KDE suck when their VM is misconfigured
10:38 Company: (shoutout to the flatpak people for shipping asahi without drivers for ages)
10:38 zamundaaa[m]: jadahl: That "if you know about it" is the annoying part, reviewers generally don't know or don't care enough, so you still see screenshots without blur
10:40 Company: I would really throw up a notification on first start with llvmpipe these days
10:40 zamundaaa[m]: Hmm, I wonder if we should have a notification about that when Plasma is used in a VM. Tell people to turn on 3D acceleration for example...
10:40 Company: GTK would like that, too
10:40 Company: this can also happen after distro updates
10:40 Company: when something screwed up the drivers
10:42 Company: MrCooper: while you're here, I remember 3 things I wanted your opinion on
10:42 Company: MrCooper: 1. Should GTK pick lavapipe or llvmpipe? Any preference?
10:44 Company: MrCooper: 2. the recent GTK fix that speeds up llvmpipe - I didn't want to backport that to 4.16 - is that okay with you or would you want that speedup?
10:44 MrCooper: 1. not in particular, happy to defer to airlied there
10:45 Company: MrCooper: 3. I noticed that lavapipe advertises dmabufs but llvmpipe does not - is that something that llvmpipe should do?
10:45 MrCooper: 2. doesn't seem critical, worst case the old GL renderer can be forced?
10:45 MrCooper: 3. what does "llvmpipe does not" mean exactly?
10:46 MrCooper: i.e. what's missing?
10:46 Company: we eglQueryDmabufFormats() or whatever that's called and llvmpipe doesn't seem to return any
10:47 Company: GTK has a debug var that prints all of the formats, and I noticed it doesn't print any
10:47 Company: while it did print them for Vulkan with lavapipe
10:48 MrCooper: offhand does seem like llvmpipe should return something there if it supports dma-buf
10:50 Company: GDK_DEBUG=dmabuf LIBGL_ALWAYS_SOFTWARE=1 gtk4-widget-factory should print them when playing a video but it doesn't seem to
10:50 Company: GDK_DEBUG=dmabuf LIBGL_ALWAYS_SOFTWARE=1 MESA_VK_DEVICE_SELECT=10005:0 gtk4-widget-factory for the full joy of software rendering
10:52 Company: oh
10:52 Company: GTK's fault
10:53 Company: the LINEAR special case doesn't print them
10:53 Company:fixes
10:53 MrCooper: *phew* :)
10:58 Company: okay, so: GTK in 4.18 is gonna pick lavapipe now, because everyone thinks Vulkan is cooler than GL
10:58 Company: 4.16 picks llvmpipe
11:04 MrCooper: sounds good
11:05 MrCooper: anyway, even rubber ducks need to eat sometimes, bbl
11:07 Company: MrCooper: oh, as for (2), yes, you can force the GL renderer and unless you do fractional scaling that's gonna work fine
11:49 alyssa: j;j
12:15 bbrezillon: zmike: when you get a chance, could have a look at the zink fix in https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31439?
13:05 zmike: bbrezillon: seems like maybe the frontend should do this?
13:06 bbrezillon: zmike: why?
13:06 bbrezillon: and how?
13:07 zmike: let me wake up a little more
13:07 zmike: again
13:07 bbrezillon: ah, frontend is the state tracker, not anv
13:07 bbrezillon: sorry
13:07 bbrezillon: well, yes, if the frontend can do it for you, why not
13:13 tzimmermann: jfalempe, i updated the series for the drm client module
14:09 MrCooper: can a GL fragment shader somehow determine how many bits per component the draw buffer has?
14:11 soreau: yea, just send it a uniform ;)
14:11 MrCooper: right, I mean without
14:14 glehmann: no
14:18 soreau: what if it's 10-10-10-2? the shader still uses traditional vec4's (with 0-1 range)?
14:20 MrCooper: yep, fragment shader always operates on one float per component
14:21 MrCooper: another question is: are those float values rounded to the nearest value in the draw buffer, or to the floor / ceiling?
14:33 tzimmermann: sima, mripard, mlankhurst, can we please revert drm-misc-fixes to f9e7ac6e2e99 ("drm/panthor: Don't add write fences to the shared BOs"). someone dump that entire patchset of 31 patches into -misc-fixes. https://patchwork.freedesktop.org/series/139038/ this is NOT acceptable
14:33 sima: ack
14:34 sima: and yeah most of that isn't for -fixes indeed
14:34 tzimmermann: sima, do i have credentials to do that?
14:34 sima: tzimmermann, I guess you'll also do a reply why this isn't ok and all that?
14:34 sima: I think so
14:34 sima: dim push -f
14:35 tzimmermann: i'll send a note
14:35 tzimmermann: thanks
14:35 sima: if the dim push -f doesn't work you might need to enable force pushes in gitlab temporarily, but as maintainer you can do that
14:36 sima: tzimmermann, https://gitlab.freedesktop.org/drm/misc/kernel/-/settings/repository -> Protected Branches
14:36 tzimmermann: sima, i never tried. i'll let you know how it went
14:36 sima: need to enable force push there temporarily
14:36 glehmann: MrCooper: not 100% clearly defined, round to nearest even is recommended
14:37 sima: tzimmermann, and maybe disable drm-misc-fixes for developers for now, until there's another patch
14:37 glehmann: but for example, with rgba8, amd will first rtz convert to fp16, and then round to nearest even
14:37 sima: otherwise they'll just go ahead and push again
14:37 MrCooper: glehmann: wow
14:37 sima: unless you need to rebase anyway, but then I'd still disable developer push while you do all this
14:41 tzimmermann: sima, worked nicely. thanks a lot
14:42 sima: tzimmermann, did you push a rebase? otherwise there's good chances someone will repush the patches again, but if it's a rebase then the force-push protection stops that
14:43 tzimmermann: sima, not sure. i did a reset to the last good commit and force-pushed that. where should i have rebased?
14:44 sima: if it's not a rebase I'd block developer push rights until someone complains (or until you've rolled to -rc2 next monday)
14:44 sima: tzimmermann, only if someone else pushed more patches on top of the ivpu patches that you wanted to keep
14:44 sima: it if all was ivpu then no option for a rebase and the reset was the right hting
14:45 tzimmermann: i'll block dev access for now to prevent this. but the ivpu series has only been there for an hour or so. hopefully no one picked it up yet
14:46 tzimmermann: mripard, mlankhorst ^ FYI
14:49 sima: yeah, but better safe than sorry and with gitlab we can now easily prevent people pushing stuff when they shouldn't
14:49 mripard: tzimmermann: thanks for the notice and dealing with this
14:49 sima: I guess if you have a fix you can push that and then unblock
14:50 mripard: it's not the first time the ivpu fixes have been a bit suspicious, but it's never been that bad either :)
14:50 tzimmermann: mripard, please unblock if someone need to push to -misc-fixes
14:50 mripard: ack
14:51 sima: I guess we could also do an mr now, I land it into drm-fixes and then you fast forward and unblock
14:51 sima: but feels a bit like overkill
14:52 sima: I'm still around for a bit more than 1h before I head out for dinner at my brother's place, if you do want to do it
14:53 sima: tzimmermann, ^^
14:53 tzimmermann: sima, i'll leave it for now
14:53 sima: ack
14:54 sima: I'll fast-forward drm-fixes just in case
15:04 MrCooper: glehmann: Vulkan 3.10.2 says "Implementations should round to nearest" tough
15:04 glehmann: should is not must
15:05 MrCooper: right, as nearest isn't nearest even :)
15:08 glehmann: I think d3d11 has some weird 0.6ULP requirement, so it's not just Vulkan which is vague
15:12 MrCooper: nice stack of turtles there
15:14 MrCooper: can of worms might be more like it
15:24 DemiMarie: MrCooper: consider yourself lucky you aren't on the browser WebGPU teams.
15:25 MrCooper: heh
15:26 DemiMarie: There is a shader fuzzer that found a bunch of vulnerabilities in various compilers.
15:27 oftcseesnot: nkyeakd please do not hang yourself , the testestosterone you got from is not so bad fucker , you can make it, me i am in troubles homeless bum quasimodo gloria brother can kill me , and lauras crankgangsters! So it's called a position set, if you fix a position of the power you only have 1024 states actually 32 but since the other operand has 32 round robin states is 1024, it's that you
15:27 oftcseesnot: can calculate that yourself. And sure even if i am wrong i definitely will kill myself, cause you are quasimodo tranny fuckers. https://www.go4expert.com/articles/c-program-calculate-combinations-n-digit-t25380/ you are deluded thinking i kill myself cause of shit like this you pull.
15:29 K900: Man this guy is persistent
15:29 kisak: oftcseesnot: that language is not welcome here. Please go away.
15:29 K900: I mean I know it's a mental health thing but man
15:29 K900: Please get help
15:55 mlankhorst: sima: I can push some fixes to drm-misc-fixes now, don't see any ivpu there?
15:56 sima: mlankhorst, yeah
15:56 mlankhorst: error: dst ref refs/heads/drm-misc-fixes receives from more than one src
15:56 sima: once you've done please enable developer push access again, since with the new patches developers wont be able to accidentally repush the ivpu patches thanks to the force-push prevention
15:57 DavidHeidelberg: K900: it's a bot, it spam multiple channels like this
15:57 sima: hm that sounds like a client side set issue
15:57 K900: DavidHeidelberg: I don't think it's a bot, it's a little too specific
15:58 mlankhorst: I've pushed manually
15:59 sima: mlankhorst, pls update gitlab settings now
15:59 mattst88: DavidHeidelberg, K900: incredibly, this is not a bot. this is who we refer to as "joss"
15:59 sima: K900, it's not a bot, but persistent issue we have to deal with unfortunately
15:59 sima: mattst88, did you also ping oftc admins?
15:59 K900: sima: I'm aware, yeah
15:59 sima:forgot the right nick to ping again
16:00 sima: K900, oops meant DavidHeidelberg
16:02 DavidHeidelberg: interesting, I seen spam looking bit like this, even tailored to the discussion of channel in other Matrix/IRC rooms too, but maybe this one is powered by "Human"
16:08 mlankhorst: I think I fixed it
16:17 Ermine: sima: I guess you need to ping dwfreed
17:04 MrCooper: DavidHeidelberg: yep, he's been doing it for many years
17:17 DavidHeidelberg: oh, well, kinda unusual hobby, but.. what can we do :D
17:17 vsyrjala: apparently i need to cc 60 people. anyone know what the max is until something start to get angry?
17:47 DragoonAethis: vsyrjala: 60 should still be safe if it's mostly text, going above 100 people gets risky
18:03 vsyrjala: did some heavy handed trimming and ended up with a bit over 20. should hopefull cover everyone well enough via lists/etc.
18:36 humbleisnogo: I understand that your folks will eventually hang themselves, but for that you have to accuse yourself, cause i never abused you so far. You were the gangsters i did the science , it never changes , it's just scientists are not gimme people despite you bragging how you chopped me under full narcosis, i come back at all of you, you are extremely retarded and with your help they erased my
18:36 humbleisnogo: fraud diagnosis from digital reciept, cause you never understood how crazy scum you are. I was little late on the compression results more on testing it, but it was rather well timed , i managed to get affected by your scam and started with little mistakes on allocation though linear algebra seemed to be correct, until i trimmed it correct entirely, this was a biggest achievement of my
18:36 humbleisnogo: life so far, i investigated that every day. And i have medals from 10 different sports events before already, but to perform so relentlessly under pressure i consider this as my biggest title. execution and encoder and decoder was already finished , but now i can store millions of positioned sets in the same variable finally, as it all went through sigma status. I care nothing anymore as
18:36 humbleisnogo: to how you hang yourself , cause what you did was your fault and very much not allowed in the world.
18:37 bl4ckb0ne: dwfreed ^
18:38 uriah: I also write poetry sometimes
19:09 humbleisnogo: I enjoyed star talk alot via facebook, cornell people the students of carl sagan had great insights as to how some parts of universe work, but i did not agree on one particular part that touched my soul enough, that we are particles the same ones all, which would state after they put anchor on me, knocked off my teeth, butchered me, humiliated me and my so called girlfriend screwed such
19:09 humbleisnogo: biggest scum while leaving me to die with broken nose alone in the state of homeless person, tried to kill me in many ways, i would not put my dick into such apparatus soaking of the cum of biggest trash in the world in millions of years dudes, once she or her cranks or lesbian humiliators show their faces again, a lot of brutal actions will be committed by me personally and i have small
19:09 humbleisnogo: teams as well consisting of one of the strongest people on the planet to treat them. But overall star talk people were smart as hell compared to me.
19:25 humbleisnogo: essentially the last algorithm is not so good at all, so what it does is iterate over all numerals of 32bit and more, store them to arrays in excess, so it already swaps to disk at the middle of the 20bit worth of combination arrays, yes that is not at all intelligent way to doing such things. but if i had 20gb of memory it would succeed prolly fast for 64bit combinations printing,
19:25 humbleisnogo: classical wrong way to think, but such mentally people exist.
19:27 humbleisnogo: but in reality how compilers and performant systems work is 4 instructions are only needed to be exposed + and - and their compressed intrinsics which also base on hw + and -
19:28 humbleisnogo: you just need to get some memory attached some you know unaligned or aligned bytes and though bitwise ops are cheap they do not have any value
19:30 humbleisnogo: bitwise barrel shift circuit has a great demo, it's cool they can tap into hw gates and fix some issues in hw , with godfather algorithm, but any other instructions does not serve well than only + and -.
19:32 humbleisnogo: the way intrinsics work, descends into the subject of science around LOGICS, mutual inclusion and mutual exclusion are the keywords, i am educated so i know those lessons luckily secondary school had it
19:34 humbleisnogo: so as + and - is called arithmetic , intrinsics of compressed + and - are more like arithmetic and logics mixed, where the last logics is the key to success, its even 80 percent logics based solution.
19:46 humbleisnogo: so yes, i am not wrong in my design or views at all, there is nothing i should hang myself for , i say the very neat but a delicate subject is quantum computer worth of performance lives in any asic ever produced, which is something that should settle you down more.
19:46 austriancoder:is looking for someone to review a src/mesa/main/formats.csv related change: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3418/diffs?commit_id=1d194e9316989cae8a00ab4062de84c332b7dc01
19:48 humbleisnogo: i am not saying i am smarter than those people who made the manufacturing of computers possible, i am saying through all my second half of life i specialized on this subject, and when i do that , i get the results normally, and the results are this time around bit shocking but slightly expected in the past so.
20:06 humbleisnogo: now for some reason i do not want to see russians and ukranians duying over this war, i do not know ukranians that well to insult them to be same as russians , though it is not insult, i consider russians one of the most intelligent humans in the whole world, such tragedy must end, now it goes like more that nato wants to accept ukranians and they will smash russians and it's clear to
20:06 humbleisnogo: me, this is the vote against moral people and betrayal at the most intelligent nation or one of them in the whole world. Could not they agree in some other ways but it smells really that like they told jewish country is built to the ukraine after all, so more like money talks there.
20:11 humbleisnogo: and there is another thing to mention, gods or humanly gods they are not at the top of the hierarchy they are between the lines and gather the strongest people from both parts , if you want to fight with the gods there is no parts important , but it starts with the loss cause when we make ditinguished lines between + and - the best people are at gods side, so gods are between the lines
20:11 humbleisnogo: on earth, not that they are almighty ontop of everyone imagined.
20:13 humbleisnogo: in other words, they are so important that they are among the best from many clans or at least two sides.
20:18 humbleisnogo: so hence i have came out alive countless of times cause i always have the best people to defend me, and it has hence failed to stop me that way with evil, but that is not being immortal at all
22:04 HdkR: Does anything use `DRM_IOCTL_SYNCOBJ_EVENTFD` yet?
22:10 JoshuaAshton: Gamescope
22:10 HdkR: Ah, cool
22:10 JoshuaAshton: https://github.com/ValveSoftware/gamescope/blob/master/src/Timeline.cpp#L151
22:11 zamundaaa[m]: HdkR: I think every compositor that supports the syncobj explicit sync protocol uses it
22:15 HdkR: Nice. Just scanning Linux uapi changes and noticed it again