#panfrost on 2020-02-04 — irc logs at freenode.irclog.whitequark.org

2019-09-06 11:20 alyssa changed the topic of #panfrost to: Panfrost - FLOSS Mali Midgard & Bifrost - Logs https://freenode.irclog.whitequark.org/panfrost - <daniels> avoiding X is a huge feature

00:00 alyssa has joined #panfrost

00:22 <alyssa> Figured out my disasm issue, but thinking I need to refactor more disasm code to make this not awful

00:26 stikonas has quit [Remote host closed the connection]

00:43 <urjaman> ... for some reason editing fields in libreoffice calc makes Xorg take 100% CPU

00:44 <urjaman> i suppose i should update panfroste before further investigation but like ... odd.

00:45 <urjaman> nothing in dmesg, just Xorg hogging CPU and libreoffice being laggy-af

00:46 <urjaman> i suppose i'll do my budget on the big computer as usual :P

00:53 <alyssa> :/

00:53 * alyssa tries to fix disasm harder

00:53 <alyssa> Definitely making progress but there are... a *lot* of combinations

01:11 raster has quit [Quit: Gettin' stinky!]

01:12 nerdboy has quit [Ping timeout: 265 seconds]

02:35 nerdboy has joined #panfrost

02:52 buzzmarshall has quit [Read error: Connection reset by peer]

02:53 chewitt has joined #panfrost

03:03 chewitt has quit [Quit: Zzz..]

03:51 nerdboy has quit [Ping timeout: 240 seconds]

04:28 chewitt has joined #panfrost

04:43 <HdkR> https://www.cnx-software.com/2020/02/04/hardrock64-sbc-rockchip-rk3399-processor-launch-in-april/ Wow. Only $35 is quite nice

05:14 nerdboy has joined #panfrost

05:48 vstehle has quit [Ping timeout: 265 seconds]

05:57 megi has quit [Ping timeout: 240 seconds]

06:00 vstehle has joined #panfrost

06:44 _whitelogger has joined #panfrost

06:53 <chewitt> that's dirt cheap

06:53 <chewitt> finally one of the vendors grasps the idea of removing expensive emmc flash to drive the costs down

06:54 <chewitt> bonus credit if it can fit in the Kodi flirc case which acts as a passive heatsink

07:08 guillaume_g has joined #panfrost

07:37 NeuroScr has quit [Quit: NeuroScr]

07:54 <anarsoul> alyssa: do you also have q3a stuck with panfrost in GLSL_InitGPUShaders for quite a while with opengl2 renderer?

07:57 <anarsoul> looks like it spends a lot of time in compiler

08:18 pH5 has joined #panfrost

08:40 icecream95 has joined #panfrost

08:46 yann|work has quit [Ping timeout: 265 seconds]

08:49 <icecream95> bbrezillon: That doesn't seem to fix the oopses when a program exits while there are still GPU jobs

08:51 <icecream95> Running `timeout 5 glmark2-es2 -b refract` in a loop doesn't take long to trigger an oops

09:00 <bbrezillon> icecream95: oh, interesting

09:01 <bbrezillon> so we do have non-cancelled jobs flying around when we close the FD :-(

09:11 icecream95 has quit [Ping timeout: 240 seconds]

09:15 megi has joined #panfrost

09:38 raster has joined #panfrost

09:42 yann|work has joined #panfrost

09:43 raster has quit [Ping timeout: 240 seconds]

09:45 raster has joined #panfrost

10:22 karolherbst has joined #panfrost

10:27 NeuroScr has joined #panfrost

10:32 <daniels> bbrezillon: right, nothing forces clients to wait for job completion before closing the FD - especially if they crash rather than are exited gracefully :P

10:32 <raster> either way ... you exit :)

10:33 <raster> crashing can be a quite fast way to achieve that :)

10:35 <HdkR> FD_CLOEXEC is a good flag for apps that want to die quickly :D

10:37 <bbrezillon> daniels: well, the kernel should at least wait for all jobs to complete (or kill them somehow) before freeing the resources attached to panfrost_file :P

10:38 <bbrezillon> and yes, it's not userspace responsibility, but I thought it was all handled by the drm_sched helpers

10:38 <bbrezillon> but apparently we have a problem

10:39 <bbrezillon> BTW, I couldn't reproduce the issue

10:39 <bbrezillon> so it's hard to tell

10:41 <daniels> bbrezillon: yeah, other drivers don't delay client exit but instead just let the job live longer than the client

10:46 <bbrezillon> daniels: we could add refcounting to panfrost_file, but I'd still like to understand how we can end up with in-flight jobs after calling drm_sched_entity_destroy()

10:48 <bbrezillon> etnaviv and v3d don't seem to refcount their file_priv objs

10:48 <bbrezillon> but the iommu context is refcounted

10:52 raster has quit [Ping timeout: 240 seconds]

10:53 <bbrezillon> so yes, maybe jobs are still being executed after drm_sched_entity_destroy() returns

10:56 MastaG has joined #panfrost

10:57 <MastaG> Goooooood morning

10:57 <MastaG> I wanted to give panfrost a shot on my odroid xu3 (exynos soc with mali t-628 gpu)

10:58 NeuroScr has quit [Quit: NeuroScr]

10:58 <MastaG> So I built mesa with a small modification to pan_screen.c adding the 0x620 case

10:59 <MastaG> furthermore I rebuilt libdrm with the exynos driver enabled

11:00 <MastaG> gnome shell seems to start with some minor corruption (happens when moving the mouse cursor) but it's really snappy

11:00 <HdkR> Sounds about right from previous reports of Midgard 1st gen testing

11:00 <HdkR> `Nice work with panfrost. I did some runs on MALI T620(Odroid XU4) and there are some artefacts. for example pure weston and glmark2. Are there any change that we see some "love" for T620 :)`

11:01 <MastaG> however when attempting to run glxgears it will throw a segfault with: DRM_IOCTL_MODE_CREATE_DUMB failed: permission denied

11:02 <MastaG> @HdkR Great, we're on the same level then I guess, did you also get the segfault?

11:03 <MastaG> afaik it seems to originate from libdrm

11:03 <HdkR> I don't know what that problem is

11:05 <HdkR> Could be something trying to create a dumb buffer from the rendernode, which isn't allowed

11:07 <daniels> bbrezillon: when the drm_file gets torn down then all the CPU VAs associated with it get torn down so you don't need to worry about those, but you do need to keep the GPU job + GPU VAs + referenced BOs alive

11:14 megi has quit [*.net *.split]

11:14 alyssa has quit [*.net *.split]

11:14 Lyude has quit [*.net *.split]

11:14 ezequielg has quit [*.net *.split]

11:14 MoeIcenowy has quit [*.net *.split]

11:14 bshah has quit [*.net *.split]

11:14 warpme_ has quit [*.net *.split]

11:14 cyrozap has quit [*.net *.split]

11:14 lvrp16 has quit [*.net *.split]

11:14 austriancoder has quit [*.net *.split]

11:14 afaerber has quit [*.net *.split]

11:14 marex-cloud has quit [*.net *.split]

11:14 EmilKarlson has quit [*.net *.split]

11:14 flacks_ has quit [*.net *.split]

11:14 mmind00 has quit [*.net *.split]

11:14 andrey-konovalov has quit [*.net *.split]

11:14 phh has quit [*.net *.split]

11:14 mani_s_ has quit [*.net *.split]

11:14 guillaume_g has quit [*.net *.split]

11:14 Yardanico has quit [*.net *.split]

11:14 steev has quit [*.net *.split]

11:14 daniels has quit [*.net *.split]

11:14 narmstrong has quit [*.net *.split]

11:14 rellla has quit [*.net *.split]

11:17 Ke has quit [*.net *.split]

11:17 thefloweringash has quit [*.net *.split]

11:17 lvrp16 has joined #panfrost

11:17 marex-cloud has joined #panfrost

11:17 afaerber has joined #panfrost

11:17 mani_s_ has joined #panfrost

11:17 mmind00 has joined #panfrost

11:17 austriancoder has joined #panfrost

11:17 EmilKarlson has joined #panfrost

11:17 phh has joined #panfrost

11:17 andrey-konovalov has joined #panfrost

11:18 hopetech has quit [*.net *.split]

11:18 Depau has quit [*.net *.split]

11:18 tlwoerner has quit [*.net *.split]

11:18 cowsay has quit [*.net *.split]

11:18 xantoz has quit [*.net *.split]

11:18 robink has quit [*.net *.split]

11:18 rellla has joined #panfrost

11:18 megi has joined #panfrost

11:18 Lyude has joined #panfrost

11:18 alyssa has joined #panfrost

11:18 warpme_ has joined #panfrost

11:18 ezequielg has joined #panfrost

11:18 MoeIcenowy has joined #panfrost

11:18 bshah has joined #panfrost

11:18 cyrozap has joined #panfrost

11:19 Yardanico has joined #panfrost

11:19 daniels has joined #panfrost

11:19 guillaume_g has joined #panfrost

11:19 narmstrong has joined #panfrost

11:19 steev has joined #panfrost

11:19 steev has quit [Max SendQ exceeded]

11:20 marex-cloud has quit [Ping timeout: 242 seconds]

11:20 lvrp16 has quit [Ping timeout: 242 seconds]

11:20 austriancoder has quit [Ping timeout: 242 seconds]

11:20 ric96 has quit [Ping timeout: 256 seconds]

11:21 daniels has quit [Ping timeout: 244 seconds]

11:21 cowsay has joined #panfrost

11:22 Depau has joined #panfrost

11:22 steev has joined #panfrost

11:22 daniels has joined #panfrost

11:23 hopetech has joined #panfrost

11:23 lvrp16 has joined #panfrost

11:23 ric96 has joined #panfrost

11:23 austriancoder has joined #panfrost

11:24 xantoz has joined #panfrost

11:55 chewitt has quit [Read error: Connection reset by peer]

11:56 chewitt has joined #panfrost

11:59 raster has joined #panfrost

12:25 buzzmarshall has joined #panfrost

12:35 raster has quit [Ping timeout: 268 seconds]

12:41 raster has joined #panfrost

12:45 warpme_ has quit [Quit: Connection closed for inactivity]

12:50 <bbrezillon> daniels: something like http://code.bulix.org/velvdy-1120523 ?

12:51 <bbrezillon> I really thought GPU jobs were cancelled (as in HW-cancelled) in that case

12:57 raster has quit [Quit: Gettin' stinky!]

13:00 alpernebbi has joined #panfrost

13:01 <tomeu> bbrezillon: guess we could actually tell the hw to cancel those jobs, but I'm worried about cases in whcih the GPU is hung

13:02 <tomeu> or maybe most correct would be to do both things (cancel all jobs asynchronously, and make sure the MMU doesn't clean up before time)

13:03 <bbrezillon> anyway, I can't reproduce it locally, so I'll wait for icecream95 to test this patch before posting it

13:08 <alyssa> anarsoul|c: dunno, I haven't played with q3a myself

13:09 <alyssa> MastaG: "gnome shell seems to start with some minor corruption (happens when moving the mouse cursor) but it's really snappy" huh, nice :)

13:09 <alyssa> MastaG: "it will throw a segfault with: DRM_IOCTL_MODE_CREATE_DUMB failed: permission denied" this seems like it might be unrelated to T620 issues..?

13:10 <alyssa> tomeu: ^

13:13 <alyssa> HdkR: https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3689

13:20 alpernebbi has quit [Ping timeout: 252 seconds]

13:27 <MastaG> Hi Alyssa.. I've been googling on this ioctl a bit.. but I couldn't make up what's causing it..

13:29 <MastaG> perhaps I could try to set DRM_RENDER_ALLOW on the ioctl in the kernel

13:36 <bbrezillon> tomeu: here it is => http://code.bulix.org/uwv79t-1120552

13:36 <bbrezillon> and "Memory manager not clean during takedown." makes me think we're doing something wrong

13:38 <bbrezillon> well, my patch should address that too, since the drm_mm obj is teared down when the MMU context has been released

14:19 tlwoerner has joined #panfrost

14:20 raster has joined #panfrost

14:45 <alyssa> no bug like missing `break`

14:46 warpme_ has joined #panfrost

15:04 raster has quit [Quit: Gettin' stinky!]

15:11 nerdboy has quit [Ping timeout: 265 seconds]

15:24 <tomeu> bbrezillon: robher: I'm hitting this warn: https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/panfrost/panfrost_mmu.c#L600

15:24 <tomeu> any ideas?

15:24 <tomeu> https://gitlab.freedesktop.org/tomeu/mesa/-/jobs/1540711

15:25 <tomeu> guess it could be related to what bbrezillon is looking at, but wonder if we aren't masking error information by having a warn and then trying to map

15:32 <robmur01> tomeu: I guess it might at least help for that WARN to include what the unexpected exception type was

15:33 <robmur01> and 'fixing' it by trying to map something that's clearly already mapped doesn't seem too helpful either

15:45 <robmur01> That said, x4 in the splat looks like the fault status for an execute permission fault, which would seem to fit

15:46 raster has joined #panfrost

16:40 yann|work has quit [Ping timeout: 268 seconds]

16:56 indy has quit [Ping timeout: 240 seconds]

17:17 raster has quit [Quit: Gettin' stinky!]

17:45 pH5 has quit [Quit: bye]

17:46 <daniels> bbrezillon: i haven't done a super-thorough review but yeah that would be the general idea I suppose

17:47 <daniels> MastaG: about the CREATE_DUMB failure, I would guess that's related to trying to use GBM on a non-master device node ...

17:54 alpernebbi has joined #panfrost

18:00 yann|work has joined #panfrost

18:15 pH5 has joined #panfrost

18:35 warpme_ has quit [Quit: Connection closed for inactivity]

18:35 raster has joined #panfrost

18:38 nerdboy has joined #panfrost

19:00 pH5 has quit [*.net *.split]

19:00 steev has quit [*.net *.split]

19:00 afaerber has quit [*.net *.split]

19:00 EmilKarlson has quit [*.net *.split]

19:00 mmind00 has quit [*.net *.split]

19:00 andrey-konovalov has quit [*.net *.split]

19:00 phh has quit [*.net *.split]

19:00 mani_s_ has quit [*.net *.split]

19:02 steev has joined #panfrost

19:02 pH5 has joined #panfrost

19:02 afaerber has joined #panfrost

19:02 mmind00 has joined #panfrost

19:02 phh has joined #panfrost

19:02 mani_s_ has joined #panfrost

19:02 EmilKarlson has joined #panfrost

19:02 andrey-konovalov has joined #panfrost

19:42 warpme_ has joined #panfrost

19:59 buzzmarshall has quit [Remote host closed the connection]

20:05 * alyssa doesn't see how shared buffers are supposed to work on mali compute shaders

20:06 <alyssa> AFAICT they're treated identically to SSBOs, but I'm trying to wrap my head how that's conformant

20:06 <alyssa> see also: http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.101897_0200_00_en/qfm1570448885817.html

20:08 <alyssa> Although I guess maybe memory being shared is more a property of coherence rather than how it's backed

20:12 <alyssa> unless they're playing games with the cache~

20:20 <alyssa> ---Yeah, I still don't see how this is conformant :\

20:23 <alyssa> ...Oops, I think i was using opengl wrong. Ahhh :p

20:24 <alyssa> Well, there goes an hour :<

20:30 <alyssa> Well, this is finally resolving a major mystrery about the hardware that's been here since day #1

20:36 <alyssa> Looks a lot like thread local storage.

20:36 <alyssa> Having major deja vu from spilling witchcraft about two months ago

20:41 stikonas has joined #panfrost

20:42 * alyssa feels like she's lookinig into sombra's door... i blame the hex...

20:43 robert_ancell has joined #panfrost

20:45 guillaume_g has quit [Quit: Konversation terminated!]

20:50 indy has joined #panfrost

21:04 alpernebbi has quit [Quit: alpernebbi]

21:12 buzzmarshall has joined #panfrost

22:05 NeuroScr has joined #panfrost

23:42 pH5 has quit [Quit: -_-]

23:45 warpme_ has quit [Quit: Connection closed for inactivity]