Open-Source NVIDIA Driver Is Still Sour For Some GPUs

[AJenbo]

Pretty sure i also only have a m64 left. I also have a PCI 6200 and PCIe 460, as well as access to a 210, plus some random bits i don't remember what are.

[imirkin]

Actually I was thinking more in terms of software setup. For hardware I am pretty much locked to what i have laying around, but in fact i do belive i have one or two working TNT2 cars

Which TNT2 cards? The M64's or the "original" ones? I have a M64 one, but not the plain (neither does anyone else afaik). Those cards don't get a whole lot of testing, although we'll happily fix issues if they're reported (e.g. an issue with VBIOS script parsing was recently fixed on early TNT cards).

I think I will run threw piglets ons a month and swap threw the GPUs I have lying around connected to a VGA monitor. If i do crash in to any issues with a card i will stop testing and instead focus on bisecting/reporting the issue. So a semi automatic test.

That would be fantastic! I'm not aware of anyone running through the test-suite with any level of regularity. If you have both nv50- and nvc0-era cards, that'd be even better. I wouldn't worry too much about testing on multiple similar-generation cards (not sure what your arsenal contains).

[imirkin]

But then later found an upstream report, https://bugs.freedesktop.org/show_bug.cgi?id=54437

There seems to have been next to no visible progress related to this bug. I suspect this is either a lack of resources in terms of people or access to a GTX 580 to debug. A couple different people have added dmesg logs to the FDO bug report, is there enough info provided there? Can I add anything else? I still have access to a GTX 580 and am keen to help in anyway to see it work with nouveau.

Unfortunately at present the only known workaround is to use the blob's pgraph firmware. There is an extractor script which grabs it for you now, so you don't need to do a mmiotrace. One of the people on that bug who has a NVC8 has been looking at the issue on and off, trying to figure out where the nouveau pgraph fw goes wrong, but while a few bugs were unearthed, they had no effect on the problem itself.

[leif81]

To anyone who is seeing these sorts of issues with nouveau -- please report them! If we don't know about it, we can't fix it.

See the Bugs wiki page for some basic information that bug reports should have. We do try to address them. In case of a regression like these, a bisect within drivers/gpu/drm/nouveau shouldn't be too time-consuming either and should vastly increase the chances the bug will be resolved.

I realize that Michael doesn't have the time/inclination/whatever to help fix these, but if you -- as a user -- are experiencing these types of problems, help us help you

I've seen a similar "failed to idle channel" gpu hang as Michael, but on my GTX 580, Michael saw it on a GTX 650.

I originnally filed a bug here, https://bugzilla.redhat.com/show_bug.cgi?id=860477

But then later found an upstream report, https://bugs.freedesktop.org/show_bug.cgi?id=54437

There seems to have been next to no visible progress related to this bug. I suspect this is either a lack of resources in terms of people or access to a GTX 580 to debug. A couple different people have added dmesg logs to the FDO bug report, is there enough info provided there? Can I add anything else? I still have access to a GTX 580 and am keen to help in anyway to see it work with nouveau.

[devius]

Some of that corruption is beautiful. I wonder if that's Nvidia's firmware engineers from the past communicating with us.

Maybe it's the software! It's become self-aware and decided to devote itself to art

[stalkerg]

Results of Michael as the paintings of Malevich.

[benmoran]

Just wanted to say thanks to the Nouveau devs for their work, in case any of you guys are reading this. I have a bit older discreet Nvidia card in my workstation at work, and the desktop compositing and basic 3D performance is spot on. It works great with hot-plugging multiple monitors also. The Proprietary driver would hang X about once a day on average, but the open source driver has always been smooth and pleasant. I use all AMD hardware for my personal machines, but I'm sure glad Nouveau exists for the ones I don't own. Cheers guys.

[AJenbo]

But really, any specific bug reports will always be appreciated -- you don't have to have the ultimate GPU collection to do that. Join us at #nouveau on freenode if you want to talk more specifics!

Actually I was thinking more in terms of software setup. For hardware I am pretty much locked to what i have laying around, but in fact i do belive i have one or two working TNT2 cars

I think I will run threw piglets ons a month and swap threw the GPUs I have lying around connected to a VGA monitor. If i do crash in to any issues with a card i will stop testing and instead focus on bisecting/reporting the issue. So a semi automatic test.

[imirkin]

I actually consider setting up a system to test GPU's on a semi regulare basis. Any suggestions on what would be the minimal (especially in terms of maintainable) set up that would give you guys the best information?
(I'm mostly used to Ubuntu, but I use GIT on a daily basis)

Well, for nv50-family, the differences tend to occur at nv50, nv84, nv92, nv94, nv98, nva0, nva3 chips (note -- marketing names have little correlation with the actual chipset). But there are also lots of other factors, like the types of connectors used, the specific scripts in the VBIOS (e.g. recently tracked down a regression that caused HDMI to sometimes fail to modeset on some NV96's due to a vbios command that wasn't quite working right). 3d-wise, nv84+ are basically all the same (within nv50), except some later chips have some additional features. Apparently a bunch of people with nva3-8 have major problems with gpu lockups, but it works just fine for others. Someone recently ran piglit on the original nv50, and apparently it has a lot of problems.

I'm less well versed in fermi+ differences -- 3d-wise fermi and kepler are pretty similar, it's largely the same until gk110/gk208 which use a different shader isa. On the kernel support end nvc0+ are also pretty similar, except for pgraph configuration, which is at least a little different on basically every nvcX card.

For older-than nv50, the differences tend to run in families, i.e. all nv40's are pretty similar, all nv30's are pretty similar. This is less true as you get older and older, e.g. nv17 introduced stuff that wasn't in nv11. And nv11 has stuff that wasn't in nv10/nv15. But I doubt you were talking about GeForce4 and older...

But really, any specific bug reports will always be appreciated -- you don't have to have the ultimate GPU collection to do that. Join us at #nouveau on freenode if you want to talk more specifics!

[AJenbo]

To anyone who is seeing these sorts of issues with nouveau -- please report them! If we don't know about it, we can't fix it.

See the Bugs wiki page for some basic information that bug reports should have. We do try to address them. In case of a regression like these, a bisect within drivers/gpu/drm/nouveau shouldn't be too time-consuming either and should vastly increase the chances the bug will be resolved.

I realize that Michael doesn't have the time/inclination/whatever to help fix these, but if you -- as a user -- are experiencing these types of problems, help us help you

I actually consider setting up a system to test GPU's on a semi regulare basis. Any suggestions on what would be the minimal (especially in terms of maintainable) set up that would give you guys the best information?
(I'm mostly used to Ubuntu, but I use GIT on a daily basis)