Radeon GPU Recovery To Hit Linux 2.6.34 Kernel

[curaga]

I wonder, how well does it recognize a hang? Maybe a magic sysrq key to force a gpu reset could be useful?

[drag]

Erm, it's just a kernel. Doesn't the POSIX portion fall under GNU?

Both. POSIX is a specification for application environment to promote source-code level compatibility between multiple operating systems.

User applications and shell support is just part of it. You have things like POSIX file system access (read/write methods and such.), POSIX file permissions (user/group/world, read/write/execute), POSIX IPC (fifos, sockets, signals (like kill -9, HUP, etc) and all sorts of stuff like that.

POSIX basically describes what a 'unix' OS needs to support to be mostly compatible with other unix systems.

A something like a non-POSIX file system would be something like Google File System were you need to specifically program your application to talk to it.

Of course you do not have to be Unix proper to be POSIX. You just have to implement compatibility in the kernel, file system, and in a handful of userland applications.

Linux is a big example of this.

Another big example is Windows NT kernel has POSIX compatibility so if you add Microsoft's SFU (services for Unix)(Windows Server stuff) or SUA (Subsystem for UNIX-based Applications)(Windows Vista, Windows 7) you can turn Windows into a unix OS. (note the lowercase) One of the neat things about NT kernel is that it is able to have multiple personalities. The Win32 environment originally was a personality for Widnows 9x compatibility that was used in Windows 2000 and XP. POSIX is another personality that NT kernel supports. It can probably support others, but those are the only two I know.

This is different from Cygwin. Cygwin tries to make a Linux-like environment that is built on top of Win32. Like how Wine creates a Win32 API on top of Linux's regular environment. Microsoft's SUA bypasses the Win32 stuff and run directly on top of the kernel.

Of course Linux implements POSIX, but Linux applications tend to use a lot of GNU, X Windows, and Linux-specific stuff that goes beyond POSIX. Mos commercialt Unix OSes, in fact, implement Linux compatibility layers; since Linux is now very common in the "Enterprise". AIX and Solaris both have Linux API compatibility layers to run Linux binaries.

Microsoft's SFU was based on OpenBSD for a lot of things. I am not sure of SUA, but I expect that it's going to be OpenBSD, too.

[bridgman]

I think the hang detect logic looks at the ring buffer command queue and checks to see if the GPU is still consuming commands (ie updating its read and write pointers). If there are commands queued and the GPU doesn't seem to be executing them any more then something in the acceleration pipe is probably hung.

[d4ddi0]

hang recovery will be welcome indeed

I know we pay a price for living on the cutting edge with th eradeon driver,

but it is still kind of embarrassing to be hard rebooting an OS that supposedly has uptimes measured in years.

My last couple of hard-lockups have been KDE compositing, and I finally just disabled it. I'll be very happy if the video card is limited to "just" crashing X.

[Neuro]

I know we pay a price for living on the cutting edge with th eradeon driver,

but it is still kind of embarrassing to be hard rebooting an OS that supposedly has uptimes measured in years.

My last couple of hard-lockups have been KDE compositing, and I finally just disabled it. I'll be very happy if the video card is limited to "just" crashing X.

Oh, it doesn't crash X. It marely blanks the screen for a second and prints some messages in messages.log. I'm running a patched 2.6.34-rc1.

I heavily commented on the bug https://bugzilla.kernel.org/show_bug.cgi?id=15276 and I have to say that Jerome has done one heck of a great work! The patch does wonders and KDE compositing is now rock-solid. I get the "blanking" only testing the special case of Present Windows, flash games work without gliches, so does Inkscape.

I really really hope this gets into mainline 2.6.34. This is really a must-have fix.

[korpenkraxar]

I really really hope this gets into mainline 2.6.34. This is really a must-have fix.

Yup. KMS + KDE was horrible without it for over a month until these patches emerged.

[pingufunkybeat]

Interesting. I've never had a problem with KDE and vanilla kernels.

Is it chip-specific?

[Luzipher]

After all GPU recovery is not going in for 2.6.34. Check this thread on the lkml: http://lkml.org/lkml/2010/3/30/5 . Seems to have caused problems with r300 series cards, Jerome might do a smaller patch version for .34 which only affects r6xx/r7xx though.

[wswartzendruber]

Eh, isn't the merge window closed?

[Yfrwlf]

Eh, isn't the merge window closed?

Not if you're special.

Seriously though, if it is really something that really really should not be detrimental to anything, perhaps including things like that after the merge window closure would still be okay.

Anyway, I'm really happy to see the graphics issues causing hard locks problem being addressed.

but it is still kind of embarrassing to be hard rebooting an OS that supposedly has uptimes measured in years.

+1 to the embarrassment. Let the days of rock-solid Linux desktops begin.