Announcement

Collapse
No announcement yet.

LOTS and LOTS of GPU fault logs from Linux 4.1.4 with Radeon HD 7770

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • LOTS and LOTS of GPU fault logs from Linux 4.1.4 with Radeon HD 7770

    While debugging my problems with Serious Sam 3: BFE on Debian stretch with Mesa 10.6.3, I came across an alarmingly large number of the following log lines from the radeon driver in the kernel:

    [22097.775805] radeon 0000:05:00.0: GPU fault detected: 147 0x03e84401
    [22097.775807] radeon 0000:05:00.0: VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x0FF0081F
    [22097.775808] radeon 0000:05:00.0: VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x08044001
    [22097.775809] VM fault (0x01, vmid 4) at page 267388959, read from TC (68)

    I checked to see how many there were and there are indeed a LOT:

    egrep "GPU fault|VM_CONTEXT1|VM fault" /var/log/kern.log | wc -l
    25808741

    Prior to booting my system a few days ago with a freshly compiled Linux 4.1.4 kernel and the Mesa 10.6.3 stack, I had never seen these messages before under any circumstance, especially when gaming or otherwise stressing the graphics card. I've already checked and the 4.1.5 kernel doesn't have any patches applied to the radeon driver yet, so I'm hoping that this is indeed a kernel bug and not a sign of a defective memory chip or a flaky power supply. I use lm-sensors to monitor CPU and GPU temps and neither has spiked alarmingly yet

    lspci:

    05:00.0 VGA compatible controller: Advanced Micro Devices, Inc. [AMD/ATI] Cape Verde XT [Radeon HD 7770/8760 / R7 250X]

    glxinfo:

    OpenGL renderer string: Gallium 0.4 on AMD CAPE VERDE
    OpenGL core profile version string: 3.3 (Core Profile) Mesa 10.6.3
    OpenGL core profile shading language version string: 3.30

    Does anyone have any advice on how I should proceed here? I'm going to check elsewhere to see if these log lines have been reported elsewhere, but quick Google searches haven't turned up anything yet.

  • #2
    Once at least mesa packages is at 10.6.4 or later, just rebuild at against llvm 3.7 if they don't.

    Comment


    • #3
      Mesa 10.6.4 contains critical fixes for your GPU. Please update.

      Comment


      • #4
        There are debian laboratory ready for 10.6.4, you can took it from there:

        http://anonscm.debian.org/cgit/pkg-xorg/lib/mesa.git

        Just change to llvm 3.6 in control and rules files... make sure you use gcc 4.9 as buiding with default gcc 5.2 does not work, etc...

        Also do not try llvm 3.7 as 10.6.4 can't be build aganst 3.7 without patching Not sure why it is not maded possible for 10.6 branch, Marek? Yeah just does not handle R600 to AMDGPU renaming. At least those 3 are needed, but maybe i miss some:

        http://cgit.freedesktop.org/mesa/mes...a1c5088d4775e1
        http://cgit.freedesktop.org/mesa/mes...98895f32ec25a2
        http://cgit.freedesktop.org/mesa/mes...a1c5088d4775e1
        Last edited by dungeon; 08-14-2015, 02:49 PM.

        Comment


        • #5
          But yeah TheOneKEA, even if you have all newest stack, Serious Sam 3 might lockup sometimes. So make sure you have disabled menu simulations before starting the game with "+men_bUseMenuSimulation 0".

          The Talos Principe also can made GPU hangs easely, but there solution is different... you don't need that switch nor to switch off menu simulation and also there VRAM detection is fine. There Croteam introduce "+mem_iOptimizeCopy = 8" cvar, so just lower down that to 4 or 0 and it will be stable... That is what i think happens with those Serious engine games, too much optimizations by default and some drivers like mesa simply hangs often on that - well not sure why
          Last edited by dungeon; 08-14-2015, 03:09 PM.

          Comment


          • #6
            Nobody has cherry-picked the llvm 3.7 fixes to make mesa 10.6 work with that. I suggest you stay with llvm 3.6 if you want to use mesa 10.6.

            Comment


            • #7
              He, he, Debian uses llvm 3.5 still with mesa 10.6.x... i guess probably because mesa can't build against llvm 3.6 on all architectures, 3.6 is real bad in that sense

              Comment


              • #8
                Originally posted by dungeon View Post
                He, he, Debian uses llvm 3.5 still with mesa 10.6.x... i guess probably because mesa can't build against llvm 3.6 on all architectures, 3.6 is real bad in that sense
                Debian has llvm-3.7 in sid. If I hadn't downgraded to jessie yesterday because sid was so badly broken I'd try it...

                Comment


                • #9
                  Originally posted by rrohbeck View Post

                  Debian has llvm-3.7 in sid. If I hadn't downgraded to jessie yesterday because sid was so badly broken I'd try it...
                  Sid is broken by design and on purpose, that show up once big transitions taking place... i just keep Sid minimal in times like this and don't worry much about

                  Comment


                  • #10
                    Originally posted by dungeon View Post
                    There are debian laboratory ready for 10.6.4, you can took it from there:

                    http://anonscm.debian.org/cgit/pkg-xorg/lib/mesa.git

                    Just change to llvm 3.6 in control and rules files... make sure you use gcc 4.9 as buiding with default gcc 5.2 does not work, etc...

                    Also do not try llvm 3.7 as 10.6.4 can't be build aganst 3.7 without patching Not sure why it is not maded possible for 10.6 branch, Marek? Yeah just does not handle R600 to AMDGPU renaming. At least those 3 are needed, but maybe i miss some:

                    http://cgit.freedesktop.org/mesa/mes...a1c5088d4775e1
                    http://cgit.freedesktop.org/mesa/mes...98895f32ec25a2
                    http://cgit.freedesktop.org/mesa/mes...a1c5088d4775e1
                    How many of the LLVM packages located at http://llvm.org/releases/download.html will I need? I'm going to start with the Clang and LLVM packages and see if that is enough to build mesa-git, but I've already tried to build LLVM once before and was stymied.

                    EDIT: I did some more research and found the LLVM 3.6 packages made available for download at http://llvm.org/apt/, so I installed those, confirmed that LLVM 3.5 was gone and that 3.6 was in my $PATH, and then used dpkg-buildpackage to start building Mesa 10.6.4.

                    EDIT 2: I was able to get the x86_64 LLVM 3.6 packages installed and compile and install Mesa 10.6.4 from the git tree, but I can't get i386 variants of Mesa 10.6.4 to compile yet, and I'm starting to encounter an apt dependency swamp that has me worried about breaking my system. Worse, by installing the x86_64 Mesa packages, I was forced to remove the i386 Mesa 10.6.3 packages, which caused the Steam client to be uninstalled. It looks like I will be busy figuring this out for a while, unless someone has some advice...
                    Last edited by TheOneKEA; 08-15-2015, 11:29 PM. Reason: Dependencies are fun!

                    Comment

                    Working...
                    X