Dear Maintainer,

*** Reporter, please consider answering these questions, where
appropriate ***

    * What led up to the situation?

      Trying to run wine-staging on a GPU/cuda compute node failed
because vulkan was not correctly configured.
      see also https://bugs.winehq.org/show_bug.cgi?id=52647
      In the course of investigation, I noticed that vulkaninfo did
always seg-fault on headless cuda/GPU machines under Debian11.


    * What exactly did you do (or not do) that was effective (or
      ineffective)?

      I've tried to install and uninstall various graphics/vulkan driver.
      Testing a machine without GPU's did not show this issue.

      Here is the output when running the debugger on vulkaninfo

     $ gdb vulkaninfo
     For help, type "help".
     Type "apropos word" to search for commands related to "word"...
     Reading symbols from vulkaninfo...
     (No debugging symbols found in vulkaninfo)
     (gdb) r
     Starting program: /usr/bin/vulkaninfo
     [Thread debugging using libthread_db enabled]
     Using host libthread_db library
"/lib/x86_64-linux-gnu/libthread_db.so.1".

     Program received signal SIGSEGV, Segmentation fault.
     0x000015554ceb23fc in ?? () from
/usr/lib/x86_64-linux-gnu/libvulkan_intel.so
     (gdb) Quit
     quit)

      The output of these commands is attached

         $ vulkaninfo
         $ vkcube
         $ vkcubepp

         $ lspci|grep VGA
         $ nvidia-smi

         $ ls -altr  /usr/share/vulkan/icd.d/
         $ cat  /usr/share/vulkan/icd.d/*

         $ dpkg -l|grep "icd\|xorg\|nvid\|cuda\|mesa\|vulkan"

         $ glxinfo

    * What was the outcome of this action?

      vulkaninfo, vkcube always ended up in a seg-fault when running on
Cuda/GPU machines.
      Doing this on machines without Nvidia GPU's does not show this issue.


    * What outcome did you expect instead?
      I understand that Cuda is proprietary technology and therefore
this platform is difficult to support.

      I'd expect that vulkaninfo does not seg-fault, and finishes
gracefully with a more or less meaning full error message.
      Ideally, I should indicate what is needed to fix this issue.

      Afterall, it should be possible to run vulkaninfo on a headless
GPU/cuda machine, that should enable also
      runing wine-staging on that machine.


*** End of the template - remove these template lines ***

Am 3/17/22 um 14:42 schrieb Debian Bug Tracking System:

attached is the log file that was promised in the initial bug report.

When checking for possible causes, I noticed that the framebuffer is now
(in Debian11) associated with the first gpu. This seems wrong as this is
a headless machine, and the nvidia-gpu's are used only as a
computational accelerator for cuda workloads, and not for visualization.

This is shown with the command
    lshw -C display

which has this entry
    logical name: /dev/fb0

The full log is attached.

When running Debian10, this entry is not shown, and vulkaninfo was working.

Is there an (easy?) way to disable /dev/fb0 ?

Further investigation shows that part of the issue is package
"mesa-vulkan-drivers"
and also on machines without nvidia-gpus.


When uninstalling  nvidia-drivers, and mesa-vulkan-drivers, vulkan-tools
reports correctly an error that no vulkan device is found
When installing mesa-vulkan-driver, and having the nvidia-driver in
place, "gdb vulkaninfo" results in this backtrace


Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
ERROR: [Loader Message] Code 0 : loader_scanned_icd_add: Could not get
'vkCreateInstance' via 'vk_icdGetInstanceProcAddr' for ICD
libGLX_nvidia.so.0
ERROR: [Loader Message] Code 0 :
/usr/lib/i386-linux-gnu/libvulkan_intel.so: wrong ELF class: ELFCLASS32
ERROR: [Loader Message] Code 0 :
/usr/lib/i386-linux-gnu/libvulkan_radeon.so: wrong ELF class: ELFCLASS32
ERROR: [Loader Message] Code 0 :
/usr/lib/i386-linux-gnu/libvulkan_lvp.so: wrong ELF class: ELFCLASS32

Program received signal SIGSEGV, Segmentation fault.
0x000015554aa4d3fc in ?? () from
/usr/lib/x86_64-linux-gnu/libvulkan_intel.so
(gdb) bt
#0  0x000015554aa4d3fc in ?? () from
/usr/lib/x86_64-linux-gnu/libvulkan_intel.so
#1  0x000015554aa4ea65 in ?? () from
/usr/lib/x86_64-linux-gnu/libvulkan_intel.so
#2  0x0000155554bef7e7 in ?? () from
/usr/lib/x86_64-linux-gnu/libvulkan.so.1
#3  0x0000155554befc22 in ?? () from
/usr/lib/x86_64-linux-gnu/libvulkan.so.1
#4  0x000015554a96fde4 in ?? () from
/usr/lib/x86_64-linux-gnu/libVkLayer_MESA_device_select.so
#5  0x0000155554bef2a3 in ?? () from
/usr/lib/x86_64-linux-gnu/libvulkan.so.1
#6  0x0000155554bf1e05 in vkEnumeratePhysicalDevices () from
/usr/lib/x86_64-linux-gnu/libvulkan.so.1
#7  0x00005555555b5639 in ?? ()
#8  0x00005555555638b6 in ?? ()
#9  0x0000155554fead0a in __libc_start_main (main=0x555555563670,
argc=1, argv=0x7fffffffe458, init=<optimized out>, fini=<optimized out>,
     rtld_fini=<optimized out>, stack_end=0x7fffffffe448) at
../csu/libc-start.c:308
#10 0x000055555556580a in ?? ()



When having mesa-vulkan-drivers installed, and nvidia drivers are
disabled (e.g. with "rmmod nvidia-drm nvidia-settings"), or on machines
without nvidia-gpu's,
vulkan-tools shows a lot of properties, but fails also with a
segmentation fault.

Running "gdb vulkaninfo", I get this backtrace:

Thread 1 "vulkaninfo" received signal SIGSEGV, Segmentation fault.
0x00007fffcd2f52bc in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_lvp.so
(gdb) bt
#0  0x00007fffcd2f52bc in ?? () from
/usr/lib/x86_64-linux-gnu/libvulkan_lvp.so
#1  0x00007fffcd2f57f1 in ?? () from
/usr/lib/x86_64-linux-gnu/libvulkan_lvp.so
#2  0x00005555555a7140 in ?? ()
#3  0x00005555555a123e in ?? ()
#4  0x0000555555564565 in ?? ()
#5  0x00007fffd6fa2d0a in __libc_start_main (main=0x555555563670,
argc=1, argv=0x7fffffffda98, init=<optimized out>, fini=<optimized out>,
     rtld_fini=<optimized out>, stack_end=0x7fffffffda88) at
../csu/libc-start.c:308
#6  0x000055555556580a in ?? ()


Though the behavior is slightly different, both libraries
libvulkan_intel.so and libvulkan_lvp.so, are part of the
mesa-vulkan-drivers package.

So this indicates that this is a bug in package mesa-vulkan-drivers.