Segfault when trying to play trace from within vktraceviewer

[eero-t]

Setup:

• HW: SKL GT2
• OS: Ubuntu 16.04 with all the updates
• SW: Latest VulkanTools and Mesa from git

Use-case:

• Record trace of "cube" demo
• View trace with traceviewer: ./vktraceviewer -o cube.vktrace
• Click "Play" button

Expected result:

• Trace plays fine, like with "vkreplay -o cube.vktrace"

Actual result:

• NULL pointer access & crash at vkGetPhysicalDeviceSurfaceSupportKHR():

  ==12998== Process terminating with default action of signal 11 (SIGSEGV)
  ==12998==  Access not within mapped region at address 0x0
  ==12998==    at 0x63104BE: xcb_send_request64 (in /usr/lib/x86_64-linux-gnu/libxcb.so.1.1.0)
  ==12998==    by 0x6310968: xcb_send_request (in /usr/lib/x86_64-linux-gnu/libxcb.so.1.1.0)
  ==12998==    by 0x631C2DE: xcb_query_extension (in /usr/lib/x86_64-linux-gnu/libxcb.so.1.1.0)
  ==12998==    by 0x1FCF7DB7: wsi_x11_connection_create (wsi_common_x11.c:75)
  ==12998==    by 0x1FCF7DB7: wsi_x11_get_connection.isra.4 (wsi_common_x11.c:120)
  ==12998==    by 0x1FCF7EE1: x11_surface_get_support (wsi_common_x11.c:311)
  ==12998==    by 0x654256D: terminator_GetPhysicalDeviceSurfaceSupportKHR (in dbuild/loader/libvulkan.so.1.0.38)
  ==12998==    by 0x6542421: vkGetPhysicalDeviceSurfaceSupportKHR (in dbuild/loader/libvulkan.so.1.0.38)
  ==12998==    by 0x471E4C: vkReplay::manually_replay_vkGetPhysicalDeviceSurfaceSupportKHR(packet_vkGetPhysicalDeviceSurfaceSupportKHR*) (vkreplay_vkreplay.cpp:2696)
  ==12998==    by 0x4C0E1B: vkReplay::replay(vktrace_trace_packet_header*) (vkreplay_vk_replay_gen.cpp:3154)
  ==12998==    by 0x463141: VkReplayReplay(vktrace_trace_packet_header*) (vkreplay.cpp:166)
  ==12998==    by 0x446887: vktraceviewer_QReplayWorker::playCurrentTraceFile(unsigned long) (vktraceviewer_QReplayWorker.cpp:322)
  ==12998==    by 0x4473D9: vktraceviewer_QReplayWorker::StepReplay() (vktraceviewer_QReplayWorker.cpp:425)
  ==12998==    by 0x44D797: vktraceviewer_QReplayWorker::qt_static_metacall(QObject*, QMetaObject::Call, int, void**) (moc_vktraceviewer_QReplayWorker.cpp:159)
  ==12998==    by 0x5EC8C00: QObject::event(QEvent*) (in /usr/lib/x86_64-linux-gnu/libQt5Core.so.5.5.1)
  ==12998==    by 0x519905B: QApplicationPrivate::notify_helper(QObject*, QEvent*) (in /usr/lib/x86_64-linux-gnu/libQt5Widgets.so.5.5.1)
  ==12998==    by 0x519E515: QApplication::notify(QObject*, QEvent*) (in /usr/lib/x86_64-linux-gnu/libQt5Widgets.so.5.5.1)

Same happens with the other test-cases I tried. What the viewer does to mess up X11 connection creation?

[davidlunarg]

We are aware of this issue.

[eero-t]

This seems to work now much better. I'm still able to segfault it by playing, pausing, continuing, stepping & stopping enough, but so far haven't been able to reproduce that under Valgrind, so I guess it's timing related issue now.

Topping/untopping the replay window and using detach seems to make it easier to crash. Enabling validation can also make it easier to crash.

As this doesn't anymore crash right away when one presses play i.e. one can actually use it, I'm fine regardless of whether you want to close this, or keep it open until it's crash free.

[davidlunarg]

vktraceviewer is alpha software in the 1.0.42.0 SDK, it still needs lots of work. We put it in the SDK so that it gets better visibility and as a preview of the functionality it will provide. We'll work on the crashes and improve it sometime soon. We'll also accept pull requests. :-)

We'll keep this issue open for now.

[eero-t]

The weird thing is that while vkreplay seems now to be able to replay things reliably, trace viewer still crashes to that swapchain issue that was a problem for vkreplay earlier (at least with Mesa/Intel).

[eero-t]

Play button is disabled in the latest git VulkanTools traceviewer, is this on purpose?

[davidlunarg]

Attempting to play a trace from vktraceviewer crashes vktraceviewer. So we disabled the play button. vktraceviewer still needs lots of work, we'll address that issue as part of that work.

[krOoze]

This looks suspiciously similar to the https://github.com/KhronosGroup/Vulkan-Tools/issues/136 backtrace.