Crash Diagnostic Layer

The Crash Diagnostic Layer (CDL) is a Vulkan layer to help track down and identify the cause of GPU hangs and crashes. It works by instrumenting command buffers with completion checkpoints. When an error is detected a dump file containing incomplete command buffers is written. Often the last complete or incomplete commands are responsible for the crash.

Building

See the associated BUILD.md file for details on how to build the Crash Diagnostic Layer for various platforms.

Runtime requirements

CDL uses the following extensions. If an extension is not present, some functionality might be disabled.

• VK_KHR_timeline_semaphore (or Vulkan 1.2) - used to track forward progress of queue submissions
• VK_NV_device_diagnostic_commands - used to track forward progress within command buffers
• VK_AMD_buffer_marker - used to track forward progress within command buffers (if the NV extension is not present) and the state of binary semaphores.
• VK_AMD_device_coherent_memory - improves the accuracy of VK_AMD_buffer_marker tracking
• VK_EXT_device_fault - allows drivers to report data related to a VK_ERROR_DEVICE_LOST, such as invalid device addresses.
• VK_EXT_device_address_binding_report - allows drivers to report lifecycle information about device address assignments, which can help identify the source of device faults.

Running

CDL can be used as an explicit or implicit layer. The loader's documentation describes the difference between implicit and explicit layers, but the relevant bit here is that implicit layers are meant to be available to all applications on the system, even if the application doesn't explicitly enable the layer. On the other hand, explicit layers are easier to use when doing application development.

Explicit Layer

To use CDL as an explicit layer, enable it with vkconfig or do the following:

1. The directory containing the file VkLayer_crash_diagnostic.json is included in the layer search path, by including it in either the VK_LAYER_PATH or VK_ADD_LAYER_PATH environment variable or by using vkconfig.
2. Include the layer name in the VK_INSTANCE_LAYERS environment variable or the ppEnabledLayerNames field of VkInstanceCreateInfo.

Implicit Layer

When using CDL as an implicit layer, it is disabled by default. To enable the layer's functionality, set the CDL_ENABLE environment variable to 1.

Registering the Layer

In order to be discovered by the Vulkan loader at runtime, implicit layers must be registered. The registration process is platform-specific and is discussed in detail in the Vulkan-Loader documentation. In all cases, it is the layer manifest (the .json file) that is registered; the manifest contains a relative path to the layer library, which can be in a separate directory.

Basic Usage

Once the layer is enabled, if vkQueueSubmit() or other Vulkan functions returns a fatal error code (usually VK_ERROR_DEVICE_LOST), a dump file of the command buffers (and other state) that failed to execute are written to disk.

Logging

CDL outputs messages at runtime to indicate it is enabled, to trace certain commands, and to report when a gpu crash has occurred. The message_severity, log_file, trace_on and trace_all_semaphores configuration settings control which messages are output and where they are sent.

VK_EXT_debug_utils and VK_EXT_debug_report extensions can also be used to receive log messages directly in the application.

Dump files

A unique log file directory is created every time an application is run with CDL enabled. The log file directories are named with the timestamp of the format YYYY-MM-DD-HHMMSS. This prevents subsequent runs from overwriting old data, in case the application is non-deterministic and producing different crashes on each run. The location of these files is platform-specific:

• Linux: ${HOME}/cdl/
• Windows: %USERPROFILE%\cdl\

The output directory for dump files can be changed using the output_path configuration setting, described below.

The name of the dump file is cdl_dump.yaml, since the file is in the YAML format. Other files, such as dumped shaders, may be present in the dump directory.

Configuration

This layer implements the VK_EXT_layer_settings extension, so it can be configured with vkconfig, programmatically, or with environment variables. See the manifest for this layer and the layer manifest schema for full details. The discussion below uses the key field to identify each setting, the env field defines the corresponding environment variable, and the label field defines the text you will see when using vkconfig.

• watchdog_timeout_mscan be set to enable a watchdog thread that monitors the time between queue submissions. If this timeout value (specified in milliseconds) is hit without a queue submission, the gpu is assumed to be crashed and a dump file is created.
• Dump files
  • output_path can be set to override the directory where log files and shader binaries are written. This can be a full path (starting with / or a drive letter) or a path relative to the application current working directory.
  • dump_queue_submissions controls which queue submissions are dumped. running causes only the submission currently executing to be dumped. pending will also dump any submissions that have not started execution.
  • dump_command_buffers controls which command buffers are dumped. running causes only the command buffer currently executing to be dumped. pending will also dump any command buffers that have not started execution. all will dump all known command buffers.
  • dump_commands controls which commands are dumped. running causes only the commands currently executing to be dumped. pending will also dump any commands that have not started execution. all will dump all commands in the command buffer.
  • dump_shaders controls if shaders are included in the dump directory. Possible values for this setting are: off - no output, on_crash - only dump the shaders that are bound at the time of a gpu crash, on_bind - dump shaders only when they are bound, and all - dump all shaders as soon as they are created.
• Logging
  • message_severity can be set to a comma-separated list of the types of messages CDL should output to the default logger. Application defined loggers should control which messages they want to recieve with the options available in the VK_EXT_debug_utils or VK_EXT_debug_report extensions.
  • log_file can be set to control where log messages are sent by the default logger. There are several special values. stderr and stdout send messages to the application console. none disables the default logger. Any other value is assumed to be an absolute or relative path to the log file.
  • trace_on can be enabled to cause important commands, such as vkQueueSubmit to be logged as they occur.
  • trace_all_semaphores enables logging messages about every vulkan command that uses semaphores.
• State tracking
  • sync_after_commands adds a pipeline barrier after every instrumented vulkan command. This will reduce performance and may cause some hangs to go away. This option currently only works when using VK_KHR_dynamic_rendering
  • instrument_all_commands can be enabled to include completion markers around every vulkan command. This may allow more accuratute fault locations at the expense of larger command buffers and reduced performance.
  • track_semaphores enables detailed semaphore state reporting in runtime logging and dump files. VK_AMD_buffer_marker is required for this feature.

Layer Details

Layer Properties

• API Version: 1.3.296
• Implementation Version: 1
• Layer Manifest: VkLayer_crash_diagnostic.json
  • File Format: 1.2.0
  • Layer Binary: libVkLayer_crash_diagnostic.so
• Platforms: WINDOWS, LINUX
• Status: ALPHA
• Number of Layer Settings: 4

Layer Settings Overview

Layer Settings Details

Watchdog timeout (ms)(ALPHA)

If set to a non-zero number, a watchdog thread will be created. This will trigger if the application fails to submit new commands within a set time (in milliseconds) and a log will be created as if the a lost device error was encountered.

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.watchdog_timeout_ms
• Environment Variable: CDL_WATCHDOG_TIMEOUT_MS
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: INT
• Setting Default Value: 30000

Output Path(ALPHA)

The directory where dump files and shader binaries are written.

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.output_path
• Environment Variable: CDL_OUTPUT_PATH
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: STRING
• Setting Default Value:

Dump queue submissions(ALPHA)

Control which queue submissions are included in the dump.

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.dump_queue_submits
• Environment Variable: CDL_DUMP_QUEUE_SUBMITS
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: ENUM
• Setting Default Value: running

Dump command buffers(ALPHA)

Control which command buffers are included in the dump.

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.dump_command_buffers
• Environment Variable: CDL_DUMP_COMMAND_BUFFERS
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: ENUM
• Setting Default Value: running

Dump commands(ALPHA)

Control which commands are included in the dump.

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.dump_commands
• Environment Variable: CDL_DUMP_COMMANDS
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: ENUM
• Setting Default Value: running

Dump shaders(ALPHA)

Control of shader dumping.

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.dump_shaders
• Environment Variable: CDL_DUMP_SHADERS
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: ENUM
• Setting Default Value: off

Message Severity(ALPHA)

Comma-delineated list of options specifying the types of log messages to be reported

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.message_severity
• Environment Variable: VK_LUNARG_CRASH_DIAGNOSTIC_MESSAGE_SEVERITY
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: FLAGS
• Setting Default Value: error

Log file name(ALPHA)

none = no logging, stderr or stdout = to the console, otherwise an absolute or relative path

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.log_file
• Environment Variable: CDL_LOG_FILE
• Platforms: WINDOWS, LINUX, MACOS
• Setting Type: STRING
• Setting Default Value: stderr

Enable Tracing(ALPHA)

All Vulkan API calls intercepted by the layer will be logged to the console.

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.trace_on
• Environment Variable: CDL_TRACE_ON
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: BOOL
• Setting Default Value: false

Enable semaphore log tracing.(ALPHA)

Semaphore events will be logged to console.

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.trace_all_semaphores
• Environment Variable: CDL_TRACE_ALL_SEMAPHORES
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: BOOL
• Setting Default Value: false

Synchronize commands(ALPHA)

Add pipeline barriers after instrumented commands.

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.sync_after_commands
• Environment Variable: CDL_SYNC_AFTER_COMMANDS
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: BOOL
• Setting Default Value: false

Instrument all commands(ALPHA)

All commands will be instrumented.

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.instrument_all_commands
• Environment Variable: CDL_INSTRUMENT_ALL_COMMANDS
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: BOOL
• Setting Default Value: false

Track semaphores(ALPHA)

Enable semaphore tracking.

Setting Properties:

• vk_layer_settings.txt Variable: lunarg_crash_diagnostic.track_semaphores
• Environment Variable: CDL_TRACK_SEMAPHORES
• Platforms: WINDOWS, LINUX, MACOS, ANDROID
• Setting Type: BOOL
• Setting Default Value: true