Release notes for the Genode OS Framework 19.08

The stated theme of this year's road map is "bridging worlds", which expresses our ambition to smoothen the practical use of Genode-based systems such as Sculpt OS. The current release pays tribute to this ambition by addressing a great number of practical concerns: How to accommodate the staggering variety of keyboard layouts out there? (Section Flexible keyboard layouts) How can the system gracefully respond when confronted with exotic USB devices? (Section Storage-stack improvements) How to set the system time from within the system? How does SNTP fit in here? (Section General system time concept) How to approach the remote administration of the system? (Section Enhanced SSH terminal) How to copy and paste text securely between mutually distrusting subsystems? (Section Clipboard) Or how to overcome the captive portal of a Hotel WiFi with Sculpt OS? (Section Disposable VM for handling captive portals) By providing answers to those questions, we believe to make Genode - and Sculpt OS in particular - generally more useful.

As another take on "bridging worlds", we continue our effort to bring the rich Sculpt OS software stack to the 64-bit ARM world, in particular to our most loved SoC family, namely NXP i.MX. Section 64-bit ARM and NXP i.MX8 reports on our progress in this direction.

Under the hood, there are a few exciting developments that will greatly reduce the effort of running existing software on Genode. In particular, Genode's (entirely optional) C runtime has gained the ability to emulate the traditional execve and fork mechanisms. (Section Consolidation of the C runtime and Noux) This will eventually alleviate the need for our present noux runtime environment to the benefits of better performance and increased flexibility.

Further highlights of Genode 19.08 are a major update of Qt5 to version 5.13 (Section Updated Qt5) and the continuation of our kernel-agnostic virtualization story (Section Virtualization).

Flexible keyboard layouts
64-bit ARM and NXP i.MX8
Base framework and OS-level infrastructure
  Tracing
  Consolidation of the C runtime and Noux
  General system time concept
  Accessing SMBIOS tables
  Clipboard
  Enhanced SSH terminal
  Storage-stack improvements
  Updated Ada/SPARK runtime
Libraries and applications
  Updated Qt5
  Virtualization
  Disposable VM for handling captive portals
Build system and tools

Flexible keyboard layouts

Genode is used worldwide in a multilingual context beyond Germany and common technical realms of English. Therefore, we had to address localized keyboard-input handling for quite some time now and introduced the input-filter component in 17.02. The component merges input streams and applies several forms of input transformations, in particular the application of keyboard layouts to supplement the input-event stream with character events.

But as we are by no means localization experts, our solution, while performing a solid job for selected layouts, also had some quirks and rough edges when it came to French or even Swiss German. First, our oversimplified notion of Caps Lock as just a pressed Shift key is plain wrong but part of all our character-generator configurations. We just missed this drawback because none of our developers uses Caps Lock regularly. Further, US English and Germany layouts work very well without dead keys, but crossing any German border (except the Austrian) is impossible without support for key sequences composing special characters. The French keyboard layout in Genode tried to alleviate the lack of compose sequences by adding an additional Circumflex modifier and character mapping, which unfortunately is not standard.

Beginning at this state of affairs, we researched common practice in international keyboard-input handling, sought a quasi-standard source for layout configurations, and addressed the drawbacks mentioned before. During our research we found out that no current implementation is void of critique and, therefore, decided to look more into X11/XKB as our open-source quasi-standard solution, but always had an eye on the proprietary world.

The handling of key events in X11/XKB happens on three layers.

Key codes

On the key-code layer, the device driver programs the keyboard and generates a stream of key-code (i.e., scan-code) events, which represent the physical location of the actual key on the keyboard.

Key symbols

These key codes are mapped to key symbols, which represent the label imprinted on the key. So, the key code producing US English Q (QWERTY keyboard) generates A on a French keyboard (AZERTY). Modifiers like Shift, AltGr, and Caps Lock are included in the key-symbol mapping. Additionally, some layouts map key codes to dead key symbols, which start the before-mentioned compose sequences. Key repeat is also implemented as key-symbol repeat actually.

Characters

On top of this stack, the key symbols are mapped to characters represented as Unicode codepoints or UTF-8 strings. The procedure obviously includes key symbols that have no character representation (e.g. Control and Alt). Key symbols forming a valid compose sequence generate characters on this level (e.g., dead-key circumflex plus e generates ê).

We limited our research to Western keyboard-input handling and only had a blink into the direction of Chinese-Japanese-Korean (CJK) and advanced input methods (IM). This simplification is supported by the fact that CJK can also be based on the mechanisms mentioned with some limitations only. Nevertheless, we do not expect to never touch this topic again.

After doing our homework of keyboard-input handling, we worked on squeezing all available layout information out of X11/XKB, which resulted in a small tool residing in tool/xkb2ifcfg. For those wondering, the name is just a silly acronym for XKB to input-filter configuration that pays tribute to the boringness of this task. After building the tool by a run of make in the tool path, it can be used as follows. Please make sure you have libxkbcommon development packages installed beforehand.

 xkb2ifcfg generate <layout> <variant> <locale>

 xkb2ifcfg generate us euro       en_US.UTF-8
 xkb2ifcfg generate de nodeadkeys de_DE.UTF-8

If the parameter combination is available, xkb2ifcfg prints a input-filer chargen configuration for the selected layout to standard output. Valid layout and variant options can be figured out from the LAYOUTS section in man 7 xkeyboard-config, where variant strings are depicted in parentheses after the layout (e.g., us(euro)). The locale option has the standard locale syntax (see /usr/share/i18n/locales). The tool needs all three parameters to gather the correct key-map and compose-sequence information. The generated chargen configurations include <map> and <key> nodes corresponding to significant modifier states and <sequence> nodes (described later). For simplicity of the generator, the <key> nodes always use the code="..." attribute, but also have a comment with the UTF-8 character appended.

 <key name="KEY_MINUS" code="0x00df"/>   <!-- ß -->

Last, we addressed the improvement of the input-filter character generator and the actual chargen configuration files in Genode. Therefore, we specified the modifier configuration assumed by the standard chargen files as <mod1> corresponds to Shift, <mod2> to Control, <mod3> to AltGr, and <mod4> to Caps Lock.

 <mod1> <key name="KEY_LEFTSHIFT"/> <key name="KEY_RIGHTSHIFT"/> </mod1>
 <mod2> <key name="KEY_LEFTCTRL"/>  <key name="KEY_RIGHTCTRL"/> </mod4>
 <mod3> <key name="KEY_RIGHTALT"/> </mod4> <!-- AltGr -->
 <mod4> <rom name="capslock"/> </mod4>

As outlined above, the <key> nodes generated by xkb2ifcfg always use the code attribute for the Unicode codepoint. Because of this and because UTF-8 also refers to codepoints, we deprecated the b0/b1/b2/b3 attributes for character definition with this release.

The chargen is also extended by the <sequence> configuration node. A sequence node permits the definition of dead-key/composing character sequences. With such sequences, the character is not generated instantly on key press but only after the sequence is completed. If an unfinished sequence can't be completed due to an unmatched character, the sequence is aborted and no character is generated. We support sequences of up to four characters at the moment.

For example, the French AZERTY keyboard layout has a dead key for Circumflex Accent ^ right of the P key (which is bracket left [ on US keyboards). When Circumflex is pressed no visible character should be generated instantly but the accent must be combined with a follow-up character (e.g., Circumflex plus a generates â).

Dead keys can be defined in the <key> nodes of any <map> by using codepoints not used for direct output, for example, Combining Diacritical Marks beginning at U+0300. The French Circumflex example can be configured like follows.

 <mod1>
   <key name="KEY_LEFTSHIFT"/> <key name="KEY_RIGHTSHIFT"/>
 </mod1>
 <map>
   <key name="KEY_Q"         code="0x0061"/> <!-- a -->
   <key name="KEY_LEFTBRACE" code="0x0302"/> <!-- dead_circumflex -->
 </map>
 <map mod1="true">
   <key name="KEY_Q"         code="0x0041"/> <!-- A -->
 </map>
 <sequence first="0x0302" second="0x0061" code="0x00e2"/> <!-- â -->
 <sequence first="0x0302" second="0x0041" code="0x00c2"/> <!-- Â -->

Fortunately, the configuration is automatically generated by xkb2ifcfg, but admittedly quite extensive. Therefore, we manually amended the chargen configurations before adding them to Genode, which also gave us the chance to apply some adjustments like follows for AltGr in Swiss German.

 <map mod1="false" mod2="false" mod3="true" mod4="false">
   <key name="KEY_1" code="0x00a6"/> <!-- ¦ (*) -->
   <key name="KEY_4" code="0x00b0"/> <!-- ° (*) -->
   <key name="KEY_5" code="0x00a7"/> <!-- § (*) -->
 </map>

Beside the advanced input methods mentioned before, there are still loose ends we are going to address in the upcoming releases. For example, the current key handling in our Qt5 back end maps localized key symbols incorrectly (think AZERTY vs. QWERTY) in combination with shortcuts like Ctrl-A.

64-bit ARM and NXP i.MX8

64-bit ARM support in our custom base-hw kernel

By introducing rudimentary Raspberry Pi 3 support on top of the Fiasco.OC kernel in the previous release, the first ARM 64-bit support has entered the Genode OS framework. We continued pursuing the ARM 64-bit path and introduce support for Raspberry Pi 3 as well as the i.MX8 evaluation kit (EVK), this time using our own base-hw kernel.

Noteworthy additions in the base-hw kernel are support for the AARCH64 system level architecture, and the use of the modern GIC v3 interrupt controller on top of the i.MX8 EVK board. In comparison to the GICv2, GICv3 adds support for more than eight CPUs, more than 1020 interrupt IDs, and offers fast register access to the CPU interface, instead of memory-mapped I/O access. Minor changes had to be made to the page-table implementation of ARMv7 with Large Physical Address Extension (LPAE) to re-use it for ARMv8. Moreover, the internal kernel API for TLB maintenance needed to be changed slightly for all ARM platforms.

We expanded our regular testing infrastructure with two AARCH64 platforms, namely Raspberry Pi 3 via Qemu and the NXP i.MX8 EVK board as physical hardware. Both platforms are driven with a single CPU core only at the moment.

Network driver for i.MX7 and i.MX8

We updated the fec network driver to version 4.16.3, which adds support for i.MX7 and i.MX8 SoCs. This makes i.MX8 a viable platform for Genode-based networking scenarios.

Enhanced packaging and test infrastructure for ARMv8

Besides the improved base-hw kernel, we enabled additional infrastructure for ARMv8 platforms. For example, noux packages - like coreutils, bash - are now available, the standard C++ library is in place, and support for Genode's port of the Linux TCP/IP stack is enabled.

Additionally, ARMv8 is now regularly tested within our nightly depot_autopilot runs.

Base framework and OS-level infrastructure

Tracing

Support for fast tracing has been built into Genode for a long time. However, the stakes to take advantage of this feature remained high because convenience functions were not in place. With the current release, we added the support for easy trace setups through a VFS plugin. The plugin is called vfs_trace and can be mounted into a Genode component as follows:

<config>
  <vfs>
    <trace ram=32MB/>
  </vfs>
</config>

This configuration will create a trace file system at the root of the VFS. The ram attribute is mandatory and determines the maximum size of all trace buffers. The file system forms a recursive directory structure that represents the parent/child relationship of running components, whereas the leaf directories represent single threads within a component. Each leaf directory currently contains three files:

enable

Start or stop the tracing of a thread by writing "true" or "false" into the file.

buffer_size

Allows for the configuration of the trace-buffer size for the thread in the usual Genode format (e.g. 5M, 512K, 1024).

trace_buffer

This read-only file contains the current content of the trace buffer. Each trace entry can only be read once, after that only new entries appear. "tail -f" can also be used to display continuous output.

As an example, tracing is started by writing true to the enable file:

 echo "true" > enable

The trace buffer can then be displayed using Unix tools like tail

 tail -f trace_buffer

which provides a continuous output.

Additionally, we have added the trace function to base/log.h that facilitates identical functionality as Genode::log

 Genode::trace("Tracepoint value: ", value);

In order to enable tracing, the parent must provide the "TRACE" service. For a real world example on Sculpt OS, please refer to this Genodians article.

With the vfs_trace plugin in place, we removed the outdated trace_fs.

Consolidation of the C runtime and Noux

On our road map, we vaguely hinted at our plan for the "consolidation" of the noux runtime, which is actually meant as a polite way of announcing that we are going to remove it. We introduced the Noux runtime in 2011 as a way to execute command-line-based GNU software directly on Genode. It has served us well over the years and is - in fact - a crucial ingredient of Sculpt OS and other system scenarios such as the Genodians.org web server. Noux supplements Genode with two valuable assets, namely a flexible and expandable virtual file system (VFS) layer, and the implementation of the Unix way to spawn applications (fork and execve).

In the meantime, noux' VFS implementation has become independent from the noux runtime and is now prominently employed by Genode's C runtime and the VFS server component. Genode's C runtime became more and more complete, alleviating the use of noux as POSIX compatibility layer except for programs that depended on a working implementation of fork and execve.

The current release fills this remaining gap in Genode's C runtime by providing fork, execve, and cousins such as wait4 and getpid as regular parts of the libc. This will eventually make noux redundant.

Note that this change does NOT make Genode reliant on POSIX. The C runtime including the Unix features are entirely optional.

As one stepping stone of this undertaking, noux applications, which previously had to be compiled for noux, have become binary compatible with the regular C runtime. So one can execute programs like bash directly as a Genode component without any friction.

There are a few collateral improvements of Genode's dynamic linker and the C runtime on the account of the new fork and execve implementation. E.g., in addition to the already supported stdin, stdout, and stderr configuration, the libc can be instructed to initialize arbitrary file descriptors as follows:

 <config>
   ...
   <libc ...>
     <fd id="3" path="/dev/log" writeable="yes" readable="no" seek="10"/>
     ...
   </libc>
 </config>

The libc-based implementation of fork and execve can be tried out via the new ports/run/bash.run script. Note that there are still a number of limitations such as the lack of signal and ioctl handling. Pipes are not supported, and shebangs (#!) are not interpreted yet. That said, once those missing pieces come into place, we can fade out the use of noux within Genode.

General system time concept

Briefly speaking, up to now there has been no notion of an overall concept of system time in Genode. Components that need to have access to some kind of real time are either configured locally, e.g., libc-based components access a configured "device" (/dev/rtc), which just might be an inline file system containing an artificial timestamp or the VFS RTC plugin, while other components query some RTC session directly. Most of the time, this session is provided by the rtc_drv on x86 machines, which is somewhat costly as reading the RTC via I/O ports takes time and is therefore done scarcely. For example, the libc will query an RTC source only once and uses this initial value to interpolate the current time. However, for executing long-running components, it will be necessary to adjust the clock to compensate for any occurring clock drift or to correct a misconfigured clock in general. In addition it is desirable to be able to use a remote time source, e.g., an NTP-server, to synchronize the system time.

To address this, we came up with the following concept:

The new "System RTC" component, located at repos/libports/src/server/system_rtc, acts as proxy for the RTC service in front of the actual RTC driver. It uses the driver to get the initial RTC value and then uses a timer session (via the timeout framework) to locally interpolate the time. In contrast to querying the RTC driver, querying the System RTC is fast.

The RTC driver and the System RTC are bundled up together in the new drivers-rtc-pc package. The runtime of this package requests two ROM modules used to update the RTC value. The first one, named system_set_rtc, is used to update the proxy component while the second one, called hw_set_rtc, is used by the RTC driver to write the value into the battery-backed RTC. A separate component, potentially accessing a remote time source, may generate these ROMs to adjust the time in the package's runtime.

The new native SNTP client at repos/libports/src/app/sntp_client is such a component. It periodically requests the current time from a given SNTP server and generates a report. The report produced by the component contains the time as UTC/GMT. Depending on the system policy, it can be used to update the time of the System RTC and/or instruct the driver to set the RTC value.

To propagate such changes to RTC values, the RTC session was enhanced by the new set signal. A client of the session can install a signal handler to adapt its own time when necessary. Based on this, the time back end of the libc was changed to instantiate a watch handler for the RTC device, which, when triggered, will cause the libc to re-read the RTC value.

This constellation should, under normal operation, allow for second to sub-second granularity updates of the overall system time and avoid drifting away from network time.

Accessing SMBIOS tables

The System Management BIOS (SMBIOS) is a specification that allows for reading management information produced by the BIOS of a system as a collection of data structures in memory. It has the potential to eliminate the need for the operating system to probe hardware for discovering present devices and their characteristics. Nowadays, the SMBIOS specification is implemented widely in PC systems, which includes modern UEFI systems as well. The data structures are referred to as tables or records by public documentation.

The new native SMBIOS decoder at os/src/app/smbios_decoder can be used on x86 to parse SMBIOS tables and report gathered information in a human-readable way. Besides general table information like number and size of structures, etc., the component supports complete parsing of SMBIOS structures of types "BIOS", "System", and "Baseboard".

The component is free from any code for acquiring an SMBIOS table through means like the bootloader or BIOS information. It expects a table to be present through a regular Genode ROM session with a smbios_table label. This way, the underlying system is required to find, select, and save the raw table on startup and create a ROM module out of it. This is currently achieved on NOVA and base-hw through an interplay of kernel, the core component, and the ACPI driver and was tested for legacy BIOSes as well as UEFI systems.

Clipboard

Genode introduced a principle copy-and-paste mechanism already four years ago. However, originally created as a part of a tech demo, the mechanism remained unused in our day to day Genode work. This changed now. We took the integration of copy-and-paste support in Sculpt OS as an opportunity to revive and refine the existing mechanism and supplement it with the features needed to make it practical for daily use. We believe that the result aligns ease of use nicely with security. The concept is described in a dedicated article at Genodians.org.

On a technical level, the existing clipboard component has received a new option that allows for dynamic information-flow policies based on user interactivity (keyboard focus, activity). When setting the config attribute match_labels="yes", the clipboard performs plausibility checks for copy and paste operations against the focus of the Nitpicker GUI server. All aspects of the clipboard policy - including information-flow policies - have become reconfigurable.

To make window-manager clients compatible with the clipboard's dynamic policy, the window manager got enhanced with the ability to proxy the interaction with the clipboard. GUI clients in turn - in particular the graphical terminal - became able to interact with the clipboard. With the <config> attribute copy="yes" specified, the terminal allows the user to select text to be reported to a "clipboard" report. The selection mode is activated by holding the left shift key. While the selection mode is active, the text position under the mouse pointer is highlighted and the user can select text via the left mouse button. Upon release of the mouse button, the selection is reported. Vice versa, with the <config> attribute paste="yes" specified, the terminal allows the user to paste the content of a "clipboard" ROM session to the terminal client by pressing the middle mouse button.

Finally, we integrated those new abilities into Sculpt OS and into several installable packages, including virtual machines, the noux-system package, and graphical Qt5-based applications.

Enhanced SSH terminal

This release paves the way for remotely managing Genode devices over SSH. Until now, only interactive SSH sessions were supported. It is now possible to execute commands from a remote SSH client. E.g., 'ssh noux@localhost -p 5555 "ls -hal /bin/"'. For non-interactive sessions, ssh_terminal requires a helper component. This component is responsible to create the environment for the command to run in. You can find an example for such a component at gems/src/test/exec_terminal. It starts noux in a sub init and executes the provided command inside of it. The new ssh_exec_channel.run script gives a demonstration on how this feature can be used.

This work is a contribution by Sid Hussmann of Gapfruit. Thanks for this great new feature!

Storage-stack improvements

The desire of one Genode developer to exchange data via Iomega ZIP drives between an Atari Falcon and Sculpt OS called for a number of small improvements across several components of the storage stack.

First, the USB-block driver has been changed to exit on an initialization failure instead of waiting for another (supported) device. This change enables the Sculpt manager to detect such conditions and release the USB device hardware by removing the driver component. Such a failed initialization may happen with exotic USB-storage devices such as ZIP drives. With the device released, however, it can be assigned to a virtual machine to access it using a guest OS with a broader support of devices.

Second, the USB-block driver received new support for issuing the SCSI START-STOP command at initialization time, thereby overcoming the ZIP-drive initialization failure.

Third, we enhanced the part-block component with the ability to parse AHDI partition schemes and detect the GEMDOS variant of FAT as used by Atari TOS.

Fourth, we enabled the Rump VFS plugin to access GEMDOS file systems. The GEMDOS variant is readily supported by NetBSD's "msdos" file-system driver. However, it must explicitly be enabled by a mount flag. Hence, we added the principle ability for passing mount flags to NetBSD file-system drivers and enabled the MSDOSFSMNT_GEMDOSFS flag based on the VFS plugin's config attribute gemdos="yes".

With these changes in place, data can now be exchanged directly between Atari-formatted disks and Sculpt OS. That said, advanced use cases such as media changes at runtime are not covered yet.

Updated Ada/SPARK runtime

Genode's Ada/SPARK runtime is developed and maintained by Componolit. Thanks for this excellent collaboration!

The updated Componolit Ada runtime 1.1.0 increases the proof coverage and cleans up the source-code structure. SPARK mode is now enabled wherever possible and unneeded abstractions have been removed. Furthermore, the 64-bit addition and subtraction have been proven to be free of runtime errors. As a new feature, the runtime now supports the use of inline assembly in Ada.

The removal of unneeded features such as the incomplete threading support for the secondary stack has greatly reduced the runtime's complexity while keeping the current functionality available. Also GNAT.IO has been removed as its implementation was incomplete and complex. A simpler replacement has been introduced with Componolit.Runtime.Debug.

Unrelated to Genode, the runtime now supports Muen and the API/ABI of the runtime has been separated from the GNAT ABI.

Libraries and applications

Updated Qt5

We updated our Qt5 port to the latest upstream version 5.13.0. Before preparing the qt5 port, please make sure to build and install the updated Qt5 host tools with the tool/tool_chain_qt5 script.

Virtualization

As follow-up of our work on the kernel agnostic virtual-machine monitor interface on x86, we added principle support to run our port of VirtualBox on Genode/Fiasco.OC. We write principle support, since we managed to get the VMM running with Fiasco.OC, but unfortunately not all features required by the VMM are available using the Fiasco.OC kernel, e.g., guest FPU registers, PDPTE registers, and task-priority support. In practice this means that the VMs with Windows and Linux come up to a certain point but will fail later whenever the guest state runs out of synchronization between VMM and hardware. In contrast, the Seoul VMM runs fine on Fiasco.OC since it does not depend on the mentioned missing features.

Our main working items have been the completion of transfer of the available guest registers and control flow synchronization improvements between VMM and Fiasco.OC kernel. Additionally, the usage of priorities for VirtualBox's pthreads in the VMM had to be disabled. Finally, some tests for VirtualBox with Genode/Fiasco.OC are enabled for nightly regular testing now.

As a side topic, we added support for using the VirtualBox CPU profile feature, which allows for presenting a different CPUID to the VM than the one of the real CPU. This can help when running Windows 7 on a Kaby Lake or newer CPU, which are considered unsupported hardware and reason enough not to receive security updates from Microsoft. The feature can be used on Genode by adding the CpuProfile attribute to the <CPU> XML node in the .vbox file, like:

 <CPU CpuProfile="Intel Core i7-5600U">

Disposable VM for handling captive portals

It is common that WiFi networks require the user to interact with a specific web page before gaining access to full network functionality. Such captive portal pages are completely individual to the accessed network and not limited in the use of common web techniques. Therefore, their handling is best be done using a fully-featured web browser like Mozilla Firefox.

This is where, in a Genode-based desktop system like Sculpt, a disposable VM for hosting a minimal browser setup becomes desirable. Its goal is to unlock a network for the native Genode surroundings with as little inconvenience as possible just to be thrown away afterwards without any side effects on the system.

Now, one could use the Firefox appliance VM of Sculpt (see the release notes or the Genodians article) for this. But this VM aims for a long-term browsing experience which, in the context of mere captive-portal handling, brings some drawbacks like a much higher RAM consumption or the required sessions for USB detection and shared folders.

Furthermore, in the captive portal VM, there's no need for managing windows or browser tabs. The one browser tab needed can always be shown in fullscreen. It is also unnecessary for the browser to maintain a content cache or remember user data. This can reduce resource consumption.

The VM we came up with is provided as package for Sculpt by Martin Stein (depot user mstein). You'll possibly need to manually add Martin's depot key and download location to your Sculpt depot directory. After enabling this user, the captive portal VM can be found in the Sculpt menu under "Depot -> mstein -> Virtual Machines -> vbox5-nova-captive-portal".

The VM is based on a TinyCore 10 Linux with Xserver, i3 WM, and a tailored Firefox browser. The package runtime doesn't need access to your file system, it merely loads some ROMs into a RAM FS, so, it will completely forget any changes made during a session. Therefore, it's also safe to simply remove an instance via the Leitzentrale component-view once you don't need it anymore. The guest additions are also included to make the VM window resizable.

Build system and tools

At Genode Labs, we have used tool/autopilot for the steering of tests in our Continuous Integration workflow for almost a decade now. This implied various improvements over the years and with the completion of our work on unified build directories it was time to amend this handy tool once again. Unified build directories support building all components for one CPU architecture in one directory saving the build server from the redundant work we previously had with board-specific directories. With the new notion of boards during builds, the definition of the target platform when integrating Genode system scenarios is now a triplet of CPU architecture, board, and kernel. This is reflected in the new -t <architecture-board-kernel> command line option, which instructs autopilot to generate a build directory for architecture and execute tests for the board-kernel combination.

 autopilot -t x86_64-pc-sel4 -t x86_64-pc-nova -r run/log

The known options for -k kernel and -p platform are still supported with the small change that the platform must now be defined as architecture-board.

 autopilot -p x86_64-pc -k sel4 -k nova -r run/log

Autopilot now also documents the hidden feature to propagate custom RUN_OPTs via the RUN_OPT_AUTOPILOT environment variable to the run tool executed. Besides that, the tool always appends RUN_OPT with –autopilot.

 RUN_OPT_AUTOPILOT="--depot-dir /data/depot" autopilot ...