From Wikipedia:
"An anti-pattern is a common response to a recurring problem that is usually ineffective and risks being highly counterproductive."
The developers of OpenBMC do not get 100% of decisions right 100% of the time. That, combined with the fact that software development is often an exercise in copying and pasting, results in mistakes happening over and over again.
This page aims to document some of the anti-patterns that exist in OpenBMC to ease the job of those reviewing code. If an anti-pattern is spotted, rather that repeating the same explanations over and over, a link to this document can be provided.
(1 paragraph) Describe how to spot the anti-pattern.
(1 paragraph) Describe the negative effects of the anti-pattern.
(1 paragraph) Describe why the anti-pattern exists. If you don't know, try running git blame and look at who wrote the code originally, and ask them on the mailing list or in Discord what their original intent was, so it can be documented here (and you may possibly discover it isn't as much of an anti-pattern as you thought). If you are unable to determine why the anti-pattern exists, put: "Unknown" here.
(1 paragraph) Describe the preferred way to solve the problem solved by the anti-pattern and the positive effects of solving it in the manner described.
The ArgumentParser object is typically present to wrap calls to get options. It abstracts away the parsing and provides a [] operator to access the parameters.
Writing a custom ArgumentParser object creates nearly duplicate code in a repository. The ArgumentParser itself is the same, however, the options provided differ. Writing a custom argument parser re-invents the wheel on c++ command line argument parsing.
The ArgumentParser exists because it was in place early and then copied into each new repository as an easy way to handle argument parsing.
The CLI11 library was designed and implemented specifically to support modern argument parsing. It handles the cases seen in OpenBMC daemons and has some handy built-in validators, and allows easy customizations to validation.
PKG_CHECK_MODULES(
[PHOSPHOR_LOGGING],
[phosphor-logging],
[],
[AC_MSG_ERROR([Could not find phosphor-logging...openbmc/phosphor-logging package required])])
The autotools PKG_CHECK_MODULES macro provides the ability to specify an "if found" and "if not found" behavior. By default, the "if not found" behavior will list the package not found. In many cases, this is sufficient to a developer to know what package is missing. In most cases, it's another OpenBMC package.
If the library sought's name isn't related to the package providing it, then the failure message should be set to something more useful to the developer.
Use the default macro behavior when it is clear that the missing package is another OpenBMC package.
PKG_CHECK_MODULES([PHOSPHOR_LOGGING], [phosphor-logging])
RDEPENDS_${PN} = "libsystemd"
Out of the box bitbake examines built applications, automatically adds runtime dependencies and thus ensures any library packages dependencies are automatically added to images, sdks, etc. There is no need to list them explicitly in a recipe.
Dependencies change over time, and listing them explicitly is likely prone to errors - the net effect being unnecessary shared library packages being installed into images.
Consult https://www.yoctoproject.org/docs/latest/mega-manual/mega-manual.html#var-RDEPENDS for information on when to use explicit runtime dependencies.
The initial bitbake metadata author for OpenBMC was not aware that bitbake added these dependencies automatically. Initial bitbake metadata therefore listed shared library dependencies explicitly, and was subsequently copy pasted.
Do not list shared library packages in RDEPENDS. This eliminates the possibility of installing unnecessary shared library packages due to unmaintained library dependency lists in bitbake metadata.
In systemd unit files:
[Service] ExecStart=/usr/bin/env some-application
Outside of OpenBMC, most applications that provide systemd unit files don't launch applications in this way. So if nothing else, this just looks strange and violates the princple of least astonishment.
This anti-pattern exists because a requirement exists to enable live patching of applications on read-only filesystems. Launching applications in this way was part of the implementation that satisfied the live patch requirement. For example:
/usr/bin/phosphor-hwmon
on a read-only filesystem becomes:
/usr/local/bin/phosphor-hwmon`
on a writeable /usr/local filesystem.
The /usr/bin/env method only enables live patching of applications. A method that supports live patching of any file in the read-only filesystem has emerged. Assuming a writeable filesystem exists somewhere on the bmc, something like:
mkdir -p /var/persist/usr mkdir -p /var/persist/work/usr mount -t overlay -o lowerdir=/usr,upperdir=/var/persist/usr,workdir=/var/persist/work/usr overlay /usr
can enable live system patching without any additional requirements on how applications are launched from systemd service files. This is the preferred method for enabling live system patching as it allows OpenBMC developers to write systemd service files in the same way as most other projects.
To undo existing instances of this anti-pattern remove /usr/bin/env from systemd service files and replace with the fully qualified path to the application being launched. For example, given an application in /usr/bin:
sed -i s,/usr/bin/env ,/usr/bin/, foo.service
OpenBMC executables that are installed in /usr/sbin. $sbindir in bitbake metadata, makefiles or configure scripts. systemd service files pointing to /bin or /usr/bin executables.
Installing OpenBMC applications in incorrect locations violates the principle of least astonishment and more importantly violates the FHS.
There are typically two types of executables:
Executables of type 1 should not be placed anywhere in ${PATH} because it is confusing and error-prone to users, but should instead be placed in a /usr/libexec/<package> subdirectory.
Executables of type 2 should be placed in /usr/bin because they are intended to be used by users and should be in ${PATH} (also, sbin is inappropriate as we transition to having non-root access).
The sbin anti-pattern exists because the FHS was misinterpreted at an early point in OpenBMC project history, and has proliferated ever since.
From the hier(7) man page:
/usr/bin This is the primary directory for executable programs. Most programs executed by normal users which are not needed for booting or for repairing the system and which are not installed locally should be placed in this directory. /usr/sbin This directory contains program binaries for system administration which are not essential for the boot process, for mounting /usr, or for system repair. /usr/libexec Directory contains binaries for internal use only and they are not meant to be executed directly by users shell or scripts.
The FHS description for /usr/sbin refers to "system administration" but the de-facto interpretation of the system being administered refers to the OS installation and not a system in the OpenBMC sense of managed system. As such OpenBMC applications should be installed in /usr/bin.
It is becoming common practice in Linux for daemons to now be moved to libexec and considered "internal use" from the perspective of the systemd service file that controls their launch.
Install OpenBMC applications in /usr/libexec or /usr/bin/ as appropriate.
The anti-pattern is for an application to continue processing after it encounters unexpected conditions in the form of error codes and exceptions and not capturing the data needed to resolve the problem.
Example C++ code:
using InternalFailure = sdbusplus::xyz::openbmc_project::Common::Error::InternalFailure;
try
{
... use d-Bus APIs...
}
catch (InternalFailure& e)
{
phosphor::logging::commit<InternalFailure>();
}
Suppressing unexpected errors can lead an application to incorrect or erratic behavior which can affect the service it provides and the overall system.
Writing a log entry instead of a core dump may not give enough information to debug a truly unexpected problem, so developers do not get a chance to investigate problems and the system's reliability does not improve over time.
Programmers want their application to continue processing when it encounters conditions it can recover from. Sometimes they try too hard and continue when it is not appropriate.
Programmers also want to log what is happening in the application, so they write log entries that give debug data when something goes wrong, typically targeted for their use. They don't consider how the log entry is consumed by the BMC administrator or automated service tools.
The InternalFailure in the Phosphor logging README is overused.
Several items are needed:
std::system_error.In the error handler:
Consider what data (if any) should be logged. Determine who will consume the log entry: BMC developers, administrator, or an analysis tool. Usually the answer is more than one of these.
The following example log entries use an IPMI request to set network access on, but the user input is invalid.
BMC Developer: Reference internal applications, services, pids, etc. the developer would be familiar with.
Example: ipmid service successfully processed a network setting packet, however the user input of USB0 is not a valid network interface to configure.
Administrator: Reference the external interfaces of the BMC such as the REST API. They can respond to feedback about invalid input, or a need to restart the BMC.
Example: The network interface of USB0 is not a valid option. Retry the IPMI command with a valid interface.
Analyzer tool: Consider breaking the log down and including several properties which an analyzer can leverage. For instance, tagging the log with 'Internal' is not helpful. However, breaking that down into something like [UserInput][ipmi][Network] tells at a quick glance that the input received for configuring the network via an IPMI command was invalid. Categorization and system impact are key things to focus on when creating logs for an analysis application.
Example: [UserInput][IPMI][Network][Config][Warning] Interface USB0 not valid.
Determine if the application can fully recover from the condition. If not, don't continue. Allow the system to determine if it writes a core dump or restarts the service. If there are severe impacts when the service fails, consider using a better error recovery mechanism.
An application uses non-standard methods on startup to indicate verbose logging and/or does not utilize standard systemd-journald debug levels for logging.
When debugging issues within OpenBMC that cross multiple applications, it's very difficult to enable the appropriate debug when different applications have different mechanisms for enabling debug. For example, different OpenBMC applications take the following as command line parameters to enable extra debug:
Along these same lines, some applications then have their own internal methods of writing debug data. They use std::cout, printf, fprintf, ... Doing this causes other issues when trying to compare debug data across different applications that may be having their buffers flushed at different times (and picked up by journald).
Everyone has their own ways to debug. There was no real standard within OpenBMC on how to do it so everyone came up with whatever they were familiar with.
If an OpenBMC application is to support enhanced debug via a command line then it will support the standard "-v,--verbose" option.
In general, OpenBMC developers should utilize the "debug" journald level for debug messages. This can be enabled/disabled globally and would apply to all applications. If a developer believes this would cause too much debug data in certain cases then they can protect these journald debug calls around a --verbose command line option.
Desire to expose a DBus interface to drive GPIOs, for example:
Platform functionality selected by GPIOs might equally be selected by other means with a shift in system design philosophy. As such, GPIOs are a (hardware) implementation detail. Exposing the implementation on DBus forces the distribution of platform design details across multiple applications, which violates the software design principle of low coupling and impacts our confidence in maintenance.
GPIOs are often used to select functionality that might be hard to generalise, and therefore hard to abstract. If this is the case, then the functionality in question is probably best wrapped up as an implementation detail of a behaviour that is more generally applicable (e.g. host power-on procedures).
Consider what functionality the GPIO provides and design or exploit an existing interface to expose that behaviour instead.
C++ code that is similar to the following:
dbus::utility::getSubTree("/", interfaces, [asyncResp](boost::system::error_code& ec, MapperGetSubTreeResult& res){ <too many lines of code> })
Inline lambdas, while useful in some contexts, are difficult to read, and have inconsistent formatting with tools like clang-format, which causes significant problems in review, and in applying patchsets that might have minor conflicts. In addition, because they are declared in a function scope, they are difficult to unit test, and produce log messages that are difficult to read given their unnamed nature. They are also difficult to debug, lacking any symbol name to which to attach breakpoints or tracepoints.
Lambdas are a natural approach to implementing callbacks in the context of asynchronous IO. Further, the Boost and sdbusplus ASIO APIs tend to encourage this approach. Doing something other than lambdas requires more effort, and so these options tend not to be chosen without pressure to do so.
Prefer to either use std::bind_front and a method or static function to handle the return, or a lambda that is less than 10 lines of code to handle an error inline. In cases where std::bind_front cannot be used, such as in sdbusplus::asio::connection::async_method_call, keep the lambda length less than 10 lines, and call the appropriate function for handling non-trivial transforms.
void afterGetSubTree(std::shared_ptr<bmcweb::AsyncResp>& asyncResp, boost::system::error_code& ec, MapperGetSubTreeResult& res){ <many lines of code> } dbus::utility::getSubTree("/xyz/openbmc_project/inventory", interfaces, std::bind_front(afterGetSubTree, asyncResp));
See also the Cpp Core Guidelines for generalized guidelines on when lambdas are appropriate. The above recommendation is aligned with the Cpp Core Guidelines.
Declaring types and functions that do not participate in a public API of a library in header files that do not live alongside their implementation in the source tree.
There's no reason to put internal headers in a parallel subtree (for example, include/). It's more effort to organise, puts unnecessary distance between declarations and definitions, and increases effort required in the build system configuration.
In C and C++, header files expose the public API of a library to its consumers and need to be installed onto the developer's system in a location known to the compiler. To delineate which header files are to be installed and which are not, the public header files are often placed into an include/ subdirectory of the source tree to mark their importance.
Any functions or structures that are implementation details of the library should not be provided in its installed header files. Ignoring this philosphy over-exposes the library's design and may lead to otherwise unnecessary API or ABI breaks in the future.
Further, projects whose artifacts are only application binaries have no public API or ABI in the sense of a library. Any header files in the source tree of such projects have no need to be installed onto a developer's system and segregation in path from the implementation serves no purpose.
Place internal header files immediately alongside source files containing the corresponding implementation.