This section of the developer's guide introduces a basic workflow for developing and working on the illumos code base. This workflow is intended and designed for those who have never worked with illumos before, though have some basic experience building programs written in the C programming language. It will take you from obtaining a copy of the illumos source tree all the way through to getting it ready for putback.
To set up and initialize the build, see the Building illumos instructions that are part of the normal docs.
As you work on illumos, builds will fail. When a nightly completes it leaves
behind two different files that describe what happened inside of the build. The
first is the nightly.log. This file contains a record of all the commands that
were executed. If your build failed, it should contain the specifics of
everything that happened. The other file is the mail_msg. This summarizes the
different phases of the build, how long they took, and any warnings or noise
encountered. It should be clean.
If a build fails, the first stop is usually the nightly.log. Every run of
nightly leaves a log in its own directory based on the time stamp of the build.
A given nightly.log may contain multiple errors. nightly tries as hard as
possible to run to the end even if there is a build failure. Because of that,
it's important to make sure you find the first error. Errors usually have the
form: *** Error code 1. The following shows how to find the logs and errors
inside of them:
illumos Developer's Guide
$ cd /ws/rm/illumos-gate
$ cd log
$ ls
log.2013-04-26.16:56 log.2013-05-01.17:02 log.2013-05-01.19:05
$ cd log.2013-05-01.19:05
$ ls
mail_msg nightly.log proto_list_i386
$ vi nightly.log
# Search for the first instance of the pattern \*\*\*
Note that if your build failed you will not generally see the file
proto_list_i386.
bldenv(1ONBLD) is a tool designed to get you into a development environment for
doing incremental build work. Once you've done your first nightly it's the
only way to work. To get started, you use the same nightly environment file
that we previously set up. Let's go ahead and use bldenv:
$ cd /ws/rm/illumos-gate
$ /opt/onbld/bin/bldenv -d ./illumos.sh
Build type is DEBUG
RELEASE is
VERSION is illumos-gate
RELEASE_DATE is May 2013
The top-level 'setup' target is available to build headers and tools.
Using /bin/bash as shell.
$
Because we had started with a debug build in nightly, we now use the -d flag
to bldenv to make sure that we continue in a debug build environment. If for
some reason we had done a normal build, we could omit the -d argument.
Before we begin building, it's rather useful to build the cscope database.
cscope is a tool which allows you to easily search the source tree for
definitions, function call locations, symbol usage, variable assignment, and
more. It can speed up the process of what might normally be a find and recursive
grep quite substantially. cscope is built and provided as a part of the build
tools. While in bldenv, it will already be in your path as a program called
cscope-fast. From the root of the tree, you can create the database and use it
via:
$ cd usr/src
$ dmake cscope.out
$ cscope-fast -dq
Various editors have integration with cscope which makes keeping around a
cscope database quite useful and makes it easier to follow code paths across the
code base.
As different components are being built, they are delivered into what we call
the proto area. Everything that you're going to want to use and test gets
installed into the proto area before it gets packaged for updating a system.
Furthermore, as you build, components look to the proto area to find what they
need. For example, all of the header files like <sys/types.h> or <stdio.h>
are installed into the proto area. When you build a command that uses one of
those header files, it uses the copy from the proto area.
You might ask yourself, why do we have the proto area. The problem that the proto area solves is that the system that you're building on might look rather different from the one that it is going to end up on. If you're adding a new interface to a library and updating a command to use that new interface, you can't use the build system's copy of that header, as it wouldn't contain the new function. The build system might not even have that header or library installed!
Furthermore, there are occasionally libraries and headers that are used to help
build the system, but are not actually shipped. For example, libpcidb.so.1
does not have a compilation symlink delivered to systems, but it does have one
in the proto area to allow commands like prtconf(8) to link against it.
One of the first questions that you have to answer is what is it I have to modify? Generally, you're working on a specific component of the kernel, some library, command, or something which spans across all three of these boundaries. Here are some questions to help guide you and help you figure out what exactly you'll need to build.
If you're changing a header file, you'll need to install the new copy into the proto area before you rebuild a library, command, or kernel module which uses it.
Sometimes when you're working with a library, you're working on bug fixes and enhancements that don't change the libraries API. However, if you're adding new functionality to it, you'll need to build and install it into the proto area before you can rebuild other libraries or commands that need it.
Here, there are multiple components that need to be modified. You need to first install the header file that both the kernel and userland share and install that into the proto area. Once that's done, it doesn't matter what order you modify the userland and kernel components, as long as they all end up in the proto area before you move onto testing.
Across most components there is a common set of three targets that are the most common and most used.
This will install the component into the proto area. It also ensures that the build is up to date first. This is probably the target that you'll run the most.
This is a special target for installing header files specifically. Generally header files need to be installed before you can use the install.
This removes all of the intermediate object files and the final components. This is useful if you want to just do a fresh, non-incremental, build of a given component.
Building userland libraries and commands is pretty straightforward. All you have
to do is go into the directory of a command and run dmake install. This will
take care of building anything that's changed, link the resulting binary or
library, and install it into the proto area.
The common workflow here is to modify the source code and compile it via dmake
install. If you're warning and error free, then you'll see the new version be
installed into the proto area. However, often times you'll run into compiler
warnings and errors. This is the time and place to fix those, don't ignore
warnings! By default, dmake builds in parallel mode, so it will try to run
multiple jobs in parallel. If you're hitting errors and it's hard to understand,
instead, use dmake -m serial install. The -m serial instructs it to only
build one thing at a time.
Here's an example of building a command:
$ # We're sitting at what we consider the source tree root, inside usr/src.
$ cd cmd/dtrace
$ vi dtrace.c
$ dmake install
$ # The new copy will now be in the proto area.
$ cd ../..
$ cd lib/libctf/
$ vi common/ctf_subr.c
$ dmake install
$ # The new copy will now be in the proto area.
Most libraries and some commands have both 32-bit and 64-bit versions. If you want to build just a specific version, you can cd into the architecture specific directory. Note that not all commands have multiple versions, some only have a single one. Say you wanted to build the 64-bit version of libumem. You would then do:
$ cd lib/libumem/amd64/
$ dmake install
$
Unless you have a specific need, you should just stick with the top level one that takes care of all the architectures.
For header files, you'll want to run dmake install_h in the appropriate
directory. For libraries, this is in the root of that library. For the kernel,
the easiest way to do this is at the root of the uts directory. In action,
this looks like:
$ cd lib/libkstat
$ vi kstat.h
$ dmake install_h
$
$ cd uts
$ vi common/sys/mac_client.h
$ vi intel/sys/cpu.h
$ dmake install_h
$
Unlike libraries and commands, the build information for these components is
kept in a different directory from the source. The majority of the kernel is in
uts/common/, while the minority is in architecture and platform specific
directories. Take for example, the driver igb(4D). The source for it lives
in uts/common/io/igb/. However, to build it you need to be in the architecture
specific directory, either intel or sparc currently. Once in there, you can
go ahead and use the same dmake install techniques that we've been using
elsewhere.
$ vi uts/common/io/igb/igb_mac.c
$ cd uts/intel/igb
$ dmake install
$
If you find yourself in doubt as to what you should be doing or you're worried
about the dependency chain, you can always just run nightly again or run dmake
install from a higher top level directory. For example, running dmake install
from inside of lib/ will make sure that every library is up to date. If you're
having trouble knowing how to build a specific component, you can look at the
Makefile in its directory, trudge through the nightly log file, or ask
around.
While writing the code is important, making sure that it is correct is even more
important. Answering the question of what testing is necessary is hard. It
depends on the component and the nature of the change. Some components have unit
test in the tree already such as zfs, mdb, and dtrace, while others do
not.
Testing should be shrink to fit. The question to ask yourself, is what testing do you need to convince not only yourself, but someone else - such as your code reviewers or integration advocate - that your change is correct. The rest of this section will cover strategies for using your new versions of the software, often referred to as 'using new bits'. Which strategies you can use depends on the nature of the changes that you've made.
Keep track of your testing. While you may not integrate your tests along side your changes, you'll need to note how you tested when you integrate your changes. Providing and making those tests available if appropriate, is always preferred.
One of the easiest ways to test your changes is to boot a test system onto your new bits. No matter what kind of change you've made, this should work. There are various ways to do this based on your distribution; however, for most folks this can be achieved via the build tool onu(1ONBLD). If you're inside of bldenv and have finished your build, the simplest way to do this is:
# onu -t new-nightly-be
#
Once this is done, you can reboot into the new build environment. You'll want to consult with your distribution for the best way to do this.
Once you have your bits installed, you can run any tests you want against a host machine.
Sometimes one chooses to test something straight from the build machine itself.
For example, if you've modified head(1) or grep(1), that might be alright.
However, this can be dangerous and lead you to wasting your time. Recall are
comments on why we use the proto area, well, now all of those can happen, but
flipped around.
If you have just modified a binary, then you may be able to use that directly from the proto area:
$ cd /ws/rm/illumos-gate/proto/root.i386-debug/root/usr/bin
$ ./head /var/tmp/foo
$
If you have modified a library and want to test a command against it, you can
take advantage of the LD_LIBRARY_PATH, LD_LIBRARY_PATH_32,
LD_LIBRARY_PATH_64, LD_PRELOAD_32, LD_PRELOAD_64, and LD_PRELOAD
environment variables to make sure that you use the new version of the library.
For the full usage of these, see the ENVIRONMENT VARIABLES section of
ld.so.1(1). You have to exercise caution when using these. Your
libc.so.1 and kernel must always match! If they don't, a lot can manage to go
wrong and be very hard to debug. It is always safest to just boot your new bits!
Depending on the nature of your change, it can be useful to understand the differences between binaries that your changes introduced. Checking the differences of the produced binaries between changes can verify that you changed what you intended to. For example, let's say that you have removed what you believe is unused code from a subsystem. If the code was indeed unused, then the binary should be almost identical to the binary from the previous build. Another use is to verify that your change did not unintentionally affect another part of the codebase.
illumos provides a tool that can verify this automatically:
wsdiff(1ONBLD). To run wsdiff,
specify -w in your NIGHTLY_OPTIONS. For wsdiff to work, it needs to be run
between two consecutive builds. Therefore, the common workflow is to first build
the gate with nightly, make your changes, and then run nightly again, but
this time with wsdiff enabled.
wsdiff isn't perfect. There are a few things can cause noise to show up in its
output. For example, the illumos DEBUG builds often carry macros like
__LINE__ that emit line numbers and DTrace probes will often have a part of
the DOF section change which encodes information about how it ws generated. As a
result, it is expected that some differences that have to do with changes in the
actual source files will pop up in such builds. You may find that wsdiff is
easier to use with non-DEBUG builds.
As you make progress, you should feel free to make incremental commits in your
local repository. This is one of the great advantages of the bring over, modify,
merge strategy: you can make your own local commits and then later on rewrite
that history. You should make commits whenever you feel like it. Think of them
as a way to track your own progress and provide snapshots that you can roll back
to in case anything goes awry. You can commit groups of files or the entire
tree. The following examples explore how to do this with git:
$ # Commit only a select few files
$ git commit cmd/mdb/common/mdb/mdb_typedef.c
$ # Commit all changes
$ git commit -a
$
If you've added files, you'll want to make sure that you've added them to the
git repository. Say you wrote the library librm. Assuming that you don't have
any build files present, you should could run the following:
$ git add -n lib/librm
$ # Check the above dry run and verify it has everything you expect
$ git add lib/librm
$ git commit -a
$
If you make commits that are in awkward places in time or have a slightly incorrect message, don't worry. Before you integrate you'll squash all of these changes down into one commit and have the chance to fix that up.
You've built your changes, made some local commits, and tested them. The hard part is done! Before you can get the change integrated there are a few steps that need to be done. These steps insure that the gate has consistent style and helps with review.
There are lots of ways that you can share your changes with other people. The preferred format by the illumos community is using the online tool gerrit at code.illumos.org. This tool allows folks to provide comments and makes it easy to see what changes between releases. For more information on getting started with gerrit see Code Review with Gerrit. It will use your ssh key that is in the bug tracker for authentication. Briefly, you will push your code to a specially named branch, which will create the review.
git push ssh://your-username@code.illumos.org/illumos-gate your-branch:refs/for/master
Will create a review for your-branch, and output its URL. Each commit in your
branch will be a separate review, though each will be grouped together, and each
needs a Change-Id line, as described in the Gerrit documentation (the error
message should you not do this will tell you how to enable a hook to create them
automatically, should you want to).
As part of your work, you'll want to find reviewers for your change. The best reviewers are those who have expertise in the area that you care about. In several cases the original authors of various parts of the system are in the illumos community. So if you were working on ZFS, you'd want to make sure that the core ZFS folks take a look at it by sending a review request to the ZFS mailing list. Much like testing, review scales to fit. The more complex the change, the more important that someone who is familiar with the subsystem and its inner workings reviews it. A longer discussion of code review is in the Integrating Changes section.
illumos has a standard commit format. That is:
<bug id> <synopsis>
<bug id 2> <synopsis 2>
<Reviewer n>
<Reviewer n+1>...
<Advocate>
An example of this is:
3673 core dumping is abysmally slow
3671 left behind enemy lines, agent LWP can go rogue
3670 add visibility into agent LWP's spymaster
Reviewed by: Keith M Wesolowski <keith.wesolowski@joyent.com>
Reviewed by: Joshua M. Clulow <jmc@joyent.com>
Reviewed by: Robert Mustacchi <rm@joyent.com>
Reviewed by: Richard Lowe <richlowe@richlowe.net>
Reviewed by: Garrett D'Amore <garrett@damore.org>
Reviewed by: Eric Schrock <eric.schrock@delphix.com>
Approved by: Richard Lowe <richlowe@richlowe.net>
In a given change, there may be one or more bugs fixed. Each bug corresponds to an entry at http://illumos.org/issues/. If you have logically separate changes, they should be done as separate commits. A rule of thumb is that if a single bug fix or RFE can stand on its own, it probably should. You should mention Reviewers in the commit message and don't worry about the 'Approved By' line, the advocate will take care of that for you.
In addition to the commit format, a given change should only be one commit. If
you already have one commit, you can modify the message by using the following
git command:
$ git commit --amend
$
If you have multiple commits, you should use git rebase to squash them down
into just one commit. As a part of a rebase, you will have the option to change
the commit message. To do this, you should run:
$ git rebase -i origin
$
Once you have, your tested change and your reviewers in line, there's just one last stop.
The last step is to run git pbchk. git pbchk runs several tests on the
source base including checking the commit message, style, copyright checks and a
few others. To see the full list see
git-pbchk(1ONBLD). With one
exception, copyright, you should fix all of the issues listed. Whether or not
you choose to put a copyright message for your modified code is up to you (and
whomever actually owns the copyright).
Congratulations! The changes that you've been working on should now be good to go. There is a lot to still explore and understand. Consider taking some time and reading it through or just going to a specific area based on your interest.
Specific places to go include: