This project provides packaging of the Intel Tofino SDE for the Nix package manager.
Disclaimer: The SDE for the Tofino series of P4-programmable ASICs is currently only available under NDA from Intel. The users of this repository are assumed to be authorized to download and use the SDE.
The Nix packaging of the SDE provides two basic functionalities: a shell in which P4 programs and control-plane programs can be developed interactively and a system to create packages for complete P4-based applications ready for deployment.
One of the goals of this work was to make the shell easy enough to use without requiring any understanding of Nix at all.
To use the system for packaging your own P4 programs clearly requires a fairly good understanding of the package manager. Explaining how Nix works in detail is out of scope of this documentation. Nevertheless, the text attempts to explain certain key concepts and provides links to more in-depth documentation on Nix to help the novice user.
This documentation does not explain how to create a complete service based on a P4 program ready for deployment by an end user, since this would involve program-specific items and mechanisms outside the scope of the Nix packages themselves, i.e. the packages will only make up one part of such a deployment.
Table of Contents
- Motivation
- Prerequisites
- Baseboard and Platform Support
- P4 Program Development with the SDE Shell
- The SDE Nix Package
- Packaging a P4 Program
- Packaging a Control-Plane Program for BfRuntime
- Using the Packages with a Nix Profile
- Deployment Models
The first questions the reader will probably ask are "what is Nix" and "what's the benefit of using it for the SDE"? The third question will most likely be "why should I have to read such a long README, clearly this is way too complicated"?
Those questions are justified and we'll make an attempt to answer them up front to provide a motivation to read on :)
Nix is a package manager
that uses a functional approach to managing packages and dependencies.
It uses a specialized functional language
(called the Nix expression language) to describe how packages are
built from source and how they relate to each other as build- and
run-time dependencies. It has the distinctive feature of storing all
packages in separate directories in a dedicated location called the
Nix store (usually /nix/store
). The path names of these
directories are of the form
/nix/store/<hash>-<name>-<verision>
The <hash>
is a cryptographic hash calculated over all of the inputs
to the package (e.g. packages it depends on, configure an build
options etc.). This mechanism overcomes most of the limitations of
standard package managers that use the file-system hierarchy standard
(/bin
, /lib
etc.) and imprecise dependency specifications prone to
suffer from the dreaded "DLL hell".
The result is that Nix provides a very high degree of reproducibility for software packages. This means that, given the specifications of the packages as a Nix expression, anyone using that specification will produce the exact same packages. This property is especially useful for distributing embedded systems which should work reliably and be robust against accidental changes of run-time dependencies.
Additional features are strict pinning of the exact versions of dependencies and the peaceful coexistence of arbitrary versions of the same package. Even though it follows a source-based deployment model by nature, it has a powerful mechanism called binary cache to provide various deployment models for pre-built packages.
This leads us to the answer to the question why this is useful for the Intel/Barefoot SDE. The SDE is a fairly complex piece of software with multiple components and a large number of dependencies required to build. Most of the dependencies must be resolved through the native package manager of the system on which the SDE is going to be built, imposing severe restrictions on the choice of build environments. Each release of the SDE is only certified to support specific versions of selected Linux distributions.
The Nix packaging of the SDE removes these restrictions by decoupling the build- and run-time dependencies from the native package manager of the underlying system. The advantages of this are obvious:
- Free choice of build- and run-time systems
- The work for supporting a new SDE version has to be done only once
- The SDE packages are guaranteed to work on every Linux system on which Nix can be installed
- No interference of dependencies with any other package (including different versions of the SDE itself)
The free choice of OS for the run-time system includes platforms with
a Tofino ASIC. The SDE Nix package has been run successfully on
ONL (based on Debian 9 and 10), on a plain
mion-based system as well as stock
Debian distributions. The only restriction is that the kernel must be
supported by the bf-drivers
component of the SDE. The Nix package
provides a facility to support any kernel on which the modules
provided by bf-drivers
can be compiled
successfully.
The Nix packaging for P4 programs only includes those features of the SDE which are required at run-time. This is important because it is currently legally forbidden to distribute certain components of the SDE (like the compiler or source code) to third parties which have no contractual relationship with Intel.
This leaves the final question "why is the documentation so darned long when this is supposed to be so simple and powerful"? The answer is that as with most powerful systems, the learning curve is rather steep. This is especially so in this case because Nix differs from other packaging systems the reader might be familiar with in a profound manner. The documentation assumes little to no familiarity with Nix, which requires a lot of space to explain various methods and concepts. A README directed at someone with a firm understanding of Nix would only take a fraction of the space :)
As a regular user, execute (or download and verify the script if you don't trust the site)
$ bash <(curl -L https://nixos.org/nix/install) --daemon
and proceed as instructed. This should work on any Linux distribution
because no support of the native package manager is required for the
installation (except for the presence of some basic commands like
curl
or rsync
).
Download the bf-sde
archive for the desired version of the SDE from
the Intel website (requires registration and NDA). Please verify that
the sha256
sums are as follows
File | sha256 |
---|---|
bf-sde-9.1.1.tar | be166d6322cb7d4f8eff590f6b0704add8de80e2f2cf16eb318e43b70526be11 |
bf-sde-9.2.0.tar | 94cf6acf8a69928aaca4043e9ba2c665cc37d72b904dcadb797d5d520fb0dd26 |
bf-sde-9.3.0.tgz | 566994d074ba93908307890761f8d14b4e22fb8759085da3d71c7a2f820fe2ec |
bf-sde-9.3.1.tgz | 71db320fa7d12757127c7da1c16ea98453f4c88ecca7853c73b2bd4dccd1d891 |
bf-sde-9.3.2.tgz | 8c637d07b788491b7a81896584be5998feadb7014b3ff42dc37d3cafd5fb56f8 |
bf-sde-9.4.0.tgz | daec162c2a857ae0175e57ab670b59341d39f3ac2ecd5ba99ec36afa15566c4e |
bf-sde-9.5.0.tgz | 61d55a06fa6f80fc1f859a80ab8897eeca43f06831d793d7ec7f6f56e6529ed7 |
bf-sde-9.5.1.tgz | 472d10360c30b21ba217eb3bc3dddc4f54182f325c7a5f7ae03e0db3cceba1b0 |
bf-sde-9.5.2.tgz | 60f366438c979f0b03d62ab997922e90e2aac447f3937930e3bd1af98c05d48a |
bf-sde-9.5.3.tgz | fd146282ec80c7fb2aea6f06db9cc871e00ffe3fed5d1de91ce27abdfc8c661a |
bf-sde-9.5.4.tgz | 3971b6b8400920529f0634de6d6211e709ec6e8797f66716d6c8bd31c4f030cb |
bf-sde-9.6.0.tgz | 0e73fd8e7fe22c62cafe7dc4415649f0e66c04607c0056bd08adc1c7710fd193 |
bf-sde-9.7.0.tgz | a4ca94f2d9602535c52613f9d8ad3504b55d99283a4e3dfc64de19e24d767423 |
bf-sde-9.7.1.tgz | dc0eb79b04797a7332f3995f37533a255a9a12afb158c53cdd421d1d4717ee28 |
bf-sde-9.7.2.tgz | e8cf3ef364e33e97f6af6dd4e39331221d61c951ffea30cc7221a624df09e4ed |
bf-sde-9.7.3.tgz | d45094c47b71fc7a21175436aaa414dd719b21ae0d94b66a5b5ae8450c1d3230 |
bf-sde-9.7.4.tgz | 1573577dc2718963dc45210fb9ed75255c68b75a2f219c85a70935dca90f4a16 |
bf-sde-9.8.0.tgz | 8d367f0812f17e64cef4acbe2c10130ae4b533bf239e554dc6246c93f826c12a |
bf-sde-9.9.0.tgz | c4314e76140a9a6f5d644176e0e3b0ca88f0df606b735c2c47c7cf5575d46257 |
bf-sde-9.9.1.tgz | 34f23716b38dd19cb34f701583b569b3006c5bbda184490bd70d5e5261e993a3 |
bf-sde-9.10.0.tgz | e0e423b92dd7c046594db8b435c7a5d292d3d1f3242fd4b3a43ad0af2abafdb1 |
bf-sde-9.11.0.tgz | 649cd026bc85a23f09c24d010d460d4192ae2a7e009da1f042183ca001d706b3 |
bf-sde-9.11.1.tgz | 3880d0ea8e245b0c64c517530c3185da960a032878070d80f4647f3bc15b4a9f |
bf-sde-9.11.2.tgz | e6c8cb7083b0c51fcccc5ba175889906cb596d3f05514dfe31f44a4c9102ad57 |
bf-sde-9.12.0.tgz | 5f3c41c32064909d8dab1c5f91b6a268b5c13835e5cfa48ff6ef7a526c93ad38 |
bf-sde-9.13.0.tgz | cc1c45f6a536ba0b26f3ae46a3d7b013e9d80f31b4c23c621f419dfb586d92f4 |
bf-sde-9.13.1.tgz | 82868acb6cf13ef44aa8b4674222df2b7208d4e4d78a724550229ea023a8e781 |
A BSP (Baseboard Support Package) contains code specific to the hardware configuration of one or more baseboards, each of which provides support for one or more platforms. The SDE uses a platform-agnostic API to implement functionality that depends on the baseboard.
The BSP is usually provided by the vendor of the device and has to be obtained separately. An exception to this is the reference BSP implementation provided by Intel, which supports the Tofino1-based WEDGE series of devices and the Tofino2-based AS9516-32D manufactured by EdgeCore (which is a subsidiary of Accton). The baseboard for the Tofino1 devices is called "accton", the one for the Tofino2 device "newport". This BSP is required to build the SDE while all other BSPs are optional.
The reference BSP is available from Intel, subject to the same NDA as for the SDE. The BSPs for other platforms must be obtained from the respective vendors individually. The current release supports the following BSPs (also see the section on Baseboard and Platform Support)
Baseboards | Vendor | File | Supported SDE version | sha256 |
---|---|---|---|---|
accton model |
Intel | bf-reference-bsp-9.1.1.tar |
9.1.1 | aebe8ba0ae956afd0452172747858aae20550651e920d3d56961f622c8d78fb8 |
accton model |
Intel | bf-reference-bsp-9.2.0.tar |
9.2.0 | d817f609a76b3b5e6805c25c578897f9ba2204e7d694e5f76593694ca74f67ac |
accton model |
Intel | bf-reference-bsp-9.3.0.tgz |
9.3.0 | dd5e51aebd836bd63d0d7c37400e995fb6b1e3650ef08014a164124ba44e6a06 |
accton model |
Intel | bf-reference-bsp-9.3.1.tgz |
9.3.1 | b934601c77b08c3281f8dcb235450b80316a42e2683ff29e4c9f2485fffbb51f |
accton model |
Intel | bf-reference-bsp-9.3.1.tgz |
9.3.2 | cb8c126d381ab0dbaf35645d1681c04df5c9675a7ac8231cf10eae5b1a402c9e |
accton model |
Intel | bf-reference-bsp-9.4.0.tgz |
9.4.0 | 269eecaf3186d7c9a061f6b66ce3d1c85d8f2022ce3be81ee9e532d136552fa4 |
accton model |
Intel | bf-reference-bsp-9.5.0.tgz |
9.5.0 | b6a293c8e2694d7ea8d7b12c24b1d63c08b0eca3783eeb7d54e8ecffb4494c9f |
accton model |
Intel | bf-reference-bsp-9.5.1.tgz |
9.5.1 | 34aa5bac92d33afc82cf4106173f7c364e9596c1bbf8d9dab3814f55de330356 |
accton model |
Intel | bf-reference-bsp-9.5.2.tgz |
9.5.2 | 2d544175f2ad57c9fc6a76305075540ee33253719bb3b9033d8af7dd39409260 |
accton model |
Intel | bf-reference-bsp-9.5.3.tgz |
9.5.3 | 2990fea8e4c7c1065cdcae88e9291e6dacb1462cc48526e93b80ebb832ac18d2 |
accton model |
Intel | bf-reference-bsp-9.5.4.tgz |
9.5.4 | d69264122986a66b0895c4d38bfa84f95f410f8a25649db33e07cd9cb69bdc33 |
accton model |
Intel | bf-reference-bsp-9.6.0.tgz |
9.6.0 | 88cb4b0978f23c28499faff75098f939374d9071859593353a18c2235e0be461 |
accton newport model |
Intel | bf-reference-bsp-9.7.0.tgz |
9.7.0 | 87f91540c0947edff2694cea9beeca78f95062b0aaca812a81c238ff39343e46 |
accton newport model |
Intel | bf-reference-bsp-9.7.1.tgz |
9.7.1 | 78aa14c5ec463cd4025b241e898e812c980bcd5e4d039213e397fcb6abb61c66 |
accton newport model |
Intel | bf-reference-bsp-9.7.2.tgz |
9.7.2 | d578438c44a19d2162079d9e4a4a5363a1503a64d7b05e96ceca96dc216f2380 |
accton newport model |
Intel | bf-reference-bsp-9.7.3.tgz |
9.7.3 | 33c33ab68dbcf085143e1e8d4a5797d3583fb2044152d063a61764939fa752d4 |
accton newport model |
Intel | bf-reference-bsp-9.7.4.tgz |
9.7.3 | 95cb4e81a4284cc22f0e0af9ef85ea1c0396b82bf1f64b79d8396715ddaec408 |
accton newport model |
Intel | bf-reference-bsp-9.8.0.tgz |
9.8.0 | 975fa33e37abffa81ff01c1142043907f05726e31efcce0475adec0f1a80f919 |
accton newport model |
Intel | bf-reference-bsp-9.9.0.tgz |
9.9.0 | f73aecac5eef505a56573c6c9c1d32e0fa6ee00218bc08e936fff966f8d2f87a |
accton newport model |
Intel | bf-reference-bsp-9.9.1.tgz |
9.9.1 | 481a2c5e6937f73ff9e9157fb4f313a4d72c0868b3eac94111ee79340c565309 |
accton newport model |
Intel | bf-reference-bsp-9.10.0.tgz |
9.10.0 | d222007fa6eee4e3a0441f09ed86b3b6f46df4c7d830b82b08bf6df7f88c4268 |
accton newport model |
Intel | bf-reference-bsp-9.11.0.tgz |
9.11.0 | a688b7468db32ea48c5ed83b040743b29f5beec28b3861440ff157cc6a5128ea |
accton newport model |
Intel | bf-reference-bsp-9.11.1.tgz |
9.11.1 | 37aa23ebf4f117bfc45e4ad1fbdb0d366b3bd094dd609f6ef1ec8b37ff6f2246 |
accton newport model |
Intel | bf-reference-bsp-9.11.2.tgz |
9.11.1 | f957ae2888289acc57271ad8d27e59075ddaaab723b38456382d25b8e3330331 |
accton newport model |
Intel | bf-reference-bsp-9.12.0.tgz |
9.12.0 | 60999d78e9a854e3a23b82ad0b644199e4aca5d88ad8eecea156e65faed2c2d4 |
accton newport model |
Intel | bf-reference-bsp-9.13.0.tgz |
9.13.0 | bd0ebd2bd8a08494668641fee7a7b7430d89925327c016c2a78315262097f485 |
accton newport model |
Intel | bf-reference-bsp-9.13.1.tgz |
9.13.1 | a6a3b8ab0164dfba1d97f41b33cc42f17c92925ca301d873800a075dcab6bca1 |
aps_bf2556 aps_bf6064 |
APS Networks | 9.5.0_AOT1.5.1_SAL1.3.2.zip |
9.4.0 | 2e56f51233c0eef1289ee219582ea0ec6d7455c3f78cac900aeb2b8214df0544 |
aps_bf2556 aps_bf6064 |
APS Networks | 9.5.0_AOT1.5.4_SAL1.3.4.zip |
9.5.0 | 510e5e18a91203fe6c4c0aabd807eb69ad53224500f7cb755f7c5b09c8e4525d |
aps_bf2556 aps_bf6064 |
APS Networks | 9.7.0_AOT1.6.1_SAL1.3.5_2.zip |
9.7.0 9.7.1 9.7.2 9.7.3 | 4941987c4489d592de9b3676c79cb2011a22fe329425e8876fa8ae026fc959ad |
inventec |
Inventec | bf-inventec-bsp93.tgz |
9.3.0 9.3.1 9.4.0 9.5.0 9.6.0 | fd1e4852d0b7543dd5d2b81ab8e0150644a0f24ca87d59f1369216f1a6e796ad |
inventec |
Inventec | bf-platform_SRC_9.7.0.2.1.tgz |
9.7.0 9.7.1 9.7.2 9.7.3 | 8391d5e791ae8b453711a79ed6f6d4372bd9ed6076b3ff54c649b69775b8d9c9 |
netberg |
Netberg | bf-platforms-netberg-7xx-bsp-9.7.0-220210.tgz |
9.7.0 9.7.1 9.7.2 9.7.3 | ad140a11fd39f7fbd835d6774d9b855f2ba693fd1d2e61b45a94aa30ed08a4f1 |
netberg |
Netberg | bf-platforms-netberg-7xx-bsp-9.9.0-221113.tgz |
9.9.0 9.9.1 | def63b745be735a0acfb4cb1a1f2eaeea91d0424762a9ffe04257b5659028870 |
netberg |
Netberg | bf-platforms-netberg-7xx-bsp-9.11.0-221209.tgz |
9.11.0 9.11.1 9.11.2 9.12.0 9.13.0 | 0a7bc5a9b152932dca7b9f269101a4d362ea07d87214c8ef594754a1234d7479 |
asterfusion |
Asterfusion | Github | 9.7.0 and later | Commit a5033f2 |
Execute (as any user)
$ nix-store --add-fixed sha256 <bf-sde-archive> <bf-reference-bsp-archive> <bsp-archive> ...
Note that the suffixes of the files differ between releases. The names in the tables above are exactly as they appear on the download site.
If this step is omitted, the build will fail with a message like the following
building '/nix/store/jx7is0zvkkpgv59s9hz6izmjn7qwfvh4-SDE-archive-error.drv'...
Missing SDE component bf-sde-9.4.0.tgz
Please add it to the Nix store with
nix-store --add-fixed sha256 bf-sde-9.4.0.tgz
The nix-store --add-fixed
command prints the name of the resulting
path in the Nix store, e.g.
$ nix-store --add-fixed sha256 bf-sde-9.3.0.tgz bf-reference-bsp-9.3.0.tgz
/nix/store/2bvvrxg0msqacn4i6v7fydpw07d4jbzj-bf-sde-9.3.0.tgz
/nix/store/4kiww8687ryxmx1xymi5rn5199yr5alj-bf-reference-bsp-9.3.0.tgz
As with any path in /nix/store
, these objects can only be deleted
with nix-store --delete <path>
, provided they are not referenced by
any "garbage collection roots" (in that case the command will fail).
More information on the Nix store can be found below.
$ git clone --branch <tag> https://github.com/alexandergall/bf-sde-nixpkgs.git
$ cd bf-sde-nixpkgs
Replace <tag>
with the desired release tag. See below
how to build and use the SDE for P4 program development.
Multiple vendors provide devices based on the Tofino ASIC. The differences in hardware configuration are isolated by a platform-independent API in the SDE. All platform dependent components are collected in a Baseboard Support Package (BSP).
The BSP is provided by the vendor that builds the device. Support for these BSPs must be added explicitly to the SDE package.
The SDE package uses the terms BSP, baseboard and platform in a very specific manner to precisely define which components are required to be installed on a given device:
- BSP. The software archive provided by a vendor. Each BSP provides support for one ore more baseboards.
- baseboard. A package created from a BSP, providing support for one or more platforms.
- platform. A unique identifier for a particular device
The list of BSPs and related baseboards is shown in the table
above. The model
baseboard is a pseudo-baseboard
which is a variant of the reference BSP that supports the Tofino
software emulation and is not bound to any specific hardware.
This hierarchy provides a unique mapping from platform to baseboard to BSP (but not the other way round). Each baseboard can provide support for multiple platforms. In this context, a platform denotes one particular device identified by a unique name. For practical reasons, the name space used for this is copied from the Open Network Install Environment project (ONIE). The assumption is that all vendors use ONIE for their devices and thus register a unique name in that name space.
To be precise, the platform identifier used in the SDE package
corresponds to the onie_machine
identifier from the
/etc/machine.conf
file that can be found on the ONIE installer
partition of the device. The same names are used in the
vendor-specific directories of onie/machine
in the ONIE Git
repository.
The following table shows the list of supported platforms and their mappings to a specific baseboard (this mapping is provided by https://github.com/alexandergall/bf-sde-nixpkgs/tree/master/bf-sde/bf-platforms/properties.nix)
Platform | Vendor | Baseboard |
---|---|---|
accton_wedge100bf_32x |
EdegCore | accton |
accton_wedge100bf_32qs |
EdegCore | accton |
accton_wedge100bf_65x |
EdegCore | accton |
accton_as9516_32d |
EdegCore | newport |
stordis_bf2556x_1t |
APS Netwokrs | aps_bf2556 |
stordis_bf6064x_t |
APS Netwokrs | aps_bf6064 |
inventec_d5264q28b |
Inventec | inventec |
inventec_d10064 |
Inventec | inventec |
netberg_aurora_710 |
Netberg | netberg |
asterfusion_x312p |
Asterfusion | asterfusion |
model |
model |
|
modelT2 |
model |
|
modelT3 |
model |
The model
platform is a pseudo-platform that exists in order to
support the Tofino ASIC emulator in a consistent manner. The model
baseboard uses the same BSP as reference
configured to provide stubs
for the platform-independent API of the SDE. The modelT2
and
modelT3
pseudo-platforms are identical to model
, but their
intrinsic target type is set to tofino2
and tofino3
,
respectively. The purpose of this is that one can call the
buildP4Program
function for these platforms and have the target
selected automatically, rather than using the model
platform and
overriding the target
paramete.
References to platforms and baseboards throughout this document refer to the table above.
If a BSP is not available for a particular platform, it can still be supported, provided that at least a file known as the board port map is present. A collection of these files is provided by the Stratum project.
The port mapping file is a low-level description of how the SerDes
lines of the Tofino chip are wired on a particular baseboard. It is
the bare minimum required to enable and configure the ports through
the bf_switchd
process. In this mode of operation, the SDE is built
without any BSP and hence contains no platform-dependent libraries at
all. The port map file is passed to the bf_switchd
process which
will then skip the step where it would try to locate those libraries
and use the static port configuration instead.
As a consequence, the QSFP ports are completely invisible, i.e. it is not possible to check, for example, whether a plugin is present or not or discover any of its properties including optical power levels. It is still expected that most plugins work.
Some of the parametets in the port map file concern characteristics of the electrical signals which could be specific for certain plugin types like copper wires. It is possible that those values are not suited for other types of plugins hence no guarantee can be made that all plugins will work correctly in BSP-less mode.
Due to the nature of the management interfaced used for QSFP-DD plugins (CMIS), which requires access to the modules for proper operation, BSP-less mode can not be applied to Tofino2-based platforms.
The quickest way to get started with compiling and running your P4 program in an interactive fashion is to use the SDE package in development mode. This can be done on a system having a Tofino ASIC installed or any host or VM without such hardware (as long as Nix can be installed on the platform). In the latter case, the Tofino ASIC software emulation (Tofino Model) can be used to test your programs.
The SDE shell is accessed through a command that must be installed
before it can be used. The installation is performed by building the
install
target from the Makefile
in the top-level directory of the
repository. Make sure that you followed all the prerequisite
steps before proceeding. It is sufficient to add only
the bf-sde
and bf-reference-bsp
files to the Nix store that
correspond to the SDE version that you want to build.
$ make install
installing 'sde-env-9.7.0'
...
This installs the command for the latest available version of the SDE
(9.7.0 in this example). You will find that this build finishes very
quickly and it does not yet build the actual SDE. That will happen
only when the sde-env-9.7.0
command is executed for the first time.
The reason for this is that only at that time is it known to the
system for which platform and possibly kernel it needs to build. This
process takes place only when the command is run for the first time.
After the build has finished, the user is greeted by the SDE shell
$ sde-env-9.7.0
[... lots of build output ...]
Intel Tofino SDE 9.7.0 on platform "accton_wedge100bf_32x"
Load/unload kernel modules: $ sudo $(type -p
bf_{kdrv,kpkt,knet}_mod_{load,unload})
Compile: $ p4_build.sh <p4name>.p4
Run: $ run_switchd.sh -p <p4name>
Build artifacts and logs are stored in /home/gall/.bf-sde/9.7.0
Use "exit" or CTRL-D to exit this shell.
[nix-shell(SDE-9.7.0):~]$
The first line after the output of the build process informs the user
that the hardware platform was identified to be
accton_wedge100bf_32x
. The command uses the following method to
determine the platform:
- Use the value passed with the
--platform
option - If
--platform
is not supplied, extract theonie_machine
identifier from/etc/machine.conf
- If
etc/machine.conf
does not exist, use themodel
platform as a fallback
To select a platform manually, use
$ sde-env-<version> --platform=<platform>
where <platform>
can be any of the supported platform
identifiers. Note that the existence of
/etc/machine.conf
depends on the details of the ONIE installation
process.
Within this shell, all commands shown in the introductory text are available in the search path and all SDE-specific environment variables are set to make compiling and running P4 programs straight forward.
The working directory of the SDE shell is inherited from the calling
shell as is the command search path (PATH
) unless the --pure
option is used
The --command
option can be used to execute arbitrary commands right
after the shell is started. This can be used to automate tasks that
require access to the SDE environment, e.g. to compile a P4
program. For example,
$ sde-env-9.7.2 --command "p4c some_program.p4; exit"
would attempt to compile some_program.p4
located in the directory
from which the sde-env
command is executed and then exit the shell.
To install the command for any other version, pass the version number
(e.g. 9.6.0 or 9.7.0) to the make
command by setting the VERSION
variable, e.g.
$ make install VERSION=9.6.0
installing 'sde-env-9.6.0'
...
The available versions can be displayed with the list-versions
target, e.g.
$ make list-versions
9.1.1
9.2.0
9.3.0
9.3.1
9.4.0
9.5.0
9.6.0
9.7.0
After installation, the command sde-env-<version>
(e.g. sde-env-9.7.0
) is available in the user's search path using a
Nix-specific feature called a profile
$ type -p sde-env-9.7.0
/home/gall/.nix-profile/bin/sde-env-9.7.0
The commands for different SDE versions can be installed concurrently
$ type -p sde-env-9.7.0 sde-env-9.6.0
/home/gall/.nix-profile/bin/sde-env-9.7.0
/home/gall/.nix-profile/bin/sde-env-9.6.0
For convenience, a separate make
target exists to install all
versions
$ make install-all
The command
$ make uninstall
removes all sde-env-*
commands from the search path.
To make a command available to all users, simply perform the
installation as the root
user to install the command in a global Nix
profile that is part of the search path for all users.
To compile a program, execute p4_build.sh
with the path to the P4
program to compile, e.g.
$ p4_build.sh ./my_example.p4
The build artifacts and logfiles are written to
$HOME/.bf-sde/<sde-version>
. Please use p4_build.sh --help
to see
available options.
To run the compiled program on the Tofino ASIC, make sure that the SDE environment has been started with the proper platform support enabled as described in the introduction. To run the P4 program execute
$ run_switchd.sh -p <program_name>
where <program_name>
is the base name of the path provided to the
p4_build.sh
command without the .p4
suffix. With the example of
the last section, this would be
$ run_switchd.sh -p my_example
To get usage information for run_switchd.sh
, run the command without
any arguments.
The script ends up calling the executable bf_switchd
, which programs
the ASIC according to the P4 program and provides access to the CLI.
It requires a kernel module for proper operation. The SDE currently
provides three such modules
bf_kdrv
bf_kpkt
bf_knet
Which of these should be used and what they do is outside the scope of
this text and the reader is referred to the documentation of the SDE
for more information. The modules bf_kdrv
and bf_kpkt
are
mutually exclusive.
Each module has a shell script called <module>_mod_load
(e.g. bf_kdrv_mod_load
) which loads it with the proper parameters.
These scripts must be executed as root. For example using sudo
$ sudo $(type -p bf_kdrv_mod_load)
The modules can be unloaded simply by
$ sudo rmmod <module>
or by using the commands <module>_mod_unload
with sudo
in a
similar fashion as the load commands.
The kernel modules need to be compiled for the exact same kernel which
is running on the current system. Kernels must be supported
explicitly by the SDE package. If the running
kernel is not supported, the *_mod_load
commands terminate with the
message
No modules available for this kernel (<output-of-uname-r>)
Instead of running the program on the actual ASIC, it can also be run using a register-accurate software emulation called the Tofino model. This option is available on all systems, including those having an actual ASIC.
To use the Tofino model, the shell must be started with
--platform=model
:
gall@spare-PB1:~/bf-sde-nixpkgs$ sde-env-9.7.0 --platform=model
Intel Tofino SDE 9.7.0 on platform "model"
Compile: $ p4_build.sh <p4name>.p4
Run: $ run_switchd.sh -p <p4name>
Run Tofino model:
$ sudo $(type -p veth_setup.sh)
$ run_tofino_model.sh -p <p4name>
$ run_switchd.sh -p <p4name>
$ sudo $(type -p veth_teardown.sh)
Run Tofino model with custom portinfo file:
$ sudo $(type -p veth_from_portinfo) <portinfo-file>
$ run_tofino_model.sh -p <p4name> -f <portinfo-file>
$ run_switchd.sh -p <p4name>
$ sudo $(type -p veth_from_portinfo) --teardown <portinfo-file>
Run PTF tests: run the Tofino model, then
$ run_p4_tests.sh -p <p4name> -t <path-to-dir-with-test-scripts>
Build artifacts and logs are stored in /home/gall/.bf-sde/9.7.0
Use "exit" or CTRL-D to exit this shell.
[nix-shell(SDE-9.7.0):~]$
By default, the model uses veth
interfaces to connect to the
emulated ports. These interfaces are set up with the script
veth_setup.sh
. It requires root privileges and needs to be called
with sudo
$ sudo $(type -p veth_setup.sh)
The veth
interfaces are persistent. They can be removed at any time
by executing
$ sudo $(type -p veth_teardown.sh)
To run the Tofino model, execute
$ run_tofino_model.sh -p <program_name>
The model then waits for a connection from a bf_switchd
process, which can be started exactly as for the ASIC:
$ run_switchd.sh -p <program_name>
Note that the bf_switchd
process crashes if one of the kernel
modules is loaded when running on the Tofino model (this behaviour is
present at least up to version 9.9.0). In this case, simply unload the
module and restart run_switchd.sh
.
The default mapping of emulated ports to veth
interfaces can be
overriden by passing a portinfo file to the model binary. This file
is in JSON and can contain two types of objects:
{
"PortToVeth": [
{
"device_port": <dev_port>,
"veth1": <n1>,
"veth2": <n2>
},
{
"device_port": <dev_port>,
"veth1": <n1>,
"veth2": <n2>
},
...
],
"PortToIf": [
{
"device_port": <dev_port>,
"if": <intf>
},
{
"device_port": <dev_port>,
"if": <intf>
},
...
]
}
The PortToVeth
object maps a specific device port <dev_port>
to a
pair of veth
interfaces with the names veth<n1>
and veth<n2>
,
where the device port is connected to veth<n1>
. The veth
pair must
exist exist when the model is started.
The PortToIf
object maps a specific device port <dev_port>
to the
interface named intf
. The interface named intf
must exist when the
model is started.
As a convenience, scripts are provided to create the veth
pairs
requried by the PortToVeth
object. The veth_{setup,teardown}.sh
scripts create/remove the veth
pairs required by the default
portinfo file built into the model binary. The default for the Tofino
target maps device ports 0-16 to the veth
pairs veth0/veth1
through veth31/veth32
and device port 64 to veth250/veth251
as the
CPU port. For the Tofino2 target, the default is to map device ports
8-24 and port 2 as the CPU port to the same veth
pairs.
These scripts are part of the standard SDE tooling. The Nix SDE
package provides an additional script veth_from_portinfo
which takes
a portinfo file as input. When called without any options, the script
generates the veth
pairs provided by the PortToVeth
list in the
file (the PortToIf
object is ignored). When called with the
--teardown
option, the veth
pairs are removed:
$ cat ports.json
{
"PortToVeth": [
{
"device_port":0,
"veth1":0,
"veth2":1
}
]
}
$ sudo $(type -p veth_from_portinfo) ports.json
Creating veth pair 0/1 for device port 0
$ sudo $(type -p veth_from_portinfo) --teardown /tmp/ports.json
Deleting veth pair 0/1
The portinfo file must also be supplied to the run_tofino_model.sh
script:
$ run_tofino_model -p <program> -f <portinfo_file>
The Packet Test Framework (PTF) is also available in the SDE development shell. The command
$ run_p4_tests.sh -p <program_name> -t <test_directory>
exercises the tests by executing all python scripts found in the
directory <test_directory>
. This requires that the P4 program being
tested is run on the Tofino model.
The default environment contains the Python interpreter required by
the bf-drivers
package (Python 3 for SDE 9.7.0 and newer and Python
2 for all older SDE versions), e.g. to access the bfrt_grcp
modules
provided by that package. This environment contains only the standard
Python modules as well as all the modules provided by bf-drivers
.
This interpreter is the preferred match of python
in the default
search path. All control-plane scripts should be executed with that
interpreter explicitly.
The PTF test scripts run through run_p4_tests.sh
use a separate
Python environment (but the same interpreter), which is provided by
the ptf-modules
package.
Any of the control-plane or PTF scripts requiring non-standard Python modules will fail to run. To solve this problem, the development shell can be called with two additional parameters
$ sde-env-<version> --python-modules=<module>,... --python-ptf-modules=<module>,...
For example, the standard environment doesn't know the jsonschema
Python Module
$ sde-env-9.7.0
[...]
[nix-shell(SDE-9.7.0):~]$ python -c "import jsonschema; print(jsonschema)"
Traceback (most recent call last):
File "<string>", line 1, in <module>
ImportError: No module named jsonschema
[nix-shell(SDE-9.7.0):~]$
Let's add jsonschema
to the environment
$ sde-env-9.7.0 --python-modules=jsonschema
[...]
[nix-shell(SDE-9.7.0):~]$ python -c "import jsonschema; print(jsonschema)"
<module 'jsonschema' from '/nix/store/6215k8b5pxvshkacq1dlg2m4bx3ij81a-python3-3.8.5-env/lib/python3.8/site-packages/jsonschema/__init__.py'>
[nix-shell(SDE-9.7.0):~]$
This works in a similar manner as the native Python virtual
environments (venv
) but is based on Nix.
Modules added with --python-ptf-modules
are not visible to the
Python interpteter:
$ sde-env-9.7.0 --python-ptf-modules=jsonschema
[...]
[nix-shell(SDE-9.7.0):~]$ python -c "import jsonschema; print(jsonschema)"
Traceback (most recent call last):
File "<string>", line 1, in <module>
ModuleNotFoundError: No module named 'jsonschema'
What happens in this case is that the locations of those modules are
added to the PTF_PYTHONPATH
environment variable in the shell rather
than adding the modules to the Python interpreter. The
run_p4_tests.sh
script merges PTF_PYTHONPATH
with PYTHONPATH
before starting the PTF tests. The result is that modules added with
--python-ptf-modules
are only visible to the PTF scripts.
The names of the modules must be exactly those used by Nix to identify them in the package repository. Here is a hack that can be used to get the list of avaiable modules for the Nixpkgs version used by the SDE:
$ echo -e $(nix eval '(with import ./. {}; with builtins; concatStringsSep "\n" (attrNames python3.pkgs))')
It needs to be executed in the top-level directory of the
bf-sde-nixepr
repository. To see the list for a different Python
version, replace python3
by, e.g. python2
.
By default, the shell started by a sde-env-*
command inherits the
PATH
from the calling shell. To be precise, PATH
inside the shell
consists of a number of paths from the Nix store with PATH
inherited
from the parent shell appended to it. The paths provided by Nix
include gcc
, make
, binutils
, coreutils
(most standard Unix
utilties), diff
, sed
, gep
, awk
, tar
, gzip
, bzip
,
xz
. Those will take precedence over the same commands provided by
the native package manager.
More Nix-based packages can be added to the shell by using the
--pkgs
option of the sde-env-*
command. For instance, suppose we
have a system that provides Perl 5.28 with a native package:
$ type perl
perl is /usr/bin/perl
$ perl -V:version
version='5.28.1';
By default, this version of perl
would be available in the SDE
shell. To override it with perl
from Nix, start the shell with
$ sde-env-9.7.0 --pkgs=perl
[...]
[nix-shell(SDE-9.7.0):~]$ type perl
perl is /nix/store/5fz4mi6ghnq6qxy8y39m3sbpzwr6nzaw-perl-5.32.0/bin/perl
[nix-shell(SDE-9.7.0):~]$ perl -V:version
version='5.32.0';
To remove all dependencies on packages provided by the native package
manager, the SDE shell can be started with the --pure
option. In
that case, PATH
is not inherited from the calling shell and all
commands that the user wants to have available in the SDE shell need
to be added explicitly with --pkgs
(apart from the default packages
mentioned above).
$ sde-env-9.7.0 --pure
[...]
[nix-shell(SDE-9.7.0):~]$ type perl
bash: type: perl: not found
As with the Python modules, the names passed with the --pkgs
option
must be the identifiers used by Nix for that specific package. To see
the list of available packages, execute
$ echo -e $(nix eval '(let pkgs = import ./. {}; in with builtins; concatStringsSep "\n" (attrNames pkgs))')
in the top-level directory of the bf-sde-nixpkgs
repository.
As discussed above, the --python-ptf-modules
option is used to add a
Python module that needs to be visible to the PTF scripts. Similarly,
the --ptf-pkgs
option adds regular (i.e. non-Python) packages to the
search Path of the PTF script. The bin
path of every package passed
with this option will be added to the PTF_PATH
environment variable,
which is added to PATH
by the run_p4_tests.sh
script.
For example, suppose the PTF script wants to execute the ip
command
in a subprocess. This command is part of the iproute
Nix package,
hence instantiating the SDE environment with
$ sde-env-9.7.0 --ptf-pkgs=iproute
will create the $PTF_PATH
[nix-shell(SDE-9.7.0):~/bf-sde-nixpkgs]$ echo $PTF_PATH
/nix/store/d0rc2qli8df2xbznca2rld7zl878frsd-iproute2-5.17.0/bin
The build procedure using make install
requires Internet access to
download the source code for various components and to access
pre-built Nix packages from a binary cache. If, for any reason, the
host on which the SDE is intended to be used has no or just restricted
Internet access, it is possible to build a "standalone" installer for
the SDE that does not require network access once it has been copied
to the system.
The installer is created with the standalone
target
$ make standalone
for the most recent version or
$ make standalone VERSION=<version>
for any other supported version. This will create a self-extracting
archive called sde-env-<version>-standalone-installer
in the user's
home directory. To use it, copy the file to the destination and
execute it as root, e.g.
$ sudo sde-env-9.7.0-standalone-installer
The only requirement for this to work is that nixpkgs
has been installed
on the system. After the installation, the
sde-env-<version>
command is available for all users on the system.
The SDE Nix package provides two types of services. The first is an environment to build and test a P4 program as explained in the previous chapter. This environment includes the P4 compiler, Tofino ASIC emulator, the Packet Test Framework and, optionally, support for a specific hardware platform.
The second service is a runtime environment for pre-compiled P4
programs. This environment only contains the build artifact of a P4
program and the components necessary to execute them on the Tofino
ASIC on a particular platform (including the pseudo-platform model
,
which executes the program on the ASIC software model). The use case
for this service is the deployment of P4 programs on hardware
appliances. The binary packages required for this deployment can be
distributed to users who do not have signed an NDA or software
license agreement with Intel.
How these services are instantiated is described later in this document. This section describes how they are composed of smaller components and how they relate to the SDE environment built according to the official procedure supported by Intel.
The standard procedure supported by Intel to build the SDE makes use
of a Python-based framework called P4 Studio. It builds and
installs the components contained in the SDE archive as distributed by
Intel into a directory tree pointed to by the SDE_INSTALL
environment variable. The unpacked SDE archive is referenced by the
SDE
environment variable. Many of the tools used to compile and run
P4 programs use both of these variables to locate various components
of the SDE.
The object in the SDE Nix package which comes closest to this can be built by executing
$ nix-build -A bf-sde.<version>
where <version>
is the same identifier for a particular version of
the SDE as described in the section about using the development
shell, e.g. v9_3_0
or v9_4_0
or latest
, which is an
alias of the most recent version.
The output of the command is a path in the Nix store (/nix/store
),
which is a directory containing the same objects as a build with P4
Studio, for example
$ nix-build -A bf-sde.latest
/nix/store/8wh2yi3v1vajw6g9gjylankmxafp3g3k-bf-sde-9.3.1
$ ls -l /nix/store/8wh2yi3v1vajw6g9gjylankmxafp3g3k-bf-sde-9.3.1
total 24
lrwxrwxrwx 1 root root 80 Jan 1 1970 bf-sde-9.3.1.manifest -> /nix/store/zxhfhfw4dsm8rvbn9pgrk0b84vk7ii2q-bf-tools-9.3.1/bf-sde-9.3.1.manifest
dr-xr-xr-x 2 root root 4096 Jan 1 1970 bin
dr-xr-xr-x 2 root root 4096 Jan 1 1970 include
dr-xr-xr-x 3 root root 4096 Jan 1 1970 lib
lrwxrwxrwx 1 root root 65 Jan 1 1970 pkgsrc -> /nix/store/zxhfhfw4dsm8rvbn9pgrk0b84vk7ii2q-bf-tools-9.3.1/pkgsrc
dr-xr-xr-x 4 root root 4096 Jan 1 1970 share
At this point, it is useful to explain a concept called user
environment (or just environment, for short) used by the Nix package
manager. In Nix, every package exists in a separate directory in the
Nix store, i.e. each one has its own collection of the standard Unix
directories bin
, lib
, share
, man
, include
etc. The
environment is a mechanism that makes the individual directories of
multiple packages available from a single location. This is done by
creating a new directory in /nix/store
containing a single set of
those directories and creating symbolic links in them to the
corresponding objects in the individual packages. It is then possible
to reference, say, the executables of all packages through a single
path, namely the bin
directory of the environment.
The output of nix-build -A bf-sde.latest
is exactly such an
environment. This particular environment combines some of the
packages described in the next section:
bf-syslibs
bf-drivers
bf-utils
bf-platforms.model
p4c
tofino-model
ptf-modules
ptf-utils
For example, the user environment contains the bf_switchd
command
supplied by bf-drivers
and the bfshell
command supplied by
bf-utils
$ ls -l
/nix/store/8wh2yi3v1vajw6g9gjylankmxafp3g3k-bf-sde-9.3.1/bin/bf_switchd
lrwxrwxrwx 1 root root 75 Jan 1 1970 /nix/store/8wh2yi3v1vajw6g9gjylankmxafp3g3k-bf-sde-9.3.1/bin/bf_switchd -> /nix/store/868l44v30k9b6jh83ha95r7jqgis4h4k-bf-drivers-9.3.1/bin/bf_switchd
$ ls -l
/nix/store/8wh2yi3v1vajw6g9gjylankmxafp3g3k-bf-sde-9.3.1/bin/bfshell
lrwxrwxrwx 1 root root 70 Jan 1 1970 /nix/store/8wh2yi3v1vajw6g9gjylankmxafp3g3k-bf-sde-9.3.1/bin/bfshell -> /nix/store/15fmcbjc42pmilzks40h56p8lywl3zpa-bf-utils-9.3.1/bin/bfshell
It also contains the following tools used to interact with the SDE
run_switchd.sh
bfshell
run_bfshell.sh
run_tofino_model.sh
run_p4_test.sh
p4_build.sh
It is configured to support the pseudo-platform model
, i.e. a P4
program started with run_switchd.sh
would be executed on the Tofino
ASIC emulator. One way to use the SDE environment on a particular
hardware platform is by using the SDE Shell feature.
In the SDE built in the traditional way by P4 Studio, one needs to set
SDE
and SDE_INSTALL
for these tools to work. In the Nix package,
this is done by shell wrappers around them, i.e. they work without
having to modify the caller's environment.
However, there is a problem when using p4_build.sh
to compile a P4
program. In the traditional SDE, the build artifacts are stored in
the SDE itself. This is not possible with Nix, because all packages
are immutable, i.e. once installed in the Nix store, they can never be
changed.
To solve this problem, the Nix SDE package uses slightly modified
versions of the scripts used to compile P4 programs and run them with
either run_switchd.sh
or run_tofino_model.sh
to allow the build
artifacts of P4 programs to be stored outside of SDE_INSTALL
. It
uses a new environment variable P4_INSTALL
instead to specify the
location of P4 compiler artifacts and SDE_BUILD
for the location of
the build directory.
Apart from this, the Nix package works just like the one built in the traditional manner.
In principle, one could use the output of nix-build
directly by
setting PATH
, P4_INSTALL
and SDE_BUILD
appropriately. To avoid
this inconvenience, the Nix package offers a method to easily create a
new shell in which all of these settings are created automatically,
which is exactly what the sde-env-*
command described
previously does.
The output of nix-build
is the closest thing to a binary package of
a traditional package manager. Nix uses a two-stage process when
building packages. The first stage is called a derivation. It is
the direct result of evaluating a Nix expression that contains the
build recipe of a software component. Like everything Nix produces, it
is a file stored in /nix/store
but with the suffix .drv
to
distinguish them from actual packages. It is not the package itself
but an encoding of the procedure that needs to be executed to perform
the build (i.e. it the build script to execute, make
and configure
flags etc). The process of generating the .drv
file is also called
instantiation of the derivation.
The build takes place in a second step when the instructions contained
in the derivation are executed (this is called the realization of a
derivation). The result of that step is another path in /nix/store
(sometimes referred to as the output path) which contains the
package itself.
The Nix community usually uses the terms derivation and package somewhat loosely as synonyms, because the distinction rarely matters in practice. We adopt this convention in the rest of this document.
The Nix packaging of the SDE contains many derivations and the command
nix-build -A bf-sde.<version>
merely creates a particular one as the
default.
The complete set of derivations provided by the SDE Nix package is provided in the next section.
The SDE itself consists of a set of components which are combined to
create the entities of interest to the user, namely the development
and runtime environments. There is one package per supported version
of the SDE. Each version has an attribute (we will see later what
exactly this means) called pkgs
containing a separate derivation for
each of these components. They are built automatically when the
default derivation is built (or any other derivation which depends on
them), but they can also be built explicitly if desired (though this
should not be necessary for normal use). In that case, they can
either be built all at once with
$ nix-build -A bf-sde.<version>.pkgs
Or individually, e.g.
$ nix-build -A bf-sde.<version>.pkgs.bf-drivers
In the latter case, all derivations on which the given derivation depends will be built implicitly.
The following derivations are available
bf-syslibs
bf-utils
bf-drivers
bf-drivers-runtime
bf-diags
bf-platforms
p4c
tofino-model
ptf-modules
ptf-utils
ptf-utils-runtime
bf-pktpy
(SDE 9.5.0 and later)kernel-modules
kernel-modules-baseboards
All but the last have a direct correspondence to the components
built by P4 Studio (with the exception of the ptf-modules
component,
which is split into separate packages for technical reasons).
kernel-modules
is actually not a derivation but an attribute set of
derivations, each of which provides the modules for one of the
supported kernels.
kernel-modules-baseboards
is an attribute set that contains one
attribute per supported kernel, like kernel-modules
but each
attribute is another set with one attribute per supported
baseboard. Each of these attributes is a derivation that contains the
same (baseboard-independent) kernel modules as the corresponding
attribute of kernel-modules
in addition to modules and arbitrary
files or programs that are specific for a particular baseboard.
bf-platforms
is also an attribute set of derivations with one
attribute per supported BSP.
Due to the fact that Nix packages are implemented as functions, it is possible to associate additional functionality with them which has the character of methods in an object-oriented framework (but note that the Nix expression language is not object-oriented in any sense of the term).
For instance, each of the SDE packages supported by this repository
(i.e. one for each supported version of the SDE) has a function
associated with it, which, when given the source code of a P4 program,
compiles the program and creates a new package containing a command
that runs bf_switchd
with the compiled artifacts. This is more
powerful than using the SDE as a mere build-time dependency as in
traditional package managers, because it includes the build procedure
for the P4 program itself (which could also vary between different SDE
versions).
The following is a list of these additional attributes. They can all
be accessed with the "attribute path" bf-sde.<version>.<attribute>
.
version
, type: stringpkgs
, type: attribute set of derivations (see previous section)buildP4Program
, type: functionkernelIDFromRelease
, type: functionmodulesForKernel
, type: functionallPlatforms
, type: attribute setplatforms
, type: attribute setruntimeEnv
, type: functionruntimeEnv'
, type: functionruntimeEnvNoBsp
, type: derivationbaseboardForPlatform
, type: functionsupport
, type: attribute set of functionsmkShell
, type: functionenvCommand
, type: derivationenvStandalone
, type: derivationtest
, type: attribute set of derivations
All of the attributes of type "derivation" can be built with
nix-build -A <...>
.
For the curious: non-derivation objects can be built by evaluating a Nix expression as follows (executed from the top-level directory of the repository)
$ nix eval '(with import ./. {}; bf-sde.latest.version)'
"9.5.0"
The mkShell
function is a special object that can only be used by
the nix-shell
command, for example as used implicitly by the
sde-env-*
commands built with make install
that provide a
development shell.
version
is self-explanatory and pkgs
has been described in the
previous section. The other attributes are explained in detail below.
This function is at the heart of the mechanism used to build
individual packages for the artifacts of arbitrary P4 programs. The
definition of this function can be found in
bf-sde/build-p4/default.nix
. It takes the following arguments
-
pname
: The name of the package to generate. It doesn't have to be unique and appears only in the name of the final path of the resulting package in/nix/store
. Nothing in the Nix machinery to build packages depends on this name. -
version
: The version of the package. It is combined withpname
to become part of the name of the package in/nix/store
-
p4Name
: The name of the top-level P4 program file to compile, without the.p4
extension and without any directories prepended (seepath
) -
path
: An optional path to the program file relative to the root of the source directory (seesrc
) -
execName
: optional name under which the program will appear in the finished package, defaults top4Name
. This is useful if the same source code is used to produce different programs, e.g. by selecting features via preprocessor symbols. Each variant of the program can be given a differentexecName
, which makes it possible to combine them all in the same Nix profile or user environment (which would otherwise result in a naming conflict because all programs would have the same name, i.e.p4Name
) -
target
: optional target for which to compile. Must be one oftofino
,tofino2
ortofino3
. The default is the intrinsic target for the selectedplatform
. Overriding the default only makes sense for themodel
pseudo-platform, which supports all targets (and defaults totofino
). To avoid this override, use the pseudo-platformsmodelT2
andmodelT3
instead. Note that while the P4 compiler has been supportingtofino2
for some time, the standard build scriptp4_build.sh
supportstofino2
andtofino3
only for SDE 9.7.0 and later. -
buildFlags
: optional list of strings of options to be passed to thep4_build.sh
build script, for example a list of preprocessor symbols[ "-Dfoo" "-Dbar" ]
-
platform
: optional name of the platform for which to build the P4 program. It must be one of the supported platform identifiers. The default ismodel
. The effect of this option is to include thebf-platforms
sub-package into the runtime environment that corresponds to the baseboard that supports the given platform. -
requiredKernelModule
: Thebf_switchd
program (provided by thebf-drivers-runtime
package) requires a kernel module to be present (if the P4 program is run on the ASIC rather than the Tofino model, which is the assumption here). Currently, there is a selection of three such modules calledbf_kdrv
,bf_kpkt
andbf_knet
(their function is not discussed here and the reader is referred to the documentation supplied by Intel). This optional parameter selects which of those modules is required by the P4 program. It is used by themoduleWrapper
support function associated with the package created bybuildP4Program
as explained below -
src
: A store path containing the source tree of the P4 program, typically the result of a call tofetchgit
orfetchFromGitHub
-
patches
: An optional list of patches to be applied to the source tree before building the package -
pureArtifacts
: an optional flag whether to remove files from the P4 artifacts produced by the compiler that are not needed for execution but generate dependencies on the source code of the program being compiled as well as the P4 compiler package. This option has no effect for SDEs older than 9.7.0. For 9.7.0 and later, it removes specific JSON files likesource.json
andfrontend-ir.json
. The default istrue
.
The function essentially performs
<path-to-sde>/bin/p4_build.sh ${buildFlags} <source-tree>/${path}/${p4Name}.p4
where p4_build.sh
is part of the SDE package. The build artifacts
are stored in the resulting package. If execName
is used, the
builder first creates the symbolic link
<source-tree>/${path}/${execName}.p4 -> <source-tree>/${path}/${p4Name}.p4
and then runs
<path-to-sde>/bin/p4_build.sh ${buildFlags} <source-tree>/${path}/${execName}.p4
The resulting package contains the following files, where <name>
is
either <p4Name>
if no execName
was given or execName
bin/<name>
share/p4/targets/tofino/<name>.conf
share/tofinopd/<name>/bf-rt.json
share/tofinopd/<name>/<name>.conf
share/tofinopd/<name>/pipe/context.json
share/tofinopd/<name>/pipe/tofino.bin
The package will have a derivation produced by the runtimeEnv
support function as run-time dependency. That package
contains just enough components of the full SDE to start the
bf_switchd
process with the artifacts of the compiled program on the
given platform. The bin/<name>
executable is just a shell script
which invokes run_switchd.sh
.
If platform
is one of the model platform (model
, modelT2
etc.),
the runtime environment includes the Tofino model binary and the
run_tofino_model.sh
utility script. The bin/<name>
executable, in
addition to calling run_switchd.sh
, also starts
run_tofino_model.sh
in the background and calls veth_setup.sh
to
create the veth
virtual interfaces that connect to the emulated
ports of the model. By default, the output of the tofino-model
and
bf_switchd
programs are both sent to stdout
and stderr
. If the
script is called with open file descriptors 3 and/or 4, the model will
write its stdout
and stderr
to descriptors 3 and 4, respectively.
For example, calling the script with
$ /nix/store/.../bin/<name> >switch.log 2>&1 3>model.log 3>&4
will collect all output of bf_switchd
in switch.log
and all output
of tofino-model
in model.log
.
By default, the model will use a built-in file for mapping ports to
local interfaces. To use a custom mapping file, assign the absolute
path to the file to the environment variable TOFINO_MODEL_PORTINFO
before running the bin/<name>
script. See the section about running
the Tofino model for details about the portinfo
mechanism.
The resulting package has additional
attributes just like the bf-sde.<version>
packages:
moduleWrapper
, type: functionmoduleWrapper'
, type: functionrunTest
, type: function
The moduleWrapper
function creates a
new package which contains the kernel modules for the specified kernel
and a shell wrapper around bin/<name>
from the package for which the
function is called. The argument to the function must be the kernel
release identifier produced by executing uname -r
on a running
instance of the kernel. This wrapper loads the kernel module that has
been specified with requiredKernelModule
and then calls bin/<name>
from the original package. See modulesForKernel
for details.
The function moduleWrapper'
is like moduleWrapper
, but it takes a
specific kernel module package as argument (i.e. one of the attributes
of the pkgs.kernel-modules
sub-package).
The runTest
function runs the PTF tests from test scripts in the
source tree (more documentation on this TBD).
The function takes a kernel release identifier as input and returns a
kernel ID that uniquely identifies a supported kernel. The release
identifier is a string as returned by the uname -r
command. There
are three possible results:
- There is exactly one match. The function returns the kernel ID
for the matching kernel (i.e. one of the attribute names of the
pkgs.kernel-modules
sub-package). - There is no match. The function returns the dummy kernel ID "none".
- There are multiple matches. This means that there are several
distinct kernels which produce the same release identifier
(i.e.
uname -r
results in the same string when executed on a running instance on each of the kernels). This is possible when, for example, the same kernel is compiled with different options. In this case, the user must disambiguate the choice by setting theSDE_KERNEL_ID
environment variable to the desired kernel ID.
If the environment variable SDE_KERNEL_ID
is set, it overrides the
value of the kernel release passed to the function. An error is thrown
if the given kernel ID does not exist in the set
pkgs.kernel-modules
.
The function takes a kernel release identifier as input and returns a
derivation containing the matching kernel modules. It first calls
kernelIDFromRelease and then returns the
attribute of pkgs.kernel-modules
for the resulting kernel ID.
This is an attribute set that contains one attribute for each known
platform as defined in bf-sde/bf-platforms/properties.nix
.
The subset of allPlatforms
that is supported by this SDE.
This function takes a baseboard identifier as
input and returns a derivation just like what is produced with
nix-build -A bf-sde.<version>
but with a reduced set of packages in
the environment. Instead of the full-fledged SDE, the runtime
environment only provides what's necessary to run a compiled P4
program on the given platform:
bf-syslibs
bf-drivers-runtime
bf-utils
ptf-utils-runtime
bf-platforms.<baseboard>
It also contains a reduced set of scripts
run_switchd.sh
run_bfshell.sh
(Note that the ptf-utils-runtime
package is required by
run_bfshell.sh
). If baseboard
is model
, the runtime Environment
also contains the tofino-model
package (to provide the model binary)
and the script run_tofino_model.sh
.
In order to deploy a compiled P4 program on a system, it is sufficient
to install the runtimeEnv
derivation together with the derivation
containing the build artifacts of the program.
This is like runtimeEnv
but instead of the baseboard identifier, the
function takes a platform identifier as input.
It is a convenience function that uses the
baseboardForPlatform
support function and
then calls runtimeEnv
.
This is like runtimeEnv
but does not install any of the
bf-platforms
packages in the environment. Hence, it cannot be used
to run P4 programs. It can be used, for example, to execute one of
the utility scripts like run_bfshell.sh
.
This is a function which, given a platform identifier, returns the identifier of the baseboard that provides support for it. An exception is raised if the platform identifier is not known.
The baseboard can be null
if the platform is supported in BSP-Less
Mode. There are two ways in which this can happen. The
first is to explicitly set the baseboard
attribute in
bf-sde/bf-platforms/properties.nix
to null. In that case, the
portMap
attribute must also be present and point to the port map
JSON file. In this mode, the platform is supported for any SDE
version in BSP-less mode.
If baseboard
is not null
, i.e. it contains the name of an actual
baseboard, the platform can only be supported if a BSP is available
for the given SDE version, i.e. the platform cannpt be supported for
SDE versions for which there is no BSP available. It is possible to
fall back to BSP-less mode for such a platform by adding the port map
file with the portMap
attribute. In that case,
baseboardForPlatform
will return null
and emit the message "No
native platform support for <platform>
, falling back to BSP-less
mode".
If the SDE does not have a BSP for the platform and no portMap
is
specified, baseboardForPlatform
raises an exception with the message
"Platform <platform>
not supported in SDE <SDE-version>
and no
BSP-less fallback provided".
This is an attribute set of functions that provide support for
creating releases and installers for SDE-based appliances. See
bf-sde/support/README.md
for further
information.
This function is at the heart of the SDE package when used as a
development environment. When evaluated it creates a new shell in
which the selected SDE is available and can be used to compile and run
P4 programs on-the-fly. It does not return a proper derivation and
therefore cannot be called with nix-build
. Instead it must be
evaluated through the nix-shell
command in a rather cryptic
manner. mkShell
is used by envCommand
to provide a user-friendly
way to create a developmen shell.
Standard and advanced usage of this
function through the sde-env-*
commands has been described earlier.
This derivation contains a single command called sde-env-<version>
that launches a development shell.
This derivation contains a single command called installer.sh
. It is
a self-extracting archive that contains the envCommand
derivation
together with all its recursive runtime dependencies (its "closure" in
Nix-speak) for all supported platforms and kernels. The result of
executing the installer is the same as building the envCommand
derivation locally but does not require any network access. Its
purpose is to facilitate the installation of the SDE on hosts in
restricted environments.
The installer itself requires a number of Nix packages to be present
on the target system. To avoid a catch-22 situation, the standalone
target in the top-level Makefile
of the repository
creates an additional wrapper that installs these dependencies before
calling the installer itself.
The SDE comes with a set of example P4 programs. The Nix package supports the P416 example programs to provide a means to verify the proper working of the SDE and the PTF system. The example programs can be exercises as follows.
The test
attribute is itself a set with one attribute per supported
compiler target (currently tofino
, tofino2
and tofino3
where
tofino2
is supported for versions 9.7.0 and later and tofino3
is
supported for versions 9.11.0 and later). Each target attribute set is
composed of the following attributes
programs
. A set of P4 packages, one for each example programcases
. A set of derivations, one for each example program. Each derivation runs the PTF tests of the example program inside a VM. The resulting store path contains three log filesmodel.log
,switch.log
andtest.log
containing the outputs of therun_tofino_model.sh
,run_switchd.sh
andrun_p4_tests.sh
programs, respectively. It also contains a file calledpassed
containing a Nix expression with eithertrue
orfalse
depending on whether the PTF tests have passed successfully or not.failed-cases
. The subset ofcases
for which the PTF tests failed.
The names of the programs are those of the P4 source files located in
the p4_16_programs
sub directory of the p4-examples
SDE components.
The list of supported example programs can be displayed by evaluating
a simple Nix expression, for example for the most recent SDE version
and the tofino
target
$ nix eval '(with import ./. {}; builtins.attrNames bf-sde.latest.test.tofino.programs)'
[ "bri_handle" "bri_with_pdfixed_thrift" "tna_action_profile" "tna_action_selector" "tna_alpm" "tna_bridged_md" "tna_checksum" "tna_counter" "tna_custom_hash" "tna_digest" "tna_dkm" "tna_dyn_hashing" "tna_field_slice" "tna_id
letimeout" "tna_lpm_match" "tna_meter_bytecount_adjust" "tna_meter_lpf_wred" "tna_mirror" "tna_multicast" "tna_operations" "tna_pktgen" "tna_port_metadata" "tna_port_metadata_extern" "tna_ports" "tna_proxy_hash" "tna_pvs" "tn
a_random" "tna_range_match" "tna_register" "tna_resubmit" "tna_snapshot" "tna_symmetric_hash" "tna_ternary_match" "tna_timestamp" ]
To build all porgrams and run all tests for the latest version and the
tofino
target in one go, use
$ nix-build -A bf-sde.latest.test.tofino
To select a single test
$ nix-build -A bf-sde.latest.test.tofino.programs.tna_checksum
[ ... ]
/nix/store/hbsfjmyshrmbdwsj9hldqasgrrndr5ka-tna_checksum-0
$ nix-build -A bf-sde.latest.test.cases.tna_checksum
[ ... ]
/nix/store/rg64frv378cw5v6wr6j95457hw544qrk-bf-sde-9.4.0-test-case-tna_checksum
$ cat /nix/store/rg64frv378cw5v6wr6j95457hw544qrk-bf-sde-9.4.0-test-case-tna_checksum/passed
true
Kernel modules are required to support some of the features of the Tofino ASIC, for example to expose the CPU PCIe port as a Linux network interface. The modules have to be built to match the kernel on the host on which they will be loaded.
In general, compiling a kernel module requires the presence of the
directory /lib/modules/$(uname -r)/build
, where uname -r
provides
the release identifier of the running kernel. The build
directory
is an artifact of the build procedure of the kernel itself. It
contains everything needed to compile a module that will work with
that specific kernel.
How exactly a kernel is built and how the build directory is instantiated on a system depends heavily on the native package manager of a given Linux distribution. Since one of the purposes of the Nix packaging of the SDE is to gain independence of the native package manager of any particular Linux distribution, we need a mechanism that extends this independence to the compilation of kernel modules.
This is achieved by adding an abstraction layer to bf-sde-nixpkgs
which takes a set of native packages (and possibly other inputs which
are not available from the native package manager) of a given
distribution and creates a plain build directory from them in which
the SDE kernel modules can be compiled.
The list of supported kernels is kept in the attribute set defined in
bf-sde/kernels/default.nix
. The names of the attributes serve as
identifiers for the kernel. Each attribute must be a set with the
following attributes
-
kernelRelease
, requiredThe release identifier of the kernel as reported by
uname -r
-
buildTree
, requiredA derivation containing a ready-to use build tree for the modules to be compiled in
-
buildModulesOverrides
, optionalAn attribute set used to override the arguments of the derivation defined in
bf-sde/kernels/build-modules.nix
-
patches
, optionalAn attribute set of lists of patches to be applied to the kernel module source code. The name of each attribute must be either a SDE version number (e.g.
"9.4.0"
) orall
. The list of patches to apply is obtained by joining the list ofall
with that from the attribute that matches the SDE's version.The source code is a pristine copy of the source code used to build the
bf-drivers
package. Patches applied by thebf-drivers
derivation are not available here.
The package containing the kernel modules for a particular kernel is
built by the function defined in bf-sde/kernels/build-modules.nix
.
It uses the source of the bf-drivers
package to build only the
kdrv
component of it using the kernel build tree from the
buildTree
attribute. The resulting package contains scripts to load
and unload each kernel module and the kernel modules themselves in the
directory lib/modules/${kernelRelease}/
.
The non-trivial part of this procedure is how the buildTree
attribute is constructed for each kernel. The current version of
bf-sde-nixpkgs
supports three types of systems/distributions:
- OpenNetworkLinux (ONL)
- Plain Debian
- Mion
The Nix expression in bf-sde/kernels/default.nix
includes utility
functions for each type of system to construct the buildTree
attribute.
ONL is based on Debian but it uses a different method to package the
kernel than standard Debian. It already supplies the entire build
directory in a single deb
file. The file can be found in the ONL
build directory at the location
REPO/<debian-release>/packages/binary-amd64/onl-kernel-<version>-lts-x86-64-all_1.0.0_amd64.deb
where <debian-release>
is the name of the Debian release on which
the ONL image is based (e.g. stretch
or buster
) and <version>
is
the kernel version used in that image (e.g. 4.14 or 4.19). There is
no online repository where those .deb
files could be fetched from,
which is why they are included in the bf-sde-nixpkgs
repository
itself.
Plain Debian splits the contents of the build directory across three
separate deb
files (linux-headers
, linux-headers-common
and
linux-kbuild
) and also adds some non-generic processing, which have
to be converted back to the behavior of a generic kernel build
directory. The deb
files are all available from the standard Debian
mirrors.
Mion doesn't create any kind of packages that we could use. It stores
the kernel build artifacts in the build tree
build/tmp-glibc/work-shared/<machine>/kernel-build-artifacts
, but it
also requires access to the full kernel sources. The former must be
present in the bf-sde-nixpkgs
repository as a tar archive while the
latter is fetched from the Yocto kernel repository. The git commit
must match exactly the commit for the kernel from the version of
https://github.com/NetworkGradeLinux/meta-mion-bsp.git used to build
the mion image.
Problem statement: given the source code of a P4 program, compile it for a specific version of the SDE and create a package that runs it on a Tofino ASIC on a specific platform or the Tofino model by executing a single command.
The good news is that all of the real work to accomplish this is
already part of the SDE package. The buildP4Program
support
function does exactly what the problem statement
says.
The bad news is that we now have to write our own Nix expression to call that function :)
At this point, the reader should be at least a little bit familiar
with Nix expressions and derivations. The buildP4Program
function
has already been discussed earlier. In this
chapter, we show how it can be applied to an actual P4 program as an
example. For this demonstration, we chose the
packet-broker
program.
The following sections only demonstrate the basic usage of the SDE package. To see how a full-fledged deployment can look like, the reader is referred to the official packaging of the Packet Broker.
The packet broker consists of a P4 program called packet_broker.p4
,
located in the top-level directory of the Git repository.
To create a package for it, we first create a file packet-broker.nix
in the top-level directory of the bf-sde-nixpkgs
working tree with
the following contents
{ kernelRelease }:
let
pkgs = import ./. {};
packet-broker = pkgs.bf-sde.latest.buildP4Program {
pname = "packet-broker";
version = "0.1";
platform = "accton_wedge100bf_32x";
src = pkgs.fetchFromGitHub {
owner = "alexandergall";
repo = "packet-broker";
rev = "366999";
sha256 = "1rfm286mxkws8ra92xy4jwplmqq825xf3fhwary3lgvbb59zayr9";
};
p4Name = "packet_broker";
requiredKernelModule = "bf_kpkt";
};
in packet-broker.moduleWrapper kernelRelease
This is really all it takes. The rest of this chapter gives a more detailed explanation of what is going on behind the scenes.
Let's go over the most important elements of this expression. The
first line identifies the expression as a function which takes one
argument called kernelRelease
.
The let...in
construct binds expressions to names (i.e. it creates
local variables) and makes them available in the scope of the
following expression. The value of the variable pkgs
is the result
of evaluating the Nix expression in the file default.nix
(the ./.
path is expanded to ./default.nix
automatically). This is the
standard "boiler plate" found in almost all Nix expressions. It
imports the entire package collection as a single, huge attribute
set. Each attribute in the set represents a package in the collection
(more or less). In particular, the SDE packages for all supported
versions can now be accessed through the attribute bf-sde
of the
pkgs
set.
The object pkgs.bf-sde
is itself an attribute set whose attributes
are the version numbers of all available SDE versions in the form
v<major>_<minor>_<patch>
and an attribute latest
which is an alias
for the latest version of the SDE.
We can now understand how the value of the packet-broker
variable is
created: take the newest version of the SDE (pkgs.bf-sde.latest
) and
call its function buildP4Program
with the following attribute set as
argument. At this point we could have selected any of the supported
SDE versions to build our program with (e.g. bf-sde.v9_3_1
). The
result is a derivation (which, as a Nix expression, is also an
attribute set), which is assigned to the variable packet-broker
.
This particular invocation uses a subset of the arguments expected by
the buildP4Program
function. The most important
input is the packet-broker
Git repository, which is downloaded at
build-time using the
fetchFromGithub
utility function.
The platform
argument selects the platform for which to build the
program (accton_wedge100bf_32x
in this case). This determines which
BSP has to be included in the runtime environment.
The requiredKernelModule
argument indicates to the function that
this P4 program requires the bf_kpkt
kernel module to be present
when the program is started. However, the resulting package does not
contain that module. As explained earlier, this is because the module
must be compiled for the system on which the program will be run
rather than the kernel on which the package is built.
In typical Nix-style, the package created by buildP4Program
has the
capability to create a new package containing the required kernel
module based on itself. For that purpose, it has an attribute named
moduleWrapper
, very much like buildP4Program
is an attribute of
the SDE package. That attribute is a function and has been described
earlier. The function takes a kernel release string
as argument.
In this example we assume that the package is built on the same system
on which we want to use it. Therefore, we can roll these two steps
into one by calling moduleWrapper
immediately. The packet-broker
package becomes a run-time dependency of the final package
automatically.
To perform the actual build, simply pass our Nix expression to nix-build
$ nix-build packet-broker.nix
error: cannot auto-call a function that has an argument without a default value ('kernelRelease')
Well, that was to be expected. We need to somehow pass the local
kernel release to the function in test.nix
as an argument. That is
the purpose of the --argstr
option of nix-build
(it treats its
argument as a Nix string without having to quote it, as opposed to
using the --arg
option):
$ nix-build packet-broker.nix --argstr kernelRelease $(uname -r)
[ ... ]
/nix/store/n3mvk07nl25allcjm5vrm7yfnsma5zsz-packet_broker-module-wrapper
The user of the package doesn't have to understand anything described in this section. This is purely for the enjoyment of the curious reader.
Let's see what's inside the package we just created
$ ls -lR /nix/store/n3mvk07nl25allcjm5vrm7yfnsma5zsz-packet_broker-module-wrapper
/nix/store/n3mvk07nl25allcjm5vrm7yfnsma5zsz-packet_broker-module-wrapper:
total 4
dr-xr-xr-x 2 root root 4096 Jan 1 1970 bin
/nix/store/n3mvk07nl25allcjm5vrm7yfnsma5zsz-packet_broker-module-wrapper/bin:
total 4
-r-xr-xr-x 1 root root 898 Jan 1 1970 packet_broker-module-wrapper
It's a single shell script that loads the bf_kpkt
kernel module if
it's not already loaded and then starts the actual P4 program. Simply
executing this script is enough to launch the P4 program on the Tofino
ASIC. The last line
exec /nix/store/2aj27ji70991ij6g3pbvizczfcrazdw3-packet-broker-0.1/bin/packet_broker "$@"
references the package packet-broker
, which occurred as an
intermediary step during the evaluation of packet-broker.nix
. It has
now become a run-time dependency of the packet_broker-module-wrapper
package. We can see all the immediate run-time dependencies with
$ nix-store -q --references /nix/store/n3mvk07nl25allcjm5vrm7yfnsma5zsz-packet_broker-module-wrapper
/nix/store/0kcx6s8gxysnygd8kxa502xfdfm1n28y-gnugrep-3.4
/nix/store/a3fc4zqaiak11jks9zd579mz5v0li8bg-bash-4.4-p23
/nix/store/2aj27ji70991ij6g3pbvizczfcrazdw3-packet-broker-0.1
/nix/store/n599lhxiidv6fpiz43y2mld2nwnscc5s-kmod-27
/nix/store/d8jfwymqiiylcf3dl2pvj8ldx4c1jcnk-bf-sde-9.5.0-kernel-modules-4.19.0-16-amd64
/nix/store/g9qsf6rcy467dxa6gxdh4sw8wm5p6alg-gawk-5.1.0
Apart from the utilities required by the shell script and the
packet-broker
package, we can see an additional package containing
the kernel modules just for the local system (which happens to be a
Debian system in this example)
$ ls -l /nix/store/d8jfwymqiiylcf3dl2pvj8ldx4c1jcnk-bf-sde-9.5.0-kernel-modules-4.19.0-16-amd64/lib/modules/4.19.0-16-amd64/
total 16384
-r--r--r-- 1 nobody nogroup 34480 Jan 1 1970 bf_kdrv.ko
-r--r--r-- 1 nobody nogroup 62936 Jan 1 1970 bf_knet.ko
-r--r--r-- 1 nobody nogroup 16646848 Jan 1 1970 bf_kpkt.ko
Let's dig a bit deeper into the dependency tree. The immediate
dependencies of the packet-broker
package are
$ nix-store -q --references /nix/store/2aj27ji70991ij6g3pbvizczfcrazdw3-packet-broker-0.1
/nix/store/a3fc4zqaiak11jks9zd579mz5v0li8bg-bash-4.4-p23
/nix/store/w1m9bhgz32mwrfwx813krf85icknn3i7-packet_broker-artifacts-0.1
/nix/store/z2sngj8cl351chvsxxwd3pi6kp2nbg9g-bf-sde-accton-runtime-9.5.0
We have finally found the actual run-time environment provided by the
SDE package. This derivation is the result of calling the
runtimeEnv'
support function with the argument
accton_wedge100bf_32x
, which maps the platform identifier to the
baseboard identifier accton
(provided by the reference BSP). It's
dependencies are
$ nix-store -q --references /nix/store/z2sngj8cl351chvsxxwd3pi6kp2nbg9g-bf-sde-accton-runtime-9.5.0
/nix/store/hcrd7nv1ggayp4mw663aba0qb77s9mil-bf-sde-accton-runtime-env-9.5.0
/nix/store/4wgc1596a5i43wc1gny1djk3ndxsia9q-bf-tools-runtime-9.5.0
This is an example of a user environment introduced
earlier, a kind of meta-package. This means that it
provides no contents of its own. It merely collects the bin
, lib
etc. directories from the packages on which it depends in a single
hierarchy with symbolic links. In this case, the environment contains
two packages: bf-sde-accton-runtime-env-9.5.0
is itself an
environment and bf-tools-runtime-9.5.0
provides the runtime utility
scripts
$ ls -l /nix/store/4wgc1596a5i43wc1gny1djk3ndxsia9q-bf-tools-runtime-9.5.0/bin/
total 8
-r-xr-xr-x 1 root root 519 Jan 1 1970 run_bfshell.sh
-r-xr-xr-x 1 root root 825 Jan 1 1970 run_switchd.sh
These scripts are wrapped inside shell scripts that set up the enviornment, e.g.
$ grep SDE /nix/store/4wgc1596a5i43wc1gny1djk3ndxsia9q-bf-tools-runtime-9.5.0/bin/run_switchd.sh
export SDE='/nix/store/hcrd7nv1ggayp4mw663aba0qb77s9mil-bf-sde-accton-runtime-env-9.5.0'
export SDE_INSTALL='/nix/store/hcrd7nv1ggayp4mw663aba0qb77s9mil-bf-sde-accton-runtime-env-9.5.0'
The bf-sde-accton-runtime-env-9.5.0
package contains the runtime
version of the SDE itself, which, in turn, is composed of the set of
packages
$ nix-store -q --references /nix/store/hcrd7nv1ggayp4mw663aba0qb77s9mil-bf-sde-accton-runtime-env-9.5.0
/nix/store/4qnl1kirw5jh5jh3ndxhwyvvzx9bwawx-bf-sde-misc-components-9.5.0
/nix/store/jl2pqcxp8qq8j3ljkkjrbd842ncc1vxq-bf-syslibs-9.5.0
/nix/store/hjfrx08hd4az8wn1zc0zvgixxza30y05-bf-utils-9.5.0
/nix/store/p8iky1fwj6q5gbk519ccvba9vlsb6636-bf-platforms-accton-9.5.0
/nix/store/xz3v5fqdgpil50faa1963agjmb9swj1k-ptf-utils-9.5.0
/nix/store/y70c74xfh5gnxaaask1hf4pzlcgjlbkk-bf-drivers-runtime-9.5.0
This is how everything comes together in the end. It can't be stressed
enough that all of this is done automatically when nix-build packet-broker.nix
is executed. Missing dependencies, for example,
cannot happen with Nix.
For illustration purposes, let's see how we can modify the build recipe to create a version that runs on the Tofino ASIC emulation instead of actual hardware. We need to make only a few modifications
- Select
model
as the target platform - Remove
requiredKernelModule
- Don't build a wrapper
The following expression does this:
let
pkgs = import ./. {};
packet-broker = pkgs.bf-sde.latest.buildP4Program {
pname = "packet-broker";
version = "0.1";
platform = "model";
src = pkgs.fetchFromGitHub {
owner = "alexandergall";
repo = "packet-broker";
rev = "366999";
sha256 = "1rfm286mxkws8ra92xy4jwplmqq825xf3fhwary3lgvbb59zayr9";
};
p4Name = "packet_broker";
};
in packet-broker
After building
$ nix-build packet-broker.nix
[...]
/nix/store/x61bsdzhz21axlpxgrf2q2kn6yrkr59f-packet-broker-0.1
it can be started just like before. The script creates veth
interfaces to provide access to the virtual ports, starts the Tofino
model in the background and finally launches bf_switchd
.
BfRuntime is a customized version of
p4runtime. The SDE contains
p4runtime
as third-party software in the bf-drivers
source package
but it is not built by default and it is not included in the Nix
package. In practice, BfRuntime is the primary interface between the
control- and data-plane components.
The bf-drivers
package contains the Python bindings derived from the
Google protobuf
specification of BfRuntime. It also contains a Python library called
bfrt_grpc
to be used by gRPC clients. This chapter explains how to
build a package for control-plane code which depends on bfrt_grpc
.
There are two things that the packaging mechanism needs to take care
of. The first is obvious: the location of bfrt_grpc
must be added
to the module search path in order for the control-plane code to be
able to import the module. The second is a bit more intricate: the
current bfrt_grpc
code (at least up to SDE 9.5.0) requires Python
2.7. Therefore, any code using the library must also be restricted to
that version. The packaging should make sure that this condition is
satisfied (this also implies that the program doesn't make use of any
features not available in Python 2.7)
To illustrate how such a package could look like we make use of the packet-broker once again.
The code for the broker's control-plane is located in the
control-plane
directory. The configd.py
script is the main program intended to
run as a daemon. An additional program brokerctl
connects to the
daemon to interact with the control-plane from the command line
(e.g. to initiate a reload of the configuration, see the
documentation
for details). The configd.py
script needs the jsonschema
and
ipaddress
modules as well as the bfrt_grpc
module provided by the
bf-drivers
package at runtime. The daemon also uses a configuration
file in JSON and a schema to validate it.
Here is the Nix expression we're going to use to create the package
let
pkgs = import ./. {};
bf-drivers-runtime = pkgs.bf-sde.latest.pkgs.bf-drivers-runtime;
python = bf-drivers-runtime.pythonModule;
in python.pkgs.buildPythonApplication {
pname = "packet-broker-configd";
version = "0.1";
src = pkgs.fetchFromGitHub {
owner = "alexandergall";
repo = "packet-broker";
rev = "366999";
sha256 = "1rfm286mxkws8ra92xy4jwplmqq825xf3fhwary3lgvbb59zayr9";
};
propagatedBuildInputs = [
bf-drivers-runtime
] ++ (with python.pkgs; [ jsonschema ipaddress ]);
preConfigure = ''cd control-plane'';
postInstall = ''
mkdir -p $out/etc/packet-broker
cp config.json schema.json $out/etc/packet-broker
'';
}
We store it in the file configd.nix
, again in the top-level
directory of bf-sde-nixpkgs
. The pkgs
and src
elements are the
same as for the P4 package. The line
bf-drivers-runtime = pkgs.bf-sde.latest.pkgs.bf-drivers-runtime;
selects the bf-drivers-runtime
package from the newest SDE
version. That package used a specific Python interpreter to build the
bfrt_grpc
module, as mentioned in the introduction. It actually
makes that interpreter available through an attribute called
pythonModule
. The assignment
python = bf-drivers-runtime.pythonModule;
picks it up to invoke the package build procedure later on. This is
how we satisfy the constraint that our control-plane code uses the
correct Python version if it imports the bfrt_grpc
module.
The rest of the expression is really just the application of the
standard Nix tooling for
Python and cannot be
covered here in detail. In a nutshell, it uses setuptools
with the
bdist_wheel
method to create the scripts and modules specified by
setup.py
.
Note that the buildPythonApplication
in relation to the Python
package referenced by python
works very much like the
buildP4Program
function in relation to the SDE package as we've seen
in the previous chapter, i.e. it creates a package (our control-plane
program) based on another package (a specific Python interpreter).
One thing to note is how run-time dependencies of the package are
declared with the propagatedBuildInputs
attribute. Normally,
run-time dependencies are not specified explicitly for derivations in
Nix (they are determined automatically when the package is built).
However, this doesn't work with Python modules. What happens here,
essentially, is that Nix creates a Python environment containing the
modules specified by porpagatedBuildInputs
and arranges for the
application (the configd.py
script in this case) to be executed in
that environment. For more information on this, refer to the section
on specifying
dependencies
and Nix pill
20.
This works exactly the same as for the P4 package
$ nix-build configd.nix
[ ... ]
/nix/store/ap8gfdykxdx7154asxyphhrqizjfbz47-packet-broker-configd-0.1
The build procedures for the P4 and control-plane packages detailed in the previous chapters are sufficient to make them usable. For instance, the packet-broker P4 program can be started simply by executing
$ /nix/store/mkcbykv8rd0giwkc9q58q27hijn09rjn-packet_broker-module-wrapper/bin/packet_broker-module-wrapper
This is true for any Nix package: they can all be used directly from
/nix/store
. We could also use those paths directly in a systemd
unit file to create a service. However, Nix offers a better mechanism
on top of the bare packages which also offers additional benefits.
This mechanism is called a profile. It is very similar to the user environment we have already seen multiple times. In fact, a profile is simply a collection of user environments organized as a sequence of profile generations.
Let's see what that means with our packet broker example. First we are
going to merge configd.nix
with packet-broker.nix
in a file
pb.nix
:
{ kernelRelease }:
let
pkgs = import ./. {};
src = pkgs.fetchFromGitHub {
owner = "alexandergall";
repo = "packet-broker";
rev = "366999";
sha256 = "1rfm286mxkws8ra92xy4jwplmqq825xf3fhwary3lgvbb59zayr9";
};
bf-sde = pkgs.bf-sde.latest;
version = "0.1";
packet-broker = bf-sde.buildP4Program {
pname = "packet-broker";
platform = "accton_wedge100bf_32x";
inherit version src;
p4Name = "packet_broker";
requiredKernelModule = "bf_kpkt";
};
bf-drivers-runtime = bf-sde.pkgs.bf-drivers-runtime;
python = bf-drivers-runtime.pythonModule;
configd = python.pkgs.buildPythonApplication {
pname = "packet-broker-configd";
inherit version src;
propagatedBuildInputs = [
bf-drivers-runtime
] ++ (with python.pkgs; [ jsonschema ipaddress ]);
preConfigure = ''cd control-plane'';
postInstall = ''
mkdir -p $out/etc/packet-broker
cp config.json schema.json $out/etc/packet-broker
'';
};
in {
packet-broker = packet-broker.moduleWrapper kernelRelease;
inherit configd;
}
The new expression evaluates to a set with attributes packet-broker
and configd
. This allows us to create both packages with a single
invocation of nix-build
$ nix-build pb.nix --argstr kernelRelease $(uname -r)
/nix/store/ap8gfdykxdx7154asxyphhrqizjfbz47-packet-broker-configd-0.1
/nix/store/9j1bpjif9qcnzzy4jrh9bdk17vwp9r4c-packet_broker-module-wrapper
Next we're going to create a profile containing these two packages. This could be done by any user, but a good choice for a global installation is to create it as root
# nix-env -f pb.nix -p /nix/var/nix/profiles/packet-broker -i -r --argstr kernelRelease $(uname -r)
building '/nix/store/dly3kq5nsaz9sxqz62hfrvn7hgwcd4q2-user-environment.drv'...
created 6 symlinks in user environment
A profile can be located anywhere but unless it's somewhere underneath
/nix/var/nix/profiles
, it won't become a garbage collection
root. So, what have we got now?
$ ls -l /nix/var/nix/profiles/packet-broker*
lrwxrwxrwx 1 root root 20 Jun 11 16:24 /nix/var/nix/profiles/packet-broker -> packet-broker-1-link
lrwxrwxrwx 1 root root 60 Jun 11 16:24 /nix/var/nix/profiles/packet-broker-1-link -> /nix/store/9ij1vlfn43dvch3ddwzn7j40djrbnwkm-user-environment
This nicely collects the artifacts of the packages in a single location. It also provides easy rollback to previous versions.
Nix creates a new generation of the profile each time we execute
nix-env -i
with a new version of the packages. Generations are never
deleted automatically. As a side-effect, any package referred by a
profile is protected from deletion, irrespective of whether the
generation of the profile is the current one or not.
What this means is the following. Nix treats packages like a memory management system using a garbage collector. It keeps a list of garbage collection roots and treats every package as being alive which is referenced by such a root (directly or indirectly).
The command nix-collect-garbage
(which can be called by any user)
deletes all packages from /nix/store
which are not alive and
nix-store --delete <path>
deletes a specific package unless it is
alive or is a dependency of another package. Those are actually the
only ways to remove anything from the Nix store.
One way to create a garbage collection root is with nix-build
. In
our example, when we executed, for example, nix-build packet-broker.nix
, you might have noticed that there appears a
symbolic link called result
in the current directory pointing to the
package
$ nix-build packet-broker.nix --argstr kernelRelease $(uname -r)
/nix/store/xyv33rlrk46yw9ix1kfdjr4i09s6j2bj-packet_broker-module-wrapper
gall@spare-PB1:~/bf-sde-nixpkgs$ ls -l result
lrwxrwxrwx 1 gall gall 72 Apr 27 14:51 result -> /nix/store/xyv33rlrk46yw9ix1kfdjr4i09s6j2bj-packet_broker-module-wrapper
This link is registered as a garbage collection root. So an attempt to delete it fails
$ nix-store --delete /nix/store/xyv33rlrk46yw9ix1kfdjr4i09s6j2bj-packet_broker-module-wrapper
finding garbage collector roots...
0 store paths deleted, 0.00 MiB freed
error: cannot delete path '/nix/store/xyv33rlrk46yw9ix1kfdjr4i09s6j2bj-packet_broker-module-wrapper' since it is still alive
The second way to create a garbage collection root is through profiles. Every generation of a profile is automatically registered as a root. In our case
$ nix-store --gc --print-roots | grep packet-broker
/nix/var/nix/profiles/packet-broker-1-link -> /nix/store/d29aiqfr74rcdghvp2p3f3aqxj2lccmd-user-environment
Generations of a profile can be manipulated with the
--list-generations
, --switch-generation
and --delete-generations
sub-commands of nix-env
.
If storage is a concern, you should make sure to remove profile
generations which are no longer needed and run nix-collect-garbage
to remove unneeded packages from /nix/store
.
I this section, the term deployment specifically refers to the
mechanism of instantiating the packages required for a particular
service in /nix/store
. In the packet broker example, this would be
the packages packet_broker-module-wrapper
and configd
and all
their recursive runtime dependencies.
In theory, a pure source deployment of all packages is possible but
clearly not practical. Nix uses a concept called substitution to
make use of pre-built packages. Before a package is actually built,
Nix first creates a description of the build process called
derivation. The derivation contains, among other things, the
so-called output path of the package, which is the location in
/nix/store
where the final package will be stored, for example
/nix/store/bsz947wniwh1wrwb5dn45h6kgvs8wssa-packet_broker-module-wrapper
This information is know before the package is built. Nix can be
configured with URLs pointing to servers, each of which provides a
/nix/store
with pre-built packages called binary caches. If this
is done, the build process will check whether the output path already
exist on any of these caches before building the package. If found,
it simply fetches the package from the remote host and adds it to the
local /nix/store
. This process is known as substitution, because
the nature of Nix guarantees that if the hash in the store path on the
remote system is the same as we expect to build locally, the packages
must be identical.
Any standard installation of Nix uses at least one such binary cache,
usually https://cache.nixos.org. This cache serves all packages built
from official releases of the nixpkgs
collection.
This mechanism essentially turns the source-deployment into binary deployment.
The bf-sde-nixpkgs
repository uses one of these nixpkgs
releases
as the base system for the SDE packages. This can be seen in default.nix
:
{ overlays ? [], ... } @attrs:
let
nixpkgs = (fetchTarball https://github.com/NixOS/nixpkgs/archive/20.09-1181-gfee7f3fcb41.tar.gz);
in import nixpkgs ( attrs // {
overlays = import ./overlay.nix ++ overlays;
})
In this case, it uses commit fee7f3
on the 20.09
release branch.
As a consequence, most of the packages can be substituted from the
standard binary cache. However, all SDE-specific packages as well as
those whose build recipes are overridden by bf-sde-nixpkgs
(e.g. to
change build options or up- or down grade versions) are not available
from the cache and need to be built from source. This limitation can
be mitigated by building a separate binary cache for these packages.
This is really a hybrid between source and binary deployment.
This is the most direct deployment model and the one we have been
using in the packet broker example up to now. In this model, we start
with a specific version of the the bf-sde-nixpkgs
Git repository and
build the desired packages either with nix-build
(or nix-env
in
case we want to create a Nix profile as well). As mentioned in the
introduction to this section, this is really just a source deployment
with regard to the SDE packages (and modified standard packages). All
standard packages are fetched from the binary cache.
Apart from the time and hardware resources it takes to build the SDE from source, this model also poses a legal problem, because it requires access to the SDE itself, which is currently only possible by entering an NDA with Intel. This would make it impossible for third parties to use the packages.
Nix makes pure binary deployment of packages very easy due to its
declarative nature and strict dependency management. Key to this is
the concept of the closure of a package. The closure is the
complete set of recursive dependencies of the package. In our
example, the closure of the packet-broker
package can
be obtained with
$ nix-store -qR $(nix-build pb.nix --argstr kernelRelease $(uname -r))
[...]
/nix/store/jqnprhfrsbl2girajpwhcv45qd8ij5lv-procps-3.3.16
/nix/store/yvl77j0zv2jdyblsi4h5ai0zr4q8l9kw-packet-broker-0.1
/nix/store/xyv33rlrk46yw9ix1kfdjr4i09s6j2bj-packet_broker-module-wrapper
This closure contains around 100 packages. It sounds like a lot but
remember that it contains the indirect dependencies as well.
Furthermore it is completely self-contained: there are no dependencies
outside /nix/store
. Also, many of the packages in the closure are
shared by other packages. The closure has an extremely useful
property: if it is copied to /nix/store
on any system with Nix
installed, the package is guaranteed to work.
Nix provides a method to extract the entire closure as a single file:
$ nix-store --export $(nix-store -qR $(nix-build pb.nix --argstr kernelRelease $(uname -r))) >closure
To install it to /nix/store
on another system, copy closure
and
execute
# nix-store --import <closure
Note that this has to be done by the root
user unless the closure is
digitally signed (which is not covered here).
Due to the manner in which the SDE packages have been constructed, this closure does not contain any components which would fall under the NDA with Intel, hence it can be safely distributed to third parties.
Also note that the Nix expression language is not used when installing
the closure on the target system, i.e. the Nix expression that defines
our application (from the pb.nix
file in this example) is not needed
at all.
Explicit distribution of the closure as described in the previous chapter may work in single-instance or small-scale deployments but could become somewhat unpractical on a larger scale. Nix supports a standard mechanism to use a Nix store on a remote server as a repository for pre-built packages, referred to as a binary cache. How such a cache is populated and made accessible is outside the scope of this documentation. The reader is referred to the NixOS wiki and the Cachix project.
For the purpose of SDE-based packages, it is assumed that the provider of a particular service like the packet broker, has set up such a cache and populated it with the closure discussed previously. It is important that the Nix store on that cache does not contain any of the SDE components covered by the NDA.
Once that cache is set up, the administrator creates a key-pair used to
sign the packages being downloaded from the cache. The public key is
published together with the URL where the cache can be reached, for
example https://cache.exmple.net. To use the cache, add the following
lines to /etc/nix/nix.conf
extra-substituters = https://cache.example.net
trusted-substituters = https://cache.example.net
trusted-public-keys = cache.nixos.org-1:6NCHdD59X431o0gWypbMrAURkbJ16ZPMQFGspcDShjY= cache.example.net:<public key>
Then restart the nix-daemon
server with systemctl restart nix-daemon
for the change to take effect. Note that the standard
binary cache https://cache.nixos.org is registered by default and so
is it's public key. But when we add the key of the new cache to
trusted-public-keys
, we also have to specify the key for
cache.nixos.org
to preserve the default.
Once this is done, we can execute the nix-build
and nix-env
commands discussed before to deploy the
service and Nix will fetch the missing parts of the closure from the
cache.