INSTALL

Installation instructions for HAProxy
=====================================

This is a development version, so it is expected to break from time to time,
to add and remove features without prior notification and it should not be used
in production, unless you're an experienced user and are willing to follow
weekly updates. If you are not used to build from sources or if you are not
used to follow updates then it is recommended that instead you use the packages
provided by your software vendor or Linux distribution. Most of them are taking
this task seriously and are doing a good job at backporting important fixes.

If for any reason you'd prefer to use a different version than the one packaged
for your system, you want to be certain to have all the fixes or to get some
commercial support, other choices are available at http://www.haproxy.com/.


Areas covered in this document
==============================

1) Quick build & install
2) Basic principles
3) Build environment
4) Dependencies
5) Advanced build options
6) How to install HAProxy


1) Quick build & install
========================

If you've already built HAProxy and are just looking for a quick reminder, here
are a few build examples :

  - recent Linux system with all options, make and install :
    $ make clean
    $ make -j $(nproc) TARGET=linux-glibc \
                USE_OPENSSL=1 USE_LUA=1 USE_PCRE2=1 USE_SYSTEMD=1
    $ sudo make install

  - FreeBSD and OpenBSD, build with all options :
    $ gmake -j 4 TARGET=freebsd USE_OPENSSL=1 USE_LUA=1 USE_PCRE2=1

  - embedded Linux, build using a cross-compiler :
    $ make -j $(nproc) TARGET=linux-glibc USE_OPENSSL=1 USE_PCRE2=1 \
                CC=/opt/cross/gcc730-arm/bin/gcc ADDLIB=-latomic

  - Build with static PCRE on Solaris / UltraSPARC :
    $ make TARGET=solaris CPU_CFLAGS="-mcpu=v9" USE_STATIC_PCRE2=1

For more advanced build options or if a command above reports an error, please
read the following sections.


2) Basic principles
===================

HAProxy uses a single GNU Makefile which supports options on the command line,
so that there is no need to hack a "configure" file to work on your system. The
makefile totally supports parallel build using "make -j <jobs>" where <jobs>
matches the number of usable processors, which on some platforms is returned by
the "nproc" utility. The explanations below may occasionally refer to some
options, usually in the form "name=value", which have to be passed to the
command line. This means that the option has to be passed after the "make"
command. For example :

  $ make -j $(nproc) TARGET=generic USE_GZIP=1

One required option is TARGET, it must be set to a target platform name, which
provides a number of presets. The list of known platforms is displayed when no
target is specified. It is not strictly required to use the exact target, you
can use a relatively similar one and adjust specific variables by hand.

Most configuration variables are in fact booleans. Some options are detected and
enabled by default if available on the target platform. This is the case for all
those named "USE_<feature>". These booleans are enabled by "USE_<feature>=1"
and are disabled by "USE_<feature>=" (with no value) or "USE_<feature>=0". An
exhaustive list of the supported USE_* features is located at the top of the
main Makefile. The last occurrence of such an option on the command line
overrides any previous one. Example :

  $ make TARGET=generic USE_THREAD=

In case of error or missing TARGET, a help screen is displayed. It is also
possible to display a list of all known options using "make help".

Some optional components which may depend on third-party libraries, are used
with popular tools which are not necessarily standard implementations, or are
maintained at slower pace than the core of the project, are located in the
"addons/" directory. These ones may disappear in a future version if the
product they depend on disappears or if their maintainers do not assign enough
resources to maintain them any more. For this reason they are not built by
default, but some USE_* options are usually provided for them, and their build
is routinely tested anyway.


3) Build environment
====================

HAProxy requires a working GCC or Clang toolchain and GNU make :

  - GNU make >= 3.80. Note that neither Solaris nor OpenBSD's make work with
    the GNU Makefile. If you get many syntax errors when running "make", you
    may want to retry with "gmake" which is the name commonly used for GNU make
    on BSD systems.

  - GCC >= 4.2 (up to 13 tested). Older versions can be made to work with a
    few minor adaptations if really needed. Newer versions may sometimes break
    due to compiler regressions or behaviour changes. The version shipped with
    your operating system is very likely to work with no trouble. Clang >= 3.0
    is also known to work as an alternative solution. Recent versions may emit
    a bit more warnings that are worth reporting as they may reveal real bugs.
    TCC (https://repo.or.cz/tinycc.git) is also usable for developers but will
    not support threading and was found at least once to produce bad code in
    some rare corner cases (since fixed). But it builds extremely quickly
    (typically half a second for the whole project) and is very convenient to
    run quick tests during API changes or code refactoring.

  - GNU ld (binutils package), with no particular version. Other linkers might
    work but were not tested.

On debian or Ubuntu systems and their derivatives, you may get all these tools
at once by issuing the two following commands :

  $ sudo apt-get update
  $ sudo apt-get install build-essential

On Fedora, CentOS, RHEL and derivatives, you may get the equivalent packages
with the following command :

  $ sudo yum groupinstall "Development Tools"

Please refer to your operating system's documentation for other systems.

It is also possible to build HAProxy for another system or platform using a
cross-compiler but in this case you probably already have installed these
tools.

Building HAProxy may require between 60 and 80 MB of free space in the
directory where the sources have been extracted, depending on the debugging
options involved.


4) Dependencies
===============

HAProxy in its basic form does not depend on anything beyond a working libc.
However a number of options are enabled by default, or are highly recommended,
and these options will typically involve some external components or libraries,
depending on the targeted platform.

Optional dependencies may be split into several categories :

  - memory allocation
  - regular expressions
  - multi-threading
  - password encryption
  - cryptography
  - compression
  - lua
  - device detection
  - miscellaneous


4.1) Memory allocation
----------------------
By default, HAProxy uses the standard malloc() call provided by the libc. It
may also be built to use jemalloc, which is fast and thread-safe. In order to
use it, please add "-ljemalloc" to the ADDLIB variable. You may possibly also
need to append "-lpthread" and/or "-ldl" depending on the operating system.


4.2) Regular expressions
------------------------
HAProxy may make use regular expressions (regex) to match certain patterns. The
regex engine is provided by default in the libc. On some operating systems, it
might happen that the original regex library provided by the libc is too slow,
too limited or even bogus. For example, on older Solaris versions up to 8, the
default regex used not to properly extract group references, without reporting
compilation errors. Also, some early versions of the GNU libc used to include a
regex engine which could be slow or even crash on certain patterns.

If you plan on importing a particularly heavy configuration involving a lot of
regex, you may benefit from using some alternative regex implementations such as
PCRE. HAProxy natively supports PCRE and PCRE2 (recommended), both in standard
and JIT flavors (Just In Time). The following options are available depending on
the library version provided on your system :

  - "USE_PCRE=1"         : enable PCRE version 1, dynamic linking
  - "USE_STATIC_PCRE=1"  : enable PCRE version 1, static linking
  - "USE_PCRE_JIT=1"     : enable PCRE version 1 in JIT mode
  - "USE_PCRE2=1"        : enable PCRE version 2, dynamic linking
  - "USE_STATIC_PCRE2=1" : enable PCRE version 2, static linking
  - "USE_PCRE2_JIT=1"    : enable PCRE version 2 in JIT mode

Both of these libraries may be downloaded from https://www.pcre.org/.

By default, the include and library paths are figured from the "pcre-config"
and "pcre2-config" utilities. If these ones are not installed or inaccurate
(for example when cross-compiling), it is possible to force the path to include
files using "PCRE_INC" and "PCRE2_INC" respectively, and the path to library
files using "PCRE_LIB" and "PCRE2_LIB" respectively. For example :

  $ make TARGET=generic \
    USE_PCRE2_JIT=1 PCRE2_INC=/opt/cross/include PCRE2_LIB=/opt/cross/lib


4.3) Multi-threading
--------------------
On some systems for which positive feedback was reported, multi-threading will
be enabled by default. When multi-threading is used, the libpthread library
(POSIX threading) will be used. If the target system doesn't contain such a
library, it is possible to forcefully disable multi-threading by adding
"USE_THREAD=" on the command line.


4.4) Password encryption
------------------------
Many systems provide password encryption functions used for authentication. On
some systems these functions are part of the libc. On others, they're part of a
separate library called "libcrypt". The default targets are pre-configured
based on which system needs the library. It is possible to forcefully disable
the linkage against libcrypt by adding "USE_LIBCRYPT=" on the command line, or
to forcefully enable it using "USE_LIBCRYPT=1".


4.5) Cryptography
-----------------
For SSL/TLS, it is necessary to use a cryptography library. HAProxy currently
supports the OpenSSL library, and is known to build and work with branches
1.0.0, 1.0.1, 1.0.2, 1.1.0, 1.1.1, 3.0, 3.1 and 3.2. It is recommended to use
at least OpenSSL 1.1.1 to have support for all SSL keywords and configuration
in HAProxy. OpenSSL follows a long-term support cycle similar to HAProxy's,
and each of the branches above receives its own fixes, without forcing you to
upgrade to another branch. There is no excuse for staying vulnerable by not
applying a fix available for your version. There is always a small risk of
regression when jumping from one branch to another one, especially when it's
very new, so it's preferable to observe for a while if you use a different
version than your system's defaults. Specifically, it has been well established
that OpenSSL 3.0 can be 2 to 20 times slower than earlier versions on
multiprocessor systems due to design issues that cannot be fixed without a
major redesign, so in this case upgrading should be carefully thought about
(please see https://github.com/openssl/openssl/issues/20286 and
https://github.com/openssl/openssl/issues/17627). If a migration to 3.x is
mandated by support reasons, at least 3.1 recovers a small fraction of this
important loss.

Four OpenSSL derivatives called LibreSSL, BoringSSL, QUICTLS, and AWS-LC are
reported to work as well. While there are some efforts from the community to
ensure they work well, OpenSSL remains the primary target and this means that
in case of conflicting choices, OpenSSL support will be favored over other
options.  Note that QUIC is not fully supported when haproxy is built with
OpenSSL. In this case, QUICTLS is the preferred alternative.  As of writing
this, the QuicTLS project follows OpenSSL very closely and provides update
simultaneously, but being a volunteer-driven project, its long-term future does
not look certain enough to convince operating systems to package it, so it
needs to be build locally. See the section about QUIC in this document.

A fifth option is wolfSSL (https://github.com/wolfSSL/wolfssl). It is the only
supported alternative stack not based on OpenSSL, yet which implements almost
all of its API and natively supports QUIC. At the time of writing, the vast
majority of SSL features are well supported by wolfSSL though not everything is
exposed in haproxy yet, advanced users might notice tiny differences that the
wolfSSL and HAProxy teams are working on together to address in the wolfSSL
code base. Features like ecdsa/rsa dual stack, crt-list and client auth might
not work as expected. As of November 2023, wolfSSL support is considered
experimental. This stack is not affected by OpenSSL's design issue regarding
multi-processor systems and is viewed by the HAProxy team as the most promising
mid-term solution for general deployments and QUIC deployments.

In order to enable SSL/TLS support, simply pass "USE_OPENSSL=1" on the command
line and the default library present on your system will be used :

  $ make TARGET=generic USE_OPENSSL=1

If you want to use a different version from the one provided by your system
(which is not recommended due to the risk of missing security fixes), it is
possible to indicate the path to the SSL include files using SSL_INC, and the
SSL library files using SSL_LIB. Example :

  $ make TARGET=generic \
    USE_OPENSSL=1 SSL_INC=/opt/ssl-1.1.1/include SSL_LIB=/opt/ssl-1.1.1/lib

To use HAProxy with WolfSSL, WolfSSL must be built with haproxy support, at
least WolfSSL 5.6.6 is needed, but a development version might be needed for
some of the features:

  $ cd ~/build/wolfssl
  $ ./configure --enable-haproxy --enable-quic --prefix=/opt/wolfssl-5.6.6/
  $ make -j $(nproc)
  $ make install

Please also note that wolfSSL supports many platform-specific features that may
affect performance, and that for production uses it might be a good idea to
check them using "./configure --help". Please refer to the lib's documentation.

When running wolfSSL in chroot, either mount /dev/[u]random devices into the
chroot:

  $ mkdir -p /path/to/chrootdir/dev/
  $ mknod -m 444 /path/to/chrootdir/dev/random c 1 8
  $ mknod -m 444 /path/to/chrootdir/dev/urandom c 1 9

Or, if your OS supports it, enable the getrandom() syscall by appending the
following argument to the wolfSSL configure command:

  EXTRA_CFLAGS=-DWOLFSSL_GETRANDOM=1

Building HAProxy with wolfSSL requires to specify the API variant on the "make"
command line, for example:

  $ cd ~/build/haproxy
  $ make -j $(nproc) TARGET=generic USE_OPENSSL_WOLFSSL=1 USE_QUIC=1 \
    SSL_INC=/opt/wolfssl-5.6.6/include SSL_LIB=/opt/wolfssl-5.6.6/lib

To use HAProxy with AWS-LC you must have version v1.13.0 or newer of AWS-LC
built and installed locally.
 $ cd ~/build/aws-lc
 $ cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/opt/aws-lc
 $ make -j $(nproc)
 $ make install

Building HAProxy with AWS-LC requires you to enable AWS-LC support, and specify
the path it was installed to when running make for HAPRoxy.

 $ cd ~/build/haproxy
 $ make -j $(nproc) TARGET=generic USE_OPENSSL_AWSLC=1 \
   SSL_INC=/opt/aws-lc/include SSL_LIB=/opt/aws-lc/lib

In order to link OpenSSL statically against HAProxy, first download OpenSSL
from https://www.openssl.org/ then build it with the "no-shared" keyword and
install it to a local directory, so your system is not affected :

  $ export STATICLIBSSL=/tmp/staticlibssl
  $ ./config --prefix=$STATICLIBSSL no-shared
  $ make && make install_sw

Then when building haproxy, pass that path via SSL_INC and SSL_LIB :

  $ make TARGET=generic \
    USE_OPENSSL=1 SSL_INC=$STATICLIBSSL/include SSL_LIB=$STATICLIBSSL/lib

When building with OpenSSL on some systems, you may also need to enable support
for the "libz" library, which is visible if the linker complains about function
"deflateInit()" not being found. In this case, simply append "ADDLIB=-lz" to
the command line.

It is worth mentioning that asynchronous cryptography engines are supported on
OpenSSL 1.1.0 and above. Such engines are used to access hardware cryptography
acceleration that might be present on your system. Due to API changes that
appeared with OpenSSL 3.0 and cause lots of build warnings, engines are not
enabled by default anymore in HAProxy 2.6. It is required to pass USE_ENGINE=1
if they are desired.

If for any reason you are forced to use OpenSSL 3.x and the performance is not
acceptable at all, you may want to try replacing the pthread locks that OpenSSL
uses with HAProxy's much lighter locks that are able to emulate them:

  $ make TARGET=generic \
    USE_OPENSSL=1 USE_PTHREAD_EMULATION=1

On large multi-processor systems, this may result in a performance increase of
50 to 100% on OpenSSL 3.0 depending on the level of contention, but this will
of course not recover everything. It should not be used by distro packagers as
it is a bit less observable.


4.6) Compression
----------------
HAProxy can compress HTTP responses before delivering them to clients, in order
to save network bandwidth. Two compression options are available. The first one
relies on the libslz library (http://libslz.org) that is embedded in haproxy.
It is enabled by default as it is very fast and does not keep a copy of the
contents in memory. It is possible to disable it, for example for very small
systems, by passing "USE_SLZ=" to the "make" command.

Please note that SLZ will benefit from some CPU-specific instructions like the
availability of the CRC32 extension on some ARM processors. Thus it can further
improve its performance to build with "CPU=native" on the target system, or
"CPU=armv81" (modern systems such as Graviton2 or A55/A75 and beyond),
"CPU=a72" (e.g. for RPi4, or AWS Graviton), "CPU=a53" (e.g. for RPi3), or
"CPU=armv8-auto" (automatic detection with minor runtime penalty).

A second option involves the widely known zlib library, which is very likely
installed on your system. In order to use zlib, simply pass "USE_ZLIB=1" to the
"make" command line, which will also automatically disable SLZ. If the library
is not installed in your default system's path, it is possible to specify the
path to the include files using ZLIB_INC, and the path to the library files
using ZLIB_LIB :

  $ make TARGET=generic \
    USE_ZLIB=1 ZLIB_INC=/opt/zlib-1.2.11/include ZLIB_LIB=/opt/zlib-1.2.11/lib

Zlib is commonly found on most systems, otherwise updates can be retrieved from
http://www.zlib.net/. It is easy and fast to build, and new versions sometimes
provide better performance so it might be worth using an up-to-date one.

Zlib compresses a bit better than libslz but at the expense of more CPU usage
(about 3.5 times more minimum), and a huge memory usage (~260 kB per compressed
stream). The only valid reason for uzing Zlib instead of SLZ here usually is to
deal with a very limited internet bandwidth while CPU and RAM are abundant so
that the last few percent of compression ratio are worth the invested hardware.


4.7) Lua
--------
Lua is an embedded programming language supported by HAProxy to provide more
advanced scripting capabilities. Only versions 5.3 and above are supported.
In order to enable Lua support, please specify "USE_LUA=1" on the command line.
Some systems provide this library under various names to avoid conflicts with
previous versions. By default, HAProxy looks for "lua5.4", "lua54", "lua5.3",
"lua53", "lua". If your system uses a different naming, you may need to set the
library name in the "LUA_LIB_NAME" variable.

If Lua is not provided on your system, it can be very simply built locally. It
can be downloaded from https://www.lua.org/, extracted and built, for example :

  $ cd /opt/lua-5.4.6
  $ make linux

The path to the include files and library files may be set using "LUA_INC" and
"LUA_LIB" respectively. For example :

  $ make TARGET=generic \
    USE_LUA=1 LUA_INC=/opt/lua-5.4.6/src LUA_LIB=/opt/lua-5.4.6/src


4.8) Device detection
---------------------
HAProxy supports several device detection modules relying on third party
products. Some of them may provide free code, others free libs, others free
evaluation licenses. Please read about their respective details in the
following files :

    doc/DeviceAtlas-device-detection.txt   for DeviceAtlas
    doc/51Degrees-device-detection.txt     for 51Degrees
    doc/WURFL-device-detection.txt         for Scientiamobile WURFL


4.9) Miscellaneous
------------------
Some systems have specificities. Usually these specificities are known and/or
detected and properly set for you. If you need to adjust the behaviour, here
are the extra libraries that may be referenced at build time :

  - USE_RT=1      build with librt, which is sometimes needed on some systems
                  when using threads. It is set by default on Linux platforms,
                  and may be disabled using "USE_RT=" if your system doesn't
                  have one. You may have to set it as well if you face an error
                  indicating that clock_gettime() was not found.

  - USE_DL=1      build with libdl, which is usually needed for Lua and OpenSSL
                  on Linux. It is automatically detected and may be disabled
                  using "USE_DL=", though it should never harm.

  - USE_SYSTEMD=1 enables support for the sdnotify features of systemd,
                  allowing better integration with systemd on Linux systems
                  which come with it. It is never enabled by default so there
                  is no need to disable it.


4.10) Common errors
-------------------
Some build errors may happen depending on the options combinations or the
selected target. When facing build errors, if you know that your system is a
bit special or particularly old, start from TARGET=generic, it is easier to
start from there and fix the remaining issues than trying to degrade another
target. Common issues may include:

  - clock_gettime() not found
    => your system needs USE_RT=1

  - many __sync_<something> errors in many files
    => your gcc is too old, build without threads.

  - many openssl errors
    => your OpenSSL version really is too old, do not enable OpenSSL

  - quic_conn-t.h: field 'level' has incomplete type
    => you tried to build QUIC with the legacy OpenSSL library, which does
       not support QUIC. Either disable QUIC with "USE_QUIC=" or use any
       other supported compatible library.

  - many "dereferencing pointer 'sa.985' does break strict-aliasing rules"
    => these warnings happen on old compilers (typically gcc-4.4), and may
       safely be ignored; newer ones are better on these.


4.11) QUIC
----------
QUIC is the new transport layer protocol and is required for HTTP/3. This
protocol stack is currently supported as an experimental feature in haproxy on
the frontend side. In order to enable it, use "USE_QUIC=1 USE_OPENSSL=1".

Note that QUIC is not fully supported by the OpenSSL library. Indeed QUIC 0-RTT
cannot be supported by OpenSSL contrary to others libraries with full QUIC
support. The preferred option is to use QUICTLS. This is a fork of OpenSSL with
a QUIC-compatible API. Its repository is available at this location:

     https://github.com/quictls/openssl

You can use the following instruction to build a functional QUICTLS.

  $ ./config --libdir=lib [--prefix=/opt/quictls]
  $ make
  $ make install

On a development environment, use SSL_INC and SSL_LIB when building haproxy to
point to the correct cryptographic library. It may be useful to specify QUICTLS
location via rpath for haproxy execution. Example :

  $ make -j $(nproc) TARGET=generic \
    USE_QUIC=1 \
    USE_OPENSSL=1 SSL_INC=/opt/quictls/include SSL_LIB=/opt/quictls/lib \
    LDFLAGS="-Wl,-rpath,/opt/quictls/lib"

Alternately, building against wolfSSL is supported as well, for example this
way assuming that wolfSSL was installed in /opt/wolfssl-5.6.0 as shown in 4.5:

  $ make -j $(nproc) TARGET=generic \
    USE_QUIC=1 \
    USE_OPENSSL_WOLFSSL=1 \
    SSL_INC=/opt/wolfssl-5.6.0/include SSL_LIB=/opt/wolfssl-5.6.0/lib
    LDFLAGS="-Wl,-rpath,/opt/wolfssl-5.6.0/lib"

As last resort, haproxy may be compiled against OpenSSL as follows:

  $ make TARGET=generic USE_OPENSSL=1 USE_QUIC=1 USE_QUIC_OPENSSL_COMPAT=1

Note that QUIC 0-RTT is not supported by haproxy QUIC stack when built against
OpenSSL. In addition to this compilation requirements, the QUIC listener
bindings must be explicitly enabled with a specific QUIC tuning parameter.
(see "limited-quic" global parameter of haproxy Configuration Manual).


5) How to build HAProxy
=======================

This section assumes that you have already read section 2 (basic principles)
and section 3 (build environment). It often refers to section 4 (dependencies).
It goes into more details with the main options.


5.1) Configuring the TARGET
---------------------------
To build haproxy, you have to choose your target OS amongst the following ones
and assign it to the TARGET variable :

  - linux-glibc         for Linux kernel 2.6.28 and above
  - linux-glibc-legacy  for Linux kernel 2.6.28 and above without new features
  - linux-musl          for Linux kernel 2.6.28 and above with musl libc
  - solaris             for Solaris 10 and above
  - freebsd             for FreeBSD 10 and above
  - dragonfly           for DragonFlyBSD 4.3 and above
  - netbsd              for NetBSD 8 and above
  - osx                 for Mac OS/X
  - openbsd             for OpenBSD 6.3 and above
  - aix51               for AIX 5.1
  - aix52               for AIX 5.2
  - aix72-gcc           for AIX 7.2 (using gcc)
  - cygwin              for Cygwin
  - haiku               for Haiku
  - generic             for any other OS or version.
  - custom              to manually adjust every setting

Example:
    $ make -j $(nproc) TARGET=linux-glibc

AIX 5.3 is known to work with the generic target. However, for the binary to
also run on 5.2 or earlier, you need to build with DEFINE="-D_MSGQSUPPORT",
otherwise __fd_select() will be used while not being present in the libc, but
this is easily addressed using the "aix52" target. If you get build errors
because of strange symbols or section mismatches, simply remove -g from
ARCH_FLAGS.

Building on AIX 7.2 works fine using the "aix72-gcc" TARGET. It adds two
special CFLAGS to prevent the loading of AIX's xmem.h and var.h. This is done
by defining the corresponding include-guards _H_XMEM and _H_VAR. Without
excluding those header-files the build fails because of redefinition errors.
Furthermore, the atomic library is added to the LDFLAGS to allow for
multithreading via USE_THREAD.

You can easily define your own target with the GNU Makefile. Unknown targets
are processed with no default option except USE_POLL=default. So you can very
well use that property to define your own set of options. USE_POLL and USE_SLZ
can even be disabled by setting them to an empty string or a zero. For
example :

  $ gmake TARGET=tiny USE_POLL="" USE_SLZ=0 TARGET_CFLAGS=-fomit-frame-pointer


5.2) Adding extra CFLAGS for compiling
--------------------------------------
A generic CFLAGS variable may be set to append any option to pass to the C
compiler. These flags are passed last so the variable may be used to override
other options such as warnings, optimization levels, include paths etc.

A default optimization level of -O2 is set by variable OPT_CFLAGS which may be
overridden if desired. It's used early in the list of CFLAGS so that any other
set of CFLAGS providing a different value may easily override it.

Some platforms may benefit from some CPU-specific options that will enable
certain instruction sets, word size or endianness for example. One of them is
the common "-march=native" that indicates to modern compilers that they need to
optimize for the machine the compiler is running on. Such options may be either
passed in the CPU_CFLAGS or in the CFLAGS variable, either will work though
one may be more convenient for certain methods of packaging and the other one
for other methods. Among the many possible options, the following ones are
known for having successfully been used:

  - "-march=native" for a native build
  - "-march=armv8-a+crc" for older ARM Cortex A53/A72/A73 (such as RPi 3B/4B)
  - "-march=armv8.1-a" for modern ARM Cortex A55/A76, Graviton2+, RPi 5
  - "-march=armv8-a+crc -moutline-atomics" to support older ARM with better
      support of modern cores with gcc-10+
  - "-mavx", "-mavx2", "-mavx512", to enable certain x86 SIMD instruction sets
  - "-march=i586" to support almost all 32-bit x86 systems
  - "-march=i686" to support only the latest 32-bit x86 systems
  - "-march=i386" to support even the oldest 32-bit x86 systems
  - "-mlittle-endian -march=armv5te" for some little-endian ARMv5 systems
  - "-mcpu=v9 -mtune=ultrasparc -m64" for a 64-bit Solaris SPARC build
  - "-march=1004kc -mtune=1004kc" for some multi-core 32-bit MIPS 1004Kc
  - "-march=24kc -mtune=24kc" for some single-core 32-bit MIPS 24Kc

If you are building for a different system than the one you're building on,
this is called "cross-compiling". HAProxy supports cross-compilation pretty
well and tries to ease it by letting you adjust paths to all libraries (please
read section 4 on dependencies for more details). When cross-compiling, you
just need to pass the path to your compiler in the "CC" variable, and the path
to the linker in the "LD" variable. Most of the time, setting the CC variable
is enough since LD points to it by default.

By default the build process runs in quiet mode and hide the details of the
commands that are executed. This allows to more easily catch build warnings
and see what is happening. However it is not convenient at all to observe what
flags are passed to the compiler nor what compiler is involved. Simply append
"V=1" to the "make" command line to switch to verbose mode and display the
details again. It is recommended to use this option when cross-compiling to
verify that the paths are correct and that /usr/include is never involved.

If you need to pass some defines to the preprocessor or compiler, you may pass
them all in the DEFINE variable. Example:

  $ make TARGET=generic DEFINE="-DDEBUG_DONT_SHARE_POOLS"

The ADDINC variable may be used to add some extra include paths; this is
sometimes needed when cross-compiling. Similarly the ADDLIB variable may be
used to specify extra paths to library files. Example :

  $ make TARGET=generic ADDINC=-I/opt/cross/include ADDLIB=-L/opt/cross/lib64


5.3) Adding extra LDFLAGS for linking
-------------------------------------
If a particular target requires specific link-time flags, these can be passed
via the LDFLAGS variable. This variable is passed to the linker immediately
after ARCH_FLAGS. One of the common use cases is to add some run time search
paths for a dynamic library that's not part of the default system search path:

    $ make -j $(nproc) TARGET=generic USE_OPENSSL_AWSLC=1 USE_QUIC=1 \
      SSL_INC=/opt/aws-lc/include SSL_LIB=/opt/aws-lc/lib \
      LDFLAGS="-Wl,-rpath,/opt/aws-lc/lib"

Some options require to be consistent between the compilation stage and the
linking stage. This is the case for options which enable debugging (e.g. "-g"),
profiling ("-pg"), link-time optimization ("-flto"), endianness ("-EB", "-EL"),
bit width ("-m32", "-m64"), or code analyzers ("-fsanitize=address"). These
options can be passed via the ARCH_FLAGS variable, which will be used at both
stages during the build process, thus avoiding the risk of inconsistencies. By
default, ARCH_FLAGS only contains "-g" to enable the generation of debug
symbols. For example, in order to build a 32-bit binary on an x86_64 Linux
system with SSL support without support for compression but when OpenSSL
requires ZLIB anyway :

    $ make TARGET=linux-glibc ARCH_FLAGS="-m32 -g" USE_OPENSSL=1 ADDLIB=-lz

and building with the address sanitizer (ASAN) simply requires:

    $ make TARGET=linux-glibc ARCH_FLAGS="-fsanitize=address -g"


5.4) Other common OS-specific options
-------------------------------------
Recent systems can resolve IPv6 host names using getaddrinfo(). This primitive
is not present in all libcs and does not work in all of them either. Support in
glibc was broken before 2.3. Some embedded libs may not properly work either,
thus, support is disabled by default, meaning that some host names which only
resolve as IPv6 addresses will not resolve and configs might emit an error
during parsing. If you know that your OS libc has reliable support for
getaddrinfo(), you can add USE_GETADDRINFO=1 on the make command line to enable
it. This is the recommended option for most Linux distro packagers since it's
working fine on all recent mainstream distros. It is automatically enabled on
Solaris 8 and above, as it's known to work.

If your system supports PCRE (Perl Compatible Regular Expressions), then you
really should build with libpcre which is between 2 and 10 times faster than
other libc implementations. Regex are used for header processing (deletion,
rewriting, allow, deny). Please see section 4 about dependencies to figure
how to build with PCRE support.

It is possible to add native support for SSL, by passing "USE_OPENSSL=1" on the
make command line. The libssl and libcrypto will automatically be linked with
HAProxy. Some systems also require libz, so if the build fails due to missing
symbols such as deflateInit(), then try again with "ADDLIB=-lz". Please check
section 4 about dependencies for more information on how to build with OpenSSL.

HAProxy can compress HTTP responses to save bandwidth. Please see section 4
about dependencies to see the available libraries and associated options.

If you need to pass other defines, includes, libraries, etc... then please
check the Makefile to see which ones will be available in your case, and
use/override the USE_* variables from the Makefile.


5.5) Adjusting the build error / warning behavior
-------------------------------------------------
If the ERR variable is set to any non-empty value other than "0", then -Werror
will be added to the compiler so that any build warning will trigger an error.
This is the recommended way to build when developing, and it is expected that
contributed patches were tested with ERR=1. Similarly, for developers, another
variable, FAILFAST enables -Wfatal-errors when set to non-empty except 0, and
makes the compiler stop at the first error instead of scrolling pages. It's
essentially a matter of taste.

Packagers who want to achieve the cleanest warning-free builds may be
interested in knowing that all enabled warnings are normally placed into
the WARN_CFLAGS variable. The variable contains a list of pre-established
warnings and a list of some that are dynamically detected on the compiler.
If the build environment or toolchain doesn't even support some of the basic
ones, it is then possible to just redefine them by passing the main ones in
WARN_CFLAGS (e.g. at the very least -W -Wall). Similarly, it may sometimes
be desirable not to disable certain warnings when porting to new platforms
or during code audits, or simply because the toolchain doesn't support some
of the most basic -Wno options. In this case, the list of automatic -Wno
variables is specified by variable NOWARN_CFLAGS, which is passed after
WARN_CFLAGS (i.e. it can undo some of the WARN_CFLAGS settings). Be careful
with it, as clearing this list can yield many warnings depending on the
compiler and options.

The DEP variable is automatically set to the list of include files and also
designates a file that contains the last build options used. It is used during
the build process to compute dependencies and decide whether or not to rebuild
everything (we do rebuild everything when .h files are touched or when build
options change). Sometimes when performing fast build iterations on inline
functions it may be desirable to avoid a full rebuild. Forcing this variable
to be empty will be sufficient to achieve this. This variable must never be
forced to produce final binaries, and must not be used during bisect sessions,
as it will often lead to the wrong commit.

Examples:
  # silence strict-aliasing warnings with old gcc-4.4:
  $ make -j$(nproc) TARGET=linux-glibc CC=gcc-44 CFLAGS=-fno-strict-aliasing

  # disable all warning options:
  $ make -j$(nproc) TARGET=linux-glibc CC=mycc WARN_CFLAGS= NOWARN_CFLAGS=

  # enable -Werror and -Wfatal-errors to immediately stop on error
  $ make -j$(nproc) TARGET=linux-glibc ERR=1 FAILFAST=1

  # try to restart the build where it was after hacking an include file, to
  # check if that was sufficient or not:
  $ make -j$(nproc) TARGET=linux-glibc ERR=1 DEP=


5.6) Enabling a DEBUG build
---------------------------
The DEBUG variable is used to extend the CFLAGS and is preset to a list of
build-time options that are known for providing significant reliability
improvements and a barely perceptible performance cost. Unless instructed to do
so by some project developers, or trying to save the last ounce of performance,
these options should not be changed. Among the usable ones are:
  - -DDEBUG_STRICT: enable some runtime assertions at key places in the code.
    The goal is to emit a warning or stop the program if certain expected
    conditions are not met, and whose violation will result in a misbehaving
    process due to memory corruption or other significant trouble, possibly
    caused by an attempt to exploit a bug in the program or a library it relies
    on. The option knows 3 values: 0 (disable all such assertions, not
    recommended), 1 (enable all inexpensive assertions, the default), and
    2 (enable all assertions even in fast paths). Setting the option with no
    value corresponds to 1, which is the recommended value for production.

  - -DDEBUG_STRICT_ACTION: indicates how to react to a check violation. There
    are 3 types of checks: BUG (condition that is known to have serious
    consequences), WARN (warning about a highly suspicious condition which the
    process may recover from, but whose unknown cause may also have serious
    consequences), CHECK (verification whether a condition that developers now
    consider impossible still happens). The variable takes a value from 0 to 3,
    that adjusts the behavior on these 3 violations:

               BUG       WARN      CHECK
         0     warn      warn      warn
         1     stop      warn      warn
         2     stop      stop      warn
         3     stop      stop      stop

    The default value is 1, which is the best balance for production in that it
    will do its best to prevent a known bogus process from running away, but
    will let it run if it believes it can recover. Users running the process in
    sensitive environments (finance etc) may prefer to run at level 2 to make
    sure to stop any detected anomaly before it may have an impact. Level 3
    should only be used at the request of developers. In any case, any emitted
    warning should be reported to developers.

  - -DDEBUG_MEMORY_POOLS: this enables by default extra controls around memory
    allocation that will help detect coding errors such as double-frees and
    freeing a bad memory location. It will also detect earlier risks of memory
    overflows, which may have security implications. The cost is extremely low
    (less than 1% increase in memory footprint). This is equivalent to adding
    "-dMtag" on the command line. This option is enabled in the default build
    options and may be disabled with -DDEBUG_MEMORY_POOLS=0.

  - -DDEBUG_DONT_SHARE_POOLS: this will keep separate pools for same-sized
    objects of different types. Using this increases the memory usage a little
    bit but further reduces the risk of memory management related bugs and will
    lead to more accurate traces in case of error. It is equivalent to adding
    "-dMno-merge" on the command line. It is not enabled in the default build
    options.

  - -DDEBUG_POOL_INTEGRITY: this will enable runtime detection and stopping of
    a class of bugs known as "use after free", which consists in modifying a
    memory area after freeing it while it was reused for something else. This
    option is quite powerful but such bugs are fortunately extremely rare, and
    it will cause a measurable performance degradation (a few percent). This is
    equivalent to adding "-dMcold-first,integrity" on the command line. This
    option is not enabled by default but users running development versions on
    moderate performance sites in order to participate to reliability testing
    are encouraged to use it, in combination with -DDEBUG_DONT_SHARE_POOLS and
    -DDEBUG_MEMORY_POOLS, as this could catch dangerous regressions.

As such, "-DDEBUG_STRICT -DDEBUG_MEMORY_POOLS" is implicit and recommended for
production. For security sensitive environments, it is recommended to use
"-DDEBUG_STRICT_ACTION=2 -DDEBUG_DONT_SHARE_POOLS". When testing new versions
or trying to nail a bug down, use "-DDEBUG_STRICT=2 -DDEBUG_STRICT_ACTION=2 \
-DDEBUG_DONT_SHARE_POOLS -DDEBUG_POOL_INTEGRITY". Finally in order to minimize
memory usage by disabling these integrity features, it is also possible to use
"-DDEBUG_STRICT=0 -DDEBUG_MEMORY_POOLS=0".


5.7) Summary of the Makefile's main variables
---------------------------------------------

The following variables are commonly used:
  - TARGET        platform name, empty by default, see help
  - CC            path to the C compiler, defaults to "cc"
  - LD            path to the linker, defaults to "$CC"
  - CFLAGS        CFLAGS to append at the end, empty by default
  - LDFLAGS       LDFLAGS to append at the end, empty by default
  - ARCH_FLAGS    flags common to CC and LD (-fsanitize, etc). Defaults to "-g"
  - OPT_CFLAGS    C compiler optimization level. Defaults to "-O2"
  - WARN_CFLAGS   list of autodetected C compiler warnings to enable
  - NOWARN_CFLAGS list of autodetected C compiler warnings to disable
  - ADDINC        include directives to append at the end, empty by default
  - ADDLIB        lib directives to append at the end, empty by default
  - DEFINE        extra macros definitions for compiler, empty by default
  - DEBUG         extra DEBUG options for compiler, empty by default
  - ERR           enables -Werror if non-zero, empty by default
  - FAILFAST      enables -Wfatal-error if non-zero, empty by default


6) How to install HAProxy
=========================

To install haproxy, you can either copy the single resulting binary to the
place you want, or run :

    $ sudo make install

If you're packaging it for another system, you can specify its root directory
in the usual DESTDIR variable.

-- end