doc: how to use QEMU to test big-endian build

[fw/openocd] / doc / manual / endianness.txt

/** @page endianness About endianness

OpenOCD has to potentially deal with different endianness between:
- the host PC endianness;
- the data endianness during communication between host and adapter;
- the target CPU endianness.

The whole OpenOCD code should be written to handle any endianness
mismatch and should run on either little and big endian hosts.

Big-endian host PC are becoming less and less common since Apple™ has
switched away from big-endian PowerPC™ in favor of little-endian intel
X86™.

The lack of commercial big-endian hosts makes hard testing OpenOCD correctness
on big-endian hosts. Running OpenOCD on low-cost commercial routers based on
big-endian MIPS is possible, but it's tricky to properly setup the system and
the cross-compiling environment.

In next sections there are two example on how to compile and test OpenOCD in an
emulated big-endian environment.


@section endianness_helpers OpenOCD API for handling endianness

Use the following OpenOCD API to handle endianness conversions:
- host endianness to/from little endian:
  - le_to_h_u64(), le_to_h_u32(), le_to_h_u16();
  - h_u64_to_le(), h_u32_to_le(), h_u16_to_le();
  - buf_get_u32(), buf_get_u64();
  - buf_set_u32(), buf_set_u64();
- host endianness to/from big endian:
  - be_to_h_u64(), be_to_h_u32(), be_to_h_u16();
  - h_u64_to_be(), h_u32_to_be(), h_u16_to_be();
- host endianness to/from target endianness:
  - target_read_u64(), target_read_u32(), target_read_u16();
  - target_write_u64(), target_write_u32(), target_write_u16();
  - target_write_phys_u64(), target_write_phys_u32(), target_write_phys_u16();
  - target_buffer_get_u64(), target_buffer_get_u32(), target_buffer_get_u24(), target_buffer_get_u16();
  - target_buffer_set_u64(), target_buffer_set_u32(), target_buffer_set_u24(), target_buffer_set_u16();
- byte swap:
  - buf_bswap32(), buf_bswap16().


@section endianness_docker Use dockers to run different endianness


Docker can run a full Linux image that includes the toolchain through QEMU
emulator.
By selecting a big-endian image, it's possible to compile and execute OpenOCD
in big-endian.
There are, so far, not many options for big-endian images; s390x is one of the
few available.

To be expanded.

User should:
- install docker;
- download the big-endian image;
- run the image in docker;
- download, in the image, the OpenOCD code to test;
- recompile OpenOCD code in the image;
- run OpenOCD binary in the image.

From https://github.com/multiarch/qemu-user-static

  @code{.unparsed}
  docker run --rm -t s390x/ubuntu bash
  @endcode


@section endianness_qemu Use buildroot and QEMU to run different endianness

QEMU User Mode Emulation is an efficient method to launch, on host's CPU,
applications compiled for another CPU and/or for different endianness.
It works either on Linux and BSD. More info available on
https://www.qemu.org/docs/master/user/index.html

With QEMU User Mode Emulation is thus possible running, on a commonly available
little-endian X86 Linux host, OpenOCD compiled for a big-endian host.

The following example will show how to use buildroot to:
- build big-endian toolchain and libraries;
- compile OpenOCD for big-endian;
- run the big-endian OpenOCD on little-endian Linux PC.

The example will use ARM Cortex-A7 big-endian only because I personally feel
comfortable reading ARM assembly during debug. User can select other CPU
architectures, as this does not impact the result.

A similar method can be used to test OpenOCD compiled for 32 vs 64 bit host.

@note
- the version of autotools locally installer in your Linux host can be
  incompatible with the version of autotools used by buildroot. This can cause
  the build to fail if buildroot has to run its autotools on a partially
  configured OpenOCD folder. Use either a clean copy of OpenOCD code in 2., or
  run "./bootstrap" in OpenOCD folder to prevent buildroot from using its own
  autotools;
- the configuration tool in 4. and 5. matches the version of OpenOCD used by
  buildroot. Some new driver could be not listed in. OpenOCD will build every
  driver that is not disabled and with satisfied dependencies. If the driver
  you plan to use is not listed, try a first build and check OpenOCD with
  command "adapter list", then try to hack the buildroot files Config.in and
  openocd.mk in folder package/openocd/ and use "make openocd-reconfigure" to
  rerun the build starting with configuration;
- using pre-built toolchains, you need 2GB of disk space for buildroot build.
  To also rebuild the toolchains you will need ~5GB and much longer time for
  the first build (it takes ~2 hour on my crap 10+ years old laptop);
- you need to install few tools for buildroot dependency, listed in
  https://buildroot.org/downloads/manual/manual.html#requirement ;
- you need to install qemu-armeb. On Arch Linux it's in package qemu-arch-extra;
  on Ubuntu/debian it's packaged in qemu-user.
  Buildroot can also be configured to build qemu for the host, if you prefer,
  by enabling BR2_PACKAGE_HOST_QEMU_LINUX_USER_MODE, but this takes longer
  compile time;
- don't use qemu-system-arm, as it emulates a complete system and requires a
  fully bootable ARM image;
- while QEMU User Mode Emulation is available for both Linux and BSD, buildroot
  only builds binaries for Linux target. This example can only be used with
  Linux hosts emulating the Linux target.


Steps to run big-endian OpenOCD on little-endian host Linux PC:

1. Get buildroot source. Today's latest version is "2022.02":
     @code{.unparsed}
     wget https://buildroot.org/downloads/buildroot-2022.02.tar.xz
     tar xf buildroot-2022.02.tar.xz
     cd buildroot-2022.02
     @endcode

2. Override the source repo for OpenOCD in order to build your own code version
   in place of the default OpenOCD release version:
     @code{.unparsed}
     echo OPENOCD_OVERRIDE_SRCDIR=/home/me/openocd.git >> local.mk
     @endcode

3. Copy default config for OpenOCD big-endian. This used:
   - ARM Cortex-A7 big-endian target,
   - external Linaro armeb toolchain (to speed up first build),
   - OpenOCD all configure options enabled.

     @code{.unparsed}
     cp $OPENOCD_OVERRIDE_SRCDIR/contrib/buildroot/openocd_be_defconfig configs/
     @endcode

4. Configure buildroot with default config for OpenOCD big-endian:
     @code{.unparsed}
     make openocd_be_defconfig
     @endcode

5. Optional, change buildroot configuration:
     @code{.unparsed}
     make menuconfig
     @endcode
   These are the options selected with default config for OpenOCD big-endian:
     @code{.unparsed}
     Target options  --->
       Target Architecture  --->
         ARM (big endian)
       Target Architecture Variant  --->
         cortex-A7
     Toolchain  --->
       Toolchain type  --->
         External toolchain
       Toolchain  --->
         Linaro armeb 2018.05
       Toolchain origin  --->
         Toolchain to be downloaded and installed
     Target packages  --->
       Hardware handling  --->
         openocd
           All adapters selected
     @endcode
   Save and exit

6. Build (and take a long coffee break ...):
     @code{.unparsed}
     make openocd
     @endcode

7. Execute big-endian OpenOCD:
     @code{.unparsed}
     cd output/target
     qemu-armeb -cpu cortex-a7 -L . usr/bin/openocd -s usr/share/openocd/scripts/ -f board/st_nucleo_f4.cfg
     @endcode

8. Optional, to rebuild after any source code modification in ${OPENOCD_OVERRIDE_SRCDIR}:
     @code{.unparsed}
     make openocd-rebuild
     @endcode

 */
/** @file
This file contains the @ref endianness page.
 */