[v3,0/2] Validating UAPI backwards compatibility

Message

John Moon March 16, 2023, 6:15 p.m. UTC

The kernel community has rigorously enforced a policy of backwards
compatibility in its UAPI headers for a long time. This has allowed user
applications to enjoy stability across kernel upgrades without
recompiling. Our goal is to add tooling and documentation to help kernel
developers maintain this stability.

We see in the kernel documentation:
"Kernel headers are backwards compatible, but not forwards compatible.
This means that a program built against a C library using older kernel
headers should run on a newer kernel (although it may not have access
to new features), but a program built against newer kernel headers may
not work on an older kernel."[1]

How does the kernel community enforce this guarantee? As we understand it,
it's enforced with thorough code review and testing. Is there any tooling
outside of this being used to help the process?

Also, could documentation on UAPI maintenance (from a developer's point
of view) be expanded? Internally, we have a set of guidelines for our
kernel developers regarding UAPI compatibility techniques. If there's
interest in supplying a document on this topic with the kernel, we'd be
happy to submit a draft detailing what we have so far as a jumping off
point.

In terms of tooling, I've attached a shell script we've been using
internally to validate backwards compatibility of our UAPI headers. The
script uses libabigail's[2] tool abidiff[3] to compare a modified
header's ABI before and after a patch is applied. If an existing UAPI is
modified, the script exits non-zero. We use this script in our
continuous integration system to block changes that fail the check.

It generates output like this when a backwards-incompatible change is made
to a UAPI header:

!!! ABI differences detected in include/uapi/linux/bpf.h from HEAD~1 -> HEAD !!!

    [C] 'struct bpf_insn' changed:
      type size hasn't changed
      1 data member change:
        type of '__s32 imm' changed:
          typedef name changed from __s32 to __u32 at int-ll64.h:27:1
          underlying type 'int' changed:
            type name changed from 'int' to 'unsigned int'
            type size hasn't changed

We wanted to share this script with the community and hopefully also
receive general feedback when it comes to tooling/policy surrounding this
issue. Our hope is that the script will help kernel UAPI authors maintain
good discipline and avoid breaking userspace.

In v3, we've changed the default behavior of the script to check _all_
UAPI headers in the tree. This is to avoid possible false negatives
brought up in the v2 review. We've also refactored how it works with
architecture-specific headers. The documentation has been updated to
reflect the changes and should serve well as a guide on how things are
working now.

Thanks for the help on v1/v2 and thanks in advance for the help on v3!

[1] Documentation/kbuild/headers_install.rst
[2] https://sourceware.org/libabigail/manual/libabigail-overview.html
[3] https://sourceware.org/libabigail/manual/abidiff.html

P.S. While at Qualcomm, Jordan Crouse <[e-mail]> authored the
original version of the UAPI checker script. Thanks Jordan!


John Moon (2):
  check-uapi: Introduce check-uapi.sh
  docs: dev-tools: Add UAPI checker documentation

 Documentation/dev-tools/checkuapi.rst | 492 ++++++++++++++++++++++
 Documentation/dev-tools/index.rst     |   1 +
 scripts/check-uapi.sh                 | 564 ++++++++++++++++++++++++++
 3 files changed, 1057 insertions(+)
 create mode 100644 Documentation/dev-tools/checkuapi.rst
 create mode 100755 scripts/check-uapi.sh


base-commit: 9c1bec9c0b08abeac72ed6214b723adc224013bf
--
2.17.1