[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

TC decision on "Merged /usr" - #914897

The Debian Technical Committee was asked in #914897 to overrule the debootstrap
maintainers regarding the "merged `/usr`" default.

The following resolution was passed:

=== Resolution ===

The Technical Committee resolves to decline to override the debootstrap

Furthermore, using its §6.1.5 "Offering advice" power, the Technical
Committee considers that the desirable solution at the time of `bullseye` is
`middle`: both directory schemes are allowed, and packages (including official
packages) can be built on hosts with either classical or "merged `/usr`"
directory schemes.

=== End Resolution ===

=== Rationale ===

## What is "merged `/usr`"

"Merged `/usr`" describes a possible future standard directories scheme in
which the `/{bin,sbin,lib*}/` directories have been made superfluous through
replacing them by symlinks to their `/usr` equivalents
The motivation to get Debian systems to converge towards such a scheme is
vastly documented elsewhere ([FDO's TheCaseForTheUsrMerge][0],
[wiki.d.o UsrMerge][1]) but can be summarized as the following points:

* having separate `/` and `/usr` filesystems has been useful in the past for
booting without initramfs onto a minimal root filesystem that carried just
enough to mount the `/usr` filesystem later in the boot process. Given the
evolution of physical hosts' capabilities, initramfs'es have been default in
Debian (and elsewhere) for a long time, and most systems no longer have an
intermediate state during boot in which they have only `/`, but not `/usr`,
mounted. Booting hosts through that intermediate state is not systematically
tested in Debian anymore.
* another use-case is to share system files from `/usr` between hosts (over a
network link) or containers (locally) which use different data or
configuration. Having all software under `/usr` (instead of spread between
`/` and `/usr`) makes the centralized update and the sharing easier.
* the packaging infrastructure to install files outside of `/usr` (e.g.
installing libs under `/lib` instead of `/usr/lib`) is not standard and
represents technical debt.
* given its status as remnant "folklore", the distinction between what
_needs_ to be shipped in `/` and what can stay in `/usr` is often interpreted
* allowing shipment of identically-named libraries or binaries in different
paths can confuse common understanding of paths precedence.

[0]: https://www.freedesktop.org/wiki/Software/systemd/TheCaseForTheUsrMerge/
[1]: https://wiki.debian.org/UsrMerge

The arguments against moving the base directories' scheme towards "merged
`/usr`" are as follows:

* there's no gain in disrupting something that is not inherently broken;
* `/{bin,sbin,lib*}/` → `/usr/{bin,sbin,lib*}/` symlinks create confusing
views of the system (`/bin/cat` and `/usr/bin/cat` are the same file), and
dpkg doesn't support this situation cleanly:
* it is possible for distributions to converge towards having all system
files in `/usr` in finite time instead of shortcutting this migration with
`/{bin,sbin,lib*}/` → `/usr/{bin,sbin,lib*}/` symlinks.

The compatibility symbolic links `/lib` → `/usr/lib` and `/lib64` →
`/usr/lib64` are required by the various CPUs' platform ABIs (for example
i386 requires `/lib/ld-linux.so.2` to resolve to glibc's `ld.so`, and amd64
requires `/lib64/ld-linux-x86-64.so.2`) so there are no plans to remove them
altogether. Similarly, removing `/bin` is not under consideration because it
would break the assumption that `/bin/sh` exists, and removing `/sbin` would
break the assumption that `/sbin/fsck.*` and `/sbin/mount.*` exist.

## "merged `/usr`" in Debian

In Debian buster, the current testing suite, "merged `/usr`" is only
considered for implementation with symlinks (there are no proposals for
simply dropping `/{bin,sbin,lib*}`) and is implemented in two main ways:

* existing hosts can be made to have a "merged `/usr`" by installing the
[usrmerge][2] package;
* new hosts get the `/{bin,sbin,lib*}/`→ `/usr/{bin,sbin,lib*}/` symlinks
by default when using debootstrap >= 1.0.102.

The usrmerge package contains a `/usr/lib/convert-usrmerge` perl executable
that runs in `postinst`, that will move the contents of `/{bin,sbin,lib*}/`
and replace these directories with symlinks when empty.

It is also possible to merge `/usr` in other ways, for example with
`debootstrap --merged-usr` or by bootstrapping into a chroot that already
contains the necessary symlinks.

[2]: https://tracker.debian.org/pkg/usrmerge

## Issues from "merged `/usr`"

Since the default change in debootstrap 1.0.102, some issues have arisen.
Due to the fact that some buster/sid hosts have the "merged `/usr`" symlinks
in place, it has been observed that some binary packages carried some traces
of these differences (notably official packages built on Debian buildd hosts
which had been resetup).
Some such differences can actually render the built packages unusable on
non-"merged `/usr`" systems.
For example, if `cat` is detected at build-time in `/usr/bin/cat` (where
coreutils ships `/bin/cat`), a binary hardcoding that path will try to use
`/usr/bin/cat` after installation, but that path doesn't exist in non-"merged
`/usr`" systems.
In order to mitigate this, debootstrap has been modified to let its "buildd"
variant be non-"merged `/usr`", the Debian buildds have been resetup and the
affected packages rebuilt.

The lesson here is that with the existence of (any of) the usrmerge and the
debootstrap default change, "merged `/usr`" Debian systems exist already, and
that packages built on hosts with such directory schemes can _potentially_ be
broken on non-"merged `/usr`" systems.
At this point, the two variants have to be supported, at least as
installation targets of Debian packages.

Two initiatives are worth mentioning at this point:
* [a patch](https://lists.debian.org/20181202212535.GC11687@gaara.hadrons.org)
has been proposed for dpkg-buildpackage to mark packages built on
"merged `/usr`" hosts with a `Build-Tainted-By: merged-usr-via-symlinks`;
* the reproducible builds team has added a "merged `/usr`" variation to their
setup, and have then
the Debian packages from unstable which had differences due to
"merged `/usr`". It seems that ~61 packages were affected by differing
builds; 32 from these have been fixed in unstable already.

## The long-term desirable situation

Various valid long-term desirable situations coexist, and while discussing
immediate countermeasures, it is useful to keep the long-term outcome that
those are most likely to produce.

These are the six possible situations at the time of bullseye (buster + 1):

* `none`:   "merged `/usr`" has been reverted
* `empty`:  "merged `/usr`" has been reverted, `/usr` is empty (but the
              mandatory files)
* `weak`:   both directory schemes are allowed, all packages only built on
              classical hosts
* `middle`: both directory schemes are allowed, all packages can be built on
* `hard`:   both directory schemes are allowed, packages can be built on
              either, official packages only built on "merged `/usr`" hosts
* `all`:    only "merged `/usr`" directory schemes are allowed, packages only
              built on "merged `/usr`" hosts

It can be summarized by the following table:

|          |     Host types that are allowed       | Are merged `/usr` |     Official packages are built on    | Packages built on … can break on the other |
| Codename | classical hosts | merged `/usr` hosts | symlinks allowed  | classical hosts | merged `/usr` hosts |   classical hosts   |  merged `/usr` hosts |
|     none |       yes       |          no         |         no        |       yes       |          no         |         yes         |          yes         |
|    empty |       yes       |          no         |         no        |       yes       |          no         |         yes         |          yes         |
|     weak |       yes       |         yes         |        yes        |       yes       |          no         |          no         |          yes         |
|   middle |       yes       |         yes         |        yes        |       yes       |         yes         |          no         |           no         |
|     hard |       yes       |         yes         |        yes        |        no       |         yes         |          no         |           no         |
|      all |       no        |         yes         |        yes        |        no       |         yes         |         yes         |           no         |

The current state of buster is `weak`.

=== End Rationale ===

Please see https://bugs.debian.org/914897 for the TC discussion on this topic.

    OdyX, for the Technical Committee

Attachment: signature.asc
Description: This is a digitally signed message part.

Reply to: