Skip to content

Check EESSI-extend install paths, explain node-local refusals, scope build secrets - #65

Merged
HaoZeke merged 3 commits into
mainfrom
fix/followups
Nov 16, 2026
Merged

HaoZeke merged 3 commits into
mainfrom
fix/followups

Conversation

@HaoZeke

@HaoZeke HaoZeke commented Nov 16, 2026

Copy link
Copy Markdown
Collaborator

A build and eb-local eb check, once EESSI-extend has loaded, that the install is where the module put it under Lmod 8.6.19: under <prefix>/versions for a build, or in the tree the install variable selects for eb-local eb. They stop with exit 2, before eb runs, when EASYBUILD_INSTALLPATH is outside that tree, or when EASYBUILD_INSTALLPATH_SOFTWARE or EASYBUILD_INSTALLPATH_MODULES is set. A leftover EESSI_USER_INSTALL next to EESSI_PROJECT_INSTALL is refused, because EESSI-extend then makes a private user install. Both warn when loading EESSI-extend sets or unsets an install variable, and a build names each install variable it clears from the caller's shell.

The refusals of a node-local mount and of an install prefix, by a build, eb-local clean and eb-local doctor, say what was refused and why, what the command then did not do, and, on the next line, what to change. A workspace that cannot be made names the error the kernel gave, where it used to say AccessDenied for every failure, and the umask is no longer left cleared when the path is too long.

build.secrets, in a configuration file or a profile, and EB_LOCAL_BUILD_SECRETS, for one build, name the secrets a build gets. The build reads only those; a secret it does not get never reaches eb. Without either, every build gets every secret, as before.

The tests pass on x86_64. The aarch64 emulation failures are those of the base.

EESSI-extend chooses its install tree from EESSI_USER_INSTALL,
EESSI_PROJECT_INSTALL, EESSI_SITE_INSTALL and EESSI_CVMFS_INSTALL as it
loads. Loaded under Lmod 8.6.19, the module EESSI's
software-layer-scripts generate installs below the directory a user or
project variable names, into the site tree for the site variable, into
EESSI's own tree for the cvmfs one, and under $HOME/eessi with none
set. With a user variable left from an earlier session beside a project
one, it makes a private user install, umask 077, in the user tree; with
the site or the cvmfs variable and another, it refuses to load. A load
hook that sets an install variable is too late for its own load and
moves the next one. Nothing checked where an install then went.

This changes behaviour:

- Once EESSI-extend has loaded, a build stops with exit 2, before eb
  runs, unless EASYBUILD_INSTALLPATH is below <prefix>/versions, and
  whenever EASYBUILD_INSTALLPATH_SOFTWARE or _MODULES is set. A site
  hook that set an install path, or an EESSI_CVMFS_REPO with a
  character such as "-", which EESSI-extend matches as a Lua pattern,
  sent the install and its logs outside the prefix eb-local checked,
  lists and cleans of secrets, while eb-local said it installed there.
- eb-local eb stops the same way when the install path is outside the
  tree the install variable selects, when that tree is not known
  because the EESSI module left EESSI_SITE_SOFTWARE_PATH or
  EESSI_SOFTWARE_PATH empty, and when EASYBUILD_INSTALLPATH_SOFTWARE
  or _MODULES is set in the shell or by the load. It passed them on to
  EasyBuild.
- Both warn when loading EESSI-extend sets or unsets an install
  variable.
- A build warns about each install variable it clears from the
  caller's shell, in the words eb-local doctor uses.
- The refusal of more than one install variable names each, with its
  value, and what EESSI-extend would do with them. It still exits 3.

The Bash 1.x script is frozen and keeps its behaviour, so the new
messages are tested against the Zig build only. The test doubles now
set the install path EESSI-extend sets.

The README says what the checks do and links to a new explanation page,
written in Org under docs/orgmode and exported to the MyST page Sphinx
builds by docs/export.el, which needs only the Markdown backend Org
ships. tests/test_docs.py checks that the two agree where Emacs is
installed.
A build, eb-local clean and eb-local doctor refused a node-local mount
or an install prefix with messages such as "allocation node-local root
must be an owned writable directory." or "installation, source and log
paths must use durable owned storage.", which named neither the path,
nor the setting, nor the reason.

Each refusal now has one form. The first line says what was refused
and why, then what the command did not do: "nothing was built",
"nothing was removed", or "a build would stop here" for --explain. The
next line, indented, says what to change: the setting, with its value
and where it was set as eb-local config shows it, or the variable,
option or command. eb-local doctor puts the same on its FAIL line,
with "builds refuse it", as its line for a refused prefix already had.

storage.jobRoot takes the node-local settings with where each was set,
and returns what to change beside the message. The checks, their order
and what they accept are unchanged; a symbolic link, a missing mount
and one that is not a directory, refused before with one message, now
each have their own.

Changed behaviour, as this item needs:
- the text of these refusals on stderr and of the doctor FAIL lines;
  exit statuses are unchanged;
- a 1.x eb-local.conf is converted to TOML line for line, so that
  eb-local config and the errors give each setting its own line, not
  one in the converted text;
- a workspace that cannot be made names the error the kernel gave,
  where makeWorkspace returned AccessDenied for every failure, and the
  umask is no longer left cleared when its path is too long.

docs/reference/refusals.md, written from Org, lists each refusal. The
README's paragraph on node-local storage moves to "Where things go"
and links to it.
Every build read every secret the configuration defines and put each in
eb's environment, so a build that used none of them still carried them
all, and one secret whose source could not be read stopped every build.

build.secrets, a list of secret names under [build] in any
configuration file or profile, and EB_LOCAL_BUILD_SECRETS, names
separated by spaces for one build, now name the secrets a build gets.
The build reads only those, after every other check as before; the
others are not read and never reach eb. An empty list names none. A
name that is not a variable a secret may set, or that no secret in
force has, is a configuration error that says where it was set.

Without either, a build gets every secret, as before. What changes:

- eb-local config shows a build.secrets row, "(every secret)" when
  nothing sets it.
- eb-local config --env prints EB_LOCAL_BUILD_SECRETS when
  build.secrets is set.
- eb-local --explain RECIPE says which secrets stay out of eb's
  environment when build.secrets leaves any out.
- eb-local doctor says FAIL, with a site file, only for an unreadable
  secret that builds get; one they do not get is a warning, and a
  readable one is info. When nothing sets build.secrets, its FAIL line
  for an unreadable secret now says to name the secrets builds need.
- The usage text says how a build's secrets are chosen.

The README and a new how-to page describe it.
@HaoZeke
HaoZeke merged commit ccb4a83 into main Nov 16, 2026
9 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant