jail, ports: Avoid set -e trap in quiet-mode "done" idiom
4cb78962 introduced `[ -n "${quiet}" ] && echo " done"` in the svn/git
checkout and update paths of install_from_vcs() (jail.sh) and ports.sh.
When quiet is unset the test is false, so the statement returns exit 1;
under set -e a bare call to the enclosing function/case arm would abort
the script right after that line runs, as seen with the identical
pattern in image.sh (PR #1378). None of these 8 occurrences are
currently a function's last statement, so they don't trip today, but
the pattern is fragile under refactoring. Use an if/fi block instead,
matching the idiom already used for the preceding header line, so the
statement always returns 0.
This closes #1387
jail: Reject a period in the new name on rename
create_jail() rejects a period in JAILNAME since jail(8) names cannot
contain one, but rename_jail() never validated NEWJAILNAME at all, so
'poudriere jail -r' could rename a jail to a name jail(8) itself would
reject. Apply the same check used at creation.
This closes #1388
pkg: Add trailing ellipsis to repository signing status messages
Each of these msg() calls is immediately followed by an unredirected
injail pkg-repo invocation whose own output streams directly to the
console right after the header, the same shape as other "action
followed by unbuffered command output" status lines that already use
a trailing "...". pkg repo has no quiet/verbose toggle in play here
(PKG_REPO_FLAGS only carries --hash/--symlink), so this is a plain
ellipsis fix, not an idiom conversion.
This closes #1381
image: Use quiet-aware msg idiom for pkgbase install status
install_world_from_pkgbase() always reported "Installing base
packages" / "Base packages installed" as two separate msg() lines,
regardless of PKG_QUIET. Under the default quiet pkg install
(-q, unless -v was passed), nothing prints between them, so treat it
like the other silent/slow operations with msg_n "...done". When -v
clears PKG_QUIET, pkg's own install output streams between the
header and completion, so keep the header on its own line via msg()
with a trailing "..." and skip the redundant "done", matching the
quiet-vs-verbose idiom already used for jail.sh's install_from_vcs().
image: Use msg_n/done idiom for ZFS image build status messages
zfs send/recv, zpool create, and zfs create/snapshot are all silent
on success and can take a noticeable amount of time (especially for
larger images), but were reported via standalone msg() calls with no
completion signal. Switch to the msg_n "...done" idiom used elsewhere
for silent, potentially slow operations.
image: Use msg_n/done idiom for install_world tar status
install_world() reported "Installing world with tar" and "Installing
world done" as two separate msg() lines, even though both the tar
pipeline and make -s are always silent on success and the operation
(copying the full world plus delete-old) can take a while. Switch to
the msg_n "...done" idiom used elsewhere for silent, potentially slow
operations.
jail: Add trailing ellipsis to make/freebsd-update status messages
Each of these msg() calls is immediately followed by an unredirected
external command (make or freebsd-update) whose own real-time output
streams directly to the console right after the header, with no
paired completion message -- the command's own output, or the
err()/msg_warn() on failure, is what signals progress/completion.
Add a trailing "..." so the header reads as in-progress rather than
a standalone statement, consistent with msg_n() callers elsewhere.
image: Use msg_n/done idiom for miniroot creation status
mkminiroot() reported "Making miniroot" as a standalone msg() even
though the function is mostly silent and can take a noticeable
amount of time (copying files, resolving shared library deps, and
running makefs/gzip). Of its steps, only makefs writes unredirected
output to stdout; gzip -9 without -v is already silent. Redirect
makefs's output to /dev/null, matching the same treatment already
used for its sibling make_esp_file(), and switch to the msg_n
"...done" idiom used elsewhere for silent, potentially slow
operations so the whole function reports as one in-progress status
line instead of a header with no matching completion.
build_port: Add trailing ellipsis to shared library dependency check
The readelf/grep output for the NEEDED library list prints directly
after this status line, so it should read as in-progress like other
such messages instead of a standalone statement.
Reject unsafe package names from ports metadata
PKGNAME is used to construct host-side package staging paths. A port
can supply traversal components and cause those paths to escape the
staging directory.
Validate package names when reading port metadata and before a build
worker uses one. This prevents path separators and dot traversal
components from reaching host filesystem operations.
(cherry picked from commit e90c7ba59dbb1b42b9e83029e88c570caf29d632)
Conflicts:
src/share/poudriere/common.sh
Issue #1363
README: point the build badge at SourceHut
Cirrus CI stopped running jobs on 2026-06-01 when Cirrus Labs joined OpenAI,
so the badge no longer renders and its link is dead. SourceHut has been
building poudriere since 2026-05-05 via .builds/freebsd.yml.
Fixes #1331
ports: Fix period check in ports tree name validation
The create-time validation matched "*:*" (colon) but the error
message says "cannot contain a period (.)", citing jail(8)'s
hierarchical dot-separated naming. jail_start()'s equivalent checks
for jailname/ptname/setname (common.sh) correctly use "*.*" -- this
was the only place with the mismatched pattern, so a period in
PTNAME was silently accepted instead of being rejected as documented.
image: Use msg_n/done idiom for ESP image creation status
make_esp_file() reported "Creating ESP image" and "ESP Image created"
as two separate complete lines via msg(), even though makefs's own
output is redirected to /dev/null and the operation can take a
noticeable amount of time. Use the msg_n "...done" idiom used
elsewhere for silent, potentially slow operations (e.g. jail removal)
so it reads as a single in-progress status line instead of two
disconnected facts.
Add missing FLAVOR to stage-qa and check-plist calls
For example, `poudriere testport -NNi devel/freebsd-gcc15 at amd64`
will correctly build the amd64 flavor, but will attempt to
stage-qa and check-plist for the aarch64 flavor because it
becomes the default flavor through alphabetic sorting.
testport/bulk -i: Pass FLAVOR when looking up WRKDIR in interactive motd
The WRKDIR line queried `make -V WRKDIR` without FLAVOR, so a flavored
port always reported its default flavor's work directory instead of
the one actually built and installed -- e.g. audio/baresip built with
FLAVOR=nox11 showed work-default. run-depends and install-package
already pass ${flavor:+FLAVOR=${flavor}} a few lines above; do the
same here.
Reproduced on misc/py-polars-runtime (FLAVOR=64 showing work-compat)
and audio/baresip (FLAVOR=nox11 showing work-default); a port whose
single auto-generated flavor happens to match the default (e.g.
security/py-biscuit-python) masks the bug, which is likely why it
went unnoticed.
Co-Authored-By: Claude Sonnet 5 <noreply at anthropic.com>
jail, ports: Fix garbled output when checking out/updating with -v
The svn/git checkout and update paths in both ports.sh and jail.sh's
install_from_vcs() used msg_n() followed by the raw command's own
output and a glued-on " done", relying on the command producing no
output of its own. With -v that assumption is wrong: quiet mode is
disabled, so svn/git's own progress output gets appended directly
onto the unterminated header line, and " done" gets glued onto the
tool's own last output line instead of starting a new one.
Use msg() instead of msg_n() when not running quiet, so the header
gets its own line before the command's output, and drop the redundant
trailing "done" for that case since the tool's own output already
signals completion. Apply the same fix to jail.sh, which had the
identical pattern for fetching/updating jail sources, and add the
existing "!! Any changes here should be considered for ... too."
cross-reference comment to the svn cases in both files so the two
code paths stay in sync going forward.
[3 lines not shown]
poudriered: retain queue while listing it
The list operation passes the global queue to send_object(), which
consumes its argument. A subsequent list request can then release the
stale queue again and terminate the daemon.
Take a temporary queue reference for the response so the global queue
remains valid.
(cherry picked from commit dbabedc99030f7d13696a8dd4382c4756c31fb9d)
poudriered: retain queue while listing it
The list operation passes the global queue to send_object(), which
consumes its argument. A subsequent list request can then release the
stale queue again and terminate the daemon.
Take a temporary queue reference for the response so the global queue
remains valid.
Do not import symlinks from non-pkg package builds
The package phase lets the jail build user write to a host-backed
/.npkg directory. build_port() imported every file and symlink from
that directory into the host package repository. An absolute symlink
changes meaning at this boundary: it points inside the jail while
mounted but at the host after import.
Only the pkg bootstrap variants need package-manager compatibility
links. Process those links before their targets and reject any that
resolve outside staging. Import only regular files for other ports,
preventing their build output from introducing host-resolved symlinks
into the repository.
Reject untrusted poudriered log paths
Poudriered allowed any client-authorized command to specify a log
path. The root daemon created its parent directories and opened the
path with O_TRUNC.
Reject log fields on client requests. Preserve configured schedule
logs, use a root-owned /var/log fallback, and refuse to follow final
symlinks when opening logs.
(cherry picked from commit 01d5de2876a23701e41957826cd8115702c41684)
build.html: update "Failed Ports" title bar to match others
a63486e introduced a new title bar style, but failed_div seems to have have been missed. This brings failed_div into line with the others.
