mirror of
https://github.com/uutils/shadow.git
synced 2026-10-16 02:48:14 -07:00
main
26
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Pierre Warnier
|
9bd1cec3d3 |
shadow-core: give the tool its own environment, not the caller's (#323)
harden_process() left the process environment as the caller exported it. The PAM stack, the NSS modules and the crypt library run inside the tool and read the environment themselves, so they read the caller's; only children were given a clean one. Every tool now replaces its environment first thing, while still single-threaded. What stays is what has a known reader here: TERM and NO_COLOR for how prompts are drawn; LANG, LANGUAGE and LC_* for the language of messages, PAM's included; VISUAL and EDITOR for vipw; and TZ, so chage -l prints dates in the caller's zone, but only when it names a zone -- glibc also takes a path there and reads that file with the process's privilege, so such a value is dropped. PATH is fixed. The rest is gone before the first line of the tool runs. The allow-list is a pure function with unit tests, and a further test re-runs the test binary as a child, calls harden_process there and reads the environment back. The deployment suite proves the claim where it matters. A small PAM module, tests/e2e/pam_envprobe.c, built into the e2e image, runs inside a real password change and writes out the environment of the process it runs in: a variable the caller exported is not there, and PATH is the fixed one. pam_exec was tried first and cannot do this job: it gives its child the PAM environment and the PAM items, never the calling process's, so a probe run through it saw nothing whether or not the tool had cleaned up. Checked against main's binary with the module: there the caller's variable and PATH both reach the PAM stack. Closes #314. |
||
|
Pierre Warnier
|
b516b87be8 |
docs: the upstream suite is not GNU, and is mostly BSD-licensed (#322)
* docs: the upstream suite is not GNU, and is mostly BSD-licensed The README called the original shadow-utils a GNU project and GPL-2.0+, and CONTRIBUTING grounded the clean-room policy on that. Neither holds, as #320 points out: shadow-maint/shadow is not a GNU project, and its tree is mostly BSD-3-Clause with GPL-2.0-or-later files among it -- su.c, vipw.c, contributed scripts, some translated manuals. Checked against Debian's copyright file for the package and Fedora's licence tag, BSD-3-Clause AND GPL-2.0-or-later, not against the source. The policy stands, on the accurate ground: this is an independent MIT implementation that copies from neither part. Copying from the BSD part would carry attribution terms this project does not take on; from the GPL part, a derivative work; sorting files by licence would mean reading them. So the rule is unchanged: do not read that tree. Every "GNU" that meant upstream is now "upstream" or "the distribution's package": README, CONTRIBUTING, Makefile, the man pages' section heading, the e2e image and suite, test comments. make test-gnu-compat is make test-upstream-compat and its script is renamed with it. The 2026-05-04 audit record gets an erratum at the top and is otherwise left as it was run. * docs: the last upstream-as-GNU mentions in the compat suite and the tool table * tests: say upstream, not GNU, in the test comments and skip messages too * tests: one more * docs: keep the README's clean-room line as it was, reworded around it |
||
|
Pierre Warnier
|
ab6b3df367 |
chage: let a user read their own aging line (#310)
`chage -l` is the one mode a user may run, on their own account, and it did not work for them: `make install` put chage at 0755, which cannot read /etc/shadow, and the multicall binary dropped its privilege before running it. The Makefile's own comment placed chage in bin "for chage -l, the one mode a user may run on their own account". The packaged binaries settle the contract: Debian ships chage 2755 root:shadow, the same as expiry; Fedora and Arch have no shadow group and ship it 4755 root. chage now follows expiry -- setgid shadow, or setuid root where the group is absent, and an applet that keeps euid 0 in the multicall binary. chage decides what the caller may do from the real uid, and --root is refused before the chroot on the same basis, so the privilege buys a user their own line and nothing else. Verified with the per-tool binaries installed in the debian image under both layouts; as an unprivileged user under each, chage -l on the own account answers, chage -l root, chage -M 5 and chage --root are refused, and the shadow field is unchanged afterwards. The deployment suite gains those four assertions for the multicall layout. |
||
|
Pierre Warnier
|
7e28940874 |
expiry, and usermod's subordinate-id options (#299)
* usermod: add and remove subordinate id ranges GNU usermod has -v/--add-subuids, -V/--del-subuids, -w/--add-subgids and -W/--del-subgids; ours had none of them, which came to light while writing the tests for newuidmap. The range helpers live in shadow_core::subid so whatever next touches /etc/subuid reads the same file the same way. The behaviour was established by running the GNU tool: a range the user already holds is not added twice; a removal trims an entry it overlaps, splits one it cuts through the middle of, deletes one it covers, and ignores a range never held; nothing checks other users' entries, since an administrator may share a range on purpose; and an invalid range -- reversed, not two numbers, above the 32-bit ids the kernel maps -- is exit 3 with the GNU wording, before any file is written. Entries are keyed by the login name given on the command line and are not renamed by -l, as GNU's are not: a range is a grant to a name. -V is GNU's letter for the deletion, so clap's own -V for --version had to go; --version stays. * expiry: check and enforce password expiration expiry is the twenty-eighth tool and the last the Scope table lists as in scope. It judges the caller's own /etc/shadow line as login would: silent when all is well, a warning inside warn_days, a forced change through PAM when the password has expired, a refusal when the account has. Shell profiles run it so a session that did not come through login -- a display manager, an ssh key -- still meets the aging policy. The rules are in shadow_core::shadow::Aging, one function over the aging fields with checked arithmetic, since every input is read from a file anyone who can write /etc/shadow chooses. Its order is login's: account expiry first, then a last change of 0 or an aged-out password, then the warning window, with a maximum age of 10000 or more meaning never as chage -l prints it. A locked password is a policy of its own and is not reported. The GNU tool is setgid shadow -- enough to read the file, no more -- and the change goes through PAM as the caller. The per-tool install ships it the same way, a new install class. The single setuid binary keeps euid 0 for it, having nothing narrower to offer. Under --prefix it reports and never changes anything, since a PAM change would act on this system's account of the same name; that is also what lets the required-change path be asserted in the integration tests without a prompt, while the e2e suite drives the prompt itself on a real user under a pty and checks the hash changed. groupmems is not coming. Debian 13 and Fedora do not ship it, and on Ubuntu 24.04 every action fails against PAM's default for want of a /etc/pam.d/groupmems: a tool one distribution ships and none can use. The Scope table says so. Verified: make check clean with the unprivileged run; 935 tests on alpine and 950 on fedora with PAM; 410 e2e assertions against a real install; 59 GNU comparisons; 35 arm64 assertions under emulation. * shadow-core: stop asking nscd to invalidate a shadow cache it never had Every passwd, chage and chpasswd ended with `nscd -i shadow`, and on a host running nscd that printed "nscd: 'shadow' is not a known database" on the user's terminal each time. nscd caches passwd, group, hosts, services and netgroup; the passwd entries it holds carry the `x` placeholder, so a password change leaves nothing of nscd's stale. Only the databases nscd knows are passed to it now. sss_cache still gets -U for a shadow change, which is right: sssd does cache what a password change affects. |
||
|
Pierre Warnier
|
aa987bb390 |
newuidmap, newgidmap: set the id maps of a user namespace (#298)
The setuid helpers that write /proc/<pid>/uid_map and gid_map: what Podman and rootless Docker call for every container they start. The kernel lets an unprivileged process map exactly one id into a namespace it created, its own; everything a container needs beyond that takes the privilege these carry, and they hand it out only within the ranges /etc/subuid and /etc/subgid grant the caller. Fedora ships them in shadow-utils; with these two, every tool the two reference distributions ship from the suite exists here. What is checked before anything is written was established by running the GNU helpers against a real user namespace, never by reading them: the operands must be a target and plain-decimal triples, a zero count is the overflow message; every outside range must sit inside one granted range -- an entry by login name or by numeric uid -- or be the caller's own id mapped once; root is exempt from neither, as newuidmap(1) says; and the target process must belong to the caller by real uid, account uid and the owner of /proc/<pid>, with all three named when they disagree. Overlapping ranges are refused by name; the kernel would refuse them with a bare EINVAL, so the message costs nothing and changes no accepted outcome. The map file is opened through a descriptor held on /proc/<pid>, so a pid recycled between the check and the write cannot redirect it, and the `fd:N` form takes such a descriptor from a caller that did its own checks. It is re-opened through /proc/self/fd rather than adopted, which keeps the tool free of unsafe. The tests do the real thing. A test user creates a namespace with unshare and the helper runs as that user, unprivileged, against it: the own-id mapping lands and the kernel's uid_map reads it back, the second write is refused by the kernel and reported, the fd form works, and every refusal is checked against the same live target. That runs on glibc and musl alike. The Debian image gains the uidmap package so tests/gnu-compat.sh can hold the helpers against GNU's. They take no --help -- any bad command line is the usage text and exit 1 -- so the comparison is on the refusals the two share rather than on help. Verified: make check clean with the unprivileged run; 912 tests on alpine and 934 on fedora with PAM; 393 e2e assertions against a real install, the kernel-visible writes included; 57 GNU comparisons; 34 arm64 assertions under emulation. |
||
|
Pierre Warnier
|
4ac1bdfe32 |
login: begin a session on the system (#297)
* shadow-core: take /etc/.pwd.lock before the per-file lock FileLock took the per-file `.lock` first and `/etc/.pwd.lock` second. Two tools working on different files at once could then meet in the classic shape: one holding passwd.lock and waiting for .pwd.lock, the other holding shadow.lock and waiting for passwd.lock's holder to let go of .pwd.lock. Neither can move; both sit out the fifteen-second timeout; both fail -- on a system that was merely busy. Sampling the lock files every 250 ms while the whole suite ran showed it plainly: a chpasswd held shadow.lock for 28 s and a useradd held passwd, group and gshadow for 14 s, each waiting on the other, and seven tests failed with "timed out" on one lock or the other. Three runs after the change: no failures, and the suite back from 35-50 s to 17 s. .pwd.lock now comes first, the order lckpwdf(3) and the GNU tools use and the only one that cannot deadlock: whoever holds it is the only process asking for per-file locks, so it never waits for anyone who is waiting for it. If the per-file lock then cannot be taken, the .pwd.lock reference is given back, so nothing is held on failure. * useradd: lock the shadow file with the others, and write it first useradd wrote /etc/passwd and the group files in one transaction, then opened /etc/shadow on its own afterwards -- under a comment claiming the first locks were still held, which they were not: commit_all releases them. A shadow lock that timed out at that point left the account half made: a passwd line with no shadow line, which a second useradd then reported as already existing. That is exactly what turned up when the login tests put the suite under load. The shadow file is now locked with passwd, group and gshadow before anything is written, and written first, for the reason pwconv writes shadow first: a failure between the writes must never leave a passwd line whose hash is nowhere. The old write_shadow_entry helper stays for its unit test only. A regression test puts a directory where the shadow file should be and checks that useradd fails with /etc/passwd untouched. Before this change the passwd line was written and the failure came after. * login: begin a session on the system login is the twenty-fifth tool and the one uutils/login was archived towards. It is what getty runs on a terminal: it prompts for a name and, through the `login` PAM service, a password; refuses after LOGIN_RETRIES failures or LOGIN_TIMEOUT seconds; opens the PAM session and takes its environment; records the login in utmp and wtmp; hands the terminal to the user; runs the login shell as the user; and when the shell exits, closes the session and records the logout. Which login to match was a question. Debian 13 and Ubuntu 25.10 take login from util-linux, but every release in service -- Ubuntu 20.04, 22.04 and 24.04, Debian 12 -- ships shadow's; the two were run side by side on real pseudo-terminals and agree on everything that matters. This implements the option set they share (-p, -f, -h, name) plus util-linux's -H and --help, which conflict with nothing. There is no -h for help: -h has been the remote host since telnetd. -r, the rlogin protocol, is refused. util-linux's vhangup is not reproduced; shadow's login, which the installed base runs, does not do it either. A wrong password and an unknown name get the same words, after the same delay. -f skips authentication and with it the auth stack, pam_nologin included, as both tools do. A missing home falls back to / with a notice when DEFAULT_HOME allows. The shell runs as a child so the session can be closed afterwards; it gets a clean signal mask, login keeps its own, and groups, gid and uid are dropped in that order in one pre_exec. shadow-core grew what login needed and nothing had: pam setcred, open_session, close_session, user and environment; utmp/wtmp recording; getpwnam through NSS; a getty-style spawn_with_controlling_tty; spawn_as_user; and a watchdog. Without the pam feature the applet says PAM is not compiled in and exits: a login that cannot authenticate is not a login, and the static musl archive is documented accordingly. The tests drive the real binary on a pseudo-terminal the way getty would, answering each prompt once it appears: autologin, the environment with and without -p, a wrong then a right password, an unknown user, the retry limit, the missing home, and the terminal handover. Writing them under the full parallel suite is what exposed the two fixes committed before this one. Verified: make check clean with the unprivileged run; 892 tests on alpine and 882 on fedora; 368 e2e assertions against a real install, including the password path on a pty; 52 GNU comparisons; 32 arm64 assertions under emulation. * docker: give the Fedora image the login PAM service Fedora ships /etc/pam.d/login in util-linux, which the test image did not install. With no service file PAM falls back to `other`, which on Fedora denies everything: `login -f` failed its account check and the password path was refused before any prompt, so every login test failed there and only there. util-linux is now installed, which is what any real Fedora has. The login tests then run through Fedora's own stack -- pam_selinux, pam_loginuid, pam_namespace -- and pass, which makes it the second PAM configuration the session code is exercised against rather than a gap. |
||
|
Pierre Warnier
|
ea2c85cf6c |
pwconv, pwunconv, grpconv, grpunconv: convert to and from shadow files (#295)
* pwconv, pwunconv, grpconv, grpunconv: convert to and from shadow files
These four complete the set both Debian and Fedora ship: every tool the two
distributions agree on now exists here. pwconv moves password hashes out of
the world-readable /etc/passwd into /etc/shadow, creating it if need be, and
reconciles the two; pwunconv merges them back and removes /etc/shadow;
grpconv and grpunconv do the same for the group files. They are one engine
with a selector, and the other three crates are the selector.
The order of writes is the substance of the change. pwconv writes shadow
before passwd: the hashes are copied into the shadow file and only then
replaced by `x`, so a failure between the two writes leaves them where they
were rather than nowhere. pwunconv writes passwd before it removes shadow:
for a moment the hashes exist twice, which is recoverable, where the other
order would have a moment with none. The removal happens while the lock on
the shadow file is still held, so no other tool can slip in between the last
read of the file and its disappearance.
What pwconv does was established by running the GNU tool:
- A hash found in passwd beside an existing shadow line is the newer of the
two -- it got there through a tool or an edit that did not know about
shadow -- so it replaces the shadow hash and the change is dated; the
aging is kept.
- A passwd line with no shadow line gets one, dated today, with the aging
from PASS_MIN_DAYS, PASS_MAX_DAYS and PASS_WARN_AGE. The hash field is
copied as it stands, an `x` included: an `x` in the hash field matches no
password, which is the honest state of an account whose hash was lost.
- A shadow line for an account no longer in passwd is dropped.
- A consistent system is left byte for byte alone.
A shadow file this run creates is made 0640 root:shadow, the layout the
distributions ship, with the `shadow` group looked up in the tree being
converted rather than through the name service: under --prefix or --root the
host's shadow group says nothing about the target's. Where the tree has no
such group the file stays 0600 root, the atomic writer's default.
The e2e suite uses the other tools as witnesses rather than only reading
files back: after pwunconv and pwconv the hash is byte for byte what it was,
passwd -S still reports the account as having a password, and after
grpunconv and grpconv sg still admits a non-member with the group password
and still refuses a wrong one.
Verified: 878 tests on debian, alpine and fedora; make check clean including
the unprivileged run; 356 e2e assertions against a real install; 51 GNU
comparisons; 31 arm64 assertions under emulation, the round trip included.
* tests: hand children their stdin as a file, not a pipe
The integration binary died on two CI runners with SIGPIPE: "write on a
pipe with no one to read". Not a failing test -- the test runner itself was
killed, part way through, and every test after it never ran.
Two things combine. The batch tools exit as soon as an input line is bad,
so a child fed "staff\n" has closed its stdin within a millisecond of
starting. And the `uumain` that `#[uucore::main]` generates resets SIGPIPE
to its default disposition every time it runs, in-process, for GNU pipeline
compatibility; several tests call `uumain` in-process -- deliberately, where
the tool would exec -- so once any of them has run, the whole test binary
has SIGPIPE at default. From then on a stdin write that lands after the
child has gone is not an EPIPE the helper could handle but a signal that
ends the process. The window is a few microseconds wide, which is why it
showed up on loaded CI runners and never locally.
The helpers now write the input to a file in the prefix tree and hand the
child that file as its stdin. A file has no reader to lose, so the class is
gone regardless of what the signal disposition is at the time. Applied to
chpasswd's helper too, which had the same shape and the same latent hazard.
Fifteen runs of the integration binary at sixteen threads after the change:
no failures. The race could not be reproduced on demand before it, so that
number is not a before-and-after; the diagnosis rests on uucore's source and
on the two CI logs.
|
||
|
Pierre Warnier
|
8575f629a8 |
vipw, vigr: edit the account files under the suite's lock (#294)
vipw and vigr are the nineteenth and twentieth tools, and the fourth and
fifth of the nine that both Debian and Fedora ship and we did not. They
hand /etc/passwd, /etc/shadow, /etc/group or /etc/gshadow to the
administrator's editor under the same lock every other tool takes, and
install the result atomically. vigr is vipw with the group file as its
default -- the GNU suite ships it as a symlink -- so it is a crate that calls
uu_vipw::run with a different default and nothing else.
The reason the tool exists is the lock, so that is what the e2e suite proves:
with the editor open on /etc/passwd, a concurrent useradd waits rather than
failing or racing, and both changes land.
Two deliberate departures from GNU, both established by running its binary:
- The edited file is parsed before it is installed. GNU installs whatever
the editor saved: a line with no colons went in, and so did
`evil::0:0:::/bin/sh`, a UID 0 account with an empty password field.
Here a file the rest of the suite could not read back is refused. The
check is structural -- pwck and grpck still judge the contents -- and
when it refuses, the edit is kept beside the file rather than discarded.
An administrator whose ten minutes of work vanished would use a raw
editor next time, which is the outcome this tool exists to prevent.
- Changes are detected by content, not by timestamp. GNU compares
modification times in whole seconds, so an edit saved within the same
second as the copy was made is silently thrown away. The first probe of
the GNU tool hit exactly that: every scripted edit came back
"unchanged" until the script was slowed down.
The working copy sits beside the target as passwd.edit, created under a
zeroed umask with the original's mode and ownership, so a copy of
/etc/shadow is never more readable than /etc/shadow. A stale copy from a
crashed run is replaced, and the editor is handed the live contents.
vipw blocks SIGINT and friends for the whole edit so a Ctrl-C aimed at the
editor cannot kill it holding the lock. That mask would be inherited by the
editor, which then could not be interrupted at all, so
shadow_core::process::spawn_with_signals_unblocked starts the child with a
clean mask -- one pre_exec in the module where unsafe is allowed.
Verified: 852 tests on debian, alpine and fedora; make check clean including
the unprivileged run; 319 e2e assertions against a real install; 47 GNU
comparisons; 24 arm64 assertions under emulation.
|
||
|
Pierre Warnier
|
5db310fc88 |
newusers: create or update accounts in batch (#293)
newusers is the eighteenth tool and the third of the nine that both Debian
and Fedora ship and we did not. One input line of seven fields produces a
whole account: the passwd record, the hashed shadow record, a group, and the
home directory with the skeleton copied in.
Every line is parsed, every name validated, every group resolved and every
password hashed before anything is written. A batch with one bad line leaves
the system exactly as it was, including for the good lines above it -- the
failure mode is "nothing happened", not "the first two hundred accounts
exist and the rest do not". That is the whole reason to reach for this tool
over a shell loop around useradd.
Home directory creation moved to shadow_core::home rather than being copied.
useradd's version forces the umask so the requested mode is exact, and hands
over ownership through a descriptor opened O_NOFOLLOW so nobody who can write
the parent can swap in a symlink between the mkdir and the chown. A second
copy of that is a second chance to get it wrong, and the mistake would be
silent. useradd now calls the shared one and its own tests still pass.
Three deliberate divergences, all established by running the GNU tool:
- An empty password field is refused before anything is written. Hashing
"" gives a valid hash that a bare Enter matches. GNU passes the empty
field to PAM, which refuses it *after* creating the account, leaving a
half-made account behind -- verified.
- A pw_gid naming a group that does not exist is refused. GNU falls back to
the user's own ID and creates no group, so the account is left pointing
at a GID that is not there, which grpck then reports.
- A missing parent directory is created, as useradd -b does. GNU's newusers
fails there, which makes it disagree with GNU's own useradd on the same
path.
--badname is implemented rather than stubbed: it drops the portability rules
on the login name while keeping the checks that stop a name from corrupting
the file or being read as an option, which is what makes the flag safe to
offer at all. It is what lets a domain-qualified name from a directory join
through.
The e2e suite re-hashes the stored field with its own salt and compares, so a
tool that wrote a well-formed hash of the wrong string would fail; the
negative control, the same check against a different password, is asserted to
fail. An earlier version of that assertion ended in `|| true` and was
therefore always green -- worth saying, because it is the failure mode these
checks exist to catch.
Verified: 828 tests on debian, alpine and fedora; make check clean, including
the unprivileged run; 297 e2e assertions against a real install; 45 GNU
comparisons; 22 arm64 assertions under emulation.
|
||
|
Pierre Warnier
|
f37939d9fc |
chgpasswd: set group passwords in batch (#292)
* chgpasswd: set group passwords in batch
chgpasswd is chpasswd's counterpart for groups: it reads group:password
lines from stdin and applies them. It is the seventeenth tool and the second
of the nine that both Debian and Fedora ship and we did not.
It follows chpasswd's rule, which is what makes a batch tool safe to run
unattended: every line is parsed, every group resolved and every password
hashed before any file is written, so a batch naming one group that does not
exist changes nothing rather than applying the lines above it and stopping.
The two account files are then written in one transaction.
Where the hash goes was established by running the GNU tool, not by reading
it. With /etc/gshadow present the hash belongs there and /etc/group keeps the
`x` placeholder -- /etc/group is world-readable. Without it the hash goes
into /etc/group, and no gshadow file is created: conjuring one up would
change how every other tool on the host reads group passwords.
Three deliberate divergences, all refusals:
- `-c NONE`, which GNU honours by storing the password as clear text in
/etc/gshadow. A readable group password is worth no more than none at
all, and `-e` already writes a field verbatim when that is what is
wanted. The message says so, or an operator would just try `-c MD5`.
- `-m` and `-c MD5`/`-c DES`, as chpasswd already refuses them.
- An empty password in plaintext mode: hashing "" yields a valid hash that
a bare Enter matches, which is a group anyone can enter rather than a
group with no password. `-e ''` remains the way to clear the field.
A field containing a colon is refused, since it would split the gshadow line
and corrupt the file. The GNU tool refuses it too and likewise leaves the
file untouched, so this is agreement, not divergence.
The e2e suite checks the tools against each other rather than only against
the files: the password chgpasswd sets is the one sg then accepts from a
non-member, and a wrong one is refused. A hash written in a format nothing
can verify would pass a file-shape assertion and fail this one.
Verified: 798 tests on debian, alpine and fedora; make check clean; 273 e2e
assertions against a real install; 44 GNU comparisons.
* tests: run the suite as an unprivileged user too
Sixteen of the chgpasswd tests assumed root. They passed locally because
every container in docker-compose.yml runs as root, and failed in the one CI
job that does not.
Worse than the two that failed outright were the ones that passed for the
wrong reason: a test asserting exit 1 got exit 1 from "permission denied"
rather than from the condition it meant to check, so it was green while
testing nothing. Only the two expecting exit 0 gave the game away.
The guard itself is what chpasswd already does -- skip_unless_root() on every
test that writes an account file. The two that only exercise argument parsing
keep running everywhere, since clap answers before the root check.
The gap that let this through was local, so the fix is too: `make
test-unprivileged` re-runs the already-built test binaries as an ordinary
user, and `make check` now does it after the root run. Removing one guard
again makes it fail with "Permission denied", which is how it was verified.
SHADOW_TEST_REQUIRE_ROOT is cleared for that run on purpose: it exists to
turn a skip into a failure when the suite *is* root, which is the opposite
case.
|
||
|
Pierre Warnier
|
651ed5c62b |
sg: run a command with a different primary group (#291)
sg is the sixteenth tool, and the first of the nine that both Debian and
Fedora ship and we did not. It runs a single command with a different
primary group: newgrp for one command instead of for a whole shell.
sg and newgrp are not merely similar. They decide who may enter a group by
the same rules and enter it the same way, and the distributions ship sg as a
symlink to the setuid newgrp binary -- one implementation, dispatched on
argv[0]. So the shared half moves into shadow_core::group_switch and both
tools call it, which is what keeps them from drifting apart. The multicall
binary already dispatches on argv[0], so it needs only the new applet.
sg execs the command rather than forking and waiting for it. The shell then
is the process, so the command's exit status reaches the caller unaltered:
`sg staff -c 'exit 7'` exits 7 and a command that cannot be run exits 127,
without sg reconstructing a wait status. This matches what the GNU tool does
and costs a process.
Writing the shared module turned up two things wrong in newgrp:
- It refused a group with no password without prompting, and prompted when
the group had one. Whether a prompt appeared therefore told any caller
which groups have passwords set, which is a list of the ones worth
attacking. Both tools now prompt either way and refuse afterwards.
- It rebuilt the supplementary group list with initgroups() alone, which
builds it from the member lists in /etc/group -- where a primary group
never appears. `newgrp staff` therefore dropped the group the caller
started in, costing them access to their own files until they left the
shell. The original group is added back, which needs setgroups(2) and so
has to happen before the setuid privilege is given up.
Both were established by running the GNU tools, not by reading them.
sg is setuid for the reason newgrp is: a caller who is not a member has to
be checked against the group password in /etc/gshadow, which is root-only.
Removing sg from the setuid list fails six assertions in the e2e suite --
including for a plain member, since initgroups() itself needs the privilege
-- and nothing else in the suite covers it.
Verified: 763 tests on debian, alpine and fedora; 255 e2e assertions against
a real setuid install, covering a member, a non-member with the group
password, a wrong password, a passwordless group and exit-status
propagation; 43 GNU comparisons; 19 arm64 assertions under emulation.
|
||
|
Pierre Warnier
|
f3a20b2043 |
tests: cover gpasswd as a group administrator (#287)
The reason gpasswd is setuid is that someone who is not root, but is named an administrator of one group, manages that group's membership. Nothing in the suite exercised a non-root caller of it, and that is the only place the omission shows: the multicall binary drops privilege for every applet not on its setuid list, so leaving gpasswd off that list produces a permission error there and nowhere else. It was off the list when the tool arrived. Twelve assertions: an administrator adds a member without being root and the change lands in both group and gshadow; a plain member may not change the membership, may not appoint administrators and may not use --root; and the administrator list is unchanged after all three refusals. Checked that the test bites by taking gpasswd back off the setuid list and rebuilding -- the administrator is then refused and the group file is untouched. Worth saying that the first attempt at that check appeared to show the fix was unnecessary; the build output had been suppressed and a stale binary was installed. |
||
 Jocelyn Le SageandPierre Warnier
|
8361bf512b |
gpasswd: added implementation of gpasswd (#282)
* gpasswd: added implementation of gpasswd
* gpasswd: rebase onto main and port to the current APIs
Rebases jlesage's implementation onto a main that moved a long way under
it, and carries it onto the interfaces that landed since the branch was
cut. The tool's behaviour is unchanged and still matches GNU on every
path probed.
What the rebase had to resolve:
- shadow-rs.spec and debian/ were revived by the merge; both were
deleted because neither could build.
- The standalone dispatch() and hand-written --list predate the single
applets() table, which now drives both. gpasswd joins the table.
- The five shadow-core parser features it asked for no longer exist.
- The fourteen [[test]] targets are one; test_gpasswd joins it.
- README, Makefile and PLATFORM-SUPPORT described the pre-#239
privilege model, in which every applet ran with euid 0.
What changed beyond the rebase:
- gpasswd keeps its privilege in the multicall build. It is setuid in
GNU and the branch marks it setuid for the per-tool install, but
SETUID_APPLETS still listed four names, so the multicall build would
have dropped privilege and no group administrator could have used it
-- silently.
- Comments in /etc/group survive. It read entries only and wrote them
back, which erases every comment, blank line and NIS compatibility
line in the file. The locked-file transaction keeps them, and commits
group and gshadow together so a bad value cannot leave the pair
disagreeing.
- The duplicated EchoGuard and password reader are gone, replaced by
shadow_core::tty, which blocks interrupts for the read so Ctrl-C at
the prompt cannot leave the terminal with echo off.
- The print macros are replaced. They panic when the stream is closed,
which a setuid-root tool must not do part way through a change.
These were the only five in the repository and no lint caught them,
so clippy::print_stdout and print_stderr are now denied
workspace-wide, with the test crate exempt.
The GNU comparison suite gains gpasswd: the unknown-group, unknown-user
and non-member exit codes, all three of which match.
CONTRIBUTING's licence list also said ISC, BSD-2-Clause, CC0-1.0, Zlib
and MPL-2.0 were acceptable, which stopped being true when deny.toml was
narrowed to the graph. It now says what deny.toml enforces.
---------
Co-authored-by: Pierre Warnier <pierre@warnier.net>
|
||
|
Pierre Warnier
|
eb940c0b46 |
--root is a real chroot in every tool, not a path prefix in eight of them (#277)
* --root is a real chroot in every tool, not a path prefix in eight of them
Five tools called chroot(2) for --root. The other eight -- groupadd,
groupdel, groupmod, grpck, pwck, useradd, userdel, usermod -- folded the
argument into the same resolver --prefix uses, so it only prepended the
directory to the files the tool itself opened.
The difference shows whenever an absolute path is read out of a record
rather than being one of the tool's own files:
useradd -R /mnt/sysroot -m alice
created the home on the *host*, at /home/alice, because the path came
back from login.defs as /home/alice and nothing prefixed it. The skeleton
was copied from the host's /etc/skel. userdel -r -R had the mirror-image
problem and resolved the home to delete against /.
All thirteen now chroot, through one helper in shadow-core rather than
five copies of it, and refuse --root for anyone but root: a setuid-root
binary pointed at a tree of the caller's choosing would read and write
account files that caller controls.
pwck and grpck have no --prefix, matching GNU, so their tests move to a
spawned process: --root cannot be exercised in-process at all now, since
a chroot in the test binary takes every later test into the tree with it,
where /tmp does not exist. That is how this was caught. The other tools'
in-process tests use -P, which is what they meant.
Verified in the deployment container against a minimal target tree: the
account, its home and the skeleton all land inside it, groupadd/groupmod/
groupdel operate on the tree's group file, pwck and grpck read the tree,
userdel -r removes the home from the tree, and the host is untouched
throughout.
Closes #270.
* pwck: name the shadow file in the bad-entry test
The test passed only the passwd file, so the shadow path defaulted to the
host's /etc/shadow. Unprivileged, that cannot be read, and the test
reported 'cannot open' (exit 3) instead of the bad entry it is about
(exit 2). It passed in the containers, which run as root, and failed on
the host runner.
It was --root that hid this: as a path prefix it pointed the default at
the fixture. --root is a chroot now, so the fixture has to name both
files.
* tests: guard the pwck tests that chroot
They spawn 'pwck --root', and chroot(2) needs root: the tool refuses the
option for anyone else. The containers run as root so they passed there,
and the host runner failed them.
|
||
|
Pierre Warnier
|
cf7ecb4a0f |
tests: one binary, spawn the real tool, and assert on what it prints (#273)
* tests: one binary, spawn the real tool, assert on what it prints
Fourteen [[test]] targets each linked all fourteen tool crates and
carried its own copy of the helpers through a #[path] include, so the
suite was compiled and linked fourteen times over. It is one target now,
the layout the rest of uutils uses.
More importantly the tests called uumain in-process and asserted on an
exit code and nothing else. Not one assertion in the suite looked at a
byte of output, so 'bit-for-bit identical to GNU' was never checked. A
shared process also let harden_process, setuid, Landlock and umask
changes leak from one test into the next.
common now offers a spawned-binary helper with assertions on the exit
code and both streams, and the tests use it. Three findings came out of
that within the hour:
- chage exited 3 for an invalid date where GNU exits 2; 3 is for an
unexpected failure, not for an argument the caller got wrong.
- passwd -l was not idempotent. Eight concurrent locks left eight
markers, and since passwd -u refuses to unlock an account that would
stay locked, an account locked twice could not be unlocked with the
tool at all. GNU leaves a single marker however many times it is
asked; two unit tests that encoded the old behaviour are rewritten
to the verified one.
- pwck's duplicate-UID test accepted either 0 or 2, so it asserted
nothing. Two accounts may deliberately share a UID (useradd -o), and
GNU reports nothing; a duplicate *name* is the fault, and that is now
what is checked, message included.
test_chage and test_chpasswd never invoked their tool at all -- they set
a field through shadow-core and read it back, testing the parser twice.
Both now drive the binary through --prefix, which those tools have since
gained, and cover the aging-field writes, the -l wording and thresholds,
the batch rules, ENCRYPT_METHOD and the flag pairs the man page forbids.
The concurrency test raced four threads through uumain against a lock
keyed on the process id, which same-process threads never contend for,
and discarded all eight return codes. It spawns eight processes now and
checks every one.
New: the multicall binary is tested for the property one table now
guarantees -- every applet listed is reachable, and the help text is
identical through the argument and the symlink.
* passwd: make -l idempotent
Locking prepended '!' unconditionally, so locking an account twice left
'!!'. One unlock then removed one marker and left the account locked --
and passwd -u refuses to unlock an account that would stay locked, so
the tool could not undo its own second lock.
GNU shadow 4.17 leaves a single marker however many times it is asked:
three locks give one '!', which one unlock removes. Unlocking a field
that already carries two markers still removes exactly one, which is
also what GNU does; that case is reachable only from a file some other
tool wrote.
Found by running eight concurrent passwd -l processes against one
account and looking at the result, which is a thing the suite could not
do while it called uumain in-process. Two unit tests asserted the old
behaviour and now assert the verified one, and the deployment suite
locks twice and checks the marker did not double.
* ci: make root-only skips visible, and test root together with pam
Roughly three quarters of the integration suite is root-only, and a
skipped test is reported ok, so a run that exercised almost nothing
looked exactly like a run that exercised everything.
SHADOW_TEST_REQUIRE_ROOT turns a skip into a failure. The Docker images
set it, since they run as root: 27 useradd tests that reported ok off-root
fail loudly there now if the suite ever stops running as root.
The matrix also runs with pam. The feature changes which code paths
exist -- chfn and chsh refuse every non-root call without it, and passwd
loses its interactive path -- so root and PAM together had never been
exercised by any job.
* fuzz: assert a real invariant, add a subid target, build in CI
The targets asserted only that parsing does not panic. For a record file
the property that matters is that a parsed entry survives being written
back: if Display can emit a line that parses differently, a tool
rewriting the file silently changes an account. Each target now checks
that the rendered line carries exactly the separators its format has --
one more would shift every following field on the next read, which is
how an injected account appears -- and that parsing it again renders
identically. /etc/subuid had no target at all.
The crate is a separate workspace nothing else builds, so it had rotted:
its manifest asked shadow-core for features that no longer exist. A
stable 'cargo check --bins' in CI catches that without nightly.
The corpus was both gitignored and tracked. It is kept: 667 accumulated
inputs are worth carrying between runs, and the main suite now replays
them through the same two properties on every pull request, so a
regression a fuzzer once found is caught on stable without a fuzzing
session.
* chage: an invalid date is a bad argument, not an unexpected failure
'chage -E 2025-02-29' exited 3, which chage(1) uses for something going
wrong that the caller did not ask for. A date the caller typed wrongly is
exit 2, which is what GNU shadow 4.17 answers -- 'chage -d -5 u' prints
"invalid date '-5'" and exits 2.
Found by the integration tests once they ran the real binary and looked
at its exit code.
* tests: guard the chage tests that need root, and let an empty flag mean off
The chage tests drive the tool against a --prefix tree, but 'chage -l' on
an account other than the caller's, and every modification, are refused
for a non-root caller before any file is opened. The host runner is not
root, so those tests failed there.
SHADOW_TEST_REQUIRE_ROOT also treated an empty value as set, so
'-e SHADOW_TEST_REQUIRE_ROOT=' turned the requirement mysteriously on
rather than off. An empty value now counts as unset, which is the
conventional reading.
* ci: do not ask Alpine for a pam build
Alpine's Rust image targets musl with crt-static, where shadow-core
refuses the pam feature at compile time on purpose: Linux-PAM loads its
modules with dlopen(3), which a static binary cannot do, so a build that
compiled would produce an authentication path that never works.
The new 'with pam' matrix step ran on all three distros and so hit that
guard. It runs on the glibc images now. The static musl job already
asserts that requesting pam there is rejected, and Alpine is in the
matrix for musl's libc behaviour, not for PAM.
|
||
|
Pierre Warnier
|
9cd5eef4ac |
Install layout, aarch64 archives, pinned images, and a pre-push hook worth keeping (#272)
* hooks: stop reproducing CI in pre-push The hook ran fmt, clippy and the test suite on Debian, Alpine and Fedora in series, on every push. That is several minutes for a push that may carry nothing: pushing the 0.2.2 tag moved one object and still blocked for over two minutes. The predictable result was that every push in the last batch of work used --no-verify, and a hook that is always skipped protects nothing while still costing time whenever someone forgets to skip it. It now runs the workspace tests once, on Debian, and returns immediately for a push that carries no new commits -- a tag, a branch deletion, or a branch the remote already has. git hands the hook one line per ref on stdin; the ranges it names are asked for a single commit, and if there are none there is nothing to test. Everything the hook dropped is already a required check on the pull request, running in parallel with retries: fmt, clippy and the tests with and without pam, MSRV, cargo-deny, the static musl archive, the three-distro matrix as root, and the end-to-end deployment image. Closes #231. * install: put the user-facing tools in bin, not sbin All fourteen binaries went to $(PREFIX)/sbin, which is not on a normal user's PATH: passwd was 'command not found' for exactly the people it is setuid for. GNU shadow splits them, verified against the Debian package -- passwd, chage, chfn, chsh and newgrp in /usr/bin, the other nine in /usr/sbin. Both install targets do the same now, and the multicall symlinks in bin point at the absolute path of the binary in sbin. chage is in bin for 'chage -l', the one mode a user may run on their own account. The deployment suite follows each tool to the directory it belongs in, and asserts that all five are reachable by name for an unprivileged user, which is the property the split exists for. make test also runs the suite with pam now. install ships binaries built with it, so a run without it covered a build nobody ships. * docker: pin the base images, drop the unused nextest download rust:latest, rust:alpine and fedora:latest meant a new toolchain or distro release landed on main with no pull request and no signal, and clippy gains lints between Rust releases. They are pinned to what the images resolve to today -- rust 1.98 on Trixie and Alpine 3.23, Fedora 44 -- and Renovate's Dockerfile manager proposes the bumps, so a toolchain change arrives as a reviewable PR that CI has run. Each image also downloaded cargo-nextest from get.nexte.st/latest on every build: unpinned, unverified, and never invoked -- nothing in the repository runs 'cargo nextest'. cargo-deny, which is used, keeps its pinned version and checksum. * deny: allow exactly the licences the graph uses The allow list carried ISC, BSD-2-Clause, CC0-1.0 and Zlib, none of which appears anywhere in the dependency graph. A list wider than the graph means a crate arriving under a new licence passes silently; the point of the check is that such an arrival is a decision someone records. It is now MIT, Apache-2.0, BSD-3-Clause (subtle, for the constant-time comparison in shadow-core::crypt) and Unicode-3.0 (tinystr and unicode-ident, via uucore's fluent stack), taken from 'cargo deny --all-features list'. Two crates carry a copyleft option in a dual licence and are accepted under the other half; that is now written down rather than left for the next reader to work out. All four checks pass with the narrower list. * ci: run the advisory workflow on a schedule only It also ran on every push touching a manifest or the lockfile, where the Cargo Deny job in ci.yml had already checked the same advisory database alongside licences and bans. Two tools reporting the same finding on the same commit is noise. What the pull request job cannot do is notice an advisory published against code nobody has touched, which is the reason to keep a scheduled run. That is now all this workflow does, plus manual dispatch. * packaging: remove the RPM spec and debian directory Neither can build. debian/ has no changelog at all, so dpkg-buildpackage stops before it starts: dpkg-buildpackage: error: cannot open file debian/changelog and shadow-rs.spec pins version 0.2.0 against a 0.2.2 workspace, points at the pre-transfer repository URL, and lists a %files layout that 'make install' does not produce -- it names the multicall binary, which only install-multicall builds, and puts every tool in sbin, which is no longer where half of them go. Shipping packaging that claims to work and does not is worse than shipping none, and this would have produced a package declaring 'Provides: passwd, login' on a system whose real passwd it replaced. Releases are the archives dist publishes; a distribution packager writes these against their own policy anyway. * release: publish an aarch64 archive The pipeline shipped x86_64 only. arm64 servers and Apple-silicon containers are common enough that a missing archive is a real gap. It builds on a native arm64 runner rather than cross-compiling, which is what makes it straightforward: shadow-core::crypt links crypt(3) and the pam feature links libpam, so a cross build would need an arm64 sysroot carrying both, while a native runner installs them from apt like any other build. dist has picked a native runner for aarch64 since 0.30. Both runners are pinned to 24.04. dist's defaults are ubuntu-22.04 and ubuntu-22.04-arm, and 22.04 enters deprecation on 2026-09-17 with the image unsupported from 2027-04-17; the pin keeps the two archives on the same generation and off an image that is about to start failing. The higher runner raises the glibc floor for the dynamically linked archives, which is what the static musl archive is there for. release.yml is regenerated by 'dist generate'; 'dist plan' now lists three archives and resolves ubuntu-24.04-arm for the arm64 job. Closes #222. |
||
|
Pierre Warnier
|
4e539357a1 |
shadow-core: remove duplicated lookups, fix login.defs radices, allocate around NSS (#271)
* shadow-core: delete the duplicated NSS entry type and the unused surface process::PwEntry was a field-by-field twin of passwd::PasswdEntry, copied across at every NSS lookup. getpwuid now returns the parser's own type: an NSS entry and a file entry describe the same account, and a second struct for it only bought a copy per call. lookup_username goes with it, and lookup_username_by_uid becomes one line. The tty module had its own signal-mask guard beside hardening's SignalBlocker, and the PAM conversation had its own controlling-terminal lookup. There is now one of each. error::ShadowResult and ShadowError::Permission had no users at all, and display_message carried an _is_error parameter it never read. harden_process returned a sanitized environment that every one of its thirteen callers discarded, which read as though the process environment had been dealt with. It has not been and is not: the in-process PAM and NSS modules still see what the caller set. The function no longer claims otherwise, and the callers no longer bind a value to throw away. Adds gid_exists, the group-side companion to getpwuid, for the allocator. * shadow-core: read login.defs numbers in the radices the man page documents login.defs(5) writes numbers 'in decimal, or in octal with a leading 0, or in hexadecimal with a leading 0x'. get_i64 parsed decimal only, which is not a rejection but a silent misreading: UID_MIN 01000 means 512 and came back as 1000, so useradd would allocate from a range the administrator had reserved. UMASK 022 and HOME_MODE 0750 have the same shape and will be read the day they are honoured. LoginDefs also gains Default, an empty configuration, which is what a caller wants when the file is missing: every lookup returns None and the tool falls back to its own default. * useradd, groupadd: never allocate an ID the name service already knows /etc/passwd is not the only source of accounts. An ID that is free in the file may belong to an LDAP, SSSD or systemd-homed user, and handing it out gives two accounts one ID -- which the kernel cannot tell apart, so the new local user ends up owning the directory user's files. Allocation now asks NSS as well as the file. Which sources count is an explicit Scope rather than a hidden default, because a --prefix or --root run manages another system's files: the name service running here describes this one, so asking it would skip IDs that are free on the target and claim none that are taken there. SysRoot::is_prefixed decides. uid_range and gid_range were the same function twice over different keys and are now one, the system ranges being the same keys with a SYS_ prefix. The allocation tests moved from next_uid to next_free. Once next_uid consults NSS its answer depends on which accounts exist on the machine running the test, and on a typical container 1000 and 1001 do. The policy is tested where it lives, and two tests assert the property that actually matters against the live system: an allocated ID is in range, not in the entries given, and unknown to NSS. * audit: write to /dev/log instead of forking logger for every event Each account event spawned /usr/bin/logger and waited for it, to write a few dozen bytes to a socket. A setuid-root tool should spawn as little as it can: the fork inherited the process state, depended on a binary being installed at a fixed path, and blocked the caller. The record now goes straight to /dev/log as an RFC 3164 datagram, and carries the terminal it came from instead of a literal '?'. Failure stays silent in every direction. A machine with no syslog daemon is a normal machine, and a record that cannot be written must not fail the operation that produced it. The deployment suite runs rsyslogd and greps the log for the creation and deletion records, so this is covered rather than assumed. Also corrects atomic's claim that its temp-file names are unique 'across threads' in a single-threaded crate; what they separate is successive writes within one process, which usermod does four of in a row. * docs: record the hardening changes and what --root really does sysroot's module documentation said every tool offering --root performs a real chroot. Five of the thirteen do; the other eight fold it into the same resolver --prefix uses, so an absolute path read out of a record resolves against / rather than the target. That is a real divergence worth its own change and is filed as #270; the documentation now says which tools do which and points there. * cargo: only gate what changes linkage; stop forcing applets in with pam Five of shadow-core's features -- shadow, group, gshadow, login-defs, subid -- gated modules that pull in no dependency. Gating them bought nothing and cost real confusion: cargo test -p shadow-core compiled a fraction of the crate and reported a pass for it. It runs 181 tests now where it ran 20, without a single one being written. uid_alloc loses its cfg(all(group, login-defs)) with them. What remains behind a feature is what actually changes the build: pam (links libpam), crypt (links libcrypt on glibc, and now pulls subtle, which nothing else uses) and landlock. In the root manifest, pam = ["passwd/pam", ...] force-enabled the three applets that can use PAM, so asking for PAM in a build that had deliberately left them out dragged them back in. The optional-dependency form pam = ["passwd?/pam", ...] applies the feature if the applet is there and does nothing if it is not. The feat_* aliases and the feat_phase* groups are gone. Every one was a rename of a feature cargo already provides -- feat_passwd = ["passwd"] -- and nothing outside the manifest had referred to them since the phases they were named for finished. Verified against all four delivery paths: make build and make build-multicall still link libpam, the static musl archive still builds and reports all fourteen applets, and a --no-default-features --features pam build now produces a binary with no applets at all and no warnings, which is the case the ? was added for. |
||
|
Pierre Warnier
|
6442faf970 |
chage/passwd/chpasswd/newgrp: aging bounds, input rules and newgrp - (#269)
* chage, passwd: one date module, GNU's -l output, and exit 1 for an unknown login
The Hinnant calendar algorithms were written out three times. They now
live in shadow-core::date, which also gained the inverse direction and
the two display forms, with the arithmetic checked.
That matters here: 'chage -l' adds lastchg + max + inactive, three values
read from a file anyone who can write /etc/shadow chooses. Plain addition
wraps in release builds and panics in debug ones. The sum is now bounded
and an unrepresentable date shows as 'never', which is what GNU displays
for a field it cannot make sense of.
Two -l rules were wrong, both verified against GNU shadow 4.17 in the
Debian image:
- A last change of 0 is passwd -e's 'must change at next login' marker,
not a date. GNU prints 'password must be changed' on all three
password lines; we printed 1970 dates for the two derived ones.
- Expiry is disabled at a maximum age of 10000 days, not 99999. -M 9999
shows a date, -M 10000 shows never.
Both tools also exited the wrong code for an unknown login: chage 15,
which chage(1) reserves for 'can't find the shadow password file', and
passwd 3, which is for an unexpected failure. GNU exits 1 for both.
chage gains --prefix, so its behaviour can be exercised without touching
the live files. The display tests, which asserted only that a string
contained '1970' and that printing did not panic, now assert the exact
output line by line.
* chpasswd: honour ENCRYPT_METHOD, reject the flag pairs the man page forbids
The default hashing scheme was hard-coded SHA-512 while ENCRYPT_METHOD in
login.defs is how a distribution picks one for the whole system. On
Debian, which sets YESCRYPT, every password set through this tool was
weaker than the one the same user would get from passwd. -c still
overrides, and an unreadable or unusable configuration falls back to
SHA-512 rather than refusing to set a password.
Two flag combinations chpasswd(8) documents as errors were accepted and
quietly did nothing: -s without -c (a rounds count with no scheme to
apply it to) and -e with -c (a scheme for input that is already hashed).
Both are now usage errors, and usage errors exit 2 as the man page says.
The batch is also resolved before it is applied: an unknown login in the
middle of a list used to leave the first half written. Membership is
looked up through an index rather than a scan per pair, which took the
critical section -- held with signals blocked -- out of quadratic time.
Also drops the setuid(0) before the write, which raised the real uid and
made caller_is_root() answer true for the rest of the process, and owns
each input line in a zeroizing buffer: every one of them holds a
password.
Adds --prefix.
* newgrp: implement '-', and stop making every shell a login shell
newgrp(1) is 'newgrp [-] [group]'. The dash was not implemented, so
'newgrp - docker' looked for a group called '-' and reported it missing.
Worse, every invocation execd a login shell (argv[0] = -sh) whether or
not the dash was given. newgrp(1) says the current environment and
working directory are kept without it, so the profile files were being
re-read on each 'newgrp docker', in an environment they had already been
applied to.
Now: without '-', a non-login shell inheriting the environment; with '-',
a login shell, chdir to the home directory, and an environment built
explicitly (HOME, SHELL, USER, LOGNAME, PATH, plus TERM and the locale
variables a real login inherits from the terminal). That needs execve,
added to shadow-core::process, because std::env::set_var is unsafe and
process-global in edition 2024.
Membership and password sources were checked against GNU shadow 4.17
rather than changed on assumption, with four groups covering the cases:
membership comes from /etc/group and the password from /etc/gshadow, and
a user listed only in gshadow's member list is not admitted. That is what
this already did, so it stays.
The duplicated EchoGuard is gone; the prompt goes through the shared
terminal helper, which blocks interrupts for the read. The env
sanitization call that computed a value and dropped it is gone too, with
a note on why newgrp deliberately does not sanitize.
The parse tests now exercise the operand grammar instead of asserting
that clap stores a string.
* shadow-core: scrub the crypt(3) password copy, identify the caller to PAM
crypt(3) takes C strings, so a copy of the password has to exist.
CString::new made that copy on the heap and freed it without zeroing:
the caller's Zeroizing<String> scrubs its own buffer and never learns
about this one. The copy is now owned in a zeroizing buffer whose
capacity is reserved exactly, since a reallocation would leave a first
copy behind unzeroed.
The PAM response cleanup zeroed with ptr::write_bytes immediately before
free(). That is a dead store, and the optimiser is entitled to delete it,
leaving the password in the heap. Volatile writes and a compiler fence
cannot be elided.
PAM_TTY and PAM_RUSER were never set, so pam_unix and pam_faillock logged
tty=? with no requesting user, which makes their records useless for
tracing a failed authentication and stops pam_faillock's per-tty handling
from working at all. Both are now set when the handle is created, best
effort: a tool with no controlling terminal must still be able to
authenticate.
* docs: rewrite the chage, chpasswd and newgrp man pages; cover the rules end to end
None of the three explained the rules that actually decide what happens:
what a last change of 0 means in the chage output and where the 'never'
threshold sits, that chpasswd takes its scheme from ENCRYPT_METHOD and
which flag pairs are errors, or what newgrp's dash does and where
membership and the group password are each read from. All three now carry
value restrictions, configuration, exit status, files and worked
examples.
The deployment suite gains thirty-five assertions against the real
binaries: the three 'password must be changed' lines, a date at -M 9999
and 'never' at -M 10000, -1 clearing a field while -5 is refused, exit 1
for an unknown login from both chage and passwd, chpasswd following
ENCRYPT_METHOD through two changes and -c overriding it, the two usage
errors, --prefix on both tools leaving the live files untouched, and
newgrp keeping the environment without '-' and reinitializing it with.
|
||
|
Pierre Warnier
|
b6c318d1d5 |
chfn/chsh: interactive mode, and fix a setuid path oracle (#268)
* shadow-core: add a shared terminal helper
Reading a line and reading a password were open-coded in pam.rs, and
every prompt outside a PAM conversation had no helper at all.
tty.rs holds all of it: prompt_line, read_password and read_stdin_line,
with the EchoGuard that restores the terminal. Ctrl-C at a password
prompt used to leave echo disabled -- the signal terminates the process
without unwinding, so the guard's destructor never ran. The read now
blocks SIGINT, SIGQUIT and SIGTSTP for its duration, which is what
readpassphrase(3) does, and restores the mask on the way out.
The prompt prefers /dev/tty and falls back to stderr and stdin when
there is no controlling terminal (ENXIO/ENODEV), the case that broke
password changes under su -c.
* chfn: prompt for fields, and check values before asking for a password
chfn(1) with no field option prompts for each field with its current
value; ours refused with 'no flags specified'. It now prompts, and an
empty answer keeps the field.
Three fixes come with it:
- CHFN_RESTRICT is enforced on the values, not on which options were
passed. Interactive input went through unrestricted, and a field the
caller may not change is now not prompted for at all.
- Values are validated before the PAM conversation, so a rejected
value no longer costs the caller a password prompt.
- mutate_passwd no longer calls setuid(0). euid 0 is all the lock and
the atomic write need; raising the real uid made caller_is_root(),
which is deliberately real-uid based, answer true for everyone.
The five sub-fields are now one Changes value with the collect,
validate, permission and apply steps on it, which is what let uumain
shrink rather than grow.
* chsh: prompt for the shell, and stop leaking path existence
Three changes to the self-service path.
chsh(1) with no -s prompts for the new shell with the current one in
brackets; ours refused with 'no shell specified'. Pressing ENTER, or
repeating the current value, is now a successful no-op.
An empty shell is accepted. passwd(5) gives the field no meaning of its
own and login falls back to /bin/sh, so -s '' selects the system default
rather than naming a program that does not exist.
The order of the two checks on a shell now matters and is documented.
chsh is setuid-root, so Path::exists answers with root's view of the
filesystem: testing existence first made the tool an oracle for paths
the caller cannot stat, one probe at a time --
chsh -s /root/.ssh/id_ed25519
answering 'is not listed in /etc/shells' for a file that exists and
'does not exist' for one that does not. For a non-root caller the
membership test now runs first and discloses nothing, /etc/shells being
world-readable. Root still bypasses the list but not existence.
Validation also moved ahead of the PAM conversation, so a refused shell
no longer costs a password prompt, and mutate_passwd no longer calls
setuid(0): euid 0 is enough, and raising the real uid made
caller_is_root() answer true for every caller.
* docs: rewrite the chfn and chsh man pages; cover self-service end to end
The man pages documented neither interactive mode, nor CHFN_RESTRICT,
nor the restricted-account rule, nor what an empty shell field means,
and chfn's still said a field option was mandatory. Both now carry
permissions, value restrictions, configuration, files and worked
examples, and chsh explains why the two shell checks run in the order
they do.
The deployment suite gains eighteen assertions against the real setuid
binaries: interactive chfn and chsh through PAM, an empty answer as a
no-op, CHFN_RESTRICT enforced on interactive input, the empty shell
field, and the oracle itself -- an existing and a missing path under
/root must give a non-root caller the same answer, with no password
prompt before it.
|
||
|
Pierre Warnier
|
382edd4dc2 |
passwd: make the interactive password change work; repair and CI the e2e suite (#251)
* passwd: make the interactive password change work Three faults kept passwd from changing a password, all found by running a real change as an unprivileged user under su -c and read from pam_unix's own debug log: - the PAM conversation only ever opened /dev/tty, which fails with ENXIO when there is no controlling terminal (su -c, a service, cron); pam_unix then reported "conversation failed ... (old) token not obtained". It now falls back to stderr for the prompt and stdin for the answer, the way getpass(3) and PAM's misc_conv do, and still prefers /dev/tty when there is one; - pam_authenticate was called against the passwd service, which defines no auth stack, so it failed before any prompt. The current password is verified by pam_unix inside pam_chauthtok, keyed on the real uid; the call is removed; - the privilege drop spanned pam_chauthtok, so pam_unix could neither run its setgid helper nor rewrite /etc/shadow. euid 0 is what the setuid bit is for; the drop is gone. Landlock is no longer applied to the PAM path: restrict_self sets no_new_privs, which strips the setgid bit from unix_chkpwd. It now wraps only the direct-file paths, with rules that let NSS, nscd and logger run. --prefix becomes root-only, like --root. * e2e: repair the deployment image and run it in CI Dockerfile.e2e ran `make install`, which since #138 installs per-tool binaries, then chmod 4755'd a shadow-rs multicall that was never installed, so the image had not built for months. It now uses install-multicall, and an `e2e` CI job builds and runs the suite on every pull request. The suite now changes a password through PAM as an unprivileged user and proves it by authenticating with the new password and failing with the old (root's own su never prompts, so the old checks proved nothing), rejects a wrong current password, has root set one without the old, checks the setuid privilege model for pwck/grpck and that /etc/shadow keeps its owner and mode. tests/pam-e2e.sh, which could not fail, is deleted. * docs: record the passwd PAM fix and the e2e CI job |
||
|
Pierre Warnier
|
881aad13c7 |
shadow-rs: drop the setuid privilege for applets that are not meant to have it (#239)
* shadow-rs: drop the setuid privilege for applets that are not meant to have it The per-tool install marks only passwd, chfn, chsh and newgrp setuid; a multicall install has to make the one binary setuid, which gave every applet euid 0. Any local user could run 'shadow-rs pwck -s' or 'grpck -s' and have /etc/passwd, /etc/shadow, /etc/group and /etc/gshadow rewritten as root. Before an applet outside that set runs, the binary now calls setuid(getuid()), so the real, effective and saved uid are all the caller's and the two layouts share one privilege model. It fails closed if the kernel does not take the privilege away. Dispatch and --list are driven by one table instead of two hand-kept lists, and an exit code outside 0..=255 no longer clamps to success. The deployment test checks that testrunner's pwck -s and grpck -s leave the files' inode and ownership untouched. * docs: the multicall binary keeps euid 0 for the four self-service tools only |
||
|
Pierre Warnier
|
9f197e0e50 |
chfn, chsh: authenticate the caller before applying a change (#235)
* shadow-core: move PrivDrop out of passwd The RAII guard that drops the effective UID for the PAM conversation lived inside uu_passwd, but it is process-level privilege handling that any setuid tool needs. Move it next to the seteuid wrapper it calls. Two changes while moving it: the error type becomes io::Result rather than PasswdError so other tools can use it, and the failure-to-restore message goes through uucore::show_error! instead of a hardcoded "passwd:" prefix, so it names whichever tool is running. * chfn, chsh: authenticate the caller before applying a change Both tools are installed setuid-root. They correctly restrict a non-root caller to their own account, checked against the real UID — but they never verify who is asking. Anyone with access to an unlocked session could change that user's login shell or GECOS field, and changing a login shell is a foothold. GNU prompts for the password here, and distributions already ship the service definitions we were ignoring: /etc/pam.d/chfn and /etc/pam.d/chsh both let root through with pam_rootok and fall back to common-auth for everyone else. CONTRIBUTING treats a difference from GNU as a bug, so this is one. Authenticate through PAM for non-root callers, reusing the pattern passwd already uses: drop to the real UID for the conversation so modules see the actual caller, then authenticate and check account validity. Without the pam feature there is no way to verify anyone, so the tools refuse rather than apply an unauthenticated change. That is a behaviour change for that build only, and the safe direction for a setuid binary. It also means every install path must enable the feature — 'make build' only enabled it for passwd, which would have shipped two setuid tools that reject all non-root use. Fixed, and the end-to-end assertion now covers all three tools instead of passwd alone. CI gained 'cargo test --features pam': the feature changes behaviour rather than only adding code, so both configurations need testing. |
||
|
Pierre Warnier
|
3de8b506ac |
build: ship binaries with PAM so passwd can change passwords (#220) (#225)
`pam` is not a default cargo feature, so every build path except the e2e
image compiled the interactive change path out of `passwd`. The tool
could report and lock accounts but not change a password:
$ passwd alice
shadow-rs: PAM support is not compiled in — cannot change password
interactively
That covered the release archive, `make install` (the default install of
14 standalone binaries) and `make install-multicall`. Only the e2e image
passed --features pam, and its assertions stop at `passwd -S`, which
works without PAM — hence nobody noticed.
Enable the feature in the release build and in both Makefile targets, and
install libpam0g-dev on the release runner. The PAM headers are already a
documented build requirement.
Add an e2e assertion that the installed passwd is linked against libpam,
so this cannot regress silently again.
Closes #220.
|
||
|
Pierre Warnier
|
24347e9913 |
usermod: add -p/--password flag for pre-hashed passwords (#114)
Ansible's user module calls `usermod -p <hash>` to set passwords. This was the most significant gap for Ansible interop. Implementation: - Add PASSWORD option constant and clap argument (-p/--password) - Write hash directly to shadow entry, update last_change to today - Follows the same lock → read → mutate → atomic_write pattern Tests: - test_set_password: basic -p flag with hash verification - test_set_password_long_flag: --password long form - test_set_password_preserves_other_fields: all 9 shadow fields intact Also updated e2e Ansible playbook to use native `user: password:` parameter instead of the chpasswd -e workaround. Fixes #114 |
||
|
Pierre Warnier
|
36a623d0dd |
review: address all 8 Copilot review comments
- Redact full command args in assert_ok failure output (password leak) - Wrap all bare useradd/chpasswd calls in assert_ok (5 test sections) - Strict symlink checks: require -L + verify target resolves to shadow-rs - Add no_log: true to Ansible password task - Redirect nscd -d stdout to /dev/null Assertion count: 97 → 117 |