pam_slinit.so — the PAM module that opens and closes slinit-logind
sessions, so a Sunlight Linux system can log users in without
pam_elogind.so.
Part of Sunlight Linux. Apache-2.0.
slinit-logind already owns org.freedesktop.login1 — sessions, seats,
users, inhibitors and sleep. What is missing is the login hook: the
thing that calls CreateSession when a user logs in. Today that is
pam_elogind.so, which ships in the elogind package, and
gnome-shell and xfce4 depend on that package. So the elogind daemon
is already switched off while the package stays mandatory.
Shipping this module does not by itself remove elogind. The
dependency has to change distro-side in slpkgs as well. Two pieces of
work, not one.
Every other Sunlight component is Go. This one is not, and the reason is not taste.
A PAM module is a shared object dlopened into login, sshd, gdm
and su. Building it with Go's -buildmode=c-shared would put the Go
runtime — and its scheduler threads and garbage collector — inside every
process that authenticates anybody. Those processes fork.
slinit already measured this in lib/slinit-mlock, where the same
question came up for an LD_PRELOAD: the Go shared object was 1.9 MB
against 17 kB for the C one, and preloading it took a trivial process
from 2.8 MB and one thread to 78 MB and two. A PAM module is loaded
far more often than a preload, into processes that are far more
sensitive.
elogind's module uses sd-bus, which comes from libsystemd or
libelogind. Linking against either would reintroduce exactly what this
module exists to remove.
libdbus-1 is already present at runtime on a Sunlight system, because
dbus-daemon links it — so pam_slinit.so adds no new runtime
dependency.
Verified against the installed rootfs rather than assumed. The built
module's direct NEEDED entries are exactly three — libdbus-1.so.3,
libpam.so.0, libc.so.6 — and in the Sunlight rootfs
libdbus-1.so.3 itself needs only libc.so.6. libsystemd is not in
that rootfs at all.
On a systemd distribution the same object does load libsystemd
transitively, because libdbus-1 is built with systemd support there.
That is a property of the host's dbus, not of this module, and it is
worth knowing before someone reads an ldd on a development machine and
concludes the opposite.
The module names the session leader by a pidfd, so a pid recycled between
the module reading it and slinit-logind acting on it cannot make another
process the leader. slinit-logind has the call from 3.16; against an
older one the call fails with UnknownMethod and the module retries with
CreateSession and the pid, as pam_systemd does against an older
logind. Both carry systemd's signatures exactly.
A session ends when its leader process exits, which slinit-logind watches
through a pidfd, or when pam_sm_close_session calls ReleaseSession.
That is systemd's contract since v258.
The file descriptor CreateSession returns is not part of it. Before v258
it was a FIFO whose closing ended the session; systemd now returns a
placeholder, and so does slinit-logind from 3.16 (earlier versions
returned a FIFO they never watched). The module closes it at once.
ReleaseSession still matters: it ends the session when the login does,
not when the leader exits, which differs for a display manager's worker
or a session that left a process behind. The module releases only a
session it created; for an existing one it does nothing at close, since
that session belongs to whoever created it.
The rules pam_systemd applies before CreateSession, so a session
looks the same to slinit-logind whichever module registered it
(src/context.c, tested by make test):
- type, class, desktop, seat and VT come from
XDG_SESSION_TYPE,XDG_SESSION_CLASS,XDG_SESSION_DESKTOP,XDG_SEATandXDG_VTNRin the PAM environment, then the process environment, then the module'stype=,class=anddesktop=arguments; - a tty containing
:is an X display;cronandsshare not ttys; a/dev/prefix is dropped; - type defaults to
x11,ttyorunspecified; class tobackgroundforunspecified, elseuser; - for a tty, seat and VT number are left to slinit-logind, which derives them from a virtual-console tty as systemd-logind does;
- for an X display with no
XDG_VTNR, the module asks the X server: it connects to the display's socket, takes the peer's pid and reads that process's controlling tty. A VT there puts the session onseat0. A display manager that setsXDG_SEATandXDG_VTNRskips this;xdmand its kind set neither and would otherwise get a session with no seat. The peer must be root or the user being logged in, because any user can bind the abstract socket name of a display nobody is serving; and the connect gives up after two seconds, so a server that has stopped accepting does not hold the login.
It then exports XDG_SESSION_ID, XDG_SESSION_TYPE, XDG_SESSION_CLASS,
XDG_SESSION_DESKTOP, XDG_SEAT, XDG_VTNR, and, when the session
belongs to the user being logged in and its runtime directory exists and
is theirs, XDG_RUNTIME_DIR — plus DBUS_SESSION_BUS_ADDRESS if a user
bus is listening at $XDG_RUNTIME_DIR/bus. A variable with no value is
removed rather than left as an earlier module set it.
A SessionBusy answer is success with no session variables exported:
there is nothing to open. systemd-logind sends it for a caller already in
a session, su and sudo above all. slinit-logind sends it from the
version that refuses nested sessions, for the same case and for a VT on
seat0 that a non-greeter session already holds; a login on such a VT
therefore goes through with no session, and none of what "active on the
console" grants. Earlier slinit-logind versions open a nested session
instead.
If the login fails after the session was created, the module releases the
session, drops the inhibitor lock and removes XDG_SESSION_ID,
XDG_SEAT, XDG_VTNR, XDG_RUNTIME_DIR and DBUS_SESSION_BUS_ADDRESS
from the PAM environment before returning the error: nothing is left
naming a session that no longer exists. Type, class and desktop stay,
since they describe what was asked for and a display manager sets them
before the module runs.
Class none, from class=none or XDG_SESSION_CLASS=none, registers
nothing and succeeds.
Checked against systemd v262 (src/login/pam_systemd.c). All of it also
happens when no session was registered.
HOMEis set from the user's passwd entry.- Inhibitor lock.
inhibit=shutdown:sleep(orXDG_SESSION_INHIBIT), withinhibit-why=(orXDG_SESSION_INHIBIT_WHY), takes a block-mode lock throughInhibitand holds it until the session closes. Failing to get it is logged and the login goes on. - Areas.
area=work(orXDG_AREA) logs the user into~/Areas/work:HOMEpoints there,XDG_AREAis set, andXDG_RUNTIME_DIRbecomes<runtime>/Areas/work, created if missing. The area must be a directory the user owns; if not, they are told and get their ordinary home.src/area.c: the runtime directory belongs to the user, so each level is opened without following symlinks and a new directory is renamed into place only once its ownership is right. - Session markers. For a
ttysession with a terminal on stdout andTERMnotdumb, the OSC 3008startsequence is written at open and itsendat close (src/osc.c). - Shell credentials.
shell.prompt.prefix,shell.prompt.suffixandshell.welcomefrom$CREDENTIALS_DIRECTORY, if the service has one, becomeSHELL_PROMPT_PREFIX,SHELL_PROMPT_SUFFIXandSHELL_WELCOME. XDG_SESSION_EXTRA_DEVICE_ACCESSis carried into the session.
Options: debug (also debug=yes|no), class=, type=, desktop=,
area=, inhibit=, inhibit-why=.
XDG_SESSION_CLASS, XDG_SESSION_TYPE, XDG_SEAT, XDG_VTNR,
XDG_AREA and XDG_SESSION_INHIBIT are read from the PAM environment
and then from the process's own, exactly as pam_systemd reads them.
That is what lets a display manager describe its session, and it is safe
only while the user being logged in cannot set them:
XDG_SESSION_CLASS=greeteris allowed to share a VT that another session holds, which slinit-logind otherwise refuses;XDG_SEATandXDG_VTNRsay which console the session is on, and being active on a console is what polkit'sallow_activerules reward;XDG_SESSION_INHIBIT=shutdownblocks the machine from powering off for as long as the session lasts.
So, in the stack that loads this module:
- no
pam_env.so user_readenv=1, and nothing else that copies user-controlled files into the PAM environment, beforepam_slinit.so; sshd_configmust notAcceptEnv XDG_*, norPermitUserEnvironment;- a service that calls PAM with these already in its own environment passes them on: that is the feature, so it must be the service's intent.
This has to be checked in Sunlight's /etc/pam.d and sshd_config when
the module is wired in; the module cannot tell a display manager's
variable from a user's.
- The bus address. The system bus address is compiled in
(
SYSTEM_BUS_ADDRESS, by default whatdbus-1.pcsays) andDBUS_SYSTEM_BUS_ADDRESSis never read. libdbus would read it, and an address can beunixexec:path=..., a program to run. libdbus ignores the variable in a setuid process, but tells one by comparing the real, effective and saved ids, so a setuid program that has made all three root and still has its caller's environment when it opens the session would run the caller's program as root. OverrideSYSTEM_BUS_ADDRESSonly to point a test build at a private bus. - Strings that are not UTF-8. D-Bus strings must be UTF-8, and libdbus
answers one that is not by calling
abort()in the process that loaded the module. A remote host name, an X display or anXDG_*value that is not valid UTF-8 is logged and left out;remotestays set for a host name that could not be sent. $XDG_RUNTIME_DIR/bus.DBUS_SESSION_BUS_ADDRESSis exported only when that name is itself a socket, not a symlink to something.
-
A failed marker.
pam_systemdfails the login when it cannot write the OSCstartsequence. This module logs it and carries on: a marker for the terminal's benefit is not worth a refused login. -
Session classes. No
user-early,user-light,background-lightoruser-incomplete. In systemd they decide whether a user service manager is started; Sunlight has none, and slinit-logind stores the class as given without acting on it. -
Release.
pam_systemdno longer readsexistingand releases whateverXDG_SESSION_IDnames. This module releases only a session it created: slinit-logind lets root release any session, and matches an existing one by leader pid. -
Varlink.
pam_systemdtriesio.systemd.Loginfirst; slinit-logind speaks D-Bus only, so the list inXDG_SESSION_EXTRA_DEVICE_ACCESS, which travels only over Varlink, is exported but not sent. -
User records. No umask, nice level, rlimits, capability sets,
TZ,LANGor default area from systemd's JSON user records. -
Resource limits.
systemd.memory_maxand its siblings are not sent as session properties.
Against pam_elogind there is one more: PAM_TTY=cron and PAM_TTY=ssh
set the class only when none was given, as pam_systemd does, where
pam_elogind overwrites it.
make # builds src/pam_slinit.so
make check # the built object: two symbols, NODELETE, three libraries
make test # session rules, OSC sequences, area directories: no PAM, no bus
make test-san # the same under ASan and UBSan
make e2e # the module in a real PAM application, against a mock daemonRequires libpam and libdbus-1 development headers (pam-devel,
dbus-devel on Sunlight/Void). make e2e also needs dbus-daemon and
python3 with dbus and gi.
make e2e (tests/e2e/run.py) loads the module into a PAM application
with a PAM configuration directory of its own, and has it talk over a
private bus to tests/e2e/mock_login1.py, which answers as
slinit-logind would and records what it was asked. Nothing runs as root,
and neither /etc/pam.d nor the machine's own bus is touched. It covers
the calls and their fallbacks, what is exported and when, the inhibitor
lock, areas, the terminal markers on a pty, the shell credentials, a
login that fails half-way, everything under "What the module does not
trust", and 300 logins in one process with a flat heap.
The module it tests is built for the run, with the private bus's address compiled in, since the address is never read from the environment.
It does not show that the real slinit-logind answers as the mock does, or that a user can log in: that takes a machine running Sunlight.
.github/workflows/tests.yml runs all of the above on every push and
pull request, with gcc and clang, the two static analysers, the build as
a packager does it, and again on musl.
Opens and releases sessions. Not yet in any /etc/pam.d file on Sunlight,
and elogind is still installed: TODO.md lists what remains.