On 3/24/23 15:52, Eric Blake wrote:
[attempting to loop in systemd folks; this started in libnbd at https://listman.redhat.com/archives/libguestfs/2023-March/031178.html - although I may have to retry since I'm not a usual subscriber of systemd-devel]
On Fri, Mar 24, 2023 at 11:32:26AM +0100, Laszlo Ersek wrote:
@@ -245,6 +245,9 @@ CONNECT_SA.START: "LISTEN_PID=", "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", &pid_ofs); SACT_VAR_PUSH (sact_var, &num_vars, "LISTEN_FDS=", "1", NULL); + if (h->sact_name != NULL) + SACT_VAR_PUSH (sact_var, &num_vars, + "LISTEN_FDNAMES=", h->sact_name, NULL); if (prepare_socket_activation_environment (&env, sact_var, num_vars) == -1)
If I'm reading this correctly, this does wipe an inherited LISTEN_FDNAMES from the environment in the case where the application linked with libnbd started life with a (different) socket activation, but now the user wants to connect to an nbd server without setting a name (default usage, or explicitly requested a name of "").
Good observation; this is a functionality gap that goes back to v1 of this patch. (I've not investigated the specifics of systemd socket activation before; but see below.)
Put another way, SACT_VAR_PUSH as written appears to be only additive for replacement purposes (if I pushed a variable, I intend to override it in the child process, so don't copy it from environ if one was previously there), but not effective for deletion purposes (I don't intend to set the variable, but if it is already set in environ, I want it omitted in the child's copy).
Is there a way to rework this so that you can pass NULL as the fourth parameter as an indication of an unset request (vs. "" when you want it set to the empty string)? At which point, you would drop the 'if (h->sact_name != NULL)', and just blindly use SACT_VAR_PUSH(, h->sact_name,). That has ripple effects earlier in the series to support those semantics.
For understanding your point, I have had to read up on systemd socket activation:
https://www.freedesktop.org/software/systemd/man/sd_listen_fds.html
Let me quote a part:
Under specific conditions, the following automatic file descriptor names are returned:
[...] "unknown" -- The process received no name for the specific file descriptor from the service manager. [...]
I've also checked the implementation of sd_listen_fds_with_names():
https://github.com/systemd/systemd/blob/main/src/libsystemd/sd-daemon/sd-dae...
...
(2) If we pass LISTEN_PID=xxx and LISTEN_FDS=1, and just "pass through" an *inherited* LISTEN_FDNAMES variable, then it will (in the general case) confuse the nbd server that we start. Namely, if LISTEN_FDNAMES has multiple colon-separated elements (more than 1), or is the empty string (= 0 elements), then sd_listen_fds_with_names() in the nbd server will fail the "n_names != n_fds" check, and return (-EINVAL). If LISTEN_FDNAMES happens to have one element, then sd_listen_fds_with_names() will succeed, but the returned name will confuse the nbd server.
...
Eww - this may be a bigger systematic issue caused by systemd itself, and we should report it there (done by adding in cc here). They did not specify LISTEN_FDNAMES at the time LISTEN_PID was first documented, so the likelihood of libnbd not being the only application that happens to leak inherited LISTEN_FDNAMES through to the child process is non-zero, where this sort of bug will bite more than one client of systemd socket activation. And it is this sort of backwards-incompatibility caused by the systemd extension that they will need to be more careful of avoiding if they ever add any future LISTEN_* environment variables.
I wonder if this is what the "unset_environment" parameter of sd_listen_fds() and sd_listen_fds_with_names() exists for. If any socket-activated application (such as a libnbd client application) is *required* to call one of these functions at least once with a nonzero "unset_environment" argument (reasonably early after startup), then that eliminates the problem. For example (specifically), by the time we got to prepare_socket_activation_environment(), we'd find (or leak!) no inherited LISTEN_* variable (so the "conflict handling" logic could be removed altogether from prepare_socket_activation_environment()) . The above-linked spec <https://www.freedesktop.org/software/systemd/man/sd_listen_fds.html> writes, "If the unset_environment parameter is non-zero [...] the variables are no longer inherited by child processes." So, if a libnbd client application (a) is socket activated, and (b) plans on calling nbd_connect_systemd_socket_activation() itself, then this application *MAY* already be responsible for passing a nonzero "unset_environment" argument, when it calls sd_listen_fds() or sd_listen_fds_with_names() early on. I'm emphasizing "may", because the spec does not explain what use case "unset_environment" is specifically meant for. Laszlo