procd/plug, branch master

hotplug: reload rule files when any of them changes

2026-09-27T16:12:22Z

Parsed rule files are cached for the lifetime of procd, so editing /etc/hotplug.json, or adding a fragment to a directory it includes, only takes effect after a reboot. Watch the directory holding every file the rule engine loads, set up from rule_handle_file() so the watched set follows the include graph and procd needs no path of its own, and drop the cache when an entry matching the requested name changes. The watch sits on the directory, so it survives a rule file replaced by rename, which is how both a package manager and an editor write one. Reloading is a free and an init: json_script parses a file on first use, so the cost falls on the next uevent and is paid once however many files an installation touches. A directory that is absent takes no watch and needs none: an unmatched glob is never cached, so it is retried on every uevent. Signed-off-by: Daniel Golle

hotplug: support glob patterns in rule file names

2026-09-25T23:20:28Z

There is no way for a package to extend the hotplug rules short of editing the single /etc/hotplug.json it shares with everyone else. json_script already caters for this: its handle_file callback may return a chain of files linked through ::next, which the interpreter runs in order, but no caller ever built one. Build that chain from glob(3) whenever the requested name contains a wildcard, so an "include" of /etc/hotplug.json.d/*.json runs every fragment in sorted order. json_script_get_file() caches by the requested name and ignores a failed avl_insert(), so the head of the chain carries the pattern as its key; keying it by its own path would reparse and leak every fragment on every uevent. Signed-off-by: Daniel Golle

hotplug: release the JSON object after parsing a rule file

2026-09-25T23:20:28Z

rule_handle_file() converts the parsed json-c object into a blobmsg buffer but never drops its reference, so the object tree stays allocated for the lifetime of procd. Only one rule file is ever loaded, so a single object is leaked rather than one per uevent. Fixes: 54108c71df64 ("fix hotplug") Signed-off-by: Daniel Golle

coldplug: fix missing header include

2026-02-09T11:31:44Z

makedev() is defined in header sys/sysmacros.h which needs to be included to not break the build. Fixes: 58d7aaa ("initd/coldplug: create /dev/null before running udevtrigger") Signed-off-by: Daniel Golle

initd/coldplug: create /dev/null before running udevtrigger

2026-02-09T11:09:58Z

When procd_coldplug() runs, it unmounts /dev and mounts a fresh empty tmpfs before forking udevtrigger to populate device nodes via hotplug events. Since udevtrigger runs asynchronously, there is a race window between the fresh mount and when the "null" device uevent is processed. If any code executes a shell redirect to /dev/null during this window (e.g., from hotplug handlers or other event processing), the shell creates /dev/null as a regular file. When the null device uevent is later processed, makedev()'s mknod() fails silently with EEXIST, leaving /dev/null as a regular file permanently. This causes all subsequent redirections to /dev/null to append to the file instead of discarding output, eventually filling up the tmpfs. Fix by explicitly creating /dev/null immediately after mounting the fresh tmpfs, before any other code can run. Signed-off-by: Steev Klimaszewski

hotplug: send event notifications

2025-08-07T13:44:21Z

Sent to per-subsystem ubus objects Signed-off-by: Felix Fietkau

udebug: add support for logging via udebug

2023-11-28T11:48:13Z

Signed-off-by: John Crispin Signed-off-by: Felix Fietkau

procd: clean up /dev/pts mounts

2022-01-11T16:23:25Z

The default mode is already 600, no need to specify it. Access times are also irrelevant. Signed-off-by: Rui Salvaterra

procd: mount /dev with noexec

2022-01-11T16:23:05Z

/dev is writable. Allowing execution inside it makes it a possible attack vector. Kees Cook recently sent a kernel patch [1] in order to mount /dev as noexec and nosuid for systems which rely on CONFIG_DEVTMPFS_MOUNT=y to create/populate /dev, which isn't our case (it's procd's responsibility). Add noexec to the /dev mount flags, since we already use nosuid (and keep the coldplug flags symmetric, while at it). This carries the risk of breaking very old, pre-KMS graphics drivers [2], but it shouldn't be a problem for systems built in the last ~15 years. The vast majority of our targets doesn't have a GPU, anyway. :) [1] https://lore.kernel.org/all/YcMfDOyrg647RCmd@debian-BULLSEYE-live-builder-AMD64/ [2] https://lore.kernel.org/all/CAPXgP12e5LpN6XVxaXOHhH=u8XXN==2reTaJDCoCk4tP4QduDQ@mail.gmail.com/ Signed-off-by: Rui Salvaterra

hotplug.c: set nl_pid to zero

2021-01-26T10:16:18Z

With the current solution where nl_pid is set through getpid we run into problems when running procd in a different PID namespace (e.g. container). The PID number inside the active PID namespace will be set which doesn't match the global PID. Therefore, procd will never receive any netlink messages. By setting nl_pid to zero the kernel will assign the global PID automatically and fixes the issue. Signed-off-by: Stefan Eichenberger Acked-by: John Crispin