On systems with a very large RLIMIT_NOFILE, calling close() in a loop
from 3 to getdtablesize() effects an enormous number of system calls.
There are better alternatives. Both BSD and Linux have the closefrom()
system call that closes all file descriptors with indices not less than
a specified minimum. Have start-stop-daemon call closefrom() on systems
where it's implemented, falling back to the old loop elsewhere.
Likewise, calling fcntl(i, F_SETFD, FD_CLOEXEC) in a loop from 3 to
getdtablesize() raises a similar performance concern. Linux 5.11 and
onward has a close_range() system call with a CLOSE_RANGE_CLOEXEC flag
that sets the FD_CLOEXEC flag on all file descriptors in a specified
range. Have supervise-daemon utilize this feature on systems where it's
implemented, falling back to the old loop elsewhere.
1364e6631c exempted the write end of the
synchronization pipe from the close() loop in the child process, but
this is unnecessary, as the pipe is opened with O_CLOEXEC, and the child
process calls execvp() soon after the close() loop, with the intervening
code not needing the pipe. Indeed, the pipe only needs to remain open in
the child process until after the call to setsid(), which occurs well
before the close() loop. So, eliminate the needless carve-out from the
close() loop, in preparation for introducing closefrom().
As described in "Why nice levels are a placebo and have been for a very
long time, and no one seems to have noticed"[1], the Linux kernel in its
default configuration on many Linux distributions autogroups tasks by
session ID and "fairly" allocates CPU time among such autogroups. The
nice levels of tasks within each autogroup are only relative to
other tasks within the same autogroup. Effectively, this means that the
traditional nice level is rendered moot for tools like start-stop-daemon
and supervise-daemon, which start each daemon in its own session and
thus in its own autogroup. Linux does provide a means to change the
niceness of autogroups relative to each other, so let's have start-stop-
daemon and supervise-daemon make use of this feature where available so
that -N,--nicelevel/SSD_NICELEVEL will actually do what the user
intends. On systems where autogroups are not supported or are disabled,
this commit introduces no change in behavior.
Note that the setsid() call in the child process of start-stop-daemon is
moved to much earlier. This is necessary so that the new process will be
assigned to a new autogroup before the autogroup nicelevel is set. To
avoid inadvertently acquiring /dev/tty as the controlling terminal of
the new session after setsid() has given up the controlling terminal
inherited from the parent process, tty_fd is opened before the call to
setsid().
[1] https://www.reddit.com/r/linux/comments/d7hx2c/why_nice_levels_are_a_placebo_and_have_been_for_a/
This fixes#542.
While running `rc-service start docker` on Gentoo,
I found that the command does not start the service 90% of the time,
with an enigmatic 'service crashed' message.
The root cause of this is apparently rc-service spawning a pty,
running start-stop-daemon inside that pty, and exitting,
before start-stop-daemon child process calls setsid(),
which results in the child process being killed with SIGHUP (SI_KERNEL).
Theoretically this bug was present ever since the file was created in
5af58b4514 ("Rewrite the core parts in C. We now provide...")
(or even before that), but it should have been only a minor issue before
45bd125dcc ("Use a pty for prefixed output instead of pipes for...").
Not sure why nobody has had the issue so far (it has been present for
almost 15 years).
As here setsid() is the last call before execve(), the most natural
locking mechanism is vfork(), as it gives back control to parent
process only after execve() or process termination.
So this way the bug can be fixed by adding a single letter. :-)
Another way to ensure this would be using an O_CLOEXEC file descriptor
or some custom lock, which would need to be released not before setsid().
Fixes: 5af58b4514 ("Rewrite the core parts in C. We now provide...")
Fixes#532.
- drop old build system
- move shared include and source files to common directory
- drop "rc-" prefix from shared include and source files
- move executable-specific code to individual directories under src
- adjust top-level .gitignore file for new build system
This closes#489.
