Re: First time caller to the show - am I understanding the fifo trick correctly?

From: Ellenor Bjornsdottir <>
Date: Wed, 25 Aug 2021 12:19:54 +0000

Personally, I'm ~banging my head against the wall by~ writing a fork of FreeBSD init which, instead of moving to its read_ttys stage, achieves what s6-l-i does with spawning runcom2 and (s6-)svscan (but does so in a sufficiently generic manner that you don't have to do the blocking fifo trick during runcom2 if you don't want to).

On 25 August 2021 11:57:11 UTC, Laurent Bercot <> wrote:
>>Forgiving privilege separation failures and minor grammatical mistakes, does it look as if I understand the fifo trick's application in practice?
> Hi Ellenor,
> Yes, I think you have the right idea.
> The goal here is to redirect s6-svscan's own stdout and stderr to
>the stdin of the catch-all logger process, so that the supervision
>tree's messages, and the messages from every service that lacks a
>dedicated logger, go to the catch-all logger instead of /dev/console.
>(Because /dev/console is a terrible default place to send logs and
>should only be used for very critical messages such as kernel panics,
>or, in our userland case, for catch-all logger failures.)
> The problem is that we want the catch-all logger to run as a service
>under the supervision tree, so the s6-log process does not exist yet
>when we exec into s6-svscan: it will be spawned later as a grandchild
>of s6-svscan (with an s6-supervise intermediary). So we cannot use an
>anonymous pipe for this.
> We use a fifo instead: we can redirect init's stdout and stderr to
>a fifo, and later on, when the catch-all logger starts, we can
>instruct it (in its run script) to read from the fifo.
> But the Unix fifo semantics say that we *cannot* open a fifo for
>writing while there is no reader: open() would either block (default
>flags) or return -1 ENXIO (with O_NONBLOCK). So the "fifo trick" is:
>1. open the fifo for reading
>2. open it for writing, which now works
>3. close the reading end
>At this point, any write() to the fifo will fail with -1 EPIPE. That is
>not a problem per se, except it will also generate a SIGPIPE, so in
>order to avoid crashing and burning, it is important to ignore SIGPIPE
>at the very least - or, better, to make sure that no process writes to
>the fifo until the catch-all logger is up. This is the case for
>and s6-supervise, so our system structure is safe; but we need to make
>sure that no other process starts before the catch-all logger is up,
>else they will just eat a SIGPIPE and die.
> In the s6-l-i model, s6-svscan is executed as soon as possible, on a
>very minimal supervision tree that only contains the catch-all logger
>and a few other essential "early services" (such as the shutdown daemon
>and an early getty). All the rest of the initialization is done in
>"stage 2 init", which is a script run as a child of s6-l-i's.
>So the end of the "fifo trick" uses the Unix fifo semantics as a
>synchronization mechanism:
>4. fork
>5. In the child, close our fd to the fifo
>6. In the child, open the fifo for writing once again,
> *without* O_NONBLOCK.
> This last open() will block until the fifo has a reader. That
>ensures the child will only resume once the parent has completed
>its work and executed into s6-svscan, and the supervision tree has
>started and the catch-all logger is running. Then the child can exec
>into stage 2 init and perform the rest of the work with the guarantee
>that the supervision tree is operational and all the stdout and stderr
>messages go to the catch-all logger by default.
> To see exactly how to implement stage 1 init and the fifo trick as
>an execline script, you can checkout (or download) any version of
>s6-l-i *prior to*; try version, downloadable from
> if you type the URL by hand, and accessible via the
>v0.4.0.1 tag in git. It is very different from what it is now, as in
>there is no sysv compatibility at all, but stage 1 should be
> A few months ago, I tried adding a few conditional compilation options
>to s6-l-i to make it work under FreeBSD, but unfortunately the
>organization of the FreeBSD init is so different from Linux's,
>especially shutdown-wise, that my attempt only succeeded in turning
>the package into an unholy plate of spaghetti. At some point in the
>future, however, a similar-but-separate s6-freebsd-init package may
>make sense.
> Laurent

Sent from my Android device with K-9 Mail. Please excuse my brevity.
Received on Wed Aug 25 2021 - 14:19:54 CEST

This archive was generated by hypermail 2.4.0 : Wed Aug 25 2021 - 14:20:30 CEST