Re: Shutdown - 25 Years Later

Jim Jackson <jj@franjam.org.uk> writes:

On 2025-04-19, Charlie Gibbs <cgibbs@kltpzyxm.invalid> wrote:

The story I heard was that typing "sync" three times takes long
enough that the system has time to flush all its buffers before
you do anything else.

>
But sync shouldn't return until the buffers have flushed.
So repeats are not needed.

sync is allowed to return before the flush has completed. See
https://pubs.opengroup.org/onlinepubs/9699919799/functions/sync.html.


On Linux (or any other sensible platform), with the normal user tools,
no explicit syncs are needed at all. The normal shutdown process takes
care of the whole question. The behavior of sync is irrelevant and any
theory about needing typing sync even once is a myth.

On other platforms, the same is usually true.


The exceptions to this are:

(a) An operating system that doesn’t guarantee clean shutdown in the
    same way. (Personally I’d consider that a bug.)
(b) Directly issuing a reboot syscall (i.e. bypassing user commands like
    halt or systemctl), or interrupting power.

Explicit syncs don’t help as much as you might hope here, for a couple
of reasons:

1) As noted above, sync isn’t guaranteed to block until all data is
   flushed anyway; it’s allowed to return immediately, with any pending
   writes going on in the background. It happens that on Linux, sync
   _does_ block, so that’s not an issue for us, but on many platforms[1]
   it is not guaranteed to block.

2) Further writes may be initiated after the final call to sync (if
   nothing else, to update st_atime on /sbin/halt and its dependencies);
   no matter how many times you run it, even if you wait until the final
   sync completes, you still do not have a 100% guarantee about data
   persistence.

In short there’s no number of syncs that will guarantee your data is all
actually on disk.

[1] Including FreeBSD, Darwin, HP-UX, OSF/1, Ultrix, SunOS, Solaris, V7
    and at this point I stop looking.

--
https://www.greenend.org.uk/rjk/