[systemd-devel] [PATCHv2] install: Assume *.wants symlinks have the same name as their target for scalability.
Zbigniew Jędrzejewski-Szmek
zbyszek at in.waw.pl
Tue Dec 31 07:01:16 PST 2013
On Tue, Dec 31, 2013 at 06:40:47PM +0400, Andrey Borzenkov wrote:
> В Вс, 29/12/2013 в 18:57 +0100, Zbigniew Jędrzejewski-Szmek пишет:
> > Hi,
> > I'd like to reopen this thread.
> >
> > I have been looking at our bugs [1-10, there are others too] related to
> > enabling/disabling of units and following of symlinks. Currently this
> > is an unintuitive mess and some changes will have to be made.
> >
> > I think that the general rules should be:
> > (/lib is short for /usr/lib, /usr/local/lib, etc.,
> > .wants is short for .wants or .requires)
> >
> > 1. when given a unit, it must be possible to read information about
> > this unit without reading all other units
> >
> > 2. all symlinks in /run and /etc are considered configuration and
> > will be removed by systemctl disable
> >
> > 3. symlinks within the configuration directories must have
> >
> > (a) the same symlink and target name, or
> >
> > (b) the symlink can be a instance and the target a template with
> > the same stem and type
> >
> > (c) in /lib, the name can be different, but the type must match, and this
> > provides an alias
> >
> > 4. systemd must follow the same logic as systemctl and must enforce
> > the rules.
> >
> > This means that:
>
> If I read this correctly, the end effect is that every systemd unit file
> must be found under standard systemd directories. Right?
It can either be found as /etc/systemd/system/unit.unit or /lib/systemd/system/unit.file,
or a symlink from /etc/systemd/system/unit.unit or /lib/systemd/system/unit.file to
the unit file must exist.
> > 1. when linking a unit with 'systemctl enable /path/to/file',
> > the unit must be linked both into /etc/systemd/system/ and
> > /etc/systemd/systemd/whatever.wants/. This is what systemctl enable
> > currently does, but we allowed only the second part done
> > manually. systemd would be changed to ignore all files in .wants/,
> > and it would be changed to ignore symlinks in .wants if the unit
> > cannot be loaded by name otherwise.
> >
>
> What is semantic of having symlink in .wants to different unit file that
> would normally be found by systemd. I.e. having both /lib/foo.service
> and /etc/foo.service will use /etc version; but what if there is symlink
> that points into /lib?
Currently we ignore the directory in the symlink destination. I'm pretty sure
we should continue to do that.
> > 2. symlinking to units to provide an alias is allowed only in /lib.
> > When systemctl disable is executed, all symlinks in /etc and /run will
> > be nuked, but /lib will be left alone. I think this is current behaviour.
> >
> > 3. .wants directories and .d directories will only be honored for
> > the main unit name. The problem that we have currently is that
> > if the alias unit is not pulled in by anything, this extra configuration
> > is ignored. But even 'systemctl status' loads the unit, and loads
> > this extra configuration. systemd would be changed to warn about this
> > and ignore those extra dirs.
> >
> > 4. support for aliases will be provided on a best-effort basis: If
> > the alias is done by symlinking in /lib, it is always known.
> > If the alias is provided by Alias= configuration directive, it
> > is only known if the unit was already loaded.
> >
> > There's a bug that 'systemctl disable' removes too much stuff for
> > instance units. I can't find the report atm, but this is a bug that
> > should be fixed too.
> >
> > Zbyszek
> >
> > [1] https://bugs.freedesktop.org/show_bug.cgi?id=72611
> > systemctl is-enabled scales poorly with many units (this one)
> > [2] https://bugs.freedesktop.org/show_bug.cgi?id=63621
> > systemctl enable doesn't work for Alias names
> > [3] https://bugs.freedesktop.org/show_bug.cgi?id=72154
> > "systemctl enable named.service" yields error messages used for legacy sysinit scripts
> > [4] https://bugzilla.redhat.com/show_bug.cgi?id=1014311
> > RFE: allow systemctl enable work on symlinked units
> > [5] https://bugzilla.redhat.com/show_bug.cgi?id=864298
> > systemctl is-enabled reports 'static' incorrectly (it appears)
> > [6] https://bugzilla.redhat.com/show_bug.cgi?id=955379
> > systemctl enable fails for user-defined service files in /etc/systemd/system
> > [7] https://bugs.freedesktop.org/show_bug.cgi?id=65255
> > some units can be started when the file is just copied into .wants/
> > [8] https://bugs.freedesktop.org/show_bug.cgi?id=54560
> > systemd does not follow symlinks that point outside of
> > /etc/systemd/system or /usr/lib/systemd/system
> > [9] https://bugzilla.redhat.com/show_bug.cgi?id=1002806
> > runlevel command returns "unknown"
> > [10] https://bugzilla.redhat.com/show_bug.cgi?id=815835
> > /sbin/runlevel reports "unknown" where it shouldn't
> >
> >
> >
> > On Fri, Dec 13, 2013 at 01:03:30PM -0800, david at davidstrauss.net wrote:
> > > From: David Strauss <david at davidstrauss.net>
> > >
> > > ---
> > > src/shared/install.c | 7 ++++++-
> > > 1 file changed, 6 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/src/shared/install.c b/src/shared/install.c
> > > index 17e8a75..512e3df 100644
> > > --- a/src/shared/install.c
> > > +++ b/src/shared/install.c
> > > @@ -419,10 +419,15 @@ static int find_symlinks_fd(
> > > r = q;
> > >
> > > } else if (de->d_type == DT_LNK) {
> > > - _cleanup_free_ char *p = NULL, *dest = NULL;
> > > + _cleanup_free_ char *p = NULL, *dest = NULL, *no_inst = NULL;
> > > bool found_path, found_dest, b = false;
> > > int q;
> > >
> > > + /* Skip symlinks with a different name the target unit */
> > > + no_inst = unit_name_replace_instance(basename(de->d_name), "");
> > > + if (!streq(no_inst, name))
> > > + continue;
> > > +
> > > /* Acquire symlink name */
> > > p = path_make_absolute(de->d_name, path);
> > > if (!p)
> > > --
More information about the systemd-devel
mailing list