[systemd-devel] preset activation on first boot takes too long
Dimitri John Ledkov
dimitri.j.ledkov at intel.com
Fri May 15 02:30:55 PDT 2015
On 14 May 2015 at 18:00, Lennart Poettering <lennart at poettering.net> wrote:
> On Thu, 14.05.15 17:13, Dimitri John Ledkov (dimitri.j.ledkov at intel.com) wrote:
>> I'm looking at bootcharts and it seems like first boot preset
>> activation takes too much time...
>> So at the moment, we iterate all units, then iterate through presets
>> until we find a match and act upon it.
>> However, most distros have "disable *" as their last setting, or don't
>> use presets at all.
>> Furthermore looking at a fully featured system (e.g. my ubuntu laptop):
>> * 158 files do not have install section
>> * 89 have an install secion
>> Also it seems odd to have all of this in the pid one critical path ->
>> e.g. these things are being parsed before anything happens.
>> Thus I wonder if the presets should be moved into e.g. a generator
>> that will do the following on first boot only:
> Hmm, generators so far were strictly something that would generate
> output in the generator unit paths we pass to them, and nowhere
> else. These directories would then be lifecycled by the daemon
> runtime, and flushed on daemon reload. The generators would then be
> invoked with new, empty generator unit directories on daemon reload.
> The preset logic otoh creates persistent changes in /etc, that will
> only and exlcusively be run if /etc is unpopulated. They logic is not
> rerun on daemon reload, and nothing is ever flushed.
> I am pretty sure this should stay that way. We shouldn't turn
> generators into more than what they are right now. They should not
> have persistent effect.
> However, I am all open for optimizing the codepaths here. Maybe we can
> cache the preset files in memory or so. And/or we could move the
> preset code into its own binary, which we fork off explicitly before
> running the generators, and then wait for after the generators (so
> that the binary runs in parallel to the generators). Alternatively we
> could even run it as thread from PID 1 in parallel to waiting for the
> generators, as long as it shares zero state with the rest of PID
> 1. (That said, I am usually not too fond of threads, and especially
> not in PID 1.)
>> * parse .preset files
>> * construct list of things to enable
>> * enable all the units in that list
> OK, so this would mean caching the preset file, and then doing pretty
> much the same as before? I am fine with that. (But not as a generator,
> as mentioned above)
>> This should cut I/O and processing time at first boot by a bit, since
>> only the units to be activated will be parsed.
> Hmm? not following here. are you suggesting to use the preset file
> list as base list of units to enable and then search for them in the
> file system without ever enumerating unit files in the fs? This will
> not really work, as the list contains glob expressions, and more
> complicated ones than just "disable *". For examples things like
> "enable avahi-daemon.*" which enables both the socket and the service
> in one step. But it could also be "enable foo-*.service", for a
> project "foo", that has many different services it wants to enable
> with a single line... And crazy people could even use more complex
> Hence, I am pretty sure the list of unit files enumerated from disk
> needs to be used as basis, and then checked against the preset file,
> not the other way round.
> One thing you could use for optimizing: a TODO list item has been for
> a while to change the first-boot preset logic to operate in a purely
> additive way: don't bother with disabling any services then (in most
> cases there will be nothing enabled at that time anyway), but only
> operate on the "enable" cases. Doing this would allow you to avoid
> loading the unit files for the "disable" lines, as you you don't care
> for their [Install] section then.
Right. I used the term "generator" too flamboyant here indeed, as this
optimised preset parsing cannot ever be a generator-spec compliant.
I think optimisations above will work, especially given that we don't
honour aliases ahead of time
(that is bar.service unit with Alias=foo.service is not enabled, with
"enable foo.service" preset)
This is really good, as I think pre-parsing presets will help a lot here.
Open Source Technology Center
Intel Corporation (UK) Ltd. - Co. Reg. #1134945 - Pipers Way, Swindon SN3 1RJ.
More information about the systemd-devel