Re: Can s6 be enough?: was s6-ps

[Laurent Bercot]

When Debian acquires a properly working s6-rc package, the answer to my
question degenerates to "why not?" But for now, for the Debian person
who only installs via package, s6-rc is out of the question, so my
question was, isn't s6 itself good enough?


 Then maybe you should ask Debian instead, "why aren't you packaging
s6-rc?"

 It's literally a distribution's *job* to package software that may
be useful to users. And instead, you are bending your mind to try
and find justifications for users not to use some software, for the
simple reason that it's not packaged by a distribution. That logic
is 100% backwards.

 You want to help users? Package s6-rc for Debian. Also, call out
Debian on their prejudice and disingenuous use of policies that
prevent people from making correct execline and s6 packages.

--
 Laurent

Re: s6-ps

[Colin Booth]

On Sun, Jan 06, 2019 at 07:30:21AM +, Laurent Bercot wrote:
> 
>  - The execline library is required to build a few of s6's utilities
> (typically: s6-ftrig-listen).
>  - The execlineb binary is required for use of the !processor directive
> in s6-log.
>  - Some of the other binaries provided by execline are used by
> s6 utilities, for instance s6-fdholder-store.
>
Ah, good point. I was misremembering and thought s6-log processed the
command string itself but used libexecline to teach it how, similar to
s6-ftrig-listen's dependency.

Looking a bit closer, the one that's going to break the hardest for
people using s6 solely as a supervisor is s6-notifyoncheck since you can
inline a check program with the -c option. I'm not sure how many people
go that route who aren't already going to have execline installed, but
it's something that will definitely cause some ill will.
> 
>  "Recommends" is a mistake. Without execline, parts of the s6 suite
> of programs will break.
>
> --
>  Laurent
> 

-- 
Colin Booth

Re: Can s6 be enough?: was s6-ps

[Brett Neumeier]

On Sat, Jan 5, 2019 at 2:30 PM Steve Litt  wrote:

> So what do you all think? Is s6 a useful init system without s6-rc?
>

My 0.02 USD -- based on my experience of setting up a simple GNU/Linux
distribution from the ground up using s6, s6-rc, and s6-linux-init...

- s6-rc provides useful functionality: it is really handy, when defining
the way that the system should start up, to have bundles and oneshots; it
is also really handy to be able to start up or shut down groups of
processes via bundles.
- The cost of using s6-rc is negligible. As installed on my x86_64 system
with documentation, it consumes around 576 *kilobytes* of storage space. It
compiles and installs in substantially less than a minute. Learning how to
craft s6-rc service definition directories is no more difficult than
learning how to craft s6 servicedirs.
- You don't lose any capability provided by s6 if you also use s6-rc. You
can send whatever signals you want to the supervised processes by using
s6-svc directly.

So ... costs ~= 0, benefits > 0, to me the question of whether s6 is useful
_without_ s6-rc is kind of pointless.

I'm inclined to turn the question around: what leads you to want to avoid
s6-rc? Is there some other system that provides more benefits at lower cost?

Cheers!

Brett

-- 
Brett Neumeier (bneume...@gmail.com)

Re: Can s6 be enough?: was s6-ps

[Laurent Bercot]

Eliminate dependency on udevd from oneshot startup scripts.


 One example among others:
Kernel events are used to automatically load dynamic kernel modules.
Say you need to mount a filesystem of a type that's not known in your
core kernel, but you have a module for that. Either you manually
modprobe everything in your initial oneshots (which is possible, but
less flexible and more painful), or you just spawn udevd first and
can then mount everything you need.
The latter is clearly the intended use case - that's what automatic
module loading is for. But you can only do that by making oneshots
depend on the udevd longrun.

Don't organize your system around the tools you have. You shouldn't
have to jump through hoops to accommodate mediocre tools/architecture
or distributor laziness; it's a losing game in the end, and the user
is the one who loses.
Instead, get the tools you need so your system works as painlessly
as possible for the user. In some cases, you won't need a service
manager; in most cases, you will.

--
Laurent

Re: s6-ps

[Laurent Bercot]

 - The execline library is required to build a few of s6's utilities
(typically: s6-ftrig-listen).
 - The execlineb binary is required for use of the !processor directive
in s6-log.
 - Some of the other binaries provided by execline are used by
s6 utilities, for instance s6-fdholder-store.

 "Recommends" is a mistake. Without execline, parts of the s6 suite
of programs will break.

--
 Laurent

Re: s6-ps

[Colin Booth]

On Sat, Jan 05, 2019 at 06:00:57PM -0300, Guillermo wrote:
> El sáb., 5 ene. 2019 a las 7:46, Jonathan de Boyne Pollard escribió:
> >
> > * https://packages.debian.org/sid/s6
> 
> And buster too. Which has the side effect of making it automatically
> available in Devuan ceres and beowulf :)
> 
> * https://pkginfo.devuan.org/cgi-bin/d1pkgweb-query?search=s6_2.7=any
> 
> I don't know why execline is a 'recommends' though. It should be a
> 'depends', I believe.
> 
> G.
libexecline is a runtime requirement for s6-log if linked dynamically.
It's also a compile time requirement for roughly the same reason.
Assuming a static (or mostly static) compilation of the s6 suite,
execline is correctly marked as a recommendation.

Looking at the package in Unstable, zhs has split libexecline from
the execline binary suite and dynamically compiled things. Therefore,
installing s6 forces installation of libexecline, libskarnet, and libs6,
and recommends the execline bundle.

Cheers!

-- 
Colin Booth

Re: s6-ps

[Steve Litt]

On Sat, 5 Jan 2019 18:00:57 -0300
Guillermo  wrote:


> I don't know why execline is a 'recommends' though. It should be a
> 'depends', I believe.

I set up a s6 supervisor using all shellscripts and no execline. You
don't absolutely need execline.

 
SteveT

Steve Litt 
January 2019 featured book: Troubleshooting: Just the Facts
http://www.troubleshooters.com/tjust

Re: Can s6 be enough?: was s6-ps

[Laurent Bercot]

Everybody appreciates the preceding two features, but personally, I
don't think they're absolutely necessary. Runit has neither, yet it
works just fine for most things.


It really depends on what "most things" are.
Small, server-only appliances? sure.
Distributions where you can start all the oneshots first and all
the longruns later? sure, but there aren't many of those. Typically
udevd is a daemon you want to bring up very early, before some
oneshots; so in the runit model, what do you do with udevd?

Void Linux, which uses runit as its init system, has an answer to
that: it doesn't supervise udevd.
Don't you think that's ugly? Because I think that's ugly.



So I'm thinking that although process ordering and intermixing are
features we'd everyone likes, all but the most discerning among us
could get along without them and therefore init with s6 alone.


Well yes. You can even get along with no supervision at all.
Most people got along with sysvinit and sysv-rc just fine. The
question is, do you want to? What do you expect from your init
system? what functionality do you want to have, what functionality
do you want your distribution to provide?



 My opinion is one can boot just fine with s6 alone, as long as
you're willing to forego startup ordering and intermixing of longruns
and oneshots, which by definition a runit fan is.


Yes, it can be done. As a matter of fact, it's completely trivial to
take a Void Linux distribution and replace runit with s6. It works
flawlessly. It just hasn't been done in the official Void distribution
yet because the Void maintainers have other priorities, and runit
works well enough for them.

s6-rc and anopa exist for the people who are *not* willing to forego
intermixing of longruns and oneshots. It all depends on what your
priorities are, of what the aim of your distribution is.



So what do you all think? Is s6 a useful init system without s6-rc?


Terminology, please! s6 isn't an init system per se. It's a process
supervision system; and s6-rc is a service manager, aka machine state
manager. Those are _components_ of an init system. You can build a
complete, full-featured init system with s6 and s6-rc. You can also
build an init system with s6 alone (and some scripting as
s6-linux-init does), but it won't include a service manager. It's
your call, really: I provide the mechanisms, distributions provide the
policy. When they don't utterly give up on their duties, obviously.

If you haven't watched it yet, please take a look at this video:
https://archive.fosdem.org/2017/schedule/event/s6_supervision/
It's only 15 minutes long and explains all the terminology.

--
Laurent

Re: s6-ps

[Guillermo]

El sáb., 5 ene. 2019 a las 7:46, Jonathan de Boyne Pollard escribió:
>
> * https://packages.debian.org/sid/s6

And buster too. Which has the side effect of making it automatically
available in Devuan ceres and beowulf :)

* https://pkginfo.devuan.org/cgi-bin/d1pkgweb-query?search=s6_2.7=any

I don't know why execline is a 'recommends' though. It should be a
'depends', I believe.

G.


G.

Can s6 be enough?: was s6-ps

[Steve Litt]

On Sat, 5 Jan 2019 10:46:29 +
Jonathan de Boyne Pollard 
wrote:


> s6 and s6-rc are actually ports/packages in FreeBSD and s6 is a
> package in Debian.  Alas, the Debian world has not yet caught up with
> the other toolsets, and the third-party Debian packaging for
> s6-linux-utils and the others has yet to make it into Debian's own
> repository.

This brings up a question I've had for a long time. Three years ago I
constructed an init consisting of Suckless Init as PID1 and s6 as the
supervisor, which was called at the end of a simple rc file forked by
Suckless Init. It worked just fine, without s6-rc or any other s6
products.

From discussions with Laurent I got the idea that the main added
functionalities of s6-rc are:

1) Process ordering during boot
2) Ability to intermix longruns with oneshots

Everybody appreciates the preceding two features, but personally, I
don't think they're absolutely necessary. Runit has neither, yet it
works just fine for most things. When I first switched to runit, with
its indeterminate ordering, I figured that with the dependencies I
built into the run scripts,  boot would resemble a field of mousetraps
throwing ping pong balls. But that wasn't the case: runit brought
everything up quickly and without trouble.

So I'm thinking that although process ordering and intermixing are
features we'd everyone likes, all but the most discerning among us
could get along without them and therefore init with s6 alone.

This is more than theoretical. As you remember, Debian's runit package
manager wrote to ask about a minor problem with too many .u files and
was contemplating changing some code, but nobody knew the code
intimately, and runit is pretty much unmaintained at this point. So I
suggested s6, which apparently has a Debian package but not one for
s6-rc. My opinion is one can boot just fine with s6 alone, as long as
you're willing to forego startup ordering and intermixing of longruns
and oneshots, which by definition a runit fan is.

So what do you all think? Is s6 a useful init system without s6-rc?

Thanks,
 
SteveT

Steve Litt 
January 2019 featured book: Troubleshooting: Just the Facts
http://www.troubleshooters.com/tjust

Re: s6-ps

[multiplexd]

Hi all,

(Apologies to JdePB, I accidentally replied to him directly instead of
to the supervision list.)

On Sat, Jan 05, 2019 at 10:46:29AM +, Jonathan de Boyne Pollard wrote:
> Alas, the Debian world has not yet caught up with the other
> toolsets, and the third-party Debian packaging for s6-linux-utils and the
> others has yet to make it into Debian's own repository.

>
> * https://github.com/multiplexd/s6-packaging
>

Thanks for the advertising, Jonathan :-). I would, however, like to note
that getting that packaging into the official Debian repos is not an
immediate goal for me, as I don't have the time to commit to being a
Debian package maintainer for the forseeable future.

Cheers,
--multiplexd.

Re: s6-ps

[Jonathan de Boyne Pollard]

That said, I would probably write my own portable tool if it came to 
that, as my gripes with |ps| are not to everyone's tastes.


 *

   It is the 21st century.  The FreeBSD, OpenBSD, and NetBSD |ps|
   programs all use |-| as the command-line option introducer, and have
   done since /at least/ BSD 4.4 Lite in 1994
   . 
   A "BSD personality" that gives me the interface from the 1980s is

   not /actually/ a BSD personality that gives me |ps| as it exists and
   as I use it on the BSDs.  This is neither compatibility nor
   interoperability.

 *

   The defaults are wrong.  Much of my GUI is processes that run in
   dæmon context.  By default, I /want/ to see processes in sessions
   that have no controlling terminal.  The |-x| option is the wrong way
   around.

 *

   The defaults are wrong.  Most of the time I want to see all users'
   processes.  The |-a| option is the wrong way around.  (Yes, most of
   the time I want to "see processes from other users" in Windows, too.)

 *

   Indeed, even in the unusual case where I am interested in only my
   own processes, my own processes /include/ set-UID processes spawned
   by processes running as my account, such as |urxvtd|, which is /my/
   Unicode rxvt dæmon.

 *

   The defaults are wrong.  Most of the time I do not want to see
   kernel processes.

 *

   When I use both the |comm| and |args| columns, I don't need an extra
   copy of the former put in brackets in the latter.

 *

   When I use the |comm| and |args| columns with |f|, I don't need both
   columns to have the tree diagram added to them, especially as it
   usually then obscures what is in the actual column, that I wanted to
   see.  Notice how FreeBSD |ps| only puts the tree diagram in if
   |args| is the last column.  Do that.

 *

   Your "BSD personality" should be giving me |-d| anyway, not |f|, as
   that is the actual option on the BSDs.  I don't want a tool with
   Multiple Personality Disorder and Jack-of-All-Trades option parsing,
   though.  I want a |ps| tool that is self-consistent and logical,
   with /one/ personality across all of the platforms that it works on,
   and /one/ way of doing most things (albeit allowing both short and
   long command-line options).

 *

   In fact, give me the tree diagram as its own, first class, column,
   as well.

 *

   My terminal emulators speak UTF-8 and have switchable fonts. Their
   grave accent sometimes really looks like a grave accent; and their
   minus sign sometimes really looks like a minus sign.  Give me proper
   line drawing characters.  It's ridiculous giving me a "BSD
   personality" from the 1980s and yet a TUI that does not even do TUI
   line drawing as well as it was done on MS/PC/DR-DOS in the 1980s.

 *

   The defaults are wrong.  It is the 21st century.  I haven't used a
   terminal with only the choices of 80 or 132 columns in years.  Stop
   giving me 132 columns as the primary alternative behaviour.

 *

   The Single UNIX Specification has standardized the |COLUMNS|
   environment variable as a way for me to tell you how many columns I
   want, for many years, now.  Don't give me idiosyncratic command-line
   option mechanisms of your own. Don't give me "It's unspecified.". 
   Give me the |COLUMNS| environment variable.


 *

   The defaults are wrong.  I'm piping your output through |less|,
   which knows how to sideways scroll.  Release your right margin when
   standard output is not a terminal device.  I do not need the final
   column to be of "undefined" width.  I need it to be defined, as
   having unlimited width.  For any column type, not just one or two of
   them.

 *

   Indeed, if I am piping output give me the option of releasing the
   width restrictions on all columns, not just the last one. I might
   want to have both the |env| and |args| columns, with neither truncated.

 *

   Don't give me |-u| , |-v| , |-l| , |-f| , |-j| and tell me that they
   are "user", "virtual", "long", "full", and "jobs", especially if you
   aren't going to give me long options or aren't even going to
   document what those are.  Give me at the least a |--format| option
   that takes |user|, |virtual|, |long|, |full|, and |jobs|.  Indeed,
   give me a |--format| option that has /all/ of the shorthands from
   |procstat|.  (But that, unlike |procstat|, can combine with |-o| and
   itself rather than give me a mutually-exclusive choice amongst
   things none of which are /quite/ what I want.)

 *

   In fact, give me just |--output| for selecting columns individually
   and |--format| for adding predefined handy sets of columns, and
   allow them to combine as I see fit.

 *

   Give me |--no-headers| too.

 *

   Give me the option of TAB rather than SPC as column separator whilst
   you are at it.  Then I can run it through tools that understand
   tables-of-one-line-per-record-with-TAB-separated-fields text. 

Re: s6-ps

[Jonathan de Boyne Pollard]

Laurent Bercot:
I just wasn't aware of people besides me actually using s6-ps. :) 


I used it as an example of not being a mess.  (-:

s6 and s6-rc are actually ports/packages in FreeBSD and s6 is a package 
in Debian.  Alas, the Debian world has not yet caught up with the other 
toolsets, and the third-party Debian packaging for s6-linux-utils and 
the others has yet to make it into Debian's own repository.


* https://svnweb.freebsd.org/ports/head/sysutils/s6/

* https://svnweb.freebsd.org/ports/head/sysutils/s6-rc/

* https://packages.debian.org/sid/s6

* https://github.com/multiplexd/s6-packaging

You could, in theory, make s6-ps not Linux-specific.  It would require 
interposing a layer that abstracted reading the information out of the 
kernel, though.  It's perhaps not worth it, because (for example) the 
FreeBSD ps is not nearly as encumbered by the mess that results from 
multiple personalities and Jack-of-All-Trades option parsing, albeit 
that it still has confusing column headings and does not have an env column.


* https://unix.stackexchange.com/questions/492513/

Re: s6-ps

[Laurent Bercot]

It occurs to me that s6-ps would gain from having an "exe" column.


 That functionality:
 - isn't available in procps. I modelled the available columns from
procps' ones. Although that doesn't mean much, because there are many
more columns available now in recent versions of procps than there are
in s6-ps.
 - requires an extra system call per process: a readlinkat() on "exe".
I tried to limit the functionality to what can be read in the "stat"
file. Using "comm" as well was unavoidable, but annoying.

 That said, if there's demand for it, I'm not opposed to adding a
column for exe. I just wasn't aware of people besides me actually using
s6-ps. :)

--
 Laurent

s6-ps

[Jonathan de Boyne Pollard]

* https://unix.stackexchange.com/a/492346/5132

It occurs to me that s6-ps would gain from having an "exe" column.