[Buildroot] Interesting circular dependency, how to solve?

[Buildroot] Interesting circular dependency, how to solve?

[Thomas Petazzoni]

Hello all,

Today, Paul (in Cc) reported on IRC an interesting problem: on a
configuration with systemd enabled (and acting as udev provider),
pciutils would fail to build due to libudev.h being missing.

After some investigation, "make graph-depends" revealed the origin of
the problem: a very interesting circular dependency. Here is what make
graph-depends thinks of the situation:

Recursion detected for  : systemd
which is a dependency of: udev
which is a dependency of: pciutils
which is a dependency of: linux
which is a dependency of: cryptodev-linux
which is a dependency of: openssl
which is a dependency of: libcurl
which is a dependency of: systemd

So, starting from the bottom:

 - When libcurl is enabled, systemd depends on it

 - When OpenSSL is enabled, obviously, libcurl will use it for SSL
   support

 - When cryptodev-linux is enabled, OpenSSL will depend on it to use
   crypto accelerators supported in the kernel via cryptodev-linux.

 - cryptodev-linux being a kernel module, it depends on linux

 - linux by itself (the kernel) does not depend on pciutils, but the
   linux tool "cpupower" (managed in linux-tool-cpupower) depends on
   pciutils

 - pciutils depends on udev when available

 - udev is provided by systemd.

So we have a dependency loop. This was not detected in the autobuilders
because it requires the linux package to be enabled to trigger.

I'm not sure where to break the dependency loop here. The real problem
is that cryptodev-linux does *NOT* require cpupower, it only needs the
kernel itself. Does this means that our current architecture for
packaging "tools" whose source code is part of the kernel as part of
the "linux" package is wrong ?

Do you have some suggestions or ideas on how to solve this ? The only
two solutions that I can find is:

 - Arbitrarily break this dependency loop by removing some
   functionality (but I'm not sure where)

 - Moving the cpupower build logic outside of the linux package.

The other question is how we can improve our autobuilder testing to
detect such situations.

Thanks,

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

[Buildroot] Interesting circular dependency, how to solve?

[ANDY KENNEDY]

> 
>  - When cryptodev-linux is enabled, OpenSSL will depend on it to use
>    crypto accelerators supported in the kernel via cryptodev-linux.

If I understand you correctly, this builds TWO things, right?

If this is the case, make it separate packages to build the
kernel module then the userspace package.

Would that remove the dependency loop?

Andy

[Buildroot] Interesting circular dependency, how to solve?

[Thomas Petazzoni]

Hello,

On Wed, 6 Jul 2016 17:54:01 +0000, ANDY KENNEDY wrote:

> >  - When cryptodev-linux is enabled, OpenSSL will depend on it to use
> >    crypto accelerators supported in the kernel via cryptodev-linux.  
> 
> If I understand you correctly, this builds TWO things, right?
> 
> If this is the case, make it separate packages to build the
> kernel module then the userspace package.

It builds a kernel module, so it has to depend on "linux", and it
installs cryptodev.h, which OpenSSL needs. I am not sure splitting this
package into two is really a good idea, though it would indeed solve
the problem: all what OpenSSL needs at build time is the cryptodev.h
header file. The kernel module can be built separately, and is only at
runtime.

So, yes, one possibility is to break cryptodev-linux into two packages.

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

[Buildroot] Interesting circular dependency, how to solve?

[ANDY KENNEDY]

> -----Original Message-----
> From: Thomas Petazzoni [mailto:thomas.petazzoni at free-electrons.com]
> Sent: Wednesday, July 06, 2016 1:20 PM
> To: ANDY KENNEDY
> Cc: buildroot at uclibc.org; paul.ashford at zurria.co.uk
> Subject: Re: [Buildroot] Interesting circular dependency, how to solve?
> 
> Hello,
> 
> On Wed, 6 Jul 2016 17:54:01 +0000, ANDY KENNEDY wrote:
> 
> > >  - When cryptodev-linux is enabled, OpenSSL will depend on it to use
> > >    crypto accelerators supported in the kernel via cryptodev-linux.
> >
> > If I understand you correctly, this builds TWO things, right?
> >
> > If this is the case, make it separate packages to build the
> > kernel module then the userspace package.
> 
> It builds a kernel module, so it has to depend on "linux", and it
> installs cryptodev.h, which OpenSSL needs. I am not sure splitting this
> package into two is really a good idea, though it would indeed solve
> the problem: all what OpenSSL needs at build time is the cryptodev.h
> header file. The kernel module can be built separately, and is only at
> runtime.
> 
> So, yes, one possibility is to break cryptodev-linux into two packages.

You _could_ do it via a hidden option that is selected by the other
option.  Then, it is still "one" package, however, the make
dependency tree would see it as two.

Andy

[Buildroot] Interesting circular dependency, how to solve?

[Yann E. MORIN]

Thomas, Paul, All,

On 2016-07-06 17:56 +0200, Thomas Petazzoni spake thusly:
> Today, Paul (in Cc) reported on IRC an interesting problem: on a
> configuration with systemd enabled (and acting as udev provider),
> pciutils would fail to build due to libudev.h being missing.
> 
> After some investigation, "make graph-depends" revealed the origin of
> the problem: a very interesting circular dependency. Here is what make
> graph-depends thinks of the situation:
> 
> Recursion detected for  : systemd
> which is a dependency of: udev
> which is a dependency of: pciutils
> which is a dependency of: linux
> which is a dependency of: cryptodev-linux
> which is a dependency of: openssl
> which is a dependency of: libcurl
> which is a dependency of: systemd

Hmmm... make itself should have spit a warning of its own. However, I
could not see a way to have make error out; it is a mere warning... :-(

> So, starting from the bottom:
> 
>  - When libcurl is enabled, systemd depends on it
> 
>  - When OpenSSL is enabled, obviously, libcurl will use it for SSL
>    support
> 
>  - When cryptodev-linux is enabled, OpenSSL will depend on it to use
>    crypto accelerators supported in the kernel via cryptodev-linux.
> 
>  - cryptodev-linux being a kernel module, it depends on linux
> 
>  - linux by itself (the kernel) does not depend on pciutils, but the
>    linux tool "cpupower" (managed in linux-tool-cpupower) depends on
>    pciutils
> 
>  - pciutils depends on udev when available
> 
>  - udev is provided by systemd.

Indeed. :-/

> So we have a dependency loop. This was not detected in the autobuilders
> because it requires the linux package to be enabled to trigger.

Maybe we could have some autobuilders use our defconfigs as starting
points for randpackageconfig? Most (all?) of our defconfigs have a
kernel enabled, so we could probably catch such a situation...

> I'm not sure where to break the dependency loop here. The real problem
> is that cryptodev-linux does *NOT* require cpupower, it only needs the
> kernel itself. Does this means that our current architecture for
> packaging "tools" whose source code is part of the kernel as part of
> the "linux" package is wrong ?
> 
> Do you have some suggestions or ideas on how to solve this ? The only
> two solutions that I can find is:
> 
>  - Arbitrarily break this dependency loop by removing some
>    functionality (but I'm not sure where)
> 
>  - Moving the cpupower build logic outside of the linux package.

I think this is the best solution. We could make the linux-tools be
built in their own package, vampirising the sources from the kernel (or
even building in there) by way of carefully crafted _PATCH_DEPENDENCIES
(but I'm afraid patch dependencies are still used to build the
dependency graph anyway... :-/ ).

I will look at it...

> The other question is how we can improve our autobuilder testing to
> detect such situations.

See above: have a randcom defconfig be picked and use it as a starting
point. We have almost all architectures covered, except maybe a few
corner cases for exotic stuff like bfin or arc...

Regards,
Yann E. MORIN.

> Thanks,
> 
> Thomas
> -- 
> Thomas Petazzoni, CTO, Free Electrons
> Embedded Linux, Kernel and Android engineering
> http://free-electrons.com
> _______________________________________________
> buildroot mailing list
> buildroot at busybox.net
> http://lists.busybox.net/mailman/listinfo/buildroot

-- 
.-----------------.--------------------.------------------.--------------------.
|  Yann E. MORIN  | Real-Time Embedded | /"\ ASCII RIBBON | Erics' conspiracy: |
| +33 662 376 056 | Software  Designer | \ / CAMPAIGN     |  ___               |
| +33 223 225 172 `------------.-------:  X  AGAINST      |  \e/  There is no  |
| http://ymorin.is-a-geek.org/ | _/*\_ | / \ HTML MAIL    |   v   conspiracy.  |
'------------------------------^-------^------------------^--------------------'

[Buildroot] Interesting circular dependency, how to solve?

[Yann E. MORIN]

Thomas, Andy, All,

On 2016-07-06 20:20 +0200, Thomas Petazzoni spake thusly:
> On Wed, 6 Jul 2016 17:54:01 +0000, ANDY KENNEDY wrote:
> > >  - When cryptodev-linux is enabled, OpenSSL will depend on it to use
> > >    crypto accelerators supported in the kernel via cryptodev-linux.  
> > 
> > If I understand you correctly, this builds TWO things, right?
> > 
> > If this is the case, make it separate packages to build the
> > kernel module then the userspace package.
> 
> It builds a kernel module, so it has to depend on "linux", and it
> installs cryptodev.h, which OpenSSL needs. I am not sure splitting this
> package into two is really a good idea, though it would indeed solve
> the problem: all what OpenSSL needs at build time is the cryptodev.h
> header file. The kernel module can be built separately, and is only at
> runtime.
> 
> So, yes, one possibility is to break cryptodev-linux into two packages.

But then we would probably have to do it for other packages that build
both a kernel module and userland stuff, no?

I'm not sure I like that... :-/

I think the best option would be to break at the linux-tools.

Regards,
Yann E. MORIN.

-- 
.-----------------.--------------------.------------------.--------------------.
|  Yann E. MORIN  | Real-Time Embedded | /"\ ASCII RIBBON | Erics' conspiracy: |
| +33 662 376 056 | Software  Designer | \ / CAMPAIGN     |  ___               |
| +33 223 225 172 `------------.-------:  X  AGAINST      |  \e/  There is no  |
| http://ymorin.is-a-geek.org/ | _/*\_ | / \ HTML MAIL    |   v   conspiracy.  |
'------------------------------^-------^------------------^--------------------'

[Buildroot] Interesting circular dependency, how to solve?

[Yann E. MORIN]

Thomas, Paul, All,

On 2016-07-07 18:33 +0200, Yann E. MORIN spake thusly:
> On 2016-07-06 17:56 +0200, Thomas Petazzoni spake thusly:
> > Today, Paul (in Cc) reported on IRC an interesting problem: on a
> > configuration with systemd enabled (and acting as udev provider),
> > pciutils would fail to build due to libudev.h being missing.
> > 
> > After some investigation, "make graph-depends" revealed the origin of
> > the problem: a very interesting circular dependency. Here is what make
> > graph-depends thinks of the situation:
> > 
> > Recursion detected for  : systemd
> > which is a dependency of: udev
> > which is a dependency of: pciutils
> > which is a dependency of: linux
> > which is a dependency of: cryptodev-linux
> > which is a dependency of: openssl
> > which is a dependency of: libcurl
> > which is a dependency of: systemd
> 
> Hmmm... make itself should have spit a warning of its own. However, I
> could not see a way to have make error out; it is a mere warning... :-(

Indeed it does catch it. But not until it is time for it to actuall
build openssl:

    make[1]: Circular /home/ymorin/dev/buildroot/O/build/openssl-1.0.2h/.stamp_configured <- cryptodev-linux dependency dropped.
    >>> openssl 1.0.2h Downloading

DAng. Why doesn't it detect it at the beginning of the build? And why
can't we tell make to consider it an error rather than a warning?

Damn... :-/

Regards,
Yann E. MORIN.

-- 
.-----------------.--------------------.------------------.--------------------.
|  Yann E. MORIN  | Real-Time Embedded | /"\ ASCII RIBBON | Erics' conspiracy: |
| +33 662 376 056 | Software  Designer | \ / CAMPAIGN     |  ___               |
| +33 223 225 172 `------------.-------:  X  AGAINST      |  \e/  There is no  |
| http://ymorin.is-a-geek.org/ | _/*\_ | / \ HTML MAIL    |   v   conspiracy.  |
'------------------------------^-------^------------------^--------------------'

[Buildroot] Interesting circular dependency, how to solve?

[Thomas Petazzoni]

Hello,

On Thu, 7 Jul 2016 19:00:24 +0200, Yann E. MORIN wrote:

> But then we would probably have to do it for other packages that build
> both a kernel module and userland stuff, no?
> 
> I'm not sure I like that... :-/
> 
> I think the best option would be to break at the linux-tools.

Agreed.

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

[Buildroot] Interesting circular dependency, how to solve?

[Thomas Petazzoni]

Hello,

On Thu, 7 Jul 2016 18:33:22 +0200, Yann E. MORIN wrote:

> Maybe we could have some autobuilders use our defconfigs as starting
> points for randpackageconfig? Most (all?) of our defconfigs have a
> kernel enabled, so we could probably catch such a situation...

My idea for this topic is more something like randomly, enabling the
Linux kernel package and depending on the selected architecture, use a
given defconfig (with a table of architecture <-> defconfig) match in
the autobuild-run script.

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

[Buildroot] Interesting circular dependency, how to solve?

[Yann E. MORIN]

Thomas, All,

On 2016-07-07 21:26 +0200, Thomas Petazzoni spake thusly:
> On Thu, 7 Jul 2016 19:00:24 +0200, Yann E. MORIN wrote:
> > But then we would probably have to do it for other packages that build
> > both a kernel module and userland stuff, no?
> > 
> > I'm not sure I like that... :-/
> > 
> > I think the best option would be to break at the linux-tools.
> Agreed.

OK, done locally, I've been a little bit side-tracked, so it is not
completely tested yet. will spin it tomorrow...

Regards,
Yann E. MORIN.

-- 
.-----------------.--------------------.------------------.--------------------.
|  Yann E. MORIN  | Real-Time Embedded | /"\ ASCII RIBBON | Erics' conspiracy: |
| +33 662 376 056 | Software  Designer | \ / CAMPAIGN     |  ___               |
| +33 223 225 172 `------------.-------:  X  AGAINST      |  \e/  There is no  |
| http://ymorin.is-a-geek.org/ | _/*\_ | / \ HTML MAIL    |   v   conspiracy.  |
'------------------------------^-------^------------------^--------------------'