On Sun, Nov 01, 2009 at 06:15:50AM +0300, Alexey Tourbin wrote:
> On Sun, Nov 01, 2009 at 05:14:42AM +0300, Dmitry V. Levin wrote:
[...]
> > I've just pushed sisyphus_check 0.8.14-alt1-4-gc78c3ee that implements
> > this check.  Please have a look.
> 
> -rpm_requires=[%{requirename:shescape}" %{requireflags:depflags} "%{requireversion:shescape}"
> +rpm_requires=[%{requirename:shescape}" %{requireflags} %{requireflags:depflags} "%{requireversion:shescape}"
> 
> %{requireflags:octal} may (or may not) be more convenient here,
> possibly depending on whether awk can handle octals just as easy
> as decimals.

I doubt that octals here may appear more convenient.  GNU awk definitely
handles decimal numbers, and nothing else is going to handle this field
yet.

> +bad_prereqs='/sbin/install-info
> +/sbin/ldconfig'
> 
> +       # 2^6: RPMSENSE_PREREQ
> +       # 2^14: RPMSENSE_FIND_REQUIRES
> +       bad=$(printf %s "$rpm_requires" |
> +               awk '$2>=64 && $2!=16384 {print $1}' |
> +               egrep -x "$bad_prereqs" |
> +               sort -u)
> 
> Since %{requireflags} is a bitmap, numeric comparison looks puzzling.

It is result of optimization, so it may look puzzling and even a bit
confusing.  The idea behind this optimization is simple: any high value flag
except RPMSENSE_FIND_REQUIRES means that the requirement has a PREREQ sense.

> What are exactly the rules we'd like to follow?  Basically, what we want
> to do is to forbid the use of /sbin/install-info in %post-like scripts.
> For such a script, the dependency on /sbin/install-info is generated
> automatically.  But usually it is also accompanied with manual PreReq
> dependency on %__install_info.  We should cover both cases -- for
> example, one can forget to remove the accompanying PreReq dependency
> while removing the scripts.
> 
> Now to the requireflags.  Older rpmbuild implementation (the one we use)
> always sets PREREQ bit along with e.g. SCRIPT_POST (please check if this
> is actually true).  Thus, to handle both cases (automatic scriptlet
> dependencies and manual prereqs) it would suffice to perform the test
> as simple as
>     
> 	bad = (requireflags & PREREQ);
> 
> Newer rpmbuild implementation does not set PREREQ bit along with
> scriptlet bits, so the test should be extended to the scriptlet bits:
> 
> 	bad = (requireflags & (PREREQ | SCRIPT_POST | SCRIPT_...));

Prerequisites are not limited to RPMSENSE_PREREQ and RPMSENSE_SCRIPT_*
bits.  There are also RPMSENSE_INTERP and RPMSENSE_TRIGGER* which should
be treated the same way as other PREREQs.  As result, an appropriate awk
expression grows like this:
awk 'and($2,2^6+2^8+2^9+2^10+2^11+2^12+2^13+2^16+2^17+2^18)>0 {print $1}'

Of course it could be pre-calculated to
awk 'and($2,474944)>0 {print $1}'

Does it look clearer than
awk '$2>=64 && $2!=16384 {print $1}'
?

> Another idea looming around is to prohibit manual dependencies
> on /sbin/install-info altogether.  I'm not quite sure if it is a good
> idea, since it goes far beyond scriptlets and prereqs.  After all,
> you name it "bad_prereqs", so it must be a prereq.

I was uncertain about this, so I made a check for /sbin/install-info and
/sbin/ldconfig requirements in Sisyphus.  I found only one manual
non-PREREQ requirement for /sbin/ldconfig (in
libreadline-java-0.8.0-alt2_12jpp5) and no manual non-PREREQ requirements
for /sbin/install-info.  So I decided to leave the check as is.
Maybe I should reconsider this.

> Also consider adding 'info-install' to bad_prereqs (since they might
> use bare 'install-info' command, not /sbin/install-info).

You mean the info-install package?  Yes, that makes sense.


-- 
ldv