I was reminded of this issue last night while sorting out
a problem with invoking
rpmbuild -ba --target i386 foo.spec
RPM CLI options are order dependent.
The invocation I gace above works perfectly. What did *NOT* work
perfectly is this invocation
rpmbuild -ba --define '_topdir /some/path' --target i386 foo.spec
And the "fix" (gud enuf for mezzanine, not generally) that was
place was to change the invocation to
rpmbuild -ba --target i386 --define '_topdir /some/path' foo.spec
Yes, permuting --define and --target options changes rpmbuild behavior.
Another CLI option that is order dependent (and much more easily fixed
than --target) is this
rpm -v --showrc
while this does not
rpm --showrc -v
None of the above is new behavior, the order dependence has been
in all versions of rpm since forever.
But perhaps its time (finally!) to attempt a better solution.
The problem with --showrc (and --querytags, and --eval, mebbe 1 or 2
is that the processing (and exit) happens immediately, terminating
other option processing. Treating the handful of options that
do immediate processing-and-exit as an additional "other" mode
(like --install/--erase/--query/--verify/...) of operation is easy
But --target fixing will be much harder because of the macro
side effect. Because macros are included in paths that contain macro
there is an intrinsic chicken <-> egg problem.
I've tried several approaches to remove order dependence from RPM CLI
Here is what has already been tried and why it doesn't work:
1) Add a --predefine option that adds a macro definition but doesn't
trigger reading macro definitions.
This "works" iff users are forced to specify --predefine before other
options that need the definition.
2) Multi-pass option processing, and rewinding options.
Way back when, RPM actually searched all options for --verbose/-v,
set the verbosity, rewound the options, and set about processing
with verbosity pre-initialized. Life was good.
However, rewinding popt option processing is no longer simple. There
are popt aliases and execs, and memory leaks not only from argv
substitutions, but also from popt table entries that malloc memory
while saving option args. And then there are other additional side
effects because of added "features" while string encodings are
converted on the fly.
All solvable, but extremely fragile to maintain properly. Which
is why rpm ceased to rewind its options years ago, it really isn't
*THAT* ugly that -v must be specified before whatever depends on
verbosity imho, ymmv.
3) (currently implemented) lazy macro initialization.
There's only a handful of options that care whether macros are
fully initialized (or not), the most important of which is
--define. Macros are stacks, so all macro initialization must
be completed before --define is pushed, or a later value from
macro initialization will be pushed on top of the --define processing.
The flaw with --define <-> --target is that each has side effects, and
so are order dependent.
There are many ways to solve the ordering problem, its no harder than
a precedence for each option, and then processing options in a
predefined order, determinism is what is needed.
And it certainly isn't the end of the world to just document that --
be used before any --define can be used, but that really doesn't solve
issue, does it?
How much should I work at fixing the above problem?
If left to my devisings, I'm likely to wire a precedence into some
bits in the opt->argInfo field in popt option tables and use that
hint to sort the argv array handed to popt before/while option
that fixing precedence ordering in popt will only delay the rpm fixing
introduces Yet Another Version pre-requisite building rpm itself. And
so one gets
to solve the precedence ordering problem twice in two different ways, an
evolutionary process to software development that takes years to
73 de Jeff
Received on Wed Dec 17 15:55:33 2008
- application/pkcs7-signature attachment: smime.p7s