Re: rm -rf: avoid ext3/4 O(N^2) performance hit, further librarify remove.c

[Jim Meyering]

Jim Meyering <address@hidden> wrote:
> Here are three patches:
> The first relaxes "make distcheck" so it'll pass even after
> the 2nd goes in (it introduces several c99 stmt-after-decl).
>
> The 2nd patch does the same thing for rm that I did for fts:
>
>   http://thread.gmane.org/gmane.comp.lib.gnulib.bugs/14808
>
> i.e., this change makes rm process dir entry names in sorted
> inode order, when that's sensible.
>
> Adding the test to exercise/exhibit the fix was interesting.
> I classified it as "expensive", so it doesn't run by default:
> it creates and removes 400,000 files.  The removal takes about
> 10s on a reasonably modern system.  If it takes more than 1 minute,
> that's deemed a failure.  Without the patch, the removal takes almost
> 6 *minutes* with a fast system.  If you want to run just
> that one test, do this:
>
>   make -C tests TESTS=rm/ext3-perf RUN_EXPENSIVE_TESTS=yes VERBOSE=yes
>
> The 3rd patch fixes remove.c to be closer to library-ready,
> by removing uses of xmalloc and by adjusting the existing obstack-
> using code so that obstack allocation doesn't provoke an exit.

I've pushed those patches with some small additional changes.
Thanks to those who gave feedback.

I'd like to put a test release on alpha.gnu.org next week,
so if you have any pending problem reports or patches, post
them to the list soon.