On Wed, Oct 14, 2026 at 08:03:50PM +0300, Jarkko Sakkinen wrote: > On Wed, Oct 14, 2026 at 07:56:32PM +0300, Jarkko Sakkinen wrote: > > On Wed, Oct 14, 2026 at 06:35:50PM +0200, Eric Biggers wrote: > > > On Wed, Oct 14, 2026 at 07:01:38PM +0300, Jarkko Sakkinen wrote: > > > > This is a kunit test for lib/assoc_array.c, which maintains a shadow > > > > model > > > > across assoc_array operations, and checks its integrity after every > > > > operation. > > > > > > > > I wrote a draft version so that I could properly test > > > > 'assoc-array-fixes' > > > > tag at Linus' tree, and spent today rewriting a new more polished > > > > version > > > > with as good description as I could possibly could on how the seeded > > > > test > > > > keys were cherry-picked. Along the way I also tweaked the numbers to be > > > > a bit less messier set than in my original version. > > > > > > > > The transcripts that follow were executed with this command-line: > > > > > > > > ./tools/testing/kunit/kunit.py run --raw_output assoc_array > > > > > > > > With 'assoc-array-fixes' commits and its subsequent tag iterations all > > > > eight tests pass, and without them all of them fail. > > > > > > > > Signed-off-by: Jarkko Sakkinen <[email protected]> > > > > > > Would this effort (and Linus's effort) be better spent on converting > > > keyrings to use an rhashtable? > > > > It did cross my mind when I reviewed and tested those fixes but you > > are asking something that I would need at least a few hours to have > > any opinion. > > > > We do bugs and QA to the end-of-life-cycle. We neither can predict > > exact time how such major overhaul would take, if we even want to > > do it. I don't know. > > > > My brain is on this and I'm really bad at multitasking :-) > > If migration was done probably two-phase approach is what should > sought after. I.e. keep the API and change the internals. > > After that it is time for re-consideration because from that point > there are multiple (API) paths forward.
rhashtable is the best pick I think. xarray not so much given the data topology. Br, Jarkko
