On Mon, 5 Mar 2007 02:50:31 +0100 Adrian Bunk [off-list ref] wrote:

This email lists some known regressions in 2.6.21-rc2 compared to 2.6.20
that are not yet fixed in Linus' tree.

We seem to have broken an unusually large amount of stuff this time.

partial post-mortem:

- The ACPICA merge landed in -mm super-late: basically it was in mainline
  a week afterwards and saw only a single -mm release.

  Part of the reason for this short period in -mm was that ACPICA had its
  paws all over x86_64 code and conflicted badly with significant changes
  in the x86_64 tree.

  That happens sometimes.  But when it does, the mess lands in my lap
  rather than in the laps of the perpetrators.

  Lesson: keep the code well-factored so that different subsystems don't
  soil each others' kennels.

- The hrtimers/dynticks stuff is simply hard: timekeeping, low-level x86,
  even APICs.  These are areas in which things break a lot, so churning it
  was inevitably going to cause problems.

  Lesson: none, I think.  Low-level x86 support is just hard, and
  changing it breaks things.


So that accounts for _some_ of the damage, but I wonder if there's more to
it than that.