From: William Lee Irwin III <wli@holomorphy.com>

The original pgd/pmd slabification patches had a critical bug on
non-PAE where both modifications of pgd entries to remove pagetables
attached for non-PSE mappings back to a PSE state and modifications of
pgd entries to attach pagetables to bring PSE mappings into a non-PSE
state were not propagated to cached pgd's. PAE was immune to it owing
to the shared kernel pmd.

The following patch vs. 2.5.69 restores the slabification done to cache
preconstructed pagetables with the proper propagation of conversions
to and from PSE mappings to cached pgd's for the non-PAE case.

This is an optimization to reduce the bitblitting overhead for spawning
small tasks (for larger ones, bottom-level pagetable copies dominate)
primarily on non-PAE; the PAE code change is largely to remove #ifdefs
and to treat the two cases uniformly, though some positive but small
performance improvement has been observed for PAE in one of mbligh's
posts. The non-PAE performance improvement has been observed on a box
running a script-heavy end-user workload as a large long-term profile
hit count reduction for pgd_alloc() and relatives thereof.

I would very much appreciate outside testers. Even though I've been
able to verify this boots and runs properly and survives several cycles
of restarting X on my non-PAE Thinkpad T21, that environment has never
been able to reproduce the bug. Those with the proper graphics hardware
to prod the affected codepaths into action are the ones best suited to
verify proper functionality. There is also some locking introduced; if
some performance verification on non-PAE SMP i386 targets (my SMP
targets unfortunately all require PAE due to arch code dependencies)
that also have the proper hardware could be done, that would help
determine whether alternative locking schemes that competed against
the one shown here are preferable (in particular, the ticket-based
scheme mentioned in the comments).