Lock breaking patch for the fully preemptible linux kernel
Or, a low-latency patch for preempt-kernel

This patch attempts to further lower latency under Linux by
eliminating the worst-case latencies under the preempt-kernel
patch by "breaking" long-held locks inside the kernel.  It is
intended for use by those with hightened response needs that
are affected by certain locks, primarily in the VM and VFS
code.

While the preemptible kernel eliminates the need for specific
scheduling points throughout the kernel, preemption must not
occur while a lock is held.

This patch borrows from Ingo Molnar and Andrew Morton's
low-latency patches.  In particular, much credit is due to
Andrew whose low-latency patch is the basis for much of this code.
Thanks goes to preempt-kernel users who reported preempt-stats
results and William Irwin for some good testing.

	Use
Patch.  Enable CONFIG_PREEMPT and CONFIG_LOCK_BREAK from
`Processor Options'.  Compile.

	Debugging
Setting DEBUG_LOCK_BREAK in include/linux/lock_break.h will
define the debug_lock_break(n) macro to report instances where
the expected lock depth does not equal the actual.  If the lock
depth is higher than expected, we may not be dropping enough locks.
If it is 0, we are wasting our time since preemption can occur.

	Enjoy,

	Robert Love
	rml@tech9.net