This is the first of three patches which reduce the amount of
kmap/kunmap traffic on highmem machines.

The workload which was tested was RAM-only dbench.  This is dominated
by copy_*_user() costs.

The three patches speed up my 4xPIII by 3%

The three patches speed up a 16P NUMA-Q by 100 to 150%

The first two patches (copy_strings and pagecache reads) speed up an
8-way by 15%.  I expect that all three patches will speed up the 8-way
by 40%.

Some of the benefit is from reduced pressure on kmap_lock.  Most of it
is from reducing the number of global TLB invalidations.


This patch fixes up copy_strings().  copy_strings does a huge amount of
kmapping.  Martin Bligh has noted that across a kernel compile this
function is the second or third largest user of kmaps in the kernel.

The fix is pretty simple: just hang onto the previous kmap as we we go
around the loop.  It reduces the number of kmappings from copy_strings
by a factor of 30.