Start virtualizing real mmap() functionlity in the kernel.  The first
signs of me thinking about this already showed up as some unused header
files earlier, this fleshes things out some more.  No actual filesystem
code yet..

This also removes my old simple math emulator, and introduces the new
and much improved one from Bill Metzenthen.

Bill originally wrote it for the djgpp suite (DJ Delories gcc port to
DOS extenders).  It was much more accurate and well designed than my
hackish one, and I was happy to throw my old code away.  The new math
emulator also did things that I had never bothered with, notably the
more complex i387 functions (exponentials and trig).

I also fixed the static maximum memory limit: we now generate the kernel
page tables dynamically rather than having a 16M or 32M static limit.

SCSI updates: removable media support (which also implies re-reading the
partition table etc)

[Original announcement below]

patch-2 is >150kB compressed, as it contains several big changes. Most
notable are:

 - the new FPU-emulator by Bill Metzenthen.  It's bigger than the old
   one, but thanks to it, linux fpu emulation is no longer a quick hack,
   but a real emulator: it does all the 387(486) math instructions, and
   does them much faster than the old emulator + the soft library.

   The new math-emulator means that a separate soft-float library is no
   longer needed.  It also makes even a non-coprocessor system pretty
   useful for limited math-calcs - the complex functions are much faster
   when they no longer have to be calculated using simple functions, and
   even the simpler instructions that my old emulator handled are faster
   using the new one.

   The size of the new emulator may mean that people who have little
   RAM, but do have a coprocessor should probably recompile the kernel
   with the emulator disabled.

 - various minor mm fixes by me: trapping kernel NULL dereferences,
   cleaning up the page table initializations and the 16MB patches, and
   various other bugfixes.  get_free_page(GFP_ATOMIC) should preserve
   the interrupt flag, so malloc() should be safe now - hopefully no
   more of the tcp/ip memory management problems.

   The NULL pointer trapping may result in errors like:
        Unable to handle kernel paging request at address C0000???
        Oops: 0000
        ..... debugging info .....

   There were several NULL pointer dereferences in the serial and tty
   drivers, which should now be fixed.  I've also fixed any other errors
   I've seen, but if there are problems in the scsi drivers or similar
   things I cannot test, I'd like to hear about them.

 - scsi driver changes by Eric Youngdale.  Preliminary support for
   removable media, and some bug-fixes.  Due to white-space problems
   with eric's patches, the scsi patches are a bit bigger than
   necessary, but they should be ok even though I had to put them in
   partly by hand (and being unable to test them...)

 - The new tcp/ip patches that were sent to the NET channel not long
   ago.  Yes, they are alpha, but so is the whole tcp/ip directory, so I
   put them in even thought they haven't been extensively tested (and
   they did have a serious problem in the ioctl code, which I fixed).

 - psaux mouse patches by Dean Troyer, as well as the mouse.wait = NULL
   patch.

Before (or after) patching, you should remove the old math-emulator (ie
"rm -rf /usr/src/linux/kernel/math") as it is no longer needed.  You
should also do a "make dep" to update dependencies: as usual, I edited
out the dependancy-changes.  Do a "make clean", edit the main (and net)
Makefiles to suit your system, and compile.

And finally: I will no longer be making the bootdisks available -
they'll be made by hlu/jwinstead and will probably be boot+root-disks
using lilo, as done on the hlu disks.  That may mean that a bootimage
won't be available at once, but most people who want to use the
absolutely newest images probably compile them themselves anyway, so
that shouldn't be a problem.

                Linus