I managed to break compilation with my last patch with a last minute edit.

that can also be used in assembly language context.

This one fixes the common CPU_FTR_COHERENT_ICACHE warning in
asm/cacheflush.h.

x86-64 shares some numa code with i386, and that didn't handle holes in
the node map properly.  Fix this.

Add accumulated bug fixes for x86-64 and some minor cleanups.

 - Add ptep_set_access_flags from i386
 - Change memory text mapping to 10MB from 40MB to make alias window
   smaller
 - Disable change_page_attr for kernel mapoing for now (wrong, but the
   other code also doesn't work)
 - Add .bss.page_aligned and align interrupt/exception stacks (idea from
   Matt Mackall)
 - Protect sysenter MSR setup against missing registers. 
 - Set mce tolerance level to 1 (panic instead of deadlock)
 - Fix cross 4GB bug in find_first_bit (Michael Matz)
 - Add missing memory clobbers in bitops.h (Suresh B.  Siddha)
 - Check kernel size at boot up
 - Set boot cpu online at boot to fix early printk
 - Fix EFER bit definitions (Dave Jones)
 - Fix comments in 32bit syscall table.
 - Quieten all 32bit syscalls that are unimplemented in 32bit. 
 - Various MCE fixes suggested by Eric Morton, David Boles et.al.
   More to come.
 - Quieten some printks
 - Fix cpu online check in msr driver
 - Lower polling interval for machine checks to 5 minutes to avoid
   overrunning the registers. 
 - Make memcpy_{from,to}io more compatible with i386 (Roland Dreier)

sparse to warn about implicit typing.

The following patch makes the vram boot option actually be recognized
and its value assigned to the vram variable.
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

	netfilter set/getsockopt annotated.  That's a continuation of previous
patch - by now the only remaining place where setsockopt()/getsockopt()
userland pointers are not traced is vcc_[sg]etsockopt() and down into the
ATM drivers.  Everything from that area in net/* should be done by now.

	That's pretty much it as far as non-trivial splitting counts; remaining
patches in net/* are independent, tty and console stuff is also not hard to
split (not much of it, anyway) and what remains is arch-dependent code for
alpha and amd64.  Of course there'll be more stuff coming, but as far as I'm
concerned the rest of my pile is managable - and finally below 200Kb.  That'll
go tomorrow - right now I'm going down and getting some sleep...

This fixes sparse warnings for the user write method, and the UART
ioctl functions which copy data to/from userspace.

into ppc970.osdl.org:/home/torvalds/v2.6/linux

Patch from Nicolas Pitre

... since this is not supported at the moment and would fail to build.

Patch from Nicolas Pitre

Patch from Nicolas Pitre

Correcting my own screw-up, found by Ian Campbell.

into ppc970.osdl.org:/home/torvalds/v2.6/linux

Patch from Nicolas Pitre

	optval (and in case of getsockopt - optlen) made __user, changes
percolated down into the instances.

Virtually all the architecture specific memory.h files were defining
the same constants and v<->p translations the same way.  Move them to
asm/memory.h

	ethtool_ioctl() annotated; it would go in the first compat_ioctl
chunk (trivial ones) if it wouldn't be right after SIOCGIFCONF mess; was
easier to split in that order...

	Switched to compat_alloc_user_space(), killed kmalloc and set_fs(),
annotated, somewhat cleaned up.

	Similar to previous - annotated, killed helper, made the thing
type-clean, cleaned up.

	annotated, killed useless (and type-buggering - void ** is not
assignment-compatible with foo **) helper, cleaned up.

	Annotated, cleaned up, wholesale copying of structure with
subsequeunt casts from hell replaced with get_user()/put_user() on
the fields we really need.

	blkpg ioctls handling switched to compat_alloc_user_space() (instead
of messing with set_fs()) and annotated.

	Whoever had done the translation apparently never heard of
big-endian boxen.  Replaced the kludge with get_user()/put_user() on
64bit values.  Note that
struct raw_config_request
{
        int     raw_minor;
        __u64   block_major;
        __u64   block_minor;
};
has __u64 on _all_ platforms and it is used as an integer in raw.c code.
The only reason why translation is needed at all is alignment for 64bit
values; block_major and block_minor themselves should be passed as-is.

	Breakage fixed, entire area annotated.

	Easy parts of compat_ioctl fixes - pure __user annotation.
In one place (SG_IO handling) replaced dereference of user pointer
with put_user() (why whoever had noticed that just put a FIXME instead
of obvios fix is a mistery...)

a bunch of trivial ->read() and ->write() instances (and their procfs
relatives) __user-annotated.  All chunks are independent, but IMO not
worth splitting.

Independent minor bits caught by sparse:

 - paride.h mixing void and int in ? :, used always in a void context
   ide-iops.c return insw() - insw is void()
 - scsi/constants.c uses undefined macros in #if; added #define to 0 in
   case that used to leave it undefined
 - usb/host/hcd.h: fixed-point arithmetics in constant
 - fs/exec.c: missing UL on a large constant
 - fs/locks.c: #if where #ifdef should've been
 - fs.h: missing UL on MAX_LFS_FILESIZE in 64bit case

	fs/*, fs/proc/*, fs/nfsd/*, fs/autofs4/* 

The only interdependent group is the fs/autofs4/* stuff.

rest of pointers in sg ioctl arguments made __user.

scsi_ioctl() and scsi_ioctl_send_command() switched to __user *;
scsi/scsi_ioctl.c annotated.

scsi_cmd_ioctl() switched to __user *, block/scsi_ioctl.c annotated.

copying moved into sg_io()