This patch fixes the following issues:

- fix a nasty corner case by which a task could inherit psr.pp/psr.up
  that could be stale by the time it runs. psr.pp/psr.up bits must be
  cleared on fork and inherited on execve(). Test case provided by
  John Levon.

- mask interrupts when locking the pfm stat structure. We don't want
  to take PMU interrupts in the middle of this kind of critical
  section.

- fix a bug in pfm_close() for context which use asynchronous
  notifications.  You cannot call fasync_helper() with interrupts
  disabled because it reenables them. There was no real danger of
  moving the call outside our critical section.

- fix a bug in in pfm_flush_pmds() where you can leave the freeze bit
  set in the pmc saved state.

- updated dump_pmu_state()

- cleanup the BUG_ON() usage

his seems to be broken for ia32 on ia64 and possibly other emulation
architectures as well. The problem comes from
fs/namei.c:set_fs_altroot() using path_lookup() instead of a
path_walk() relative to "/" (which is the case for 2.4).

The drivers/char/mem.c change was accepted, so here's the
ia64-specific part.

gas will match anything with a prefix
'.data.' as a data special section.  Thus to stop unnecessary warning
about changing attributes, we use the .data section standard
attributes aw when creating gate-data.o.

I stumbled on a couple trivial bugs in ia64 numa/discontig support.
The first just sets the default number of nodes to something reasonable
for a generic kernel, otherwise it's really easy to start walking over
your initdata (more error checking should probably be added).  The
second fixes a memcpy to a physical address.

Without this patch, if a signal handler tried to access TLS data (via %gs), things break, because the GS descriptor is zero.

To be compatible with i386, we shouldn't be touching the segment descriptors
before getting into signal handlers.

	was inconsistent and in places still assumed there is both a GDT and a TSS
	shared page, but the latter was removed a long time ago).

	underflows.  In my case, the underflows occurred with the
	first memory descriptor which got trimmed down to a size of 0.
	Due to the underflow, this descriptor ended up covering the entire
	address-range which in turn caused Bad Things to happen with the
	X server.

- converts hdr_cur_pos into hdr_cur_offs to avoid leaking kernel
  addresses to userland. As a consequence hdr_last_pos is gone.

- correct buffer saturation mode bug by which the hdr_count would
  be reset, leading to no visible samples collected.

Newer versions of strace manipulate the syscall arguments and to make this
work for ia32 processes, we need to reload the syscall args after
doing the syscall-trace callback.

	enable_irq() sequences).

Tony pointed out (thanks Tony) that in -test8 kobject_set_name() will
try to free a kobject's k_name field if it's non-NULL, so we need to
zero it out in case kmalloc() gave us recycled memory or we'll try to
kfree a bogus area.

into tiger.hpl.hp.com:/data1/bk/lia64/to-linus-2.5

This fixes a problem in EFI memory map trimming. For example,
here's part of the memory map on my i2000:

    mem00: type=4, attr=0x9, range=[0x0000000000000000-0x0000000000001000) (0MB)
    mem01: type=7, attr=0x9, range=[0x0000000000001000-0x0000000000088000) (0MB)
    mem02: type=4, attr=0x9, range=[0x0000000000088000-0x00000000000a0000) (0MB)
    mem03: type=5, attr=0x8000000000000009, range=[0x00000000000c0000-0x0000000000100000) (0MB)
    mem04: type=7, attr=0x9, range=[0x0000000000100000-0x0000000004000000) (63MB)
    mem05: type=2, attr=0x9, range=[0x0000000004000000-0x00000000049ba000) (9MB)
    mem06: type=7, attr=0x9, range=[0x00000000049ba000-0x000000007ec0b000) (1954MB)
    ...

There's a hole at 0xa0000-0xc0000, so we should ignore all the WB memory
in that granule.  With 16MB granules, the existing code trims like this
(note the 4K page at 0x0 should have been ignored, but wasn't).

Populate topology directories correctly now that NUMA kernels work.

They were added by Torben to workaround ServerWorks driver problems
(fixed by previous patch), but depending on BIOS can be dangerous on
other chipsets and it is always better to fix specific driver.

Removal of these options was acked by Torben.

If PIO mode should be auto-tuned xferspeed argument for svwks_tune_chipset()
is equal to 255 (0xFF).  It is then passed to ide_rate_filter() which matches
desired mode with chipset capabilities.  Since 255 is greater than any of the
values used for transfer modes, ide_rate_filter() will always return the
highest mode supported by both device/chipset (which sometimes should not be
used ie. when host is a simplex one) and the wrong mode will be set.

Thanks to Torben Mathiasen and Tomas Szepe for testing it.

The original change turned out to be too restrictive, it catches devices
that don't need the work around.

This fixes that, thanks to Bart for informing me of the screwup.

into kernel.bkbits.net:/home/davem/net-2.5