This patch changes the I/O SAPIC code to cache the low 32 bits of the mask
word in kernel memory.  This greatly speeds up mask_irq() and unmask_irq().

Normally, these operations are not on the speed-critical path of the
kernel but with certain devices drivers (including users-level device-
drivers) they can become performance-critical.

        - fix bug in pfm_unload(), not allowed when not on correct CPU for
          system-wide
        - some perf/cleanup in overflow handler
        - fix reset_pmds to be on a per PMD basis on counter overflow rather
          than global
        - remove timing debug code on messages
        - no kernel info leak on PFM_END_MSG
        - remove double-store on reg_flags for pfm_write_pmcs, pfm_write_pmds
        - on restart reset_pmds is 0 by default
        - cleanup useless macros
        - cleanup some debug prints
        - added ability to remove debug code
        - streamlined sys_perfmonctl(), pfm_read_pmds(), pfm_write_*()
        - added current->tgid to default format sample header by using one
          reserved field

The problem was that IDE-disks on machines with IDE harddisks,
memory above 4GB and no hardware I/O TLB would go BUG_ON() in
blk_queue_bounce_limit() because the IDE-controller could only
address 4GB of memory and that was much less than BLK_BOUNCE_ISA
(which is equal to ISA_DMA_THRESHOLD).  Note that the problem does
NOT trigger with CD-ROMs, which always uses the software I/O TLB
(and hence bounce-buffers) on such machines.

The best fix seems to be to simply lower ISA_DMA_THRESHOLD to 4GB-1 since
that will ensure that the IDE block layer allocates memory with GFP_DMA,
which will minimize bounce buffering.

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

This patch fixes a bug which could cause do_softirq() to be called
at the wrong time (from do_IRQ()) or without pre-emption protection.

into kernel.bkbits.net:/home/davem/net-2.6

into nuts.davemloft.net:/disk1/BK/net-2.6

into ppc970.osdl.org:/home/torvalds/v2.5/linux

of redundancy, and to help prevent future errors.

This makes GregKH happy by removing the device name from the
device.bus_id field (and replacing it with a "name" sysfs attribute).

It also renames the parent device from "vdevice" to "vio", making the
/sys/bus and /sys/devices hierarchies consistent.

into ppc970.osdl.org:/home/torvalds/v2.5/linux

Add documentation for the above two functions.  Allow a NULL name to
be passed to amba_request_regions.  In this case, we use the driver's
name associated with the device.

This allows the decompressor to automatically detect the geometry of
the cache while booting to ensure correct behaviour over a wider
number of ARM CPUs.

This optimises the underlying string-long read function for little
endian platforms.

This gives greater information on the cause of the bad IRQ, allowing
the cause to be more effectively traced.

The old sysctl PM interface has been replaced by the driver models
interface.  This has been gone for a while in the -rmk tree and no
one complained, so there's no point keeping it around.

- use dev_dbg for device-centric debugging messages
- use pr_debug for general debugging messages
- use dmapools instead of pcipools
- use NULL rather than 0 for NULL pointers
- use enum dma_data_direction rather than int
- use DMA_* direction definitions rather than PCI_DMA_*
- only check for sane DMA direction on mapping functions, but
  check that DMA direction matches when unmapping/syncing.

Thanks to Adrian Bunk.

Remove kernel 2.0 #ifdef's from arm code, and move MODULE_* to
the end of the file.  Add an appropriate MODULE_LICENSE().

This patch makes it possible to select the hugetlbfs page size at boot-time,
rather than at compile-time.  We believe we have resolved all the remaining
issues.  All critical speed path has been taken care of, i.e., vhpt handler
and context switch.  There should be no performance penalty with this dynamic
hugetlb page size feature.

We would like to thank Jack Steiner for his initiative on this feature
and his initial cool working patch.

Currently, sinit_max_init_timeo is treated as the maximum timeout
value for INIT retransmissions and the INIT process is aborted when
the timeout reaches this value. But as per the SCTP sockets API draft,
sinit_max_init_timeo only limits the timeout. The INIT process should
be aborted only after MAX_INIT_RETRANSMITS.

Add an iseries defconfig and update the pseries one.

CONFIG_VETH isnt used anymore so we lets get rid of it.

Fix a compile error in the ppc64 pci hotplug code.

ppc64 tlb flush rework from Paul Mackerras

Instead of doing a double pass of the pagetables, we batch things
up in the pte flush routines and then shoot the batch down in
flush_tlb_pending.

Our page aging was broken, we never flushed entries out of the ppc64
hashtable. We now flush in ptep_test_and_clear_young.

A number of other things were fixed up in the process:

- change ppc64_tlb_batch to per cpu data
- remove some LPAR debug code
- be more careful with ioremap_mm inits
- clean up arch/ppc64/mm/init.c, create tlb.c

Set the ks bit on userspace segments otherwise the kernel can read/write
into userspace mprotected pages.

This allows us to use more than 2Gb of RAM on a PowerMac G5.  You can
also boot with iommu=off to disable it, but then you get back the 2Gb
RAM limitation.

Make IDE advertise only 128 entries of SG table on archs with an IOMMU.

The current IOMMU implementations of pci_map_sg() may produce segments
that don't match the boundary requirements of IDE, thus causing the
driver to break them up.  The BIO is supposed to account for that,
however, it cannot account for a pci_map_sg producing a segment of the
requested size, but with incorrect alignement, thus we may still try to
break up the list in more entries than is supported by the HW.

A similar fix already went in libata.  The "real" long term fix will be
to move the boundary requirements to struct device so that pci_map_sg()
can respect them when producing the sglist.  In the meantime, this
band-aid works around the problem.

Lots of things renamed, sillicaps killed, stuffs moved around and common
code properly extracted from implementation specific code, new
allocator, etc...  The code is overall a lot simpler, faster, less prone
to fail, and a lot more manageable.  I didn't use "bk mv", there is no
need to keep the old history attached to the new file.

This adds a hack to /dev/mem (along with the other ones already there)
to prevent mapping cacheable of the IO hole.

Without this, XFree blows up on machines with enough memory to go past
the IO hole.  It also tries to prevent memory from beeing mapped
uncached.  Cache paradoxes are evil and can kill the CPU.

The necessary page_is_ram() call was added by the previous patch doing
the proper IO hole accounting.

On machines with an IO hole (like Apple G5 with more than 2Gb of RAM,
but also some pSeries) we failed to properly account for the real
amount of physical RAM and inform the zone allocator of our hole size.

During the process, I included Anton slaughtering of the guard page we
had in the first 256Mb kernel segment, thus allowing this segment to be
mapped with large pages as it should be.

This removes the unused common_regs_m6 and add one more register to be
cleared, according to latest XFree code from ATI.

Fix a stupid warning in PowerMac SMP build on ppc64