In get_sectorsize(), we can already sleep in scsi_wait_req().
In get_capabilities(), we can use GFP_KERNEL because it is only called
from sr_attach() which already does a GFP_KERNEL allocation.

This GFP_DMA can be GFP_KERNEL because sd_init_onedisk calls sd_spinup_disk
calls scsi_wait_req which can sleep.

dito

get it to actually compile cleanly again, switch to scsi_add_host,
remove host list walking.

This includes the following items:

 - a new wakeup feature: SMART_WAKE_CHILD (now unconditional) [this is a
   reworked version of the wakeup changes from Andrea's tree.]

 - scheduler parameter tunings (CHILD_PENALTY, MAX_TIMESLICE,
   STARVATION_LIMIT, MAX_SLEEP_AVG)

 - activate_task() splitup to separate the interactivity-active variant
   from the task-queue variant => this fixes our statistics in some cases.

 - finer-grained preemption: if priority is equal then the task with the
   longer pending timeslice wins.

 - make it possible to disable sync-wakeups via SYNC_WAKEUPS (default:1)

 - a naming cleanup (sleep_timestamp => last_run)

 - fix up STARVATION_LIMIT usage so that a value of 0 is recognized as 'no
   starvation limit'.

It also merges Linus' "priority boost backmerging"

that is actually showing interactive behaviour (sleeping) and the
process that wakes up the interative process get a bonus.

This means that processes that help interactive processes (usually
system daemons, and X11 in particular) will themselves get some bonus.

The waker only gets whatever is left over from the "real" interactive
user, to avoid any bad behaviour.

into home.transmeta.com:/home/torvalds/v2.5/linux

pcibios_claim_console_setup() doesn't look behind PCI-PCI bridges.

Noticed by Anton Blanchard.

Ivan.

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

(While "signed long" is the same in practice, it is technically
incorrect, and the new anal type-checker complains).

into home.transmeta.com:/home/torvalds/v2.5/linux

Use ->dev instead of ->intf.dev in struct cpufreq_policy. Introduced whilst
cleaning up part of the cpufreq code. 

During startup, sysfs won't succeed in registering with itself, because
it's not yet inititialized. This was causing an unnecessary error on boot.

- Make it compare the name in struct platform_dev to the driver's name.

From Dominik Brodowski

Credits also to Mitsuru Kanda <kanda@karaba.org>,
YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>,
and Kunihiro Ishiguro.

into nuts.ninka.net:/home/davem/src/BK/net-2.5

The below has been in -mm for a while, and has been tested on my UP
and 2-way machines.

OProfile was completely unsafe on UP - a spinlock is no protection
against the NMI arriving and putting data into the buffer. Pretty stupid
bug. This fixes it by implementing reader/writer windows in the buffer
and removing the lock altogether. This patch was originally done by Will
Cohen.

It also fixes the oops Dave Hansen saw on 2.5.62 SMP

I had my list empty checks reversed if aborting and bus device reset
failed.  The condition that causes the error handler to run is still
unknown.

Now that the agpgart support is in, this can be enabled.

/jeff

into nuts.ninka.net:/home/davem/src/BK/sparc-2.5

Many places use inappropriate types for comparing jiffies,
int for example.  Fix those places up.

 - Cleanup Makefile.
 - Fix potentional crash in dv1394 module unload.
 - Why does ARM's memcpy not return void*? Oh well, work around it in
  eth1394 by not expecting standard behavior.

into penguin.transmeta.com:/home/penguin/torvalds/repositories/kernel/linux

 - Inode and dcache Hash table only needs half the memory/cache because
   of using hlists.
 - Simplify dcache-rcu code.  With NULL end markers in the hlists
   is_bucket is not needed anymore.  Also the list walking code
   generates better code on x86 now because it doesn't need to dedicate
   a register for the list head.
 - Reorganize struct dentry to be more cache friendly.  All the state
   accessed for the hash walk is in one chunk now together with the
   inline name (all at the end)
 - Add prefetching for all the list walks.  Old hash lookup code didn't
   use it.
 - Some other minor cleanup.

into home.transmeta.com:/home/torvalds/v2.5/linux

into osdl.org:/home/mochel/src/kernel/devel/linux-2.5-core

Originally from Nikita Danliov. 

- Add struct subsystem to struct file_system_type. 
- Add fs_subsys_init() to explicitly call sysfs_init() very early.
- Register base fs subsystem then, too.
- Register filesystems' embedded subsystem when registered with fs core.

Resulting tree looks like:

# tree /sys/fs/
/sys/fs/
|-- bdev
|-- devpts
|-- eventpollfs
|-- ext2
|-- ext3
|-- futexfs
|-- iso9660
|-- pipefs
|-- proc
|-- ramfs
|-- rootfs
|-- sockfs
|-- sysfs
`-- tmpfs

- For matching registered platform devices with platform drivers.

- Also, make sure platform_bus is initialized early, with the rest of the 
  core stuff.

This ensures that sysfs is initialized very early, before anything can 
possibily registered with it. 

It happens early enough that all filesystems can registered with sysfs 
(besides sysfs itself), which is coming next.

- Call driver_init() from init/main.c::do_basic_setup(). 

This ensures that all the driver model subsystems are initialized before
any drivers or devices can be registered. 

It nearly frees up the core and postcore initcall levels, making them 
available for other kernel code to use freely.