Realize that *_empty_fifo() and *_fill_fifo() are always called
under the card lock, so no additional protection when reading/writing
fifos is needed.

Same thing which just happened to hscx_irq.c - Use the struct
bc_hw_ops instead of macros and inlining.

Each driver for HSCX based cards defined some hardware access macros
and #include'd hscx_irq.c. Instead, we now add the needed routines to
the corresponding bc_hw_ops and use regular function calls. It's a slight
bit less efficient, but shouldn't be noticable.

Using {isac,hscx}_{read,write} instead of the low-level
readreg/writereg simplifies the code a bit.

Preparing for un-inlining hscx_irq.c, we add operations needed
to access the B-Channel FIFO's, similar to what we already have
for the D-Channel.

Basically only a cosmetic patch, with much renaming and a little
rearranging of code.

Just a simple search and replace, switch to the more commonly
used "u8" type.

In many cases, readreg()/writereg() users can be simplified by
passing struct IsdnCardState instead of individual register
addresses.

Abstract the D-Channel access operations into a struct dc_hw_ops.

As a preparation for moving the B-Channel access functions into
an ops struct, introduce inline helper functions for using them.

The D-channel access functions are only used in w6692.c itself,
so no need to go through function pointers.

The D-channel access functions are only used in hfc_2bds0.c itself,
so no need to go through function pointers.

Instead of having the B-Channel callbacks directly in struct
IsdnCardState, use an struct bc_hw_ops.

As a preparation for moving the B-Channel access functions into
a ops struct, introduce inline helper functions for using them.

The callbacks to read/write B-Channel registers are only used in
w6692.c itself, so we can just call the appropriate functions
directly instead of going through the function pointers.

The B-Channel callbacks are only used when a common chipset driver
is used by multiple hardware drivers. Remove the dummy/unused
functions for the other cases.

Make hisax compile for CONFIG_ISAPNP not set.

The recent PnP API changes broke HiSax completely if CONFIG_ISAPNP is
set, now it at least compiles again w/o ISAPnP support.

sh_link is a 32-bit field on 64 bit archs, so a bad choice to store
module section offsets, and also INIT_OFFSET_MASK is out of range.

So switch to using sh_entsize instead.

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

arch/alpha/kernel/ to the config options that need them.  Fix
a few build problems for XLT and RX164.

By a suggestion from David Mosberger: use this new version of
compat_jiffies_to_clock_t, which is more appropriate on 64-bit hosts
that don't have overflow issues.

It does better rounding and will not fail if COMPAT_USER_HZ > HZ.

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

into tetrachloride.(none):/mnt/stuff/kernel/2.5/agpgart

(based upon 2.4 patch from Richard Baverstock <beaver@gto.net>)

Right now its the only VIA AGP 3.0 chipset. At some point in
time it may become via-agp3.c, but until then its cleaner to
seperate this from the VIA AGP 2.0 driver.

The 2.0 driver will still work on a KT400 with a 2.0 compliant graphic
card present. If the 2.0 driver finds the bridge in 3.0 mode, it aborts.
Likewise, the 3.0 driver will abort if the bridge is in 2.0 mode.

Confused yet?

Currently, only the K8 GART needs this, but this is cleaner than
exporting the agp_reserved_memory variable.

intel_agp_init() must not be declared static for explicit early
initialization to work (ie i810fb).