WRT 2.4, Ed Vance put this fix into 2.4:

 According to the data sheet, the OX16PCI952 has port 0 in BAR0 and
 port 1 in BAR1. (unlike the OX16C954, BTW) The 921600 base baud value
 is your oscillator rate divided by the default oversample rate, 16.

Since 2.6 PCI IDs have been updated to use the real device ID for
this part, we really should include this fix as well.

If a serial port is hung up using vhangup() with HUPCL cleared (so
DTR isn't dropped), and then re-opened, we don't want to drop the
DTR line.  This occurs during the login process, and dropping DTR
would cause a modem to hang up.

I didn't yet use one of my AFAVLAB boards with 2.6.x until now.  The
upper 4 ports are not detected at all.  I suppose the bug was
introduced while porting the driver from 2.4.x.

Please consider applying the following patch.  It also adds support
for a new 8 port board called P030.

The current serial console code only works for ports that are either
defined in SERIAL_PORT_DFNS (and set up by serial8250_isa_init_ports())
or registered by early_serial_setup().

On ia64, SERIAL_PORT_DFNS is empty because we discover everything
via ACPI and PCI.  And we only use early_serial_setup() for one port
described by the HCDP firmware table.

This patch against 2.6.4-rc1 makes it work for any valid port.  If we
don't know about the port early, we just return -ENODEV from the
setup() function, which leaves the serial console disabled.  After the
driver has found all the ports, we try to register the serial console
again if it hasn't been enabled already.

Keith Owens noticed that the first version of this patch broke some
serial console setups because many early serial ports are registered
with "type == PORT_UNKNOWN".  So this version tests "port->ops"
instead, and Keith has confirmed that this works for him.

This fixes two errors in fixup_exception() for PnP BIOS faults:
 - Check for the correct segments used for the BIOS
 - Fix asm constraints so that EIP and ESP are properly reloaded

Free the request queues on the rd_init() error recovery path.

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

Noticed by me, fixed by Jens.

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

This fixes a typo in the list of PCI IDs in radeonfb, which in turn
fixes detection of some mobility models.

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

Patch from Deepak Saxena

The ARM makefile currently assumes that the compiler being used 
defaults to little-endian builds unless the big-endian options
are provided. This may not always be the case (my toolchains default
to BE since that's what I mostly work with) so this patch forces
endianess options to be passed to the toolchain. W/O this,
building without CONFIG_CPU_BIG_ENDIAN on a big-endian toolchain 
will still result in a big-endian kernel.

Patch from Ian Campbell

__arch_strncpy_from_user needs to be exported if you build the
framebuffer console driver as a module. 

Patch from Deepak Saxena

 - inode leak on d_alloc_root() failure

 - inode leak on d_alloc_root() failure
 - sanitized the cleanup logics

 - inode leak on d_alloc_root() failure

 - inode leak on d_alloc_root() failure

 - inode leak on d_alloc_root() failure

 - inode leak on d_alloc_root() failure

 - inode leak on d_alloc_root() failure

 - inode leak on d_alloc_root() failure

 - inode leak on d_alloc_root() failure
 - unchecked result of d_alloc_root() leading to oops in fs/super.c

 - unchecked result of d_alloc_root() leading to oops in fs/super.c

 - inode leak if d_alloc_root() fails.

 - oops if inode allocation fails.

 - useless dput(NULL) on failure exit (would be a double-free if
   we ever got there with non-NULL dentry).

 - leak of root inode if d_alloc_root() fails.

gatgetfs fill_super:
 - double-free of inode on last two failure exits
 - leak on the last failure exit
 - touching addresses near 0 on last two failure exits

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

New revision of the Sigmatel irda driver.  This version has much
better performance and doesn't drop frames in FIR mode.
Tested with both UHCI and EHCI/OHCI against nsc-ircc on laptop.
Thanks to Martin for additional testing and feedback.

* Receiver:
	- changed from interrupt to bulk URB.  Queue's one bulk
	  urb per USB tick (1ms).
	- FIR unpack now uses get_unaligned and cpu32_le
	  instead of explicit shifts
	- FIR unpack copies small frames
* Transmitter:
	- pack data into io buffer.  Since irda is half duplex
	  can use same buffer for transmit and receive.
	- use xchange and wait_event for synchronization
* Other:
	- use USB format messages rather than IRDA (ugly)
	- clean up USB startup
	- reset device on network open to clear up stuck state
	- handle FIFO errors
	- disconnect cleanup (from viro)
	- don't DMA onto stack when reading fifo status

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

It was missing an "s" from "uses_dvo"

This adds a "timbase" entry in /proc/cpuinfo like p/iSeries that
provides the CPU timebase frequency.  It is using by a all sort of
performance analysis tools we are hacking in house.

It also remove a useless bit about the l2 cache that was copied over
from ppc32.