The needed fix winds up breaking SG, checksumming, and other stuff
in the process.

Ouch.

* Dynamic interrupt coalescing
* mii lib support
* dynamic IPG support (disabled by default)
* jumbo frame fix
* vlan fix
* rx irq coalescing fix

into redhat.com:/garz/repo/net-drivers-2.5

in 8139too and pci-skeleton drivers.

* use pci_{request,release}_regions for PCI devices
* use alloc_etherdev (fixes race)
* propagate error returns from pci_xxx function errors

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

 - JFFS2 bugfixes and performance improvements
 - Support for 64-bit flash arrangements
 - Optimise for linear mappings of flash, without out-of-line access functions
 - New map drivers
 - Updated NAND flash support, new board drivers
 - Support for DiskOnChip Millennium Plus and INFTL translation layer
 - Clean up all translation layers with a single blkdev helper library.
 - Fix races in MTD device registration/deregistration
 - Add support for new flash chips
 - Clean up partition parsing code

More detailed comments in per-file changelogs.

Anybody who still thinks K&R makes sense should just be shot.

into flint.arm.linux.org.uk:/usr/src/linux-bk-2.5/linux-2.5-pcmcia

socket->base is unmapped in yenta_close(), which is called by
cardbus_remove().  The value of socket->base is not changed to
NULL, so it becomes invalid.

Then cardbus_remove() calls class_device_unregister(), which calls
pcmcia_unregister_socket(), which it turn tries to access memory
space of the socket.

Restore 2.4 behavior when setting the single step breakpoints.

Make the "addr" arg to test_bit "const" to prevent flood of compile
warnings in networking code.

This adds a cache of lastest accessed cluster to inode for sequential
access.

The following is 500M file of FAT-to-FAT copy test, this may be a most
different case in usual operations, because maximum readahead window
flush the all caches.

512 bytes blocksize, 4096 bytes cluster size.

linux-2.5.69-bk12
	root@devron (a)[1232]# time cp file file1

	real    7m58.900s
	user    0m0.267s
	sys     6m44.258s

linux-2.5.69-bk12+patch
	root@devron (a)[1576]# time cp file file1

	real    2m44.309s
	user    0m0.270s
	sys     0m28.631s

This shifts the data position caches from module to per-superblock, and
cleanups.

This removes the obsolete Documentation/filesystems/fat_cvf.txt.

This fixes the return value of ioctl() for enables using the same way as
readdir().

put/get_user() return code check patch from John R R Leavitt
<jrrl@steampunk.com>

This adds large partition (> 128GB) support to FAT.

This silences at least one compile-time warning... :-)

This is needed to match the output of the C compiler.

This header ends up getting included by uClibc (though nothing in it is
used), so this protection is necessary to avoid problems with kernel-only
typedefs.

This is to define `struct file'; apparently some include-file change
removed a previous implicit include.

These registers are now saved in a difference place, but the old code
was inadvertently left in.

	fs/inode.c assumes that any ->delete_inode() will call clear_inode().
procfs instance doesn't.  It had passed unpunished for a while; cdev changes
combined with ALSA creating character devices in procfs made it fatal.

	Patch follows.  It had fixed ALSA-triggered memory corruption here -
what happens in vanilla 2.5.70 is that clear_inode() is not called when
procfs character device inodes are freed.  That leaves a freed inode on
a cyclic list, with obvious unpleasantness following when we try to traverse
it (e.g. when unregistering a device).

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

This allows drivers to resize their tag depth at run-time.

Argh.  Missing initialization in char_dev.c - it's definitely
responsible for crap on unload.  Load side appears to be something else,
though...

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