Again, we don't need to sync indirects as we dirty them because
we run a commit if IS_SYNC(inode) prior to returning to the
caller of write(2).

Writing a 10 meg file in 0.1 meg chunks is sped up by, err,
a factor of fifty.  That's a best case.

At present, when mounted synchronously or with `chattr +S' in effect,
ext2 syncs the indirect blocks for every new block when extending a
file.

This is not necessary, because a sync is performed on the way out of
generic_file_write().  This will pick up all necessary data from
inode->i_dirty_buffers and inode->i_dirty_data_buffers, and is
sufficient.

The patch removes all the syncing of indirect blocks.

On a non-write-caching scsi disk, an untar of the util-linux tarball
runs three times faster.  Writing a 100 megabyte file in one megabyte
chunks speeds up ten times.

The patch also removes the intermediate indirect block syncing on the
truncate() path.  Instead, we sync the indirects at a single place, via
inode->i_dirty_buffers.  This not only means that the writes (may)
cluster better.  It means that we perform much, much less actual I/O
during truncate, because most or all of the indirects will no longer be
needed for the file, and will be invalidated.

fsync() and msync() still work correctly.  One side effect of this
patch is that VM-initiated writepage() against a file hole will no
longer block on writeout of indirect blocks.  This is good.

problem.

In ext3_new_block() we increment i_blocks early, so the
quota operation can be performed outside lock_super().
But if the block allocation ends up failing, we forget to
undo the allocation.  

This is not a serious bug, and probably does not warrant
an upgrade for production machines.  Its effects are:

1) errors are generated from e2fsck and

2) users could appear to be over quota when they really aren't.

The patch undoes the accounting operation if the allocation
ends up failing.

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

This simple check was first suggested by Andrew Morton.  Pretty basic -
whines if a task exits with a nonzero preempt_count value.

I put an identical check in the 2.4 preempt-kernel patch and - sure
enough - it was found that XFS essentially disables preemption as it
destroys data structures containing locks without first unlocking.  The
SGI folks are working on that.

Anyhow, its a quick and clean solution to debugging potential problems.
Patch is against 2.5.7, please apply.

	Robert Love

Hopefully this is all of them..

fix the build on ia32.

Author: Dave Miller

dev->rmem_{start,end} in skfp and smctr drivers.

Updated the CREDITS, Config.help, Config.in, and Makefile to be up to date
with the last round of USB changes.

  
   - An oopsable bug affecting unlink of interrupt
     transfers. Fix mirrors one done ages ago for ISO.
     (Original patch by Matt Hughes)
   - Better cleanup on init failure (Matthew Frederickson)
   - fixes the problem Stuart reported, where interrupt urbs
     couldn't be unlinked from their completion handlers, and it
     also makes OHCI return the correct status code for async
     unlink requests (-EINPROGRESS not zero).

added safe_serial driver
tweaks to the driver done by greg@kroah.com to get things to work on 2.5

added tiusb driver
some tweaks to the driver done by greg@kroah.com

The patch ensures that uhci.c doesn't use urb->status after the
completion callback if it doesn't need to.

fix the "small packet" problem and debug messages cleanup

removed CONFIG_USB_HIDDEV #ifdefs in the driver.

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

one more step closer to killing off ether=

once eth0 was found) - it is long since anyone shipped or built kernels
with all the ISA drivers compiled in.

This change will eliminate the need for adding "ether=...." at the boot 
prompt for a lot of users who build their own kernels and have multiple
ISA ethercards at standard (i.e. probed) I/O addresses.

Also got sick of counting zeros, so did a struct init cleanup,
i.e. {1,0,0,0,0,0,0,0,9}  -> {one:1, nine:9}

based net cards and hence these should be moved into the dev->priv
for these cards.

This patch adds rmem_start and rmem_end to dev->priv in 8390.h, and does:

        s/dev->rmem_/ei_local.rmem_/g   on all 8390 shared mem drivers.

the rmem_end field to store an I/O port address in.  As rmem_end is going
away, this does a s/rmem_end/mem_start/g since mem_start is otherwise 
unused in sb1000.

(probably as a result of copying skeleton or similar).  Removed this as 
a step in the goal to remove rmem_start/end from netdev struct entirely.

aironet4500, arlan, e2100, baycom, soundmodem

into mandrakesoft.com:/home/jgarzik/repo/net-drivers-2.5

* region resource handling
* do-while macro definitions
* error cleanup

dev->last_rx equal to jiffies.

Whitespace cleanup

Update a few stale comments

Bump version to 4.2.8-k1, use EXPORT_NO_SYMBOLS, minor changes
to ensure that reported link information is always correct, 
remove magic numbers in calls to memset.

ProcFS code updates, check for page boundaries.

change_mtu cleanup.  Allows frame sizes up to 2k on the 82542
instead of limiting the MTU to 1500, so that the 82542 can be 
used with software 802.1q VLANs.

Minor receive cleanup, queue empty buffers to the hardware in 
groups of 16 to reduce unneeded fetches and improve PCI 
efficiency.

Adaptive Inter-Frame Spacing to reduce collisions and improve 
half duplex transmit performance.