I have encountered a bug in the nicstar ATM driver for linux.
If you open a CBR TX only connection on a specific vpi/vci and later open a RX 
only connection on the same vpi/vci, the RX connection will overwrite the 
pointer to the SCQ of the TX connection. This changes the cell rate of the TX 
channel and what is worse is that when the TX connection is closed we get a 
segmentationfault and the TX part of the vpi/vci remains reserved.

The bug in the driver is that if the opened channel is not TX CBR, the driver 
assumes it is TX UBR. I have attached a patch that adds a check for TX UBR. 
The patch is against RedHat kernel 2.4.18-3. I have checked linux vanilla 
kernels 2.4.19 and 2.5.49 and not found a fix.

Merged from 2.4.x.

Syncing this driver with the 2.4.x version.

init_etherdev zeroes this memory for us, no need to do it again.
Spotted by Dave Jones.

Use of init_etherdev's second argument causes the ->priv member
to be allocated at the same time as the struct net_device
itself.  Therefore, no additional kfree() is needed for the
struct net_device ->priv member in this case.

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

After taking a second look, i just recognized that both cases (MAC adress
all-zero or not) are handled exactly the same (by duplicated code), so the
whole stuff is unnecessary.

The whole function just reduces to a simple copy loop:

This problem was identified by Olaf Kirch:

In xprt_request_init(), the XID allocation needs to be protected by a
global spinlock.

 - Only set up pipefs entries for those RPC services that actually
   need them (for the moment NFS only). Portmap, lockd,... shouldn't
   need to make upcalls.

 - Add in missing semaphore in rpc_populate().

 - Make inode/dentry variable names in rpc_depopulate/rpc_populate
   more consistent w.r.t other functions in rpc_pipe.c

 - Call shrink_dcache_parent() in order to clean up child entries
   before we rmdir().

This patch fix es the signals vs strace problem.  The signal code
assumes that anybody who is TAK_STOPPED will check signals before
resuming, and the ptrace paths didn't do that.

This is a rather large update for the starfire network driver,
implementing VLAN support, 64-bit dma_addr_t support, and NAPI support. It
also fixes a couple of show-stopper bugs in the old driver which were
biting real people out there.

I've had a few positive test results with this version, including one from
Martin Bligh who tested it on his monster SMP boxes. So I'm pretty
confident that it's mostly all right, and certainly better than what's
currently in the tree.

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

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

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

- It depends on the rest of the subsys_initcalls in acpi to be done to get
  the sleep info from the firmware. 
- It doesn't need to be available any earlier...

- Make acpi_sleep_proc_init() a late_initcall(), and not called from 
  acpi_sleep_init(). This guarantees that the acpi proc hierarchy is at 
  least there when we create the dang file.

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

- Created arch/i386/kernel/acpi/
- Split file into boot.c and sleep.c.
- Moved acpi_wakeup.S into there.

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

This converts the driver to use the new scsi_add_host() interface to add
scsi controllers.

This removes the old LinuxVersionCode() and replaces it with
KERNEL_VERSION().  Thanks go to Adrian Bunk for sending the initial
version.

This adds a few bits of code and couple #ifdefs that are needed to
enable this driver to work for 2.4 as well as 2.5.

This adds back the reboot_notifier hook that was accidentally removed in
the last set of patches.

It also adds support handle a Mode Page 8 command so the scsi layer
won't complain anymore.  And the movement to using ->device to get scsi
numbers resulted in 2 typos in debug code.  These are also fixed.

Lingering around for a while. 2.4 forward port.

The new C3s won't boot a C3 kernel as they dropped 3dnow support in
favour of SSE.  It now also has cmov though and can be scheduled as a
686 CPU.

I've a patch for gcc pending inclusion that adds the -march=c3-2 option.

Forgot to bring this bit forward when I updated the watchdog drivers.

Fixes from 2.4 + compile fixes from me.

This one has been around for ages, has Rusty's name
all over it, but wasn't from him.
48 lines of comments, and no swearing? 8-)

Mostly documentation, but also has some buffer overflow
detection in there.

from 2.4

adap_lock is a little generic for a global..

This went into 2.4.18

Whoops, VIA goofed, and it reports L2 size as 65KB.

If we have a non-SMP approved CPU, but its in a UP system,
we should allow SMP kernels.

from 2.4
only register gameport if the port is active

from 2.4
only register gameport if the port is active

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