When sending a multicast and using looping back a copy to the
local machine, the interface filter checks can be done before the
source address is specified. For an INCLUDE filter, this won't match
the allowed sources and the packets won't be delivered locally,
even when the ultimate source address chosen is in the allowed list.

The patch below fixes the filter checks for both IGMPv3 and MLDv2
to only apply when a source address is available.

Thanks to Steven Hessing for reporting the problem and providing
a test case for reproducing it.

The code directly accessed the "cpucontrol" semaphore used
for CPU hotplug. That doesn't work all that well, since the
semaphore doesn't even exist on UP.

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

into debian.org:/usr/src/kernel/linux-2.6

From: NeilBrown <neilb@cse.unsw.edu.au>

If a raid array was syncing on shutdown, it would hang on shutdown,
constantly re-entering md_enter_safemade.  This fixes it.

slipped in.

I did a s/2.5/2.6/ a while ago, as it made more sense when 2.6 appeared.
The old URL will continue to work (symlink to the new file).  If I move
this again, whack me.

A lot of the blacklists never made it forward, here's what I found
still lying around in my old 2.5 tree when I brought it up to date.

I think 2.4 has had more updates since then (and there may be
some entries languishing in vendor 2.4 trees), I'll take a peek
when I get some spare cycles.

Yet another misplaced ! by the looks..

Looks like another instance of a ! in the wrong place.

Echoing changes done in 2.4. (It now has a pci_pool_create backport).

This currently prints out the maximum number of CPUs the
kernel is configured to support, instead of the actual
number that the kernel brought up. Which results in odd
displays that look like you have more CPUs than you do.

ARM_PACKET_SIZE is 4KB. Ouch.

Grep of the tree only turned up these two uses.

Negate the expression not the register seems more sensible?

If the CPU doesn't support MTRRs, don't create a /proc/mtrr

2KB onstack allocation. Nasty.

The centaur CPU init code gets linking errors without it.

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

sizes will not overflow pagebuf lengths.

SGI Modid: xfs-linux:xfs-kern:164827a

SGI Modid: xfs-linux:xfs-kern:164517a

From: Rusty Russell <rusty@rustcorp.com.au>

Looks like an obvious typo.  Works fine if "bio" is the name of the
iterator.

From: Benjamin Herrenschmidt <benh@kernel.crashing.org>

This patch (which has been in my tree for some time now) does 2 things to
the ppc32 signal code:

- The new sys_swapcontext() syscall that we added recently (and which is
  _not_ yet used by glibc, so it's ok to change it slightly at this point,
  glibc kernel version check will limit us to 2.6.2 or 2.6.3) gets a new
  context size argument, so we can deal with future context size changes.

- When ucontext is get/set/swapped using the above syscall, the TLS (r2)
  is preserved (it's still saved/restored on signal entry & return though).

The equivalent of this patch is already in the ppc64 signal32.c emulation,
and it has no effect until glibc is updated to use the new syscall, which
should happen soon now, so please apply.

From: "H. Peter Anvin" <hpa@transmeta.com>

Fix for RAID-6 on IA-64, from Bjorn Helgaas, and verified by me.

From: Chris Meadors <clubneon@hereintown.net>

Add a missing PCI ID.

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

Add back the old i82489DX bits to use timer scaling for the old
non-integrated APIC setup. 

It's possible these bits don't need to be set on i82489DXs,
but not having this HW for testing I elected to maintain
the old behaviour on these old machines.

The cpu handling in net/core/flow.c is complex: it tries to allocate
flow cache as each CPU comes up.  It might as well allocate them for
each possible CPU at boot.

We used to write fields that were marked RESERVED and that
are apparently some old stale timer base. Stop doing that.

Verified with Mikael Pettersson, and confirmed to fix ACPI
boot-time lockups for a few people. 

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

ppc64 doesn't use drivers/Kconfig (it should) so it needs to include i2c by
hand.

From: Anton Blanchard <anton@samba.org>

We were missing the sched_balance_exec call.  Could explain some NUMA
scheduling weirdness we were seeing.

From: "H. Peter Anvin" <hpa@zytor.com>

This is purely a doc patch saying RAID-6 support is available in
mdadm-1.5.0 and the patch is no longer necessary.

From: "H. Peter Anvin" <hpa@zytor.com>

As expected, when it hit mainline I started getting real bug reports... 
the attached patch does the following:

- Fixes a few x86-64 specific bugs;

- Removes MMX and SSE-1 from x86-64 (if we have x86-64 we have SSE-2);

- Slightly astracts the integer code to make it easier to add
  architecture-specific optimizations later (e.g.  IA64 once gcc gets
  better IA64 intrinsics support);

- Resurrects the user-space testbench, and makes it not output the known
  false positive of the D+Q case (D+Q is equivalent to a RAID-5 recovery,
  so I didn't implement it in the user-space testbench.)

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

RFC 2461 requires that the source address of Neighbor Discovery messages
is an address assigned to the sending interface.

Duplicate Address Detection should also be interface specific. We don't,
for example, want a node to DoS itself just because it has two interfaces
on the same link and both happen to listen to the same multicast group. If
there is a true duplicate on the link, the interface doing DAD will notice
it anyway.

The attached patch adds a 'strict' parameter to ip6_chk_addr() and
ip6_get_ifaddr() to allow link-local protocols like ND and DAD to do
strict address checks even on addresses with greater scope than
link-local.