The check for "write was incomplete" is not atomic with the actual write,
so a second write could slip in after a partial write to a tcp socket
and this will cause problems.

So we move the locking out a bit, and test if the socket has
been deleted (SK_DEAD) between getting the lock and trying
to send.

SOCK_NOSPACE is set, so we must set it when we want to get
a ->write_space callback.

From: Hirokazu Takahashi <taka@valinux.co.jp>

sock_wspace() is used to see how much can be written to a udp socket,
but tcp_wspace must be used on a tcp socket.
This patch informs sunrpc/svcsock.c of this subtlety.

I noticed that the Nagle algorithm was disabled just recently on the
client side, while it still seems to be enabled on the server side. The
previous patch came from Chuck Lever.

Below patch disables it on the server side as well. For latency reasons,
this should be the desired behaviour NFS at both client and server.

Below patch resolves a hang where a TCP nfsd connection will hang even
though new data is received on the socket. We've seen this a few times in
our lab, but it usually happened every few weeks.

If a short record header is received, the SK_BUSY flag is never cleared,
and even though new data arrives, it will not be handled. This in turn
leads to hangs of particular clients (while others will continue to work
without problem).

I also changed the return code for that condition to be the same as for a
(regular) short read.

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

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

Add a vdbg macro for verbose debugging, and convert some
noisy debugging statements to use it.

Make the most discriminating comparison first.

The stats field is never NULL.

Replace spin_lock_irqsave/spin_unlock_irqrestore with
spin_lock_irq/spin_unlock_irq in tasklet actions, since
these are always called with local irqs enabled.

Replace spin_lock_irqsave/spin_unlock_irqrestore with
spin_lock_irq/spin_unlock_irq in routines that are only
called in process context.

Should have gone long ago.

Defend against future maintainers.

Use set_current_state (TASK_RUNNING); schedule(); instead.

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

It looks like the do_fork was converted in voyager_smp.c, but the
addition of wake_up_forked_process() was missed leading to a boot
panic.  The attached fixes it.

This patch touches the most significant places in usbcore
that need to respond to configuration change events:

  - set_interface changes altsettings; affects endpoints
    in both new and old settings (but not other endpoints).

  - set_configuration; affects every endpoint

The endpoints get disabled, flushing hardware state
reliably.  This resolves two FIXMEs here, part of one
in OHCI (ed_get) and all of one in EHCI (qh_append_tds).

Not many drivers currently use these code paths, but
now they should start to work properly.

Stefano Barbato wrote:
> Dual PIII
> kernel 2.4.21-rc2 (w/ SMP)  (2.5.69 below)
> ...
>
> I put a few printk before the BUG() and I found that the offending if() is
> this:
>         if(qh->qh_state != QH_STATE_LINKED
>                                 && qh->qh_state != QH_STATE_UNLINK_WAIT)
>
> because qh_state were QH_STATE_COMPLETING.

I got a similar SMP report recently, but without info about
which clause was failing -- which is a key clue, thanks!!

The COMPLETING state is used only while a QH is being
scanned for completed TDs.  (Think CPU-0 irq handler.)
Looking at the handful of places that call the routine
reporting the BUG(), a couple seem like they could make
trouble with multiple CPUs in the driver.

Alan Stern wrote:
> I suggest you just forget about acquiring the lock in status_dequeue() and
> simply leave it as
>
> 	del_timer_sync (&hcd->rh_timer);
> 	hcd->rh_timer.data = 0;

Hmm, so if some other URB gets queued in that window,
it'll get trashed?  Unlikely .. the clean fix would be
making the status endpoint have a real URB queue.

I combined your suggested change with two others:
(a) protect the status-unlink and control completion
    handlers against IRQs [ the cases Duncan noted]
(b) use mod_timer to retrigger the timer, instead of the
    heavy weight path.

From Marc Zyngier <mzyngier@freesurf.fr>.

into kanga.twiddle.net:/home/rth/work/linux/axp-2.5

From Ivan Kokshaysky <ink@jurassic.park.msu.ru>.

into kroah.com:/home/linux/linux/BK/gregkh-2.5

Update ppc for mingo's do_fork changes.  Together with the pmac compile
fixes this allows me to stay at the bleeding edge with my my ibook..

This updates ppc64 for the do_fork() semantics change.

I completely misread this code and there was no bug.  In fact yesterday's fix
broke it.  Revert.

struct page is undefined in this file without this on Alpha.

From: Andi Kleen <ak@suse.de>

There is some assembly code in setup.c which doesn't explicitly set its
section.  It comes after a C function which is marked __init.  The compiler
does not reset the section after such C functions.  The assembly code is
going into .text.init.

That's the wrong section.  We get oopses modprobing sctp and, with Manfred's
unmap-free-pages debug patch we get oopses modprobing e100.ko.  In
apply_alternatives().

into nuts.ninka.net:/home/davem/src/BK/net-2.5

into nuts.ninka.net:/home/davem/src/BK/net-2.5

into nuts.ninka.net:/home/davem/src/BK/net-2.5