If you perform the following commands in order:

 # cardctl eject
 # rmmod yenta_socket
 # insmod drivers/pcmcia/yenta_socket.ko
 # killall cardmgr

the rmmod ends up freeing the pcmcia_bus_socket while the wait
queue is still active.  The killall cardmgr cases the the select()
to complete, and users to be removed from the "queue" - which ends
up writing to freed memory.

The following patch adds refcounting to pcmcia_bus_socket so we
won't free it until all users have gone.  We also add "SOCKET_DEAD"
to mark the condition where the socket is no longer present in the
system.

Note that we don't wake up cardmgr when we remove sockets -
unfortunately cardmgr doesn't like receiving errors from read().
Really, cardmgr should treat EIO from read() as a fatal error
for that socket, and stop listening for events from it.

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

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

Thanks to DaveM.

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

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

Obviously almost noone uses the pdeath_signal feature, since this has gone
unnoticed for quite some time.

This patch calls the function that does the right locking for the context
of this call (inside exit_notify).  The names of the signal.c entrypoints
are a little confusing.

The spinlock was held while calling functions that could block,
while simultaneously being at all times inside the context of
module init/exit.

Thanks to DaveM.

	correctly even on 32-bit hosts. As an added bonus, it's faster,
	too.  Run "unwcheck" by default, but for now, don't let unwcheck
	errors cause the kernel build to fail.

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

This patch fixes DMA bugs on x86-64 and ia64 machines.  The driver was
using commands that only support 32bit addresses in places that could
return 64bit addresses.  One place was DMAing off the stack.  The other
place was causing problems on x86-64 machines by calling pci_map()
functions on a region allocated by pci_alloc_consistent().

This patch fixes two minor bugs.  It allows zero length write commands
through to devices.  It also prevents the writing of any  '\0'
characters at the end of version numbers to ips's /proc/scsi files.

From: "Andrew Vasquez" <andrew.vasquez@qlogic.com>

With additional changes from: "James Bottomley" <James.Bottomley@SteelEye.com>,
                              "Christoph Hellwig" <hch@infradead.org>

This is the qlogic driver version 8.00.00b7 with the ioctl and failover code stripped
out and a few associated fixes put in.

From: "Randy.Dunlap" <rddunlap@osdl.org>

The problem is that the detect function requests an IO region
of 16 bytes (at least when a command line override parameter is
used) but the release function only tries to release 8 bytes,
and this request isn't done because it doesn't match any allocated
IO region.  [NCR53C400 extensions are not enabled, so
NCR5380_region_size is 8, not 16, but the request uses
NCR5380_region_size regardless.]

Fix: save the allocated region size in instance->n_io_ports and release
   that size only;

Fixes build on some platforms.

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

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

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

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

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

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

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

  This was broken by the last commit for cdev stuff which removed the
  "int ret;" that these routines used.

  Retire our char device dispatching logic.  With the 2.6 cdev API we
  can register much smaller device number regions, so we use that instead.

On Tue, Jan 13, 2004 at 06:02:08PM +0300, Sergey Tikhonov wrote:
> The final kernel could not pass linking stage with 2.1.18f version
> (included into 2.6.1 kernel patch).
> The linker complains that :
> drivers/built-in.o(.init.text+0x8cec): In function 'sym2_probe':
> : undfined reference to 'local symbols in discarded section .exit.text'
>
> I compared arch/alpha/kernel/vmlinux.lds.S with one from the i386 tree.
> The '.exit.text' and '.exit.data'
> sections were moved out of DISCARD attributes in the i386 version. I did
> the same with alpha version
> and it helped.

Thanks for this report; you uncovered a real bug.  Actually two real
bugs; one is that parisc is not discarding the .init.text and .exit.text
sections (so I didn't notice this bug) and the other is that sym_detach
is marked __devexit, yet called from a __devinit function.

This patch gets rid of some superfluous printk's which get tedious on
machines with insane number of CPUs.  The virtual/physical address bit
info is already available via /proc/pal, so there is no need to print
it at boot- time.

Currently you can select HCDP independent of serial console.  This
doesn't seem very useful, and makes the setup_serial_hcdp() reference
unresolved if the serial driver is built as a module.

CONFIG_SERIAL_8250_CONSOLE is only selectable if the serial core is
built in (SERIAL_8250=y), so this patch makes sure we don't try to
call setup_serial_hcdp() unless it is actually built in to the kernel.

(I think we should also make HCDP selection dependent on
SERIAL_8250_CONSOLE=y in Kconfig; I'll send a separate patch
for that since it's not ia64-specific.)

      modular ACPI device drivers.

Fix system type selection to workaround menuconfig bug (select "HP",
get "HP-simulator").

I fixed the code to compile with CONFIG_ACPI_NUMA and
!CONFIG_DISCONTIG since it's actually possible to compile that
configuration as well as cleaned up the #include usage a litte.

Uses the new module_params that nobody else uses now.