[PATCH] ppc64: change the iSeries virtual device drivers to use the vio infrastructure for DMA mapping

From: Stephen Rothwell <sfr@canb.auug.org.au>

This patch changes the iSeries virtual device drivers to use the
vio infrastructure for DMA mapping instead of the PCI infrastructure.
This is a step along the way to integrating them correctly into the
driver model.

From: Stephen Rothwell <sfr@canb.auug.org.au>

This patch consolidates some of the iommu DMA mapping routines.

From: Stephen Rothwell <sfr@canb.auug.org.au>

This is just a cleanup to use enum dma_data_direction for all APIs
except the pci_dma_ ones (since they are defined generically).

Also make most of the functions in arch/ppc64/kernel/pci_iommu.c
static.

From: Stephen Rothwell <sfr@canb.auug.org.au>

This patch uses enum dma_data_direction for the vio DMA api routines.
This allows us to remove some include of linux/pci.h.

Also missed some pci_dma_mapping_error uses.

From: Anton Blanchard <anton@samba.org>

Register secondary threads in NUMA init code

From: Anton Blanchard <anton@samba.org>

Add HW PMC support to oprofile

From: Anton Blanchard <anton@samba.org>

Add PMCs to sysfs.

From: Anton Blanchard <anton@samba.org>

Add some POWER5 specific optimisations:
- icache is coherent, no need to explicitly flush
- tlbie lock no longer required

From: Anton Blanchard <anton@samba.org>

Move sysfs specific stuff into sysfs.c

From: Anton Blanchard <anton@samba.org>

Update CPU features. Remove DABR feature, all cpus have it. Add MMCRA,
PMC8, SMT, COHERENT_ICACHE, LOCKLESS_TLBIE features

From: David Engebretsen <engebret@us.ibm.com>

Put SMT threads into global interrupt queue

From: Anton Blanchard <anton@samba.org>

Create xics get_irq_server and use it in enable/disable code.

From: Paul Mackerras <paulus@samba.org>

Create and use irq_offset_up/down, get_irq_desc, for_each_irq

From: Anton Blanchard <anton@samba.org>

Fix xics irq affinity bug. We were anding with cpu_online_map but werent
using the result later on.

From: Michael Strosaker <strosake@us.ibm.com>

Add RTAS os-term call for panic on pSeries

From: Joel Schopp <jschopp@austin.ibm.com>

Add support for hotplug cpus

From: Will Schmidt <willschm@us.ibm.com>

Additional PVR value for power5 processor

From: Jake Moilanen <moilanen@austin.ibm.com>

Misc rtasd fixes for some broken firmware versions.

From: Joel Schopp <jschopp@austin.ibm.com>

Fix includes to avoid the compiler warning:
arch/ppc64/xmon/start.c: In function `xmon_readchar':
arch/ppc64/xmon/start.c:104: warning: implicit declaration of function
`xmon_printf'

From: Jake Moilanen <moilanen@austin.ibm.com>

Change the loglevel of an error log printed so it
does not goto the console.  Since error logs can
be upto 2k in size, it can spam the console.

From: Will Schmidt <willschm@us.ibm.com>

Correct comments for the offsets of fields in paca

From: Jake Moilanen <moilanen@austin.ibm.com>

The JS20 uses devfn 0 for a HT->PCI bridge.  The PHB devfn
assumption does not hold for this case.

From: Jake Moilanen <moilanen@austin.ibm.com>

Allow PCI devices to use address that happens to fall in the ISA range,
but still protect against ISA device accesses when there is not an ISA
bus.

From: Anton Blanchard <anton@samba.org>

Disable SMT snooze by default

From: Olof Johansson <olof@austin.ibm.com>

RTAS on IBM pSeries runs in real mode, so all pointers being passed in to
it need to be in low memory.  There's two places in the RAS code that
passes in pointers to items on the stack, which might end up being above
the limit.

Below patch resolves this by creating a buffer in BSS + a lock for
serialization.  There's no reason to worry about contention on the lock,
since rtas_call() also serializes on a single spinlock and this is an
infrequent code path in the first place.

From: David Gibson <david@gibson.dropbear.id.au>

On PPC64, to deal with the restrictions imposed by the PPC MMU's segment
design, hugepages are only allowed to be mapping in two fixed address
ranges, one 2-3G (for use by 32-bit processes) and one 1-1.5T (for use in
64-bit processes).  This is quite limiting, particularly for 32-bit
processes which want to use a lot of large page memory.

This patch relaxes this restriction, and allows any of the low 16 segments
(i.e.  those below 4G) to be individually switched over to allow hugepage
mappings (provided the segment does not already have any normal page
mappings).  The 1-1.5T fixed range for 64-bit processes remains.

From: Stephen Rothwell <sfr@canb.auug.org.au>

This is the iSeries virtual ethernet driver.  David Gibson has taken you
previous comments and hopefully sitisfied most of them.  The driver has
also undergone some more testing which showed up some bugs which have been
addressed.

Unfortunately, Anton is about to submit some other patches of mine which
will sightly comflict with this.  I will send a patch shortly that will
(hopefully) fix that.

From: Paul Mackerras <paulus@samba.org>

This patch from Julie DeWandel exports the symbol itLpNaca on iSeries
machines, for the use of the viodasd driver.

From: Paul Mackerras <paulus@samba.org>

This patch from Julie DeWandel makes CONFIG_VT default to N on iSeries
machines which are using the iSeries virtual console driver viocons.c.  The
VT console and the viocons code can't coexist because they use the same tty
numbers, that is, viocons supplies /dev/tty1.  Without this patch the user
has to figure out somehow that s/he has to turn on CONFIG_EMBEDDED in order
to be able to turn off CONFIG_VT, which is really very non-obvious.

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

A recent patch that cleaned up some absolute/virt translation macros forgot
one occurence, thus breaking g5 build with iommu support.

From: Olof Johansson <olof@austin.ibm.com>

A bug snuck in during the rewrite of ppc64 IOMMU code.  When a
{pci,vio}_alloc_consistent() call fails, DMA_ERROR_CODE is returned instead
of NULL.

From: David Gibson <david@gibson.dropbear.id.au>

Found this again while looking at hugepage extensions.  Haven't actually had
it bite yet - the race is small and the other bug will never be triggered in
32-bit processes, and the function is rarely called on 64-bit processes.

This patch fixes two bugs in the (same part of the) PPC64 hugepage code.
First the method we were using to free stale PTE pages was not safe with some
recent changes (race condition).  BenH has fixed this to work in the new way.
 Second, we were not checking for a valid PGD entry before dereferencing the
PMD page when scanning for stale PTE page pointers.

From: David Gibson <david@gibson.dropbear.id.au>

The PPC64 version of is_aligned_hugepage_range() is buggy.  It is supposed to
test not only that the given range is hugepage aligned, but that it lies
within the address space allowed for hugepages.  We were checking only that
the given range intersected the hugepage range, not that it lay entirely
within it.  This patch fixes the problem and changes the name of some macros
to make it less likely to make that misunderstanding again.

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

This patch fixes si_addr on some segfaults in 64 bits mode, it used to be
bogus (address not passed to do_page_fault by the asm code after a failure
to set an SLB entry).

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

Without this patch, executing an altivec instruction on an altivec capable
CPU with a kernel that do not have CONFIG_ALTIVEC set would result in a
kernel crash.

(Fix forward ported from 2.4 by John Whitney
<jwhitney-linuxppc@sands-edge.com>)

From: William Lee Irwin III <wli@holomorphy.com>

Now the scheduler text is in its own ELF section this branch is asking for
an illegal displacement.

From: William Lee Irwin III <wli@holomorphy.com>

This addresses the issue with get_wchan() that the various functions acting
as scheduling-related primitives are not, in fact, contiguous in the text
segment.  It creates an ELF section for scheduling primitives to be placed
in, and places currently-detected (i.e.  skipped during stack decoding)
scheduling primitives and others like io_schedule() and down(), which are
currently missed by get_wchan() code, into this section also.

The net effects are more reliability of get_wchan()'s results and the new
ability, made use of by this code, to arbitrarily place scheduling
primitives in the source code without disturbing get_wchan()'s accuracy.

Suggestions by Arnd Bergmann and Matthew Wilcox regarding reducing the
invasiveness of the patch were incorporated during prior rounds of review. 
I've at least tried to sweep all arches in this patch.

From: Armin Schindler <armin@melware.de>

With this patch the ISDN kernel CAPI code uses a per application workqueue
with proper locking to prevent message re-ordering due to the fact a
workqueue may run on another CPU at the same time.  Also some locks for
internal data is added.

Removed global recv_queue work, use per application workqueue.  Added
proper locking mechanisms for application, controller and application
workqueue function.  Increased max.  number of possible applications and
controllers.

The race is that con_close() can sleep, and drops the BKL while
tty->count==1.  But another thread can come into init_dev() and will take a
new ref against the tty and start using it.

But con_close() doesn't notice that new ref and proceeds to null out
tty->driver_data while someone else is using the resurrected tty.

So the patch serialises con_close() against init_dev() with tty_sem.


Here's a test app which reproduced the oops instantly on 2-way.  It realy
needs to be run against all tty-capable devices.

/*
 * Run this against a tty which nobody currently has open, such as /dev/tty9
 */

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/ioctl.h>
#include <linux/kd.h>

void doit(char *filename)
{
	int fd,x;

	fd = open(filename, O_RDWR);
	if (fd < 0) {
		perror("open");
		exit(1);
	}
	ioctl(fd, KDKBDREP, &x);
	close(fd);
}

main(int argc, char *argv[])
{
	char *filename = argv[1];

	for ( ; ; )
		doit(filename);
}

Remove the down_tty_sem() and up_tty_sem() and replace them with open-coded
up() and down().  This is an equivalent transformation.

I assume these functions were created to open the possibility of per-tty
semaphores at some time in the future.  But the code which is protected by
this lock deals with two tty's at the same time, and the next patch will need
to release the lock after the tty has been destroyed.