Adapt to "macio" driver infrastructure, cleanup power management, make some
functions static, fix IO accessors, remove local list of commands/directions,
etc...
This driver could use more work & study of the interesting ASIC bugs
documented in Apple source code...
Note that I'm removing support for the CHRP machines using the MESH as
part of the "Hydra" ASIC. Nobody sane really uses that, but if one of
the 3 remaining Longtrail users want to fix that, then move Hydra to
use the "macio" model...
Adapt to the "macio" driver infrastructure, cleanup IO accessors,
remove local list of command/directions, uses infos provided by
the SCSI layer. This driver certainly could use some more work.
The cascaded northbridge MPIC is currently dealt with a special
modified version of open_pic.c (open_pic2.c). A better mecanism
will have to be found for 2.7
Create a low-level synchronous implementation suitable for use
by the early boot platform code or other places where the
asynchronous driver isn't useable. This also exports the locks
used by the real driver to avoid collisions.
Use this new implementation to properly setup the clock chip
at boot on Apple latest laptops
ppc32: Fix release_OF_resource() function.
It would fail to properly release the resource if taken over
by the macio device resource hierarchy. This could still use
some locking, but the io resource lock isn't exported
ppc32: fix a possible race in pte_free()
Another processor could be walking the page table in the middle of the
PTE page to be freeded. Synchronize with hash_page using the lock.
On G5 machines, we remap the AGP port to bus number 0xf0. XFree
contains a hack that is unfixable at the moment for getting the
IO base which is hard coded to bus number 0 (AGP on earlier machines).
We work around this by passing the IO base of bus 0xf0 when asked
for bus 0 in pciconfig_iobase on those machines. Bus 0 is the HT
root and has no IOs, so that is safe.
Move pmac-specific PCI quirks to pmac_pci.c, update cardbus one for new
TI controller, add some for fixing up ATA & SATA controllers (switch normal
ATA to fully native mode and disable unused function on G5 K2 SATA)
I've been chasing a weird SELinux bug which shows up mostly when doing
installs of a dev-* rpm (ie. creating and overwriting lots of block
device inodes), but which I've also seen when doing mkinitrd.
It turned out not to be an SELinux problem at all, but a core VFS
S_ISBLK bug. It seems that SELinux simply widens the race window.
The code at fault is fs/fs-writeback.c:__mark_inode_dirty():
/*
* Only add valid (hashed) inodes to the superblock's
* dirty list. Add blockdev inodes as well.
*/
if (!S_ISBLK(inode->i_mode)) {
if (hlist_unhashed(&inode->i_hash))
goto out;
if (inode->i_state & (I_FREEING|I_CLEAR))
goto out;
}
The "I_FREEING|I_CLEAR" condition was added after the ISBLK/unhashed
tests were already in the source, but I can't see any reason why we'd
want the I_FREEING test not to apply to block devices. And indeed, this
results in all sorts of inode list corruptions. Simply moving the
I_FREEING|I_CLEAR test out of the protection of the S_ISBLK() condition
fixes things entirely.
The existing 2.6 kernel will reliably fail on me in about 2 seconds once
"rpm -Uvh --force dev*.rpm" starts its actual installation of the new
inodes. With the patch below I can't reproduce it at all.
Andrew Morton [Wed, 4 Feb 2004 02:54:15 +0000 (18:54 -0800)]
[PATCH] Fix ptrace in the vsyscall dso area
From: Roland McGrath <roland@redhat.com>
The #include is the part of this patch that matters, so the #ifdef below
works.
The rest of the patch removes gratuitous duplication due to some strange
aversion to concision in the presence of #ifdef, the kind that is all too
common, utterly pointless, and error prone.
Andrew Morton [Wed, 4 Feb 2004 02:53:48 +0000 (18:53 -0800)]
[PATCH] Altix update: irq fixes
From: Pat Gefre <pfg@sgi.com>
arch/ia64/sn/kernel/irq.c
Need to get the cpu from the passed in pcibr struct
Made the interrupt list static and gave it a better name - credit jes
Some lindent'isms
Took out some code that isn't used ..... yet
Andrew Morton [Wed, 4 Feb 2004 02:53:30 +0000 (18:53 -0800)]
[PATCH] Altix update: VGA, keyboard, other changes
From: Pat Gefre <pfg@sgi.com>
arch/ia64/sn/kernel/setup.c
If generic enabled legacy VGA or kbd - disable them
Slightly different check for work arounds and only do it once
If there is no klconfig info and we are in the simulator - ignore it
Update the pxm_to_nasid() routine. It failed for SP configurations and some
SMP configurations where M-bricks used pxm numbers lower that the first
c-brick.
If we don't find the cpu for pxm, search the memblks for it.
Andrew Morton [Wed, 4 Feb 2004 02:52:16 +0000 (18:52 -0800)]
[PATCH] Fix x86-64 boot problem
From: Andi Kleen <ak@muc.de>
Fix a bug introduced with the last merge that prevented booting with
CONFIG_DEBUG_INFO on on x86-64. It would corrupt registers in interrupts.
This has hit a few people, so I would consider it as a critical fix.
Andrew Morton [Wed, 4 Feb 2004 02:50:19 +0000 (18:50 -0800)]
[PATCH] CDROMREADAUDIO frames
From: Jens Axboe <axboe@suse.de>
2.6 imposes a 64 frame limit where 2.4 does not (just relies on kmalloc()
failing and limiting frames from that). That breaks at least on guys app.
With MSF adressing, it's much simpler to be able to ask for a full second
at the time, so I think we should just allow that. So bump the limit from
64 to CD_FRAMES (which is 75).
Andrew Morton [Wed, 4 Feb 2004 02:50:09 +0000 (18:50 -0800)]
[PATCH] PCI Scan all functions
From: Jake Moilanen <moilanen@austin.ibm.com>
On a ppc64 logically partitioned system, there can be a setup where function
0 of a PCI-PCI bridge is assigned to one partition and (for example) function
2 is assigned to a second partition. On the second partition, it would
appear that function 0 does not exist, but function 2 does. If all the
functions are not scanned, everything under function 2 would not be detected.
This patch allows devices that don't respond to function 0, but do respond to
other functions to be marked with a quirk and have all of their functions
scanned.
Andrew Morton [Wed, 4 Feb 2004 02:49:59 +0000 (18:49 -0800)]
[PATCH] Move cpu_vm_mask to be closer to mmu_context_t in struct mm
From: Jack Steiner <steiner@sgi.com>
The cpu_vm_mask use to be close to the mmu_context_t field in the mm
struct. Recently some large members were added between "cpu_vm_mask" and
"context". I suspect that was an oversight.
Here is a patch that puts the fields close together. This makes it likely
that both fields are in the same cache line. Since both fields are likely
to be updated at the same time, this may improve performance.
Andrew Morton [Wed, 4 Feb 2004 02:49:23 +0000 (18:49 -0800)]
[PATCH] shrink_list(): check PageSwapCache() after add_to_swap()
From: Nikita Danilov <Nikita@Namesys.COM>
shrink_list() checks PageSwapCache() before calling add_to_swap(), this
means that anonymous page that is going to be added to the swap right
now these checks return false and:
(*) it will be unaccounted for in nr_mapped, and
(*) it won't be written to the swap if gfp_flags include __GFP_IO but
not __GFP_FS.
(Both will happen only on the next round of scanning.)
Patch below just moves may_enter_fs initialization down. I am not sure
about (*nr_mapped) increase though.
Andrew Morton [Wed, 4 Feb 2004 02:49:14 +0000 (18:49 -0800)]
[PATCH] Use address hint in mmap for search
From: Andi Kleen <ak@suse.de>
When the user gave an address hint in mmap use it as starting point for the
search for !MAP_FIXED.
Currently it is only checked directly and when already used the free area
cache is used as starting point. With this change you can use mmap(4096,
....) to e.g. get the lowest free address in your address space, which is
sometimes useful. For example on x86-64 glibc wants to preferably allocate
thread local data in the first 4GB but use higher addresses when this is
not possible.
This can be a bit more costly in CPU time because it may have to skip over
more VMAs, but gives better semantics for most cases. Most programs pass
NULL as hint anyways so it won't make any difference for them.
I did it for the generic mmap and for x86-64 for now. Also minor white
space fixes for x86-64.
Andrew Morton [Wed, 4 Feb 2004 02:49:05 +0000 (18:49 -0800)]
[PATCH] rate limit nr_free_pages
From: Jes Sorensen <jes@trained-monkey.org>
nr_free_pages() is expensive, especially on large SMP machines. The patch
changes the memory overcommit code so that it only calls nr_free_pages() is
we're about to fail the allocation attempt.
Andrew Morton [Wed, 4 Feb 2004 02:48:48 +0000 (18:48 -0800)]
[PATCH] proc_check_root() locking fix
From: Maneesh Soni <maneesh@in.ibm.com>
The patch fixes locking in proc_check_root(). It brings is_subdir() call
under vfsmount_lock. Holding vfsmount_lock will ensure mnt_mountpoint
dentry is intact and the dentry does not go away while it is being checked
in is_subdir().
Andrew Morton [Wed, 4 Feb 2004 02:48:39 +0000 (18:48 -0800)]
[PATCH] is_subdir locking fix
From: Maneesh Soni <maneesh@in.ibm.com>
o The following patch fixes is_subdir() races with d_move. Due to concurrent
d_move, in is_subdir() we can end up accessing freed d_parent pointer in
case of pre-emptible kernel. To avoid this we can use rcu_read_lock() and
rcu_read_unlock().
o This also fixes the seqlock uses in is_subdir() as we need to restart the
the inner loop with the origianl new_dentry passed to the routine in case
of any rename occured while we are traversing d_parent links.
Andrew Morton [Wed, 4 Feb 2004 02:48:21 +0000 (18:48 -0800)]
[PATCH] u_int32_t causes cross-compile problems
From: Pratik Solanki <pratik.solanki@timesys.com>
I came across this C standards issue while cross-compiling the Linux kernel
with gcc on Solaris. The file gen_crc32table.c uses the non-standard type
u_int32_t. It's possible that the host machine's sys/types.h does not
define u_int32_t. The attached patch replaces u_int32_t with the POSIX
standard uint32_t and includes POSIX inttypes.h instead of sys/types.h.
Andrew Morton [Wed, 4 Feb 2004 02:48:12 +0000 (18:48 -0800)]
[PATCH] fix menuconfig choice item help display
From: Bjorn Helgaas <bjorn.helgaas@hp.com>
Anders Gustafsson <andersg@0x63.nu>
Roman Zippel <zippel@linux-m68k.org>
This patch fixes menuconfig so it can display help text for individual
choice group config entries.
Previously it would only display the help text attached to the "choice"
item. There was no way to display the help attached to individual config
entries inside the choice group. Typically, the "choice" item has no help
text, and all the useful help is attached to the individual entries, so
this was a bit of a problem.
Andrew Morton [Wed, 4 Feb 2004 02:48:03 +0000 (18:48 -0800)]
[PATCH] /proc/stat:btime fix
From: George Anzinger <george@mvista.com>,
Petri Kaukasoina <kaukasoi@elektroni.ee.tut.fi>
btime in /proc/stat does not stay constant but decreases at a rate of 15
secs/day, because we're assuming that HZ is exactly 100. Use the correct
adjustments to fix that up.
Andrew Morton [Wed, 4 Feb 2004 02:47:54 +0000 (18:47 -0800)]
[PATCH] Fine tune the time conversion to eliminate conversion errors.
From: George Anzinger <george@mvista.com>
The time conversion code is erroring on the side of a bit too small. The
attached patch forces any error to be on the high side. The current code will
convert 1 nanosecond to zero jiffies (standard says that should be 1). It also
is around 1 nanosecond late on each roll to the next jiffie.
I have done some error checks with this patch applied and get the following
errors in PPB ( Parts Per Billion):
Andrew Morton [Wed, 4 Feb 2004 02:47:35 +0000 (18:47 -0800)]
[PATCH] try reiserfs before other filesystems
reiserfs places its superblock in weird places which can result in false
positives and various printks when other filesystems probe a resierfs
filesystem.
Andrew Morton [Wed, 4 Feb 2004 02:46:42 +0000 (18:46 -0800)]
[PATCH] oprofile, typo in alpha driver
From: Philippe Elie <phil.el@wanadoo.fr>
Unless I miss something this look like a typo, one user reported to get
error from the daemon: 'Unknown event for counter 1' (alpha ev6) and the
behavior was better but not completly sane after trying this patch: he get
spurious event for counter 1 when enabling only counter 0 but rarely now.
No alpha box to test this.
Andrew Morton [Wed, 4 Feb 2004 02:46:33 +0000 (18:46 -0800)]
[PATCH] oprofile per-cpu buffer overrun
From: Philippe Elie <phil.el@wanadoo.fr>
In a ring buffer controlled by a read and write positions we can't use
buffer_size but only buffer_size - 1 entry, the last free entry act as a
guard to avoid write pos overrun. This bug was hidden because the writer,
oprofile_add_sample(), request one more entry than really needed.
Andrew Morton [Wed, 4 Feb 2004 02:46:26 +0000 (18:46 -0800)]
[PATCH] console cleanup
From: Sam Ravnborg <sam@ravnborg.org>,
Ben Collins <bcollins@debian.org>
Fix up the console makefiles and logo generation.
1) To make output look like the rest of the kernel build.
2) To avoid make utilising chained rules, and therefore issuing a 'rm
drivers/video/logo/linux_logo.c ...' during the build.
I have previously submitted a few patches for logo/Makefile, but this is the
first one that actually address the problems I have seen in a proper way.
And no, I did not like such a simple thing to look that complicated, the
other option was to list too many files or to use other types of kbuild/make
magic.
Andrew Morton [Wed, 4 Feb 2004 02:45:42 +0000 (18:45 -0800)]
[PATCH] Zero last byte of mount option page.
From: James Morris <jmorris@redhat.com>
Here's a patch which zeroes the last byte of the mount option data copied
from userspace during mount(2).
For filesystems which parse mount options as strings (the majority), lack
of a zero terminator could cause the page to be overrun. The source code
comments specify that the maximum size of the mount data is PAGE_SIZE-1, so
this patch will not affect any valid binary-formatted mount data.
projects / linux / log