Truncates can take a very long time.  Especially if there is a lot of
writeout happening, because truncate must wait on in-progress I/O.

And sys_unlink() is performing that truncate while holding the parent
directory's i_sem.  This basically shuts down new accesses to the entire
directory until the synchronous I/O completes.

In the testing I've been doing, that directory is /tmp, and this hurts.

So change sys_unlink() to perform the actual truncate outside i_sem.

When there is a continuous streaming write to the same disk, this patch
reduces the time for `make -j4 bzImage' from 370 seconds to 220.

When a throttled writer is performing writeback, and it encounters an inode
which is already under writeback it is forced to wait on the inode.  So that
process sleeps until whoever is writing it out finishes the writeout.

Which is OK - we want to throttle that process, and another process is
currently pumping data at the disk anyway.

But in one situations the delays are excessive.  If one process is performing
a huge linear write, other processes end up waiting for a very long time
indeed.  It appears that this is because the writing process just keeps on
hogging the CPU, returning to userspace, generating more dirty data, writing
it out, sleeping in get_request_wait, etc.  All other throttled dirtiers get
starved.

So just remove the wait altogether if it is just a memory-cleansing writeout.
 The calling process will then throttle in balance_dirty_pages()'s call to
blk_congestion_wait().

We do not need to pass the `wait' argument down to __sync_single_inode().
That information is now present at wbc->sync_mode.

- Fix dead comment in load_elf_interp() (Dave Airlie)

- Add some (hard-won) commentary around the early SET_PERSONALITY() in
  load_elf_binary().

- Remove dead hugetlb prototype.

- Fix some silliness in hugetlbpage.c

Patch from Pavel Machek <pavel@ucw.cz>

scsi_eh_* needs to run even during suspend because suspend does not prevent a
hard disk from reporting an error.

Patch from miles@lsi.nec.co.jp (Miles Bader)

Otherwise it won't compile.  I guesss this used to work because <asm/page.h>
was included somewhere to get the BUG macros, but now with the advent of
<asm/bug.h> that's changed.

Patch from john stultz <johnstul@us.ibm.com>

Adds some lost-tick compensation code, which handles the case where time
accounting goes wrong due to interrupts being disabled for longer than two
ticks.

This patch solves the problem by checking when an interrupt occurs if
timer->get_offset() is a value greater then 2 ticks.  If so, it increments
jiffies appropriately.

I was concerned that we'd be better off finding and fixing the misbehaving
drivers, but it turns out that the main culprits are system management cards
over which the kernel has no control.

However John has added some debug code which will drop a backtrace on the
first five occurrences which will allow us to find-and-fix bad drivers if
overruns _are_ due to Linux software.  (I disabled this - it was irritating
me.  Dave Hansen has a patch which allows it to be turned on via a kernel
boot parameter, like the x86_64 equiv).

Patch from "Randy.Dunlap" <rddunlap@osdl.org>

This fixes a few #warning's that gcc 2.96 complains about having
unmatched single-quote marks.  (warnings on #warnings)

Patch from Russell King <rmk@arm.linux.org.uk>

Have a couple of extra warnings:

fs/binfmt_elf.c: In function `create_elf_tables':
fs/binfmt_elf.c:239: warning: initialization makes integer from pointer without a cast
fs/binfmt_elf.c:249: warning: initialization makes integer from pointer without a cast

#ifndef elf_addr_t
#define elf_addr_t unsigned long
#endif

        elf_addr_t *argv, *envp;

        __put_user(NULL, argv);
        __put_user(NULL, envp);

It would therefore appear that x86 __put_user is not properly type-checking
the arguments to __put_user().

Here's a patch which fixes the warning (but doesn't fix x86's type-check
challenged __put_user implementation).

Patch from Matthew Dobson <colpatch@us.ibm.com>

When I originally wrote the patches implementing the in-kernel topology
macros, they were meant to be called as a second layer of functions,
sans underbars.  This additional layer was deemed unnecessary and
summarily dropped.  As such, carrying around (and typing!) all these
extra underbars is quite pointless.  Here's a patch to nip this in the
(sorta) bud.  The macros only appear in 16 files so far, most of them
being the definitions themselves.

Patch from Dave Olien <dmo@osdl.org>

The following patch implements retry on media errors for the DAC960 driver.
On such media errors, the DAC960 apparently doesn't report how much of the
transfer may have been successful before the error was encountered.

This type of error should be rare on healthy hardware, especially if the
disks are stripped or mirrored.  But, when large transfers are submitted to
the controller, it's especially bad to have to fail the entire transfer
because one disk sector may have been bad.

Patch from William Lee Irwin III <wli@holomorphy.com>

CONFIG_X86_NUMA no longer exists.  This changes the MAX_IO_APICS definition
to 32, where it is required to be so large on NUMA-Q in order to boot.

Patch from William Lee Irwin III <wli@holomorphy.com>

will_become_orphaned_pgrp()'s sole use is is_orphaned_pgrp().  Fold its body
into is_orphaned_pgrp(), rename __will_become_orphaned_pgrp(), and adjust
callers.  Code shrinkage plus some relief from underscore-itis.

From Anton Blanchard: remove spaces from slab cache identifiers.  Simplifies
parsing of /proc/slabinfo.

Patch from William Lee Irwin III <wli@holomorphy.com>

Okay, this one looks ugly because we're missing some of the definitions
available with which to convert to inline functions (esp. struct page).
A lot of these introduce temporaries and sort of hope names won't clash,
which might be important to whoever cares about -Wshadow.

 - node_end_pfn() evaluates nid twice
 - local_mapnr() evaluates kvaddr twice
 - kern_addr_valid() evaluates kaddr twice
 - pfn_to_page() evaluates pfn multiple times
 - page_to_pfn() evaluates page thrice
 - pfn_valid() doesn't parenthesize its argument

Patch from "Stephen D. Smalley" <sds@epoch.ncsc.mil>

This patch restores the LSM hook calls in sendfile to 2.5.59.  The hook was
previously added as of 2.5.29 but the hook calls in sendfile were
subsequently lost as a result of the sendfile rewrite as of 2.5.30.

Patch from: Joel Becker <Joel.Becker@oracle.com>

This kernel module will detect long durations when jiffies has failed to
increment, and will reboot the machine in response.

Joel says:


"Here's why Oracle wants such a thing.  We run clusters.  Imagine a two node
 cluster.  Node1 pauses completely for some reason.  There are multiple
 reasons this can happen.  A bad driver can udelay() for 90 seconds (qla used
 to do this).  zVM on S/390 can page Linux out for minutes at a time.
 Anything that causes the box to freeze.  Jiffies does *not* count during
 this, so when Node1 returns it feels that no time has passed.

 Node2, however, has been counting time.  When Node1 goes away, the Oracle
 cluster manager starts looking for it.  After a timeout, it gives up.  It
 then recovers any in-progress transactions from Node1.  After that, it
 starts new operations, modifying the data in ways that Node1 has no idea
 about (it's still out to lunch).

 When Node1 finally returns (udelay() ends, zVM pages it in, whatever), any
 I/O that it has queued or is about to queue will get sent to the disk.
 Oops, you've just corrupted your shared data.

 hangcheck-timer should catch this and reboot the box.

 This is why Oracle wants this driver.  We figure that such functionality
 would be beneficial to others as well, so we posted to l-k.  We'd all hope
 that driver writers don't udelay() for 90s, but S/390 with zVM is still
 around.  Some folks might want to notice when it happens.  I am sure other
 things exist that trigger the same symptoms."

Lots of nonsensical locking in there.

Patch from Manfred Spraul <manfred@colorfullife.com>

When slab debugging is enabled we're holding off interrupts for too long
(more than a jiffy), so reduce the alloc/free batching size when slab debug
is enabled.

From wli

A moment's reflection on the subject suggests to me it's worthwhile to
generalize pgd_ctor support so it works (without #ifdefs!) on both PAE
and non-PAE. This tiny tweak is actually more noticeably beneficial
on non-PAE systems but only really because pgd_alloc() is more visible;
the most likely reason it's less visible on PAE is "other overhead".
It looks particularly nice since it removes more code than it adds.

Touch tested on NUMA-Q (PAE). OFTC #kn testers testing the non-PAE case.

From Bill Irwin

This allocates pgd's and pmd's using the slab and slab ctors.  It has a
benefit beyond preconstruction in that PAE pmd's are accounted via
/proc/slabinfo

Profiling of kernel builds by Martin Bligh shows a 30-40% drop in CPU load
due to pgd_alloc()'s page clearing activity.  But this was already a tiny
fraction of the overall CPU time.

Patch From: Hugh Dickins <hugh@veritas.com>

Recently noticed that __GFP_HIGHIO has played no real part since bounce
buffering was converted to mempool in 2.5.12: so this patch (over 2.5.58-mm1)
removes it and GFP_NOHIGHIO and SLAB_NOHIGHIO.

Also removes GFP_KSWAPD, in 2.5 same as GFP_KERNEL; leaves GFP_USER, which
can be a useful comment, even though in 2.5 same as GFP_KERNEL.

One anomaly needs comment: strictly, if there's no __GFP_HIGHIO, then
GFP_NOHIGHIO translates to GFP_NOFS; but GFP_NOFS looks wrong in the block
layer, and if you follow them down, you find that GFP_NOFS and GFP_NOIO
behave the same way in mempool_alloc - so I've used the less surprising
GFP_NOIO to replace GFP_NOHIGHIO.

Patch from Nikita Danilov <Nikita@Namesys.COM>

read_cache_pages() is passed a bunch of pages to start I/O against and it is
supposed to consume all those pages.  But if there is an I/O error, someone
need to throw away the unused pages.

At present the single user of read_cache_pages() (nfs_readpages) does that
cleanup by hand.  But it should be done in the core kernel.

- Don't require a 160-col xterm

- Coding style consistency

Implements merging of file-backed VMA's.  Based on Andrea's 2.4 patch.

It's only done for mmap().  mprotect() and mremap() still will not merge
VMA's.

It works for hugetlbfs mappings also.

pagecache can be reclaimed via the page LRU and via prune_icache.  We
currently don't know how much reclaim is happening via each.

The patch adds instrumentation to display the number of pages which were
freed via prune_icache.  This is displayed in /proc/vmstat:pginodesteal and
/proc/vmstat:kswapd_inodesteal.

Turns out that under some workloads (well, dbench at least), fully half of
page reclaim is via the unused inode list.  Which seems quite OK to me.

A static function in a header where presumably a static inline was intended.

An implementation of posix_fadvise64().  It adds 368 bytes to my vmlinux and
is worth it.

I didn't bother doing posix_fadvise(), as userspace can implement that by
calling fadvise64().

The main reason for wanting this syscall is to provide userspace with the
ability to explicitly shoot down pagecache when streaming large files.  This
is what O_STEAMING does, only posix_fadvise() is standards-based, and harder
to use.

posix_fadvise() also subsumes sys_readahead().

POSIX_FADV_WILLNEED will generally provide asynchronous readahead semantics
for small amounts of I/O.  As long as things like indirect blocks are aready
in core.

POSIX_FADV_RANDOM gives unprivileged applications a way of disabling
readahead on a per-fd basis, which may provide some benefit for super-seeky
access patterns such as databases.



The POSIX_FADV_* values are already implemented in glibc, and this patch
ensures that they are in sync.

A test app (fadvise.c) is available in ext3 CVS.  See

	http://www.zip.com.au/~akpm/linux/ext3/

for CVS details.

Ulrich has reviewed this patch (thanks).

into penguin.transmeta.com:/home/penguin/torvalds/repositories/kernel/linux

One of the goals of the whole new modversions implementation:
export-objs is gone for good!

into shaggy.austin.ibm.com:/shaggy/bk/jfs-2.5

o Build modules with CONFIG_MODVERSIONS when just saying "make"
o Ignore generated *.ver.c files
o Fix a typo (Sam Ravnborg)
o Fix another typo (Paul Marinceu)

into tp1.ruhr-uni-bochum.de:/scratch/kai/kernel/v2.5/linux-2.5.make

Skip over the first part of __vermagic in modversioning is on: otherwise
you'll have to force it when changing from 2.6.0 to 2.6.1.

Fix the case where no CRCs are supplied (OK, but taints kernel), and
only print one tainted message (otherwise --force gives hundreds of them).

We generated the intermediate files that contain checksums for
unresolved symbols in .tmp_versions, which had the disadvantage
that is obscured what's going on during the build. Just
generate them as .ver.[co] right next to the actual objects in the
object tree.

CONFIG_MODVERSIONING was a temporary name introduced to distinguish
between the old and new module version implementation. Since the
traces of the old implementation are now gone from the build system,
we rename the config option back in order to not confuse users more
than necessary in 2.6.
 
Also, remove some historic modversions cruft throughout the tree.

into hostme.bitkeeper.com:/ua/repos/j/jfs/linux-2.5