This patch is a performance and correctness update to the direct-IO
code: O_DIRECT and the raw driver.  It mainly affects IO against
blockdevs.

The direct_io code was returning -EINVAL for a filesystem hole.  Change
it to clear the userspace page instead.

There were a few restrictions and weirdnesses wrt blocksize and
alignments.  The code has been reworked so we now lay out maximum-sized
BIOs at any sector alignment.

Because of this, the raw driver has been altered to set the blockdev's
soft blocksize to the minimum possible at open() time.  Typically, 512
bytes.  There are now no performance disadvantages to using small
blocksizes, and this gives the finest possible alignment.

There is no API here for setting or querying the soft blocksize of the
raw driver (there never was, really), which could conceivably be a
problem.  If it is, we can permit BLKBSZSET and BLKBSZGET against the
fd which /dev/raw/rawN returned, but that would require that
blk_ioctl() be exported to modules again.

This code is wickedly quick.  Here's an oprofile of a single 500MHz
PIII reading from four (old) scsi disks (two aic7xxx controllers) via
the raw driver.  Aggregate throughput is 72 megabytes/second:

c013363c 24       0.0896492   __set_page_dirty_buffers
c021b8cc 24       0.0896492   ahc_linux_isr
c012b5dc 25       0.0933846   kmem_cache_free
c014d894 26       0.09712     dio_bio_complete
c01cc78c 26       0.09712     number
c0123bd4 40       0.149415    follow_page
c01eed8c 46       0.171828    end_that_request_first
c01ed410 49       0.183034    blk_recount_segments
c01ed574 65       0.2428      blk_rq_map_sg
c014db38 85       0.317508    do_direct_IO
c021b090 90       0.336185    ahc_linux_run_device_queue
c010bb78 236      0.881551    timer_interrupt
c01052d8 25354    94.707      poll_idle

A testament to the efficiency of the 2.5 block layer.

And against four IDE disks on an HPT374 controller.  Throughput is 120
megabytes/sec:

c01eed8c 80       0.292462    end_that_request_first
c01fe850 87       0.318052    hpt3xx_intrproc
c01ed574 123      0.44966     blk_rq_map_sg
c01f8f10 141      0.515464    ata_select
c014db38 153      0.559333    do_direct_IO
c010bb78 235      0.859107    timer_interrupt
c01f9144 281      1.02727     ata_irq_enable
c01ff990 290      1.06017     udma_pci_init
c01fe878 308      1.12598     hpt3xx_maskproc
c02006f8 379      1.38554     idedisk_do_request
c02356a0 609      2.22637     pci_conf1_read
c01ff8dc 611      2.23368     udma_pci_start
c01ff950 922      3.37062     udma_pci_irq_status
c01f8fac 1002     3.66308     ata_status
c01ff26c 1059     3.87146     ata_start_dma
c01feb70 1141     4.17124     hpt374_udma_stop
c01f9228 3072     11.2305     ata_out_regfile
c01052d8 15193    55.5422     poll_idle

Not so good.

One problem which has been identified with O_DIRECT is the cost of
repeated calls into the mapping's get_block() callback.  Not a big
problem with ext2 but other filesystems have more complex get_block
implementations.

So what I have done is to require that callers of generic_direct_IO()
implement the new `get_blocks()' interface.  This is a small extension
to get_block().  It gets passed another argument which indicates the
maximum number of blocks which should be mapped, and it returns the
number of blocks which it did map in bh_result->b_size.  This allows
the fs to map up to 4G of disk (or of hole) in a single get_block()
invokation.

There are some other caveats and requirements of get_blocks() which are
documented in the comment block over fs/direct_io.c:get_more_blocks().

Possibly, get_blocks() will be the 2.6 kernel's way of doing gang block
mapping.  It certainly allows good speedups.  But it doesn't allow the
fs to return a scatter list of blocks - it only understands linear
chunks of disk.  I think that's really all it _should_ do.

I'll let get_blocks() sit for a while and wait for some feedback.  If
it is sufficient and nobody objects too much, I shall convert all
get_block() instances in the kernel to be get_blocks() instances.  And
I'll teach readahead (at least) to use the get_blocks() extension.

Delayed allocate writeback could use get_blocks().  As could
block_prepare_write() for blocksize < PAGE_CACHE_SIZE.  There's no
mileage using it in mpage_writepages() because all our filesystems are
syncalloc, and nobody uses MAP_SHARED for much.

It will be tricky to use get_blocks() for writes, because if a ton of
blocks have been mapped into the file and then something goes wrong,
the kernel needs to either remove those blocks from the file or zero
them out.  The direct_io code zeroes them out.

btw, some time ago you mentioned that some drivers and/or hardware may
get upset if there are multiple simultaneous IOs in progress against
the same block.  Well, the raw driver has always allowed that to
happen.  O_DIRECT writes to blockdevs do as well now.

todo:

1) The driver will probably explode if someone runs BLKBSZSET while
   IO is in progress.  Need to use bdclaim() somewhere.

2) readv() and writev() need to become direct_io-aware.  At present
   we're doing stop-and-wait for each segment when performing
   readv/writev against the raw driver and O_DIRECT blockdevs.

Convert ext3 to the C99 initialiser format.  From Rusty.

Alan's overcommit patch, brought to 2.5 by Robert Love.

Can't say I've tested its functionality at all, but it doesn't crash,
it has been in -ac and RH kernels for some time and I haven't observed
any of its functions on profiles.

"So what is strict VM overcommit?  We introduce new overcommit
 policies that attempt to never succeed an allocation that can not be
 fulfilled by the backing store and consequently never OOM.  This is
 achieved through strict accounting of the committed address space and
 a policy to allow/refuse allocations based on that accounting.

 In the strictest of modes, it should be impossible to allocate more
 memory than available and impossible to OOM.  All memory failures
 should be pushed down to the allocation routines -- malloc, mmap, etc.

 The new modes are available via sysctl (same as before).  See
 Documentation/vm/overcommit-accounting for more information."

Patch from Robert Love.

Attached patch implements for_each_zone(zont_t *) which is a helper
macro to cleanup code of the form:

        for (pgdat = pgdat_list; pgdat; pgdat = pgdat->node_next) {
                for (i = 0; i < MAX_NR_ZONES; ++i) {
                        zone_t * z = pgdat->node_zones + i;
                        /* ... */
                }
        }

and replace it with:

        for_each_zone(zone) {
                /* ... */
        }

This patch only replaces one use of the above loop with the new macro.
Pending code, however, currently in the full rmap patch uses
for_each_zone more extensively.

Patch from Robert Love.

This patch implements for_each_pgdat(pg_data_t *) which is a helper
macro to cleanup code that does a loop of the form:

        pgdat = pgdat_list;
        while(pgdat) {
	        /* ... */
	        pgdat = pgdat->node_next;
	}

and replace it with:

	for_each_pgdat(pgdat) {
		/* ... */
	}

This code is from Rik's 2.4-rmap patch and is by William Irwin.

Reorganise the members of struct page.

- Place ->flags at the start so the compiler can generate indirect
  addressing rather than indirect+indexed for this commonly-accessed
  field.  Shrinks the kernel by ~100 bytes.

- Keep ->count with ->flags so they have the best chance of
  being in the same cacheline.

ifdef out some operations in pte_chain_lock() which are not necessary
on uniprocessor.

Cleanup to show_free_areas() from Bill Irwin:

show_free_areas() and show_free_areas_core() is a mess.
(1) it uses a bizarre and ugly form of list iteration to walk buddy lists
        use standard list functions instead
(2) it prints the same information repeatedly once per-node
        rationalize the braindamaged iteration logic
(3) show_free_areas_node() is useless and not called anywhere
        remove it entirely
(4) show_free_areas() itself just calls show_free_areas_core()
        remove show_free_areas_core() and do the stuff directly
(5) SWAP_CACHE_INFO is always #defined, remove it
(6) INC_CACHE_INFO() doesn't use the do { } while (0) construct

This patch also includes Matthew Dobson's patch which removes
mm/numa.c:node_lock.  The consensus is that it doesn't do anything now
that show_free_areas_node() isn't there.

Patch from Bill Irwin.

It removes the custom pte_chain allocator in mm/rmap.c and replaces it
with a slab cache.

"This patch
 (1) eliminates the pte_chain_freelist_lock and all contention on it
 (2) gives the VM the ability to recover unused pte_chain pages

 Anton Blanchard has reported (1) from prior incarnations of this patch.
 Craig Kulesa has reported (2) in combination with slab-on-LRU patches.

 I've left OOM detection out of this patch entirely as upcoming patches
 will do real OOM handling for pte_chains and all the code changed anyway."

There are a few VM-related patches in this series.  Mainly fixes;
feature work is on hold.

We have some fairly serious locking contention problems with the reverse
mapping's pte_chains.  Until we have a clear way out of that I believe
that it is best to not merge code which has a lot of rmap dependency.

It is apparent that these problems will not be solved by tweaking -
some redesign is needed.  In the 2.5 timeframe the only practical
solution appears to be page table sharing, based on Daniel's February
work.  Daniel and Dave McCracken are working that.

Some bits and pieces here:

- list_splice() has an open-coded list_empty() in it.  Use
  list_empty() instead.

- in shrink_cache() we have a local `nr_pages' which shadows another
  local.  Rename the inner one.  (Nikita Danilov)

- Add a BUG() on a can't-happen code path in page_remove_rmap().

- Tighten up the bug checks in the BH completion handlers - if the
  buffer is still under IO then it must be locked, because we unlock it
  inside the page_uptodate_lock.

in the page cache, it needs to use page_cache_release()
instead of plain "put_page()".

the attached patch fixes two things:

 - a TLS related bug noticed by Arjan van de Ven: apm_init() should set up
   all CPU's gdt entries - just in case some code happens to call in the
   APM BIOS on the wrong CPU. This should also handle the case when some
   APM code gets triggered (by suspend or power button or something).

 - compilation problem

Add support for the glibc 'd_type' field in cases where we have the
READDIRPLUS file attribute information available to us in
nfs_do_filldir().

Add support for positive lookups using the READDIRPLUS cached
information. Both new lookups and lookup revalidation is supported.

Use READDIRPLUS instead of READDIR on NFSv3 directories with lengths
shorter than 8*PAGE_SIZE.

Note that inode attribute information is only updated if it is seen to
be more recent than any existing cached information.

Cache the information about whether or not the server supports
READDIRPLUS.

Cleanup for readdirplus. Allow the file attribute struct to set the
NFS_READTIME(inode) to some value other than 'jiffies'.

Cleanup for the readdirplus code. Make struct nfs_entry take pointers
to the filehandle and file attributes.

A patch by Charles Lever (Charles.Lever@netapp.com) that ensures the
PG_uptodate bit gets set if an entire page gets written by
nfs_writepage_sync().

This patch replaces the cli/sti calls in the trident.c driver with
spin_lock_irqsave/spin_unlock_irqrestore.

This patch (1/2) brings the sound/oss/trident.c driver up to date with
the driver in the 2.4-ac tree. It fixes the following bugs:

* fix wrong cast in suspend/resume (Eric Lemar via Ian Soboroff)

* fix bug where we would free with free_pages() memory allocated via
pci_alloc_consistent().

* add a missing unlock on an error path.

* rewrite the code to read/write registers of audio codecs for Ali5451
(Lei Hu)

It also does various cleanups so that the code conforms to
Documentation/CodingStyle and is nicer to work with.

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

The gdt entry is consulted only while loading its index into
the segment register.

So load_TLS_desc(next, cpu) must be called before loading
next->fs,next->gs in __switch_to().

generic spinlock implementation for downgrade_write().

into samba.org:/home/paulus/kernel/for-linus-ppc

into samba.org:/home/paulus/kernel/for-linus-ppc

Since we have a signed 32-bit time_t, the fact that y % 4 == 0
will get it wrong in 2100 is irrelevant.

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

test, since it is needed regardless.

We use the pci host lock so that we lock config space portably while
handling the CMD640 config space via our own routines to avoid pci bios
tripping CMD640 hardware stuff. We need to use this lock in order to
ensure that we lock at a portable layer. Also add the 2.4.19 fixes for
avoiding wrong probes, and the fix noted on the list.

This should do the trick for pnpbios - we load the initial gdt into each
gdt, and we load the parameters into the gdt of the cpu making the call
relying on the spinlock to avoid bouncing cpu due to pre-empt

 - Remove third argument from file_lock security op.  Whether the lock is
   blocking or not cannot make any difference to a security module!
 - Fix the call in sys_flock to pass the translated lock command, not the
   original.
 - Add a call in fcntl_setlease.  If they're going to know about two types
   of lock, let's tell them about the third too.

The patch below does two things:

(i) fixes a small bug in the new partition code
This is the final chunk s/n/slot/. I'll refrain from giving a vi script.
This is uncontroversial.

(ii) removes ancient garbage concerning disk managers
This may well be controversial.

(Long ago, when disks became larger than 500 MB, lots of tricks were
invented to keep DOS happy. Both hardware tricks and software tricks.
One of the software tricks was the invention of boot managers.
There have been many of those. The Linux kernel has had support
for two of them: OnTrack Disk Manager and EZdrive.
More precisely: there have been many versions of both OnTrack Disk Manager
and EZdrive, and the kernel had support for a few of these versions.

I think the time has come to remove the automatic support - every now
and then it bites some innocent user, and the support is not really
needed any longer, and the support is for outdated versions of these
boot managers.

No doubt it will turn out that users still exist that use some form
of this stuff, but I would prefer to support them by explicit
kernel boot parameters, rather than by code that guesses what
might be the right thing to do.

The patch below just rips out the old stuff. Depending on the screams
this might provoke I expect to add some boot parameters.)

This fixes a synchronize_irq() bug: if the interrupt is freed while an
IRQ handler is running (irq state is IRQ_INPROGRESS) then
synchronize_irq() will return early, which is incorrect.

there was another do_IRQ() bug that in fact necessiated the bad code that
caused the synchronize_irq() bug - we kept the IRQ_INPROGRESS bit set for
not active interrupt sources - after they happen for the first time. Now
the only effect this has is on i8259A irq handling - we used to keep these
irqs disabled after the first 'spurious' interrupt happened.  Now what the
i8259A code really wants to do IMO is to keep the interrupt disabled if
there is no handler defined for that interrupt source. The patch adds
exactly this. I dont remember why this was needed in the first place (irq
probing? avoidance of interrupt storms?), but with the patch the behavior
should be equivalent.

Fix PMD typo

Added irq-argument to synchronize_irq to make sound/oss/cs46xx.c
compile again.

the attached patch fixes the scheduler's migration thread startup bug that
got unearthed by Rusty's recent CPU-startup enhancements.

the fix is to let a startup-helper thread migrate the migration thread,
instead of the migration thread calling set_cpus_allowed() itself.
Migrating a not running thread is a simple and robust thing, and needs no
cooperation from migration threads - thus the catch-22 problem of how to
migrate the migration threads is solved finally.

the patch is against Rusty's initcall fix/hack which calls
migration_init() before other CPUs are brought up - this ordering is
clearly the clean way of doing migration init. [the patch also fixes a UP
compiliation bug in Rusty's hack.]

As pointed out by Andrew Morton, this fixes:
	softirq.c: In function `spawn_ksoftirqd':
	softirq.c:416: warning: statement with no effect