We're seeing heavy contention against j_list_lock on 8-way in
do_get_write_access().

We actually don't need j_list_lock in there except for one little case - the
per-bh jbd_lock_bh_state() is sufficient to protect this buffer's internal
state.

On some nice quick LVM array Ram Pai measured an overall 3x speedup from this
patch:

the script took the following time on 265mm1
 real    0m57.504s
 user    0m0.400s
 sys     7m29.867s


 and with the 2patches it took
 real 	0m19.983s
 user    0m0.438s
 sys     1m55.896s

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

From: Anton Blanchard <anton@samba.org>

Directory notify code was showing up in a dd bs=1024k from 2 raid arrays
on an emulex FC adapter:

3635     69.4896  vmlinux-2.6.5            .default_idle
332       6.3468  vmlinux-2.6.5            .__copy_tofrom_user
112       2.1411  vmlinux-2.6.5            .save_remaining_regs
76        1.4529  vmlinux-2.6.5            .scsi_dispatch_cmd
64        1.2235  vmlinux-2.6.5            .dnotify_parent
61        1.1661  vmlinux-2.6.5            .do_generic_mapping_read

We already have a sysctl to enable/disable it, the patch below uses it
in dnotify_parent. dnotify_parent disappears and idle time goes up:

4508     70.8582  vmlinux-2.6.5            .default_idle
253       3.9767  vmlinux-2.6.5            .__copy_tofrom_user
142       2.2320  vmlinux-2.6.5            .save_remaining_regs
88        1.3832  vmlinux-2.6.5            .shrink_zone
84        1.3203  vmlinux-2.6.5            .elx_drvr_unlock
75        1.1789  vmlinux-2.6.5            .scsi_dispatch_cmd
69        1.0846  vmlinux-2.6.5            .do_generic_mapping_read

Of course, to gain this small speedup isers need to know to set
/proc/sys/fs/dir-notify-enable to zero.  Nobody does that.

From: Marcelo Tosatti <marcelo.tosatti@cyclades.com>

The cyclades.c driver was marked BROKEN_ON_SMP during early 2.6.  It was
fixed later on but the tag was left in Kconfig.

The driver is not very smart wrt SMP locking, it can be improved.  There is
only one spinlock per card which guarantees command block ordering and
protects different shared data, which can be held for long periods.

_But_ the locking works reliably, so remove the BROKEN_ON_SMP tag.

From: "Martin J. Bligh" <mbligh@aracnet.com>

I'd prefer we renamed this to page_to_nid() before anyone starts using it. 
This fits with the naming convention of everything else (pfn_to_nid, etc). 
Nobody uses it right now - I grepped the whole tree.

From: Hugh Dickins <hugh@veritas.com>

Some arches refer to page->mapping for their dcache flushing: use
page_mapping(page) for safety, to avoid confusion on anon pages, which will
store a different pointer there - though in most cases flush_dcache_page is
being applied to pagecache pages.

arm has a useful mapping_mapped macro: move that to generic, and add
mapping_writably_mapped, to avoid explicit list_empty checks on i_mmap and
i_mmap_shared in several places.

Very tempted to add page_mapped(page) tests, perhaps along with the
mapping_writably_mapped tests in do_generic_mapping_read and
do_shmem_file_read, to cut down on wasted flush_dcache effort; but the
serialization is not obvious, too unsafe to do in a hurry.

Fix up the rw_swap_page_sync() gorrors by fully decoupling this function
from the VM - it is now just a helper function which reads a page from or
writes a page to swap.

From: Hugh Dickins <hugh@veritas.com>

Tracking anonymous pages by anon_vma,pgoff or mm,address needs a
pointer,offset pair in struct page: mapping,index the natural choice.  But
swapcache uses those for &swapper_space,swp_entry_t.

It's trivial to separate swapcache from pagecache with radix tree; most of
swapper_space is actually unused, just a fiction to pretend swap like file;
and page->private is a good place to keep swp_entry_t, now that swap never
uses bufferheads.

Define PG_anon bit, page_add_rmap SetPageAnon and put an oopsable address in
page->mapping to test that we're not confused by it.  Define
page_mapping(page) macro to give NULL when PageAnon, whatever may be in
page->mapping.  Define PG_swapcache bit, deduce swapper_space from that in
the few places we need it.

add_to_swap_cache now distinct from add_to_page_cache.  Separating the caches
somewhat simplifies the tmpfs swizzling in swap_state.c, now the page can
briefly be in both caches.

The rmap method remains pte chains, no change to that yet.  But one small
functional difference: the use of PageAnon implies that a page truncated
while still mapped will no longer be found and freed (swapped out) by
try_to_unmap, will only be freed by exit or munmap.  But normally pages are
unmapped by vmtruncate: this should only affect nonlinear mappings, and a
later patch not in this batch will fix that.

From: Hugh Dickins <hugh@veritas.com>

First of a batch of three rmap patches: this initial batch of three paving
the way for a move to some form of object-based rmap (probably Andrea's, but
drawing from mine too), and making almost no functional change by itself.  A
few days will intervene before the next batch, to give the struct page
changes in the second patch some exposure before proceeding.

rmap 1 create include/linux/rmap.h

Start small: linux/rmap-locking.h has already gathered some declarations
unrelated to locking, and the rest of the rmap declarations were over in
linux/swap.h: gather them all together in linux/rmap.h, and rename the
pte_chain_lock to rmap_lock.

From: Jens Axboe <axboe@suse.de>

CFQ I/O scheduler

From: Jens Axboe <axboe@suse.de>

There's a small discrepancy in when we decide to unplug a queue based on
q->unplug_thresh.  Basically it doesn't work for tagged queues, since
q->rq.count[READ] + q->rq.count[WRITE] is just the number of allocated
requests, not the number of requests stuck in the io scheduler.  We could
just change the nr_queued == to a nr_queued >=, however that is still
suboptimal.

This patch adds accounting for requests that have been dequeued from the io
scheduler, but not freed yet.  These are q->in_flight.  allocated_requests
- q->in_flight == requests_in_scheduler.  So the condition correctly
becomes

	if (requests_in_scheduler == q->unplug_thresh)

instead.  I did a quick round of testing, and for dbench on a SCSI disk the
number of timer induced unplugs was reduced from 13 to 5 :-).  Not a huge
number, but there might be cases where it's more significant.  Either way,
it gets ->unplug_thresh always right, which the old logic didn't.

From: Neil Brown <neilb@cse.unsw.edu.au>

I've made a bunch of changes to the 'md' bits - largely moving the
unplugging into the individual personalities which know more about which
drives are actually in use.

From: Jens Axboe <axboe@suse.de>

Dog slow software suspend found this one. If WB_SYNC_ALL, then you need
to mark the bio as sync as well.

This is because swap_writepage() does a remove_exclusive_swap_page() (going
to __delete_from_swap_cache -> __remove_from_page_cache) which can kill
page->mapping, thus aops->sync_page() has nothing to work with for unplugging
the address space.

From: Jens Axboe <axboe@suse.de>,
      Chris Mason,
      me, others.

The global unplug list causes horrid spinlock contention on many-disk
many-CPU setups - throughput is worse than halved.

The other problem with the global unplugging is of course that it will cause
the unplugging of queues which are unrelated to the I/O upon which the caller
is about to wait.

So what we do to solve these problems is to remove the global unplug and set
up the infrastructure under which the VFS can tell the block layer to unplug
only those queues which are relevant to the page or buffer_head whcih is
about to be waited upon.

We do this via the very appropriate address_space->backing_dev_info structure.

Most of the complexity is in devicemapper, MD and swapper_space, because for
these backing devices, multiple queues may need to be unplugged to complete a
page/buffer I/O.  In each case we ensure that data structures are in place to
permit us to identify all the lower-level queues which contribute to the
higher-level backing_dev_info.  Each contributing queue is told to unplug in
response to a higher-level unplug.

To simplify things in various places we also introduce the concept of a
"synchronous BIO": it is tagged with BIO_RW_SYNC.  The block layer will
perform an immediate unplug when it sees one of these go past.

From: Joe Thornber <thornber@redhat.com>

dm.c: remove __dm_request (merge with previous patch).

From: Miquel van Smoorenburg <miquels@cistron.nl>
      Joe Thornber <thornber@redhat.com>

This implements the queue congestion callout for DM stacks.  To make
bdi_read/write_congested() return correct information.

- md->lock protects all fields in md _except_ md->map
- md->map_lock protects md->map
- Anyone who wants to read md->map should use dm_get_table() which
  increments the tables reference count.

This means the spin lock is now only held for the duration of a
reference count increment.

Udpate:

dm.c: protect md->map with a rw spin lock rather than the md->lock
semaphore.  Also ensure that everyone accesses md->map through
dm_get_table(), rather than directly.

From: Miquel van Smoorenburg <miquels@cistron.nl>

The VM and VFS use the address_space_backing_dev_info to track the realtime
status of the device which backs the mapping.  The read_congested and
write_congested fields are used to determine whether a read or write
against that device may block.

We use this infrastructure to

a) allow pdflush to service many queues in parallel (by not getting
   stuck on any particular one) and

b) to avoid undesirable and uncontrolled latencies in places such as
   page reclaim and

c) To avoid blocking in readahead operations

The current code only supports simple disk queues (and I have a patch here
for NFS).  Stacked queues (MD and DM) don't get this information right and
problems were expected.  Efficiency problems have now been noted and it's
time to fix it.

This patch lays down the infrastructure which permits the queue
implementation to get control when someone at a higher level is querying
the queue's congestion state.  So DM (for example) can run around and
examine all the queues which contribute to the higher-level queue.


It also adds bdi_rw_congested() for code in xfs and ext2 that calls both
bdi_read_congested() and bdi_write_congested() in a row, and it was "free"
anyway.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

The rewritten qeth network driver.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

The crypto device driver for PCICA & PCICC cards, part 2.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

The crypto device driver for PCICA & PCICC cards, part 1.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

zfcp host adapter log message cleanup part 2:
 - Shorten log output.
 - Increase log level for some messages.
 - Always print leading zeroes for wwpn and fcp-lun.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

zfcp host adapter log message cleanup part 1:
 - Shorten log output.
 - Increase log level for some messages.
 - Always print leading zeroes for wwpn and fcp-lun.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

zfcp host adapter fixes:
 - Reuse freed scsi_ids and scsi_luns for mappings.
 - Order list of ports/units by assigned scsi_id/scsi_lun.
 - Don't update max_id/max_lun in scsi_host anymore.
 - Get rid of all magics.
 - Add owner field to ccw_driver structure.
 - Avoid deadlock on bus->subsys.rwsem.
 - Use a macro for all scsi device sysfs attributes.
 - Change proc_name from "dummy" to "zfcp".
 - Don't wait for scsi_add_device to complete while holding a semaphore.
 - Cleanup include files in zfcp_aux.c & zfcp_def.h.
 - Get rid of zfcp_erp_fsf_req_handler.
 - Proper link up/down handling.
 - Avoid possible NULL pointer dereference in zfcp_erp_schedule_work.
 - Remove module_exit function. Without an external release function for
   the zfcp_port/zfcp_unit objects module unloading is racy.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

DCSS block device driver changes:
 - Fix remove_store function, put_device is called too early.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

Network driver changes:
 - ctc: move kfree of driver structure after the last use of it.
 - netiucv: stay in state startwait if peer is down.
 - lcs: initialize ipm_list and unregister netdev only if it is present.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

dasd driver changes:
 - Fix check for device type in error recovery for fba devices.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

Tape driver changes:
 - Add missing break in tape_34xx_work_handler to avoid misleading message.
 - Cleanup offline/remove code.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

Common i/o layer changes:
 - Avoid de-registering a ccwgroup device multiple times.
 - Remove check for channel path objects in get_subchannel_by_schid.
   Channel patch objects are never in the bus list.
 - Avoid NULL pointer deref. in qdio_unmark_q.
 - Fix reference counting on subchannel objects.
 - Add shutdown function to terminate i/o and disable subchannels at reipl.
 - Remove all ccwgroup devices if the ccwgroup driver is unregistered.

From: Martin Schwidefsky <schwidefsky@de.ibm.com>

s390 core changes:
 - Fix _raw_spin_trylock for 64 bit.
 - Add clarification to s390 debug debug documentation.

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

The following patch consolidates redundant code in various hugetlb
implementations.  I took the liberty of renaming a few things, since the
code was all moved anyway, and it has the benefit of helping to catch
missed conversions and/or consolidations.

From: Arjan van de Ven <arjanv@redhat.com>

patch below fixes a missing \n in a printk; without this you get to see a
<4> in the middle of that line...

From: <gerg@snapgear.com>

Add support for boards that have a 64MHz clock to common Coldfire header.

From: <gerg@snapgear.com>

Add type specifier to printk calls in 68EZ328/ucdimm setup code.  Patch
original from kernel janitors.

From: <gerg@snapgear.com>

Clean up debug trace in startup code of 68EZ328 DragonEngine board.

From: <gerg@snapgear.com>

A couple of fixes for the DragonEngine sepcific setup code:

. remove cs8900 ethernet setup from here
. add type specifier to printk calls (from kernel janitors)

From: <gerg@snapgear.com>

Some fixes for the 68360 common ints management code:

. use irqreturn_t for return type of interrupt handlers
. add type field to printk calls (from kernel janitors)
. there is no loop in show_interrupts(), don't use continue

From: <gerg@snapgear.com>

Some fixes for the 68328 common ints management code:

. use irqreturn_t for return type of interrupt handlers
. clean up asm code to be gcc-3.3.x clean
. add type field to printk calls (from kernel janitors)
. there is no loop in show_interrupts(), don't use continue

From: <gerg@snapgear.com>

A number of small fixes for the Motorola 68328 setup code:

. fix interrupt routine return types to be irqreturn_t
. add type specifier to printk calls (from kernel janitors)
. rework asm code to be gcc-3.3.x clean

From: <gerg@snapgear.com>

Fixes to the Motorola ColdFire 5407 setup code:

. fix interrupt routine return types to be irqreturn_t
. add DMA base addresses array
. support compile time setting of kernel boot arguments

From: <gerg@snapgear.com>

Add type specifier to printk calls.  Patch originally from kernel janitors.