Running AIO is pinning inode in memory using file reference. Once AIO
is completed using aio_complete(), file reference is put and inode can
be freed from memory. So we have to be sure that calling aio_complete()
is the last thing we do with the inode.

CC: Christoph Hellwig <hch@infradead.org>
CC: Jens Axboe <axboe@kernel.dk>
CC: Jeff Moyer <jmoyer@redhat.com>
CC: stable@vger.kernel.org
Acked-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Since directio can work on a raw block device, and the block size of the
device can change under it, we need to do the same thing that
fs/buffer.c now does: read the block size a single time, using
ACCESS_ONCE().

Reading it multiple times can get different results, which will then
confuse the code because it actually encodes the i_blksize in
relationship to the underlying logical blocksize.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Move unplugging for direct I/O from around ->direct_IO() down to
do_blockdev_direct_IO(). This implicitly adds plugging for direct
writes.

CC: Li Shaohua <shli@fusionio.com>
Acked-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

READ is 0, so the result of the bit-and operation is 0.  Rewrite with == as
done elsewhere in the same file.

This problem was found using Coccinelle (http://coccinelle.lip6.fr/).
Signed-off-by: NJulia Lawall <julia@diku.dk>
Reviewed-by: NJeff Moyer <jmoyer@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Use the same mechanism as the block devices are using, but move the
helper functions from fs/direct-io.c into fs/inode.c to remove the
dependency on CONFIG_BLOCK.
Signed-off-by: NTrond Myklebust <Trond.Myklebust@netapp.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Fred Isaman <iisaman@netapp.com>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

With kernel 3.1, Christoph removed i_alloc_sem and replaced it with
calls (namely inode_dio_wait() and inode_dio_done()) which are
EXPORT_SYMBOL_GPL() thus they cannot be used by non-GPL file systems and
further inode_dio_wait() was pushed from notify_change() into the file
system ->setattr() method but no non-GPL file system can make this call.

That means non-GPL file systems cannot exist any more unless they do not
use any VFS functionality related to reading/writing as far as I can
tell or at least as long as they want to implement direct i/o.

Both Linus and Al (and others) have said on LKML that this breakage of
the VFS API should not have happened and that the change was simply
missed as it was not documented in the change logs of the patches that
did those changes.

This patch changes the two function exports in question to be
EXPORT_SYMBOL() thus restoring the VFS API as it used to be - accessible
for all modules.

Christoph, who introduced the two functions and exported them GPL-only
is CC-ed on this patch to give him the opportunity to object to the
symbols being changed in this manner if he did indeed intend them to be
GPL-only and does not want them to become available to all modules.
Signed-off-by: NAnton Altaparmakov <anton@tuxera.com>
CC: Christoph Hellwig <hch@infradead.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Some investigation of a transaction processing workload showed that a
major consumer of cycles in __blockdev_direct_IO is the cache miss while
accessing the block size.  This is because it has to walk the chain from
block_dev to gendisk to queue.

The block size is needed early on to check alignment and sizes.  It's only
done if the check for the inode block size fails.  But the costly block
device state is unconditionally fetched.

- Reorganize the code to only fetch block dev state when actually
  needed.

Then do a prefetch on the block dev early on in the direct IO path.  This
is worth it, because there is substantial code run before we actually
touch the block dev now.

- I also added some unlikelies to make it clear the compiler that block
  device fetch code is not normally executed.

This gave a small, but measurable improvement on a large database
benchmark (about 0.3%)

[akpm@linux-foundation.org: coding-style fixes]
[sfr@canb.auug.org.au: using prefetch requires including prefetch.h]
Signed-off-by: NAndi Kleen <ak@linux.intel.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: NStephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

In get_more_blocks(), we use dio_count to calcuate fs_count and do some
tricky things to increase fs_count if dio_count isn't aligned.  But
actually it still has some corner cases that can't be coverd.  See the
following example:

	dio_write foo -s 1024 -w 4096

(direct write 4096 bytes at offset 1024).  The same goes if the offset
isn't aligned to fs_blocksize.

In this case, the old calculation counts fs_count to be 1, but actually we
will write into 2 different blocks (if fs_blocksize=4096).  The old code
just works, since it will call get_block twice (and may have to allocate
and create extents twice for filesystems like ext4).  So we'd better call
get_block just once with the proper fs_count.
Signed-off-by: NTao Ma <boyu.mt@taobao.com>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

This doesn't change anything for the compiler, but hch thought it would
make the code clearer.

I moved the reference counting into its own little inline.
Signed-off-by: NAndi Kleen <ak@linux.intel.com>
Acked-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Add inlines to all the submission path functions. While this increases
code size it also gives gcc a lot of optimization opportunities
in this critical hotpath.

In particular -- together with some other changes -- this
allows gcc to get rid of the unnecessary clearing of
sdio at the beginning and optimize the messy parameter passing.
Any non inlining of a function which takes a sdio parameter
would break this optimization because they cannot be done if the
address of a structure is taken.

Note that benefits are only seen with CONFIG_OPTIMIZE_INLINING
and CONFIG_CC_OPTIMIZE_FOR_SIZE both set to off.

This gives about 2.2% improvement on a large database benchmark
with a high IOPS rate.
Signed-off-by: NAndi Kleen <ak@linux.intel.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Only a single b_private field in the map_bh buffer head is needed after
the submission path. Move map_bh separately to avoid storing
this information in the long term slab.

This avoids the weird 104 byte hole in struct dio_submit which also needed
to be memseted early.
Signed-off-by: NAndi Kleen <ak@linux.intel.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

A direct slab call is slightly faster than kmalloc and can be better cached
per CPU. It also avoids rounding to the next kmalloc slab.

In addition this enforces cache line alignment for struct dio to avoid
any false sharing.
Signed-off-by: NAndi Kleen <ak@linux.intel.com>
Acked-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Fix most problems reported by pahole.

There is still a weird 104 byte hole after map_bh. I'm not sure what
causes this.
Signed-off-by: NAndi Kleen <ak@linux.intel.com>
Acked-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

There's nothing on the stack, even before my changes.
Signed-off-by: NAndi Kleen <ak@linux.intel.com>
Acked-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

This large, but largely mechanic, patch moves all fields in struct dio
that are only used in the submission path into a separate on stack
data structure. This has the advantage that the memory is very likely
cache hot, which is not guaranteed for memory fresh out of kmalloc.

This also gives gcc more optimization potential because it can easier
determine that there are no external aliases for these variables.

The sdio initialization is a initialization now instead of memset.
This allows gcc to break sdio into individual fields and optimize
away unnecessary zeroing (after all the functions are inlined)
Signed-off-by: NAndi Kleen <ak@linux.intel.com>
Acked-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

This allows us to move duplicated code in <asm/atomic.h>
(atomic_inc_not_zero() for now) to <linux/atomic.h>
Signed-off-by: NArun Sharma <asharma@fb.com>
Reviewed-by: NEric Dumazet <eric.dumazet@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: NMike Frysinger <vapier@gentoo.org>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

For filesystems that delay their end_io processing we should keep our
i_dio_count until the the processing is done.  Enable this by moving
the inode_dio_done call to the end_io handler if one exist.  Note that
the actual move to the workqueue for ext4 and XFS is not done in
this patch yet, but left to the filesystem maintainers.  At least
for XFS it's not needed yet either as XFS has an internal equivalent
to i_dio_count.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Maintain i_dio_count for all filesystems, not just those using DIO_LOCKING.
This these filesystems to also protect truncate against direct I/O requests
by using common code.  Right now the only non-DIO_LOCKING filesystem that
appears to do so is XFS, which uses an opencoded variant of the i_dio_count
scheme.

Behaviour doesn't change for filesystems never calling inode_dio_wait.
For ext4 behaviour changes when using the dioread_nonlock option, which
previously was missing any protection between truncate and direct I/O reads.
For ocfs2 that handcrafted i_dio_count manipulations are replaced with
the common code now enable.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

i_alloc_sem is a rather special rw_semaphore.  It's the last one that may
be released by a non-owner, and it's write side is always mirrored by
real exclusion.  It's intended use it to wait for all pending direct I/O
requests to finish before starting a truncate.

Replace it with a hand-grown construct:

 - exclusion for truncates is already guaranteed by i_mutex, so it can
   simply fall way
 - the reader side is replaced by an i_dio_count member in struct inode
   that counts the number of pending direct I/O requests.  Truncate can't
   proceed as long as it's non-zero
 - when i_dio_count reaches non-zero we wake up a pending truncate using
   wake_up_bit on a new bit in i_flags
 - new references to i_dio_count can't appear while we are waiting for
   it to read zero because the direct I/O count always needs i_mutex
   (or an equivalent like XFS's i_iolock) for starting a new operation.

This scheme is much simpler, and saves the space of a spinlock_t and a
struct list_head in struct inode (typically 160 bits on a non-debug 64-bit
system).
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Reject zero sized reads as soon as we know our I/O length, and don't
borther with locks or allocations that might have to be cleaned up
otherwise.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

With the plugging now being explicitly controlled by the
submitter, callers need not pass down unplugging hints
to the block layer. If they want to unplug, it's because they
manually plugged on their own - in which case, they should just
unplug at will.
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Code has been converted over to the new explicit on-stack plugging,
and delay users have been converted to use the new API for that.
So lets kill off the old plugging along with aops->sync_page().
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

When using devices that support max_segments > BIO_MAX_PAGES (256), direct
IO tries to allocate a bio with more pages than allowed, which leads to an
oops in dio_bio_alloc().  Clamp the request to the supported maximum, and
change dio_bio_alloc() to reflect that bio_alloc() will always return a
bio when called with __GFP_WAIT and a valid number of vectors.

[akpm@linux-foundation.org: remove redundant BUG_ON()]
Signed-off-by: NDavid Dillow <dillowda@ornl.gov>
Reviewed-by: NJeff Moyer <jmoyer@redhat.com>
Cc: <stable@kernel.org>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Signed-off-by: NNamhyung Kim <namhyung@gmail.com>
Cc: Jiri Kosina <trivial@kernel.org>
Signed-off-by: NJiri Kosina <jkosina@suse.cz>

Fix up truncation (ssize_t->int).  This only matters with >2G
reads/writes, which the kernel doesn't permit.
Signed-off-by: NEdward Shishkin <edward@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

commit c2c6ca41 (direct-io: do not merge logically non-contiguous requests)
introduced a bug whereby all O_DIRECT I/Os were submitted a page at a time
to the block layer.  The problem is that the code expected
dio->block_in_file to correspond to the current page in the dio.  In fact,
it corresponds to the previous page submitted via submit_page_section.
This was purely an oversight, as the dio->cur_page_fs_offset field was
introduced for just this purpose.  This patch simply uses the correct
variable when calculating whether there is a mismatch between contiguous
logical blocks and contiguous physical blocks (as described in the
comments).

I also switched the if conditional following this check to an else if, to
ensure that we never call dio_bio_submit twice for the same dio (in
theory, this should not happen, anyway).

I've tested this by running blktrace and verifying that a 64KB I/O was
submitted as a single I/O.  I also ran the patched kernel through
xfstests' aio tests using xfs, ext4 (with 1k and 4k block sizes) and btrfs
and verified that there were no regressions as compared to an unpatched
kernel.
Signed-off-by: NJeff Moyer <jmoyer@redhat.com>
Acked-by: NJosef Bacik <jbacik@redhat.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: <stable@kernel.org>		[2.6.35.x]
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Move the call to vmtruncate to get rid of accessive blocks to the callers
in prepearation of the new truncate calling sequence.  This was only done
for DIO_LOCKING filesystems, so the __blockdev_direct_IO_newtrunc variant
was not needed anyway.  Get rid of blockdev_direct_IO_no_locking and
its _newtrunc variant while at it as just opencoding the two additional
paramters is shorted than the name suffix.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Filesystems with unwritten extent support must not complete an AIO request
until the transaction to convert the extent has been commited.  That means
the aio_complete calls needs to be moved into the ->end_io callback so
that the filesystem can control when to call it exactly.

This makes a bit of a mess out of dio_complete and the ->end_io callback
prototype even more complicated. 
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz> 
Signed-off-by: N"Theodore Ts'o" <tytso@mit.edu>

Filesystems with unwritten extent support must not complete an AIO request
until the transaction to convert the extent has been commited.  That means
the aio_complete calls needs to be moved into the ->end_io callback so
that the filesystem can control when to call it exactly.

This makes a bit of a mess out of dio_complete and the ->end_io callback
prototype even more complicated.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJan Kara <jack@suse.cz>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Introduce a new truncate calling sequence into fs/mm subsystems. Rather than
setattr > vmtruncate > truncate, have filesystems call their truncate sequence
from ->setattr if filesystem specific operations are required. vmtruncate is
deprecated, and truncate_pagecache and inode_newsize_ok helpers introduced
previously should be used.

simple_setattr is introduced for simple in-ram filesystems to implement
the new truncate sequence. Eventually all filesystems should be converted
to implement a setattr, and the default code in notify_change should go
away.

simple_setsize is also introduced to perform just the ATTR_SIZE portion
of simple_setattr (ie. changing i_size and trimming pagecache).

To implement the new truncate sequence:
- filesystem specific manipulations (eg freeing blocks) must be done in
  the setattr method rather than ->truncate.
- vmtruncate can not be used by core code to trim blocks past i_size in
  the event of write failure after allocation, so this must be performed
  in the fs code.
- convert usage of helpers block_write_begin, nobh_write_begin,
  cont_write_begin, and *blockdev_direct_IO* to use _newtrunc postfixed
  variants. These avoid calling vmtruncate to trim blocks (see previous).
- inode_setattr should not be used. generic_setattr is a new function
  to be used to copy simple attributes into the generic inode.
- make use of the better opportunity to handle errors with the new sequence.

Big problem with the previous calling sequence: the filesystem is not called
until i_size has already changed.  This means it is not allowed to fail the
call, and also it does not know what the previous i_size was. Also, generic
code calling vmtruncate to truncate allocated blocks in case of error had
no good way to return a meaningful error (or, for example, atomically handle
block deallocation).

Cc: Christoph Hellwig <hch@lst.de>
Acked-by: NJan Kara <jack@suse.cz>
Signed-off-by: NNick Piggin <npiggin@suse.de>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Btrfs cannot handle having logically non-contiguous requests submitted.  For
example if you have

Logical:  [0-4095][HOLE][8192-12287]
Physical: [0-4095]      [4096-8191]

Normally the DIO code would put these into the same BIO's.  The problem is we
need to know exactly what offset is associated with what BIO so we can do our
checksumming and unlocking properly, so putting them in the same BIO doesn't
work.  So add another check where we submit the current BIO if the physical
blocks are not contigous OR the logical blocks are not contiguous.
Signed-off-by: NJosef Bacik <josef@redhat.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

Because BTRFS can do RAID and such, we need our own submit hook so we can setup
the bio's in the correct fashion, and handle checksum errors properly.  So there
are a few changes here

1) The submit_io hook.  This is straightforward, just call this instead of
submit_bio.

2) Allow the fs to return -ENOTBLK for reads.  Usually this has only worked for
writes, since writes can fallback onto buffered IO.  But BTRFS needs the option
of falling back on buffered IO if it encounters a compressed extent, since we
need to read the entire extent in and decompress it.  So if we get -ENOTBLK back
from get_block we'll return back and fallback on buffered just like the write
case.

I've tested these changes with fsx and everything seems to work.  Thanks,
Signed-off-by: NJosef Bacik <josef@redhat.com>
Signed-off-by: NChris Mason <chris.mason@oracle.com>

Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Currently the locking in blockdev_direct_IO is a mess, we have three different
locking types and very confusing checks for some of them.  The most
complicated one is DIO_OWN_LOCKING for reads, which happens to not actually be
used.

This patch gets rid of the DIO_OWN_LOCKING - as mentioned above the read case
is unused anyway, and the write side is almost identical to DIO_NO_LOCKING.
The difference is that DIO_NO_LOCKING always sets the create argument for
the get_blocks callback to zero, but we can easily move that to the actual
get_blocks callbacks.  There are four users of the DIO_NO_LOCKING mode:
gfs already ignores the create argument and thus is fine with the new
version, ocfs2 only errors out if create were ever set, and we can remove
this dead code now, the block device code only ever uses create for an
error message if we are fully beyond the device which can never happen,
and last but not least XFS will need the new behavour for writes.

Now we can replace the lock_type variable with a flags one, where no flag
means the DIO_NO_LOCKING behaviour and DIO_LOCKING is kept as the first
flag.  Separate out the check for not allowing to fill holes into a separate
flag, although for now both flags always get set at the same time.

Also revamp the documentation of the locking scheme to actually make sense.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Currently the locking in blockdev_direct_IO is a mess, we have three
different locking types and very confusing checks for some of them.  The
most complicated one is DIO_OWN_LOCKING for reads, which happens to not
actually be used.

This patch gets rid of the DIO_OWN_LOCKING - as mentioned above the read
case is unused anyway, and the write side is almost identical to
DIO_NO_LOCKING.  The difference is that DIO_NO_LOCKING always sets the
create argument for the get_blocks callback to zero, but we can easily
move that to the actual get_blocks callbacks.  There are four users of the
DIO_NO_LOCKING mode: gfs already ignores the create argument and thus is
fine with the new version, ocfs2 only errors out if create were ever set,
and we can remove this dead code now, the block device code only ever uses
create for an error message if we are fully beyond the device which can
never happen, and last but not least XFS will need the new behavour for
writes.

Now we can replace the lock_type variable with a flags one, where no flag
means the DIO_NO_LOCKING behaviour and DIO_LOCKING is kept as the first
flag.  Separate out the check for not allowing to fill holes into a
separate flag, although for now both flags always get set at the same
time.

Also revamp the documentation of the locking scheme to actually make
sense.

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Zach Brown <zach.brown@oracle.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Alex Elder <aelder@sgi.com>
Cc: Mark Fasheh <mfasheh@suse.com>
Cc: Joel Becker <joel.becker@oracle.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Intel reported a performance regression caused by the following commit:

commit 848c4dd5
Author: Zach Brown <zach.brown@oracle.com>
Date:   Mon Aug 20 17:12:01 2007 -0700

    dio: zero struct dio with kzalloc instead of manually

    This patch uses kzalloc to zero all of struct dio rather than
    manually trying to track which fields we rely on being zero.  It
    passed aio+dio stress testing and some bug regression testing on
    ext3.

    This patch was introduced by Linus in the conversation that lead up
    to Badari's minimal fix to manually zero .map_bh.b_state in commit:

      6a648fa7

    It makes the code a bit smaller.  Maybe a couple fewer cachelines to
    load, if we're lucky:

       text    data     bss     dec     hex filename
    3285925  568506 1304616 5159047  4eb887 vmlinux
    3285797  568506 1304616 5158919  4eb807 vmlinux.patched

    I was unable to measure a stable difference in the number of cpu
    cycles spent in blockdev_direct_IO() when pushing aio+dio 256K reads
    at ~340MB/s.

    So the resulting intent of the patch isn't a performance gain but to
    avoid exposing ourselves to the risk of finding another field like
    .map_bh.b_state where we rely on zeroing but don't enforce it in the
    code.

Zach surmised that zeroing out the page array was what caused most of
the problem, and suggested the approach taken in the attached patch for
resolving the issue.  Intel re-tested with this patch and saw a 0.6%
performance gain (the original regression was 0.5%).

[akpm@linux-foundation.org: add comment]
Signed-off-by: NJeff Moyer <jmoyer@redhat.com>
Acked-by: NZach Brown <zach.brown@oracle.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

There seems to be a regression in direct write path due to following
commit in for-2.6.33 branch of block tree.

commit 1af60fbd
Author: Jeff Moyer <jmoyer@redhat.com>
Date:   Fri Oct 2 18:56:53 2009 -0400

    block: get rid of the WRITE_ODIRECT flag

Marking direct writes as WRITE_SYNC_PLUG instead of WRITE_ODIRECT, sets
the NOIDLE flag in bio and hence in request. This tells CFQ to not expect
more request from the queue and not idle on it (despite the fact that
queue's think time is less and it is not seeky).

So direct writers lose big time when competing with sequential readers.

Using fio, I have run one direct writer and two sequential readers and
following are the results with 2.6.32-rc7 kernel and with for-2.6.33
branch.

Test
====
1 direct writer and 2 sequential reader running simultaneously.

[global]
directory=/mnt/sdc/fio/
runtime=10

[seqwrite]
rw=write
size=4G
direct=1

[seqread]
rw=read
size=2G
numjobs=2

2.6.32-rc7
==========
direct writes: aggrb=2,968KB/s
readers	     : aggrb=101MB/s

for-2.6.33 branch
=================
direct write: aggrb=19KB/s
readers	      aggrb=137MB/s

This patch brings back the WRITE_ODIRECT flag, with the difference that we
don't set the BIO_RW_UNPLUG flag so that device is not unplugged after
submission of request and an explicit unplug from submitter is required.

That way we fix the jeff's issue of not enough merging taking place in aio
path as well as make sure direct writes get their fair share.

After the fix
=============
for-2.6.33 + fix
----------------
direct writes: aggrb=2,728KB/s
reads: aggrb=103MB/s

Thanks
Vivek
Signed-off-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Hi,

Some workloads issue batches of small I/O, and the performance is poor
due to the call to blk_run_address_space for every single iocb.  Nathan
Roberts pointed this out, and suggested that by deferring this call
until all I/Os in the iocb array are submitted to the block layer, we
can realize some impressive performance gains (up to 30% for sequential
4k reads in batches of 16).
Signed-off-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Hi,

The WRITE_ODIRECT flag is only used in one place, and that code path
happens to also call blk_run_address_space.  The introduction of this
flag, then, could result in the device being unplugged twice for every
I/O.

Further, with the batching changes in the next patch, we don't want an
O_DIRECT write to imply a queue unplug.
Signed-off-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Until now we have had a 1:1 mapping between storage device physical
block size and the logical block sized used when addressing the device.
With SATA 4KB drives coming out that will no longer be the case.  The
sector size will be 4KB but the logical block size will remain
512-bytes.  Hence we need to distinguish between the physical block size
and the logical ditto.

This patch renames hardsect_size to logical_block_size.
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>