Commit Graph

41 Commits

Author SHA1 Message Date
Al Viro 06777d308f dio: fix use-after-free
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2009-12-17 04:52:13 -05:00
Christoph Hellwig 5fe878ae7f direct-io: cleanup blockdev_direct_IO locking
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: Christoph 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: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-16 07:20:13 -08:00
Jeff Moyer 23aee091d8 dio: don't zero out the pages array inside struct dio
Intel reported a performance regression caused by the following commit:

commit 848c4dd515
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:

      6a648fa721

    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: Jeff Moyer <jmoyer@redhat.com>
Acked-by: Zach Brown <zach.brown@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-16 07:20:13 -08:00
Vivek Goyal d9449ce35a Fix regression in direct writes performance due to WRITE_ODIRECT flag removal
There seems to be a regression in direct write path due to following
commit in for-2.6.33 branch of block tree.

commit 1af60fbd75
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: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-11-26 09:46:46 +01:00
Jeff Moyer cfb1e33eed aio: implement request batching
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: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-10-28 09:29:25 +01:00
Jeff Moyer 1af60fbd75 block: get rid of the WRITE_ODIRECT flag
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: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-10-28 09:29:25 +01:00
Martin K. Petersen e1defc4ff0 block: Do away with the notion of hardsect_size
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: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-05-22 23:22:54 +02:00
Nikanth Karthikesan 4d1f9fdb61 dio: Remove code handling bio_alloc failure with __GFP_WAIT
Remove code handling bio_alloc failure with __GFP_WAIT.
GFP_KERNEL implies __GFP_WAIT.

Signed-off-by: Nikanth Karthikesan <knikanth@suse.de>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-04-15 12:10:13 +02:00
Jens Axboe aeb6fafb8f block: Add flag for telling the IO schedulers NOT to anticipate more IO
By default, CFQ will anticipate more IO from a given io context if the
previously completed IO was sync. This used to be fine, since the only
sync IO was reads and O_DIRECT writes. But with more "normal" sync writes
being used now, we don't want to anticipate for those.

Add a bio/request flag that informs the IO scheduler that this is a sync
request that we should not idle for. Introduce WRITE_ODIRECT specifically
for O_DIRECT writes, and make sure that the other sync writes set this
flag.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-04-06 08:04:54 -07:00
Dmitri Monakhov 0f64415d42 fs: truncate blocks outside i_size after O_DIRECT write error
In case of error extending write may have instantiated a few blocks
outside i_size.  We need to trim these blocks.  We have to do it
*regardless* to blocksize.  At least ext2, ext3 and reiserfs interpret
(i_size < biggest block) condition as error.  Fsck will complain about
wrong i_size.  Then fsck will fix the error by changing i_size according
to the biggest block.  This is bad because this blocks contain garbage
from previous write attempt.  And result in data corruption.

####TESTCASE_BEGIN
$touch /mnt/test/BIG_FILE
## at this moment /mnt/test/BIG_FILE size and blocks equal to zero
open("/mnt/test/BIG_FILE", O_WRONLY|O_CREAT|O_DIRECT, 0666) = 3
write(3, "aaaaaaaaaaaa"..., 104857600) = -1 ENOSPC (No space left on device)
## size and block sould't be changed because write op failed.
$stat /mnt/test/BIG_FILE
File: `/mnt/test/BIG_FILE'
Size: 0 Blocks: 110896 IO Block: 1024 regular empty file
<<<<<<<<^^^^^^^^^^^^^^^^^^^^^^^^^^^^^file size is less than biggest block idx
Device: fe07h/65031d Inode: 14 Links: 1
Access: (0644/-rw-r--r--) Uid: ( 0/ root) Gid: ( 0/ root)
Access: 2007-01-24 20:03:38.000000000 +0300
Modify: 2007-01-24 20:03:38.000000000 +0300
Change: 2007-01-24 20:03:39.000000000 +0300

#fsck.ext3 -f /dev/VG/test
e2fsck 1.39 (29-May-2006)
Pass 1: Checking inodes, blocks, and sizes
Inode 14, i_size is 0, should be 56556544. Fix<y>? yes
Pass 2: Checking directory structure
....
#####TESTCASE_ENDdiff --git a/fs/direct-io.c b/fs/direct-io.c
index af0558d..4e88bea 100644

[akpm@linux-foundation.org: use i_size_read()]
Signed-off-by: Dmitri Monakhov <dmonakhov@openvz.org>
Cc: Zach Brown <zach.brown@oracle.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 15:59:06 -08:00
Francois Cami e1f8e87449 Remove Andrew Morton's old email accounts
People can use the real name an an index into MAINTAINERS to find the
current email address.

Signed-off-by: Francois Cami <francois.cami@free.fr>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-16 11:21:32 -07:00
Nick Piggin f5dd33c494 dio: use get_user_pages_fast
Use get_user_pages_fast in the common/generic block and fs direct IO paths.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Dave Kleikamp <shaggy@austin.ibm.com>
Cc: Andy Whitcroft <apw@shadowen.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Dave Kleikamp <shaggy@austin.ibm.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Zach Brown <zach.brown@oracle.com>
Cc: Jens Axboe <jens.axboe@oracle.com>
Reviewed-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-26 12:00:06 -07:00
Christoph Lameter eebd2aa355 Pagecache zeroing: zero_user_segment, zero_user_segments and zero_user
Simplify page cache zeroing of segments of pages through 3 functions

zero_user_segments(page, start1, end1, start2, end2)

        Zeros two segments of the page. It takes the position where to
        start and end the zeroing which avoids length calculations and
	makes code clearer.

zero_user_segment(page, start, end)

        Same for a single segment.

zero_user(page, start, length)

        Length variant for the case where we know the length.

We remove the zero_user_page macro. Issues:

1. Its a macro. Inline functions are preferable.

2. The KM_USER0 macro is only defined for HIGHMEM.

   Having to treat this special case everywhere makes the
   code needlessly complex. The parameter for zeroing is always
   KM_USER0 except in one single case that we open code.

Avoiding KM_USER0 makes a lot of code not having to be dealing
with the special casing for HIGHMEM anymore. Dealing with
kmap is only necessary for HIGHMEM configurations. In those
configurations we use KM_USER0 like we do for a series of other
functions defined in highmem.h.

Since KM_USER0 is depends on HIGHMEM the existing zero_user_page
function could not be a macro. zero_user_* functions introduced
here can be be inline because that constant is not used when these
functions are called.

Also extract the flushing of the caches to be outside of the kmap.

[akpm@linux-foundation.org: fix nfs and ntfs build]
[akpm@linux-foundation.org: fix ntfs build some more]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Cc: Steven French <sfrench@us.ibm.com>
Cc: Michael Halcrow <mhalcrow@us.ibm.com>
Cc: <linux-ext4@vger.kernel.org>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Cc: Anton Altaparmakov <aia21@cantab.net>
Cc: Mark Fasheh <mark.fasheh@oracle.com>
Cc: David Chinner <dgc@sgi.com>
Cc: Michael Halcrow <mhalcrow@us.ibm.com>
Cc: Steven French <sfrench@us.ibm.com>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-05 09:44:13 -08:00
Nick Piggin 557ed1fa26 remove ZERO_PAGE
The commit b5810039a5 contains the note

  A last caveat: the ZERO_PAGE is now refcounted and managed with rmap
  (and thus mapcounted and count towards shared rss).  These writes to
  the struct page could cause excessive cacheline bouncing on big
  systems.  There are a number of ways this could be addressed if it is
  an issue.

And indeed this cacheline bouncing has shown up on large SGI systems.
There was a situation where an Altix system was essentially livelocked
tearing down ZERO_PAGE pagetables when an HPC app aborted during startup.
This situation can be avoided in userspace, but it does highlight the
potential scalability problem with refcounting ZERO_PAGE, and corner
cases where it can really hurt (we don't want the system to livelock!).

There are several broad ways to fix this problem:
1. add back some special casing to avoid refcounting ZERO_PAGE
2. per-node or per-cpu ZERO_PAGES
3. remove the ZERO_PAGE completely

I will argue for 3. The others should also fix the problem, but they
result in more complex code than does 3, with little or no real benefit
that I can see.

Why? Inserting a ZERO_PAGE for anonymous read faults appears to be a
false optimisation: if an application is performance critical, it would
not be doing many read faults of new memory, or at least it could be
expected to write to that memory soon afterwards. If cache or memory use
is critical, it should not be working with a significant number of
ZERO_PAGEs anyway (a more compact representation of zeroes should be
used).

As a sanity check -- mesuring on my desktop system, there are never many
mappings to the ZERO_PAGE (eg. 2 or 3), thus memory usage here should not
increase much without it.

When running a make -j4 kernel compile on my dual core system, there are
about 1,000 mappings to the ZERO_PAGE created per second, but about 1,000
ZERO_PAGE COW faults per second (less than 1 ZERO_PAGE mapping per second
is torn down without being COWed). So removing ZERO_PAGE will save 1,000
page faults per second when running kbuild, while keeping it only saves
less than 1 page clearing operation per second. 1 page clear is cheaper
than a thousand faults, presumably, so there isn't an obvious loss.

Neither the logical argument nor these basic tests give a guarantee of no
regressions. However, this is a reasonable opportunity to try to remove
the ZERO_PAGE from the pagefault path. If it is found to cause regressions,
we can reintroduce it and just avoid refcounting it.

The /dev/zero ZERO_PAGE usage and TLB tricks also get nuked.  I don't see
much use to them except on benchmarks.  All other users of ZERO_PAGE are
converted just to use ZERO_PAGE(0) for simplicity. We can look at
replacing them all and maybe ripping out ZERO_PAGE completely when we are
more satisfied with this solution.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus "snif" Torvalds <torvalds@linux-foundation.org>
2007-10-16 09:42:53 -07:00
NeilBrown 6712ecf8f6 Drop 'size' argument from bio_endio and bi_end_io
As bi_end_io is only called once when the reqeust is complete,
the 'size' argument is now redundant.  Remove it.

Now there is no need for bio_endio to subtract the size completed
from bi_size.  So don't do that either.

While we are at it, change bi_end_io to return void.

Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2007-10-10 09:25:57 +02:00
Zach Brown 848c4dd515 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:

  6a648fa721

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.

Signed-off-by: Zach Brown <zach.brown@oracle.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-08-20 22:50:25 -07:00
Badari Pulavarty 6a648fa721 direct-io: fix error-path crashes
Need to initialize map_bh.b_state to zero.  Otherwise, in case of a faulty
user-buffer its possible to go into dio_zero_block() and submit a page by
mistake - since it checks for buffer_new().

http://marc.info/?l=linux-kernel&m=118551339032528&w=2

akpm: Linus had a (better) patch to just do a kzalloc() in there, but it got
lost.  Probably this version is better for -stable anwyay.

Signed-off-by: Badari Pulavarty <pbadari@us.ibm.com>
Acked-by: Joe Jin <joe.jin@oracle.com>
Acked-by: Zach Brown <zach.brown@oracle.com>
Cc: gurudas pai <gurudas.pai@oracle.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-08-11 15:47:40 -07:00
Zach Brown fcb82f8835 dio: remove bogus refcounting BUG_ON
Badari Pulavarty reported a case of this BUG_ON is triggering during
testing.  It's completely bogus and should be removed.

It's trying to notice if we left references to the dio hanging around in
the sync case.  They should have been dropped as IO completed while this
path was in dio_await_completion().  This condition will also be
checked, via some twisty logic, by the BUG_ON(ret != -EIOCBQUEUED) a few
lines lower.  So to start this BUG_ON() is redundant.

More fatally, it's dereferencing dio-> after having dropped its
reference.  It's only safe to dereference the dio after releasing the
lock if the final reference was just dropped.  Another CPU might free
the dio in bio completion and reuse the memory after this path drops the
dio lock but before the BUG_ON() is evaluated.

This patch passed aio+dio regression unit tests and aio-stress on ext3.

Signed-off-by: Zach Brown <zach.brown@oracle.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-03 18:23:23 -07:00
Linus Torvalds 9a9136e270 Merge git://git.kernel.org/pub/scm/linux/kernel/git/bunk/trivial
* git://git.kernel.org/pub/scm/linux/kernel/git/bunk/trivial: (25 commits)
  sound: convert "sound" subdirectory to UTF-8
  MAINTAINERS: Add cxacru website/mailing list
  include files: convert "include" subdirectory to UTF-8
  general: convert "kernel" subdirectory to UTF-8
  documentation: convert the Documentation directory to UTF-8
  Convert the toplevel files CREDITS and MAINTAINERS to UTF-8.
  remove broken URLs from net drivers' output
  Magic number prefix consistency change to Documentation/magic-number.txt
  trivial: s/i_sem /i_mutex/
  fix file specification in comments
  drivers/base/platform.c: fix small typo in doc
  misc doc and kconfig typos
  Remove obsolete fat_cvf help text
  Fix occurrences of "the the "
  Fix minor typoes in kernel/module.c
  Kconfig: Remove reference to external mqueue library
  Kconfig: A couple of grammatical fixes in arch/i386/Kconfig
  Correct comments in genrtc.c to refer to correct /proc file.
  Fix more "deprecated" spellos.
  Fix "deprecated" typoes.
  ...

Fix trivial comment conflict in kernel/relay.c.
2007-05-09 12:54:17 -07:00
Nate Diller 01f2705daf fs: convert core functions to zero_user_page
It's very common for file systems to need to zero part or all of a page,
the simplist way is just to use kmap_atomic() and memset().  There's
actually a library function in include/linux/highmem.h that does exactly
that, but it's confusingly named memclear_highpage_flush(), which is
descriptive of *how* it does the work rather than what the *purpose* is.
So this patchset renames the function to zero_user_page(), and calls it
from the various places that currently open code it.

This first patch introduces the new function call, and converts all the
core kernel callsites, both the open-coded ones and the old
memclear_highpage_flush() ones.  Following this patch is a series of
conversions for each file system individually, per AKPM, and finally a
patch deprecating the old call.  The diffstat below shows the entire
patchset.

[akpm@linux-foundation.org: fix a few things]
Signed-off-by: Nate Diller <nate.diller@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 12:30:55 -07:00
Robert P. J. Day beb7dd86a1 Fix misspellings collected by members of KJ list.
Fix the misspellings of "propogate", "writting" and (oh, the shame
:-) "kenrel" in the source tree.

Signed-off-by: Robert P. J. Day <rpjday@mindspring.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
2007-05-09 07:14:03 +02:00
Zach Brown 5eb6c7a2ab [PATCH] dio: lock refcount operations
The wait_for_more_bios() function name was poorly chosen.  While looking to
clean it up it I noticed that the dio struct refcounting between the bio
completion and dio submission paths was racey.

The bio submission path was simply freeing the dio struct if
atomic_dec_and_test() indicated that it dropped the final reference.

The aio bio completion path was dereferencing its dio struct pointer *after
dropping its reference* based on the remaining number of references.

These two paths could race and result in the aio bio completion path
dereferencing a freed dio, though this was not observed in the wild.

This moves the refcount under the bio lock so that bio completion can drop
its reference and decide to wake all in one atomic step.

Once testing and waking is locked dio_await_one() can test its sleeping
condition and mark itself uninterruptible under the lock.  It gets simpler
and wait_for_more_bios() disappears.

The addition of the interrupt masking spin lock acquiry in dio_bio_submit()
looks alarming.  This lock acquiry existed in that path before the recent
dio completion patch set.  We shouldn't expect significant performance
regression from returning to the behaviour that existed before the
completion clean up work.

This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine.

Signed-off-by: Zach Brown <zach.brown@oracle.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Suparna Bhattacharya <suparna@in.ibm.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: <xfs-masters@oss.sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 09:57:21 -08:00
Zach Brown 8459d86aff [PATCH] dio: only call aio_complete() after returning -EIOCBQUEUED
The only time it is safe to call aio_complete() is when the ->ki_retry
function returns -EIOCBQUEUED to the AIO core.  direct_io_worker() has
historically done this by relying on its caller to translate positive return
codes into -EIOCBQUEUED for the aio case.  It did this by trying to keep
conditionals in sync.  direct_io_worker() knew when finished_one_bio() was
going to call aio_complete().  It would reverse the test and wait and free the
dio in the cases it thought that finished_one_bio() wasn't going to.

Not surprisingly, it ended up getting it wrong.  'ret' could be a negative
errno from the submission path but it failed to communicate this to
finished_one_bio().  direct_io_worker() would return < 0, it's callers
wouldn't raise -EIOCBQUEUED, and aio_complete() would be called.  In the
future finished_one_bio()'s tests wouldn't reflect this and aio_complete()
would be called for a second time which can manifest as an oops.

The previous cleanups have whittled the sync and async completion paths down
to the point where we can collapse them and clearly reassert the invariant
that we must only call aio_complete() after returning -EIOCBQUEUED.
direct_io_worker() will only return -EIOCBQUEUED when it is not the last to
drop the dio refcount and the aio bio completion path will only call
aio_complete() when it is the last to drop the dio refcount.
direct_io_worker() can ensure that it is the last to drop the reference count
by waiting for bios to drain.  It does this for sync ops, of course, and for
partial dio writes that must fall back to buffered and for aio ops that saw
errors during submission.

This means that operations that end up waiting, even if they were issued as
aio ops, will not call aio_complete() from dio.  Instead we return the return
code of the operation and let the aio core call aio_complete().  This is
purposely done to fix a bug where AIO DIO file extensions would call
aio_complete() before their callers have a chance to update i_size.

Now that direct_io_worker() is explicitly returning -EIOCBQUEUED its callers
no longer have to translate for it.  XFS needs to be careful not to free
resources that will be used during AIO completion if -EIOCBQUEUED is returned.
 We maintain the previous behaviour of trying to write fs metadata for O_SYNC
aio+dio writes.

Signed-off-by: Zach Brown <zach.brown@oracle.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Suparna Bhattacharya <suparna@in.ibm.com>
Acked-by: Jeff Moyer <jmoyer@redhat.com>
Cc: <xfs-masters@oss.sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 09:57:21 -08:00
Zach Brown 20258b2b39 [PATCH] dio: remove duplicate bio wait code
Now that we have a single refcount and waiting path we can reuse it in the
async 'should_wait' path.  It continues to rely on the fragile link between
the conditional in dio_complete_aio() which decides to complete the AIO and
the conditional in direct_io_worker() which decides to wait and free.

By waiting before dropping the reference we stop dio_bio_end_aio() from
calling dio_complete_aio() which used to wake up the waiter after seeing the
reference count drop to 0.  We hoist this wake up into dio_bio_end_aio() which
now notices when it's left a single remaining reference that is held by the
waiter.

Signed-off-by: Zach Brown <zach.brown@oracle.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Suparna Bhattacharya <suparna@in.ibm.com>
Acked-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 09:57:21 -08:00
Zach Brown 0273201e69 [PATCH] dio: formalize bio counters as a dio reference count
Previously we had two confusing counts of bio progress.  'bio_count' was
decremented as bios were processed and freed by the dio core.  It was used to
indicate final completion of the dio operation.  'bios_in_flight' reflected
how many bios were between submit_bio() and bio->end_io.  It was used by the
sync path to decide when to wake up and finish completing bios and was ignored
by the async path.

This patch collapses the two notions into one notion of a dio reference count.
 bios hold a dio reference when they're between submit_bio and bio->end_io.

Since bios_in_flight was only used in the sync path it is now equivalent to
dio->refcount - 1 which accounts for direct_io_worker() holding a reference
for the duration of the operation.

dio_bio_complete() -> finished_one_bio() was called from the sync path after
finding bios on the list that the bio->end_io function had deposited.
finished_one_bio() can not drop the dio reference on behalf of these bios now
because bio->end_io already has.  The is_async test in finished_one_bio()
meant that it never actually did anything other than drop the bio_count for
sync callers.  So we remove its refcount decrement, don't call it from
dio_bio_complete(), and hoist its call up into the async dio_bio_complete()
caller after an explicit refcount decrement.  It is renamed dio_complete_aio()
to reflect the remaining work it actually does.

Signed-off-by: Zach Brown <zach.brown@oracle.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Suparna Bhattacharya <suparna@in.ibm.com>
Acked-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 09:57:21 -08:00
Zach Brown 17a7b1d74b [PATCH] dio: call blk_run_address_space() once per op
We only need to call blk_run_address_space() once after all the bios for the
direct IO op have been submitted.  This removes the chance of calling
blk_run_address_space() after spurious wake ups as the sync path waits for
bios to drain.  It's also one less difference betwen the sync and async paths.

In the process we remove a redundant dio_bio_submit() that its caller had
already performed.

Signed-off-by: Zach Brown <zach.brown@oracle.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Suparna Bhattacharya <suparna@in.ibm.com>
Acked-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 09:57:21 -08:00
Zach Brown 6d544bb4d9 [PATCH] dio: centralize completion in dio_complete()
There have been a lot of bugs recently due to the way direct_io_worker() tries
to decide how to finish direct IO operations.  In the worst examples it has
failed to call aio_complete() at all (hang) or called it too many times
(oops).

This set of patches cleans up the completion phase with the goal of removing
the complexity that lead to these bugs.  We end up with one path that
calculates the result of the operation after all off the bios have completed.
We decide when to generate a result of the operation using that path based on
the final release of a refcount on the dio structure.

I tried to progress towards the final state in steps that were relatively easy
to understand.  Each step should compile but I only tested the final result of
having all the patches applied.

I've tested these on low end PC drives with aio-stress, the direct IO tests I
could manage to get running in LTP, orasim, and some home-brew functional
tests.

In http://lkml.org/lkml/2006/9/21/103 IBM reports success with ext2 and ext3
running DIO LTP tests.  They found that XFS bug which has since been addressed
in the patch series.

This patch:

The mechanics which decide the result of a direct IO operation were duplicated
in the sync and async paths.

The async path didn't check page_errors which can manifest as silently
returning success when the final pointer in an operation faults and its
matching file region is filled with zeros.

The sync path and async path differed in whether they passed errors to the
caller's dio->end_io operation.  The async path was passing errors to it which
trips an assertion in XFS, though it is apparently harmless.

This centralizes the completion phase of dio ops in one place.  AIO will now
return EFAULT consistently and all paths fall back to the previously sync
behaviour of passing the number of bytes 'transferred' to the dio->end_io
callback, regardless of errors.

dio_await_completion() doesn't have to propogate EIO from non-uptodate bios
now that it's being propogated through dio_complete() via dio->io_error.  This
lets it return void which simplifies its sole caller.

Signed-off-by: Zach Brown <zach.brown@oracle.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Suparna Bhattacharya <suparna@in.ibm.com>
Acked-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 09:57:21 -08:00
Andrew Morton 98c4d57dec [PATCH] io-accounting: direct-io
Account for direct-io.

Cc: Jay Lan <jlan@sgi.com>
Cc: Shailabh Nagar <nagar@watson.ibm.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Chris Sturtivant <csturtiv@sgi.com>
Cc: Tony Ernst <tee@sgi.com>
Cc: Guillaume Thouvenin <guillaume.thouvenin@bull.net>
Cc: David Wright <daw@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 09:55:41 -08:00
Ingo Molnar d8aa905b42 [PATCH] lockdep: annotate direct io
Teach special (rwsem-in-irq) locking code to the lock validator.  Has no
effect on non-lockdep kernels.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Cc: Russell King <rmk@arm.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-07-03 15:27:06 -07:00
Jens Axboe b31dc66a54 [PATCH] Kill PF_SYNCWRITE flag
A process flag to indicate whether we are doing sync io is incredibly
ugly. It also causes performance problems when one does a lot of async
io and then proceeds to sync it. Part of the io will go out as async,
and the other part as sync. This causes a disconnect between the
previously submitted io and the synced io. For io schedulers such as CFQ,
this will cause us lost merges and suboptimal behaviour in scheduling.

Remove PF_SYNCWRITE completely from the fsync/msync paths, and let
the O_DIRECT path just directly indicate that the writes are sync
by using WRITE_SYNC instead.

Signed-off-by: Jens Axboe <axboe@suse.de>
2006-06-23 17:10:39 +02:00
Eric Sesterhenn d4569d2e69 BUG_ON() Conversion in fs/direct-io.c
this changes if() BUG(); constructs to BUG_ON() which is
cleaner and can better optimized away

Signed-off-by: Eric Sesterhenn <snakebyte@gmx.de>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
2006-04-01 01:10:13 +02:00
Nathan Scott 3c674e7423 Fixes a regression from the recent "remove ->get_blocks() support"
change.  inode->i_blkbits should be used when making a get_block_t
request of a filesystem instead of dio->blkbits, as that does not
indicate the filesystem block size all the time (depends on request
alignment - see start of __blockdev_direct_IO).

Signed-off-by: Nathan Scott <nathans@sgi.com>
Acked-by: Badari Pulavarty <pbadari@us.ibm.com>
2006-03-29 09:26:15 +10:00
Badari Pulavarty 1d8fa7a2b9 [PATCH] remove ->get_blocks() support
Now that get_block() can handle mapping multiple disk blocks, no need to have
->get_blocks().  This patch removes fs specific ->get_blocks() added for DIO
and makes it users use get_block() instead.

Signed-off-by: Badari Pulavarty <pbadari@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-26 08:57:01 -08:00
Chen, Kenneth W 174e27c607 [PATCH] direct-io: bug fix in dio handling write error
There is a bug in direct-io on propagating write error up to the higher I/O
layer.  When performing an async ODIRECT write to a block device, if a
device error occurred (like media error or disk is pulled), the error code
is only propagated from device driver to the DIO layer.  The error code
stops at finished_one_bio().  The aysnc write, however, is supposedly have
a corresponding AIO event with appropriate return code (in this case -EIO).
 Application which waits on the async write event, will hang forever since
such AIO event is lost forever (if such app did not use the timeout option
in io_getevents call.  Regardless, an AIO event is lost).

The discovery of above bug leads to another discovery of potential race
window with dio->result.  The fundamental problem is that dio->result is
overloaded with dual use: an indicator of fall back path for partial dio
write, and an error indicator used in the I/O completion path.  In the
event of device error, the setting of -EIO to dio->result clashes with
value used to track partial write that activates the fall back path.

It was also pointed out that it is impossible to use dio->result to track
partial write and at the same time to track error returned from device
driver.  Because direct_io_work can only determines whether it is a partial
write at the end of io submission and in mid stream of those io submission,
a return code could be coming back from the driver.  Thus messing up all
the subsequent logic.

Proposed fix is to separating out error code returned by the IO completion
path from partial IO submit tracking.  A new variable is added to dio
structure specifically to track io error returned in the completion path.

Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Acked-by: Zach Brown <zach.brown@oracle.com>
Acked-by: Suparna Bhattacharya <suparna@in.ibm.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-25 08:23:00 -08:00
Nathan Scott 3fb962bde4 Fix a direct I/O locking issue revealed by the new mutex code.
Affects only XFS (i.e. DIO_OWN_LOCKING case) - currently it is
not possible to get i_mutex locking correct when using DIO_OWN
direct I/O locking in a filesystem due to indeterminism in the
possible return code/lock/unlock combinations.  This can cause
a direct read to attempt a double i_mutex unlock inside XFS.

We're now ensuring __blockdev_direct_IO always exits with the
inode i_mutex (still) held for a direct reader.

Tested with the three different locking modes (via direct block
device access, ext3 and XFS) - both reading and writing; cannot
find any regressions resulting from this change, and it clearly
fixes the mutex_unlock warning originally reported here:
http://marc.theaimsgroup.com/?l=linux-kernel&m=114189068126253&w=2

Signed-off-by: Nathan Scott <nathans@sgi.com>
Acked-by: Christoph Hellwig <hch@lst.de>
2006-03-15 15:14:45 +11:00
Jeff Moyer 35dc8161d0 [PATCH] fix O_DIRECT read of last block in a sparse file
Currently, if you open a file O_DIRECT, truncate it to a size that is not a
multiple of the disk block size, and then try to read the last block in the
file, the read will return 0.  The problem is in do_direct_IO, here:

        /* Handle holes */
        if (!buffer_mapped(map_bh)) {
                char *kaddr;

		...

                if (dio->block_in_file >=
                        i_size_read(dio->inode)>>blkbits) {
                        /* We hit eof */
                        page_cache_release(page);
                        goto out;
                }

We shift off any remaining bytes in the final block of the I/O, resulting
in a 0-sized read.  I've attached a patch that fixes this.  I'm not happy
about how ugly the math is getting, so suggestions are more than welcome.

I've tested this with a simple program that performs the steps outlined for
reproducing the problem above.  Without the patch, we get a 0-sized result
from read.  With the patch, we get the correct return value from the short
read.

Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Suparna Bhattacharya <suparna@in.ibm.com>
Cc: Mingming Cao <cmm@us.ibm.com>
Cc: Joel Becker <Joel.Becker@oracle.com>
Cc: "Chen, Kenneth W" <kenneth.w.chen@intel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-03 08:32:07 -08:00
Jes Sorensen 1b1dcc1b57 [PATCH] mutex subsystem, semaphore to mutex: VFS, ->i_sem
This patch converts the inode semaphore to a mutex. I have tested it on
XFS and compiled as much as one can consider on an ia64. Anyway your
luck with it might be different.

Modified-by: Ingo Molnar <mingo@elte.hu>

(finished the conversion)

Signed-off-by: Jes Sorensen <jes@sgi.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2006-01-09 15:59:24 -08:00
Nick Piggin b5810039a5 [PATCH] core remove PageReserved
Remove PageReserved() calls from core code by tightening VM_RESERVED
handling in mm/ to cover PageReserved functionality.

PageReserved special casing is removed from get_page and put_page.

All setting and clearing of PageReserved is retained, and it is now flagged
in the page_alloc checks to help ensure we don't introduce any refcount
based freeing of Reserved pages.

MAP_PRIVATE, PROT_WRITE of VM_RESERVED regions is tentatively being
deprecated.  We never completely handled it correctly anyway, and is be
reintroduced in future if required (Hugh has a proof of concept).

Once PageReserved() calls are removed from kernel/power/swsusp.c, and all
arch/ and driver code, the Set and Clear calls, and the PG_reserved bit can
be trivially removed.

Last real user of PageReserved is swsusp, which uses PageReserved to
determine whether a struct page points to valid memory or not.  This still
needs to be addressed (a generic page_is_ram() should work).

A last caveat: the ZERO_PAGE is now refcounted and managed with rmap (and
thus mapcounted and count towards shared rss).  These writes to the struct
page could cause excessive cacheline bouncing on big systems.  There are a
number of ways this could be addressed if it is an issue.

Signed-off-by: Nick Piggin <npiggin@suse.de>

Refcount bug fix for filemap_xip.c

Signed-off-by: Carsten Otte <cotte@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-29 21:40:39 -07:00
Christoph Hellwig 92198f7eaa [PATCH] pass iocb to dio_iodone_t
XFS will have to look at iocb->private to fix aio+dio.  No other filesystem
is using the blockdev_direct_IO* end_io callback.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-24 00:05:19 -07:00
Daniel McNeil 29504ff3be [PATCH] Direct IO async short read fix
The direct I/O code is mapping the read request to the file system block.  If
the file size was not on a block boundary, the result would show the the read
reading past EOF.  This was only happening for the AIO case.  The non-AIO case
truncates the result to match file size (in direct_io_worker).  This patch
does the same thing for the AIO case, it truncates the result to match the
file size if the read reads past EOF.

When I/O completes the result can be truncated to match the file size
without using i_size_read(), thus the aio result now matches the number of
bytes read to the end of file.

Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-04-16 15:25:50 -07:00
Linus Torvalds 1da177e4c3 Linux-2.6.12-rc2
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.

Let it rip!
2005-04-16 15:20:36 -07:00