We are seeing a large regression in database performance on recent
kernels. The database opens a block device with O_DIRECT|O_SYNC and a
number of threads write to different regions of the file at the same time.
A simple test case is below. I haven't defined DEVICE since getting it
wrong will destroy your data :) On an 3 disk LVM with a 64k chunk size we
see about 17MB/sec and only a few threads in IO wait:
procs -----io---- -system-- -----cpu------
r b bi bo in cs us sy id wa st
0 3 0 16170 656 2259 0 0 86 14 0
0 2 0 16704 695 2408 0 0 92 8 0
0 2 0 17308 744 2653 0 0 86 14 0
0 2 0 17933 759 2777 0 0 89 10 0
Most threads are blocking in vfs_fsync_range, which has:
mutex_lock(&mapping->host->i_mutex);
err = fop->fsync(file, dentry, datasync);
if (!ret)
ret = err;
mutex_unlock(&mapping->host->i_mutex);
commit 148f948ba8 (vfs: Introduce new
helpers for syncing after writing to O_SYNC file or IS_SYNC inode) offers
some explanation of what is going on:
Use these new helpers for syncing from generic VFS functions. This makes
O_SYNC writes to block devices acquire i_mutex for syncing. If we really
care about this, we can make block_fsync() drop the i_mutex and reacquire
it before it returns.
Thanks Jan for such a good commit message! As well as dropping i_mutex,
Christoph suggests we should remove the call to sync_blockdev():
> sync_blockdev is an overcomplicated alias for filemap_write_and_wait on
> the block device inode, which is exactly what we did just before calling
> into ->fsync
The patch below incorporates both suggestions. With it the testcase improves
from 17MB/s to 68M/sec:
procs -----io---- -system-- -----cpu------
r b bi bo in cs us sy id wa st
0 7 0 65536 1000 3878 0 0 70 30 0
0 34 0 69632 1016 3921 0 1 46 53 0
0 57 0 69632 1000 3921 0 0 55 45 0
0 53 0 69640 754 4111 0 0 81 19 0
Testcase:
#define _GNU_SOURCE
#include <stdio.h>
#include <pthread.h>
#include <unistd.h>
#include <stdlib.h>
#include <string.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#define NR_THREADS 64
#define BUFSIZE (64 * 1024)
#define DEVICE "/dev/mapper/XXXXXX"
#define ALIGN(VAL, SIZE) (((VAL)+(SIZE)-1) & ~((SIZE)-1))
static int fd;
static void *doit(void *arg)
{
unsigned long offset = (long)arg;
char *b, *buf;
b = malloc(BUFSIZE + 1024);
buf = (char *)ALIGN((unsigned long)b, 1024);
memset(buf, 0, BUFSIZE);
while (1)
pwrite(fd, buf, BUFSIZE, offset);
}
int main(int argc, char *argv[])
{
int flags = O_RDWR|O_DIRECT;
int i;
unsigned long offset = 0;
if (argc > 1 && !strcmp(argv[1], "O_SYNC"))
flags |= O_SYNC;
fd = open(DEVICE, flags);
if (fd == -1) {
perror("open");
exit(1);
}
for (i = 0; i < NR_THREADS-1; i++) {
pthread_t tid;
pthread_create(&tid, NULL, doit, (void *)offset);
offset += BUFSIZE;
}
doit((void *)offset);
return 0;
}
Signed-off-by: Anton Blanchard <anton@samba.org>
Acked-by: Jan Kara <jack@suse.cz>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit 148f948ba8 (vfs: Introduce new
helpers for syncing after writing to O_SYNC file or IS_SYNC inode) broke
the raw driver.
We now call through generic_file_aio_write -> generic_write_sync ->
vfs_fsync_range. vfs_fsync_range has:
if (!fop || !fop->fsync) {
ret = -EINVAL;
goto out;
}
But drivers/char/raw.c doesn't set an fsync method.
We have two options: fix it or remove the raw driver completely. I'm
happy to do either, the fact this has been broken for so long suggests it
is rarely used.
The patch below adds an fsync method to the raw driver. My knowledge of
the block layer is pretty sketchy so this could do with a once over.
If we instead decide to remove the raw driver, this patch might still be
useful as a backport to 2.6.33 and 2.6.32.
Signed-off-by: Anton Blanchard <anton@samba.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Jens Axboe <jens.axboe@oracle.com>
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Tested-by: Jeff Moyer <jmoyer@redhat.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Thanks Thomas and Christoph for testing and review.
I removed 'smp_wmb()' before up_write from the previous patch,
since up_write() should have necessary ordering constraints.
(I.e. the change of s_frozen is visible to others after up_write)
I'm quite sure the change is harmless but if you are uncomfortable
with Tested-by/Reviewed-by on the modified patch, please remove them.
If MS_RDONLY, freeze_bdev should just up_write(s_umount) instead of
deactivate_locked_super().
Also, keep sb->s_frozen consistent so that remount can check the frozen state.
Otherwise a crash reported here can happen:
http://lkml.org/lkml/2010/1/16/37http://lkml.org/lkml/2010/1/28/53
This patch should be applied for 2.6.32 stable series, too.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Tested-by: Thomas Backlund <tmb@mandriva.org>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Cc: stable@kernel.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Currently there is no barrier support in the block device code. That
means we cannot guarantee any sort of data integerity when using the
block device node with dis kwrite caches enabled. Using the raw block
device node is a typical use case for virtualization (and I assume
databases, too). This patch changes block_fsync to issue a cache flush
and thus make fsync on block device nodes actually useful.
Note that in mainline we would also need to add such code to the
->aio_write method for O_SYNC handling, but assuming that Jan's patch
series for the O_SYNC rewrite goes in it will also call into ->fsync
for 2.6.32.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
commit 0762b8bde9
(from 14 months ago) introduced a use-after-free bug which has just
recently started manifesting in my md testing.
I tried git bisect to find out what caused the bug to start
manifesting, and it could have been the recent change to
blk_unregister_queue (48c0d4d4c0) but the results were inconclusive.
This patch certainly fixes my symptoms and looks correct as the two
calls are now in the same order as elsewhere in that function.
Signed-off-by: NeilBrown <neilb@suse.de>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Currently we held s_umount while a filesystem is frozen, despite that we
might return to userspace and unlock it from a different process. Instead
grab an active reference to keep the file system busy and add an explicit
check for frozen filesystems in remount and reject the remount instead
of blocking on s_umount.
Add a new get_active_super helper to super.c for use by freeze_bdev that
grabs an active reference to a superblock from a given block device.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Now that we have the freeze count there is not much reason for bd_mount_sem
anymore. The actual freeze/thaw operations are serialized using the
bd_fsfreeze_mutex, and the only other place we take bd_mount_sem is
get_sb_bdev which tries to prevent mounting a filesystem while the block
device is frozen. Instead of add a check for bd_fsfreeze_count and
return -EBUSY if a filesystem is frozen. While that is a change in user
visible behaviour a failing mount is much better for this case rather
than having the mount process stuck uninterruptible for a long time.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
It has been unused since it was introduced in:
commit 520808bf20e90fdbdb320264ba7dd5cf9d47dcac
Author: Andrew Morton <akpm@osdl.org>
Date: Fri May 21 00:46:17 2004 -0700
[PATCH] block device layer: separate backing_dev_info infrastructure
So lets just kill it.
Acked-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
generic_file_aio_write_nolock() is now used only by block devices and raw
character device. Filesystems should use __generic_file_aio_write() in case
generic_file_aio_write() doesn't suit them. So rename the function to
blkdev_aio_write() and move it to fs/blockdev.c.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Create bdgrab(). This function copies an existing reference to a
block_device. It is safe to call from any context.
Hibernation code wishes to copy a reference to the active swap device.
Right now it calls bdget() under a spinlock, but this is wrong because
bdget() can sleep. It doesn't need a full bdget() because we already
hold a reference to active swap devices (and the spinlock protects
against swapoff).
Fixes http://bugzilla.kernel.org/show_bug.cgi?id=13827
Signed-off-by: Alan Jenkins <alan-jenkins@tuffmail.co.uk>
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Rename the function so that it better describe what it really does. Also
remove the unnecessary include of buffer_head.h.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
It is unnecessarily fragile to have two places (fsync_super() and do_sync())
doing data integrity sync of the filesystem. Alter __fsync_super() to
accommodate needs of both callers and use it. So after this patch
__fsync_super() is the only place where we gather all the calls needed to
properly send all data on a filesystem to disk.
Nice bonus is that we get a complete livelock avoidance and write_supers()
is now only used for periodic writeback of superblocks.
sync_blockdevs() introduced a couple of patches ago is gone now.
[build fixes folded]
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
__fsync_super() does the same thing as fsync_super(). So change the only
caller to use fsync_super() and make __fsync_super() static. This removes
unnecessarily duplicated call to sync_blockdev() and prepares ground
for the changes to __fsync_super() in the following patches.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
* 'for-linus' of git://linux-arm.org/linux-2.6:
kmemleak: Add the corresponding MAINTAINERS entry
kmemleak: Simple testing module for kmemleak
kmemleak: Enable the building of the memory leak detector
kmemleak: Remove some of the kmemleak false positives
kmemleak: Add modules support
kmemleak: Add kmemleak_alloc callback from alloc_large_system_hash
kmemleak: Add the vmalloc memory allocation/freeing hooks
kmemleak: Add the slub memory allocation/freeing hooks
kmemleak: Add the slob memory allocation/freeing hooks
kmemleak: Add the slab memory allocation/freeing hooks
kmemleak: Add documentation on the memory leak detector
kmemleak: Add the base support
Manual conflict resolution (with the slab/earlyboot changes) in:
drivers/char/vt.c
init/main.c
mm/slab.c
There are allocations for which the main pointer cannot be found but
they are not memory leaks. This patch fixes some of them. For more
information on false positives, see Documentation/kmemleak.txt.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
This reverts commit db2dbb12dc.
It apparently causes problems with partition table read-ahead
on archs with large page sizes. Until that problem is diagnosed
further, just drop the readpages support on block devices.
Signed-off-by: Jens 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: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Doing a proper block dev ->readpages() speeds up the crazy dump(8)
approach of using interleaved process IO.
Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Move some block device related code out from buffer.c and put it in
block_dev.c. I'm trying to move non-buffer_head code out of buffer.c
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The ioctls for the generic freeze feature are below.
o Freeze the filesystem
int ioctl(int fd, int FIFREEZE, arg)
fd: The file descriptor of the mountpoint
FIFREEZE: request code for the freeze
arg: Ignored
Return value: 0 if the operation succeeds. Otherwise, -1
o Unfreeze the filesystem
int ioctl(int fd, int FITHAW, arg)
fd: The file descriptor of the mountpoint
FITHAW: request code for unfreeze
arg: Ignored
Return value: 0 if the operation succeeds. Otherwise, -1
Error number: If the filesystem has already been unfrozen,
errno is set to EINVAL.
[akpm@linux-foundation.org: fix CONFIG_BLOCK=n]
Signed-off-by: Takashi Sato <t-sato@yk.jp.nec.com>
Signed-off-by: Masayuki Hamaguchi <m-hamaguchi@ys.jp.nec.com>
Cc: <xfs-masters@oss.sgi.com>
Cc: <linux-ext4@vger.kernel.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Dave Kleikamp <shaggy@austin.ibm.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Alasdair G Kergon <agk@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4: (57 commits)
jbd2: Fix oops in jbd2_journal_init_inode() on corrupted fs
ext4: Remove "extents" mount option
block: Add Kconfig help which notes that ext4 needs CONFIG_LBD
ext4: Make printk's consistently prefixed with "EXT4-fs: "
ext4: Add sanity checks for the superblock before mounting the filesystem
ext4: Add mount option to set kjournald's I/O priority
jbd2: Submit writes to the journal using WRITE_SYNC
jbd2: Add pid and journal device name to the "kjournald2 starting" message
ext4: Add markers for better debuggability
ext4: Remove code to create the journal inode
ext4: provide function to release metadata pages under memory pressure
ext3: provide function to release metadata pages under memory pressure
add releasepage hooks to block devices which can be used by file systems
ext4: Fix s_dirty_blocks_counter if block allocation failed with nodelalloc
ext4: Init the complete page while building buddy cache
ext4: Don't allow new groups to be added during block allocation
ext4: mark the blocks/inode bitmap beyond end of group as used
ext4: Use new buffer_head flag to check uninit group bitmaps initialization
ext4: Fix the race between read_inode_bitmap() and ext4_new_inode()
ext4: code cleanup
...
Anton Blanchard
Andrew Morton
Jun'ichi Nomura
Jens Axboe
Christoph Hellwig
Neil Brown
Alexey Dobriyan
Alan Jenkins
Jan Kara
Linus Torvalds
Catalin Marinas
Martin K. Petersen
Jeff Moyer
Al Viro
Nick Piggin
Takashi Sato