lkml.org 
[lkml]   [2009]   [Mar]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] issue storage device flush via sync_blockdev() (was Re: Linux 2.6.29)
Eric Sandeen wrote:
> Jeff Garzik wrote:
>> On Wed, Mar 25, 2009 at 01:40:37PM -0700, Linus Torvalds wrote:
>>> On Wed, 25 Mar 2009, Jeff Garzik wrote:
>>>> It is clearly possible to implement an fsync(2) that causes FLUSH CACHE to be
>>>> issued, without adding full barrier support to a filesystem. It is likely
>>>> doable to avoid touching per-filesystem code at all, if we issue the flush
>>>> from a generic fsync(2) code path in the kernel.
>>> We could easily do that. It would even work for most cases. The
>>> problematic ones are where filesystems do their own disk management, but I
>>> guess those people can do their own fsync() management too.
>>>
>>> Somebody send me the patch, we can try it out.
>> This is a simple step that would cover a lot of cases... sync(2)
>> calls sync_blockdev(), and many filesystems do as well via the generic
>> filesystem helper file_fsync (fs/sync.c).
>>
>> XFS code calls sync_blockdev() a "big hammer", so I hope my patch
>> follows with known practice.
>>
>> Looking over every use of sync_blockdev(), its most frequent use is
>> through fsync(2), for the selected filesystems that use the generic
>> file_fsync helper.
>>
>> Most callers of sync_blockdev() in the kernel do so infrequently,
>> when removing and invalidating volumes (MD) or storing the superblock
>> prior to release (put_super) in some filesystems.
>>
>> Compile-tested only, of course :) But it should be work :)
>>
>> My main concern is some hidden area that calls sync_blockdev() with
>> a high-enough frequency that the performance hit is bad.
>>
>> Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
>>
>> diff --git a/fs/buffer.c b/fs/buffer.c
>> index 891e1c7..7b9f74a 100644
>> --- a/fs/buffer.c
>> +++ b/fs/buffer.c
>> @@ -173,9 +173,14 @@ int sync_blockdev(struct block_device *bdev)
>> {
>> int ret = 0;
>>
>> - if (bdev)
>> - ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
>> - return ret;
>> + if (!bdev)
>> + return 0;
>> +
>> + ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
>> + if (ret)
>> + return ret;
>> +
>> + return blkdev_issue_flush(bdev, NULL);
>> }
>> EXPORT_SYMBOL(sync_blockdev);
>
> What about when you're running over a big raid device with
> battery-backed cache, and you trust the cache as much as much as the
> disks. Wouldn't this unconditional cache flush be painful there on any
> of the callers even if they're rare? (fs unmounts, freezes, unmounts,
> etc? Or a fat filesystem on that device doing an fsync?)

What exactly do you think sync_blockdev() does? :)

It is used right before a volume goes away. If that's not a time to
flush the cache, I dunno what is.

The _whole purpose_ of sync_blockdev() is to push out the data to
permanent storage. Look at the users -- unmount volume, journal close,
etc. Things that are OK to occur after those points include: power off,
device unplug, etc.

A secondary purpose of sync_blockdev() is as a hack, for simple/ancient
bdev-based filesystems that do not wish to bother with barriers and all
the associated complexity to tracking what writes do/do not need flushing.


> xfs, reiserfs, ext4 all avoid the blkdev flush on fsync if barriers are
> not enabled, I think for that reason...

Enabling barriers causes slowdowns far greater than that of simply
causing fsync(2) to trigger FLUSH CACHE, because barriers imply FLUSH
CACHE issuance for all in-kernel filesystem journalled/atomic
transactions, in addition to whatever syscalls userspace is issuing.

The number of FLUSH CACHES w/ barriers is orders of magnitude larger
than the number of fsync/fdatasync calls.

Jeff





\
 
 \ /
  Last update: 2009-03-26 00:11    [W:0.505 / U:1.308 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site