Dave Kleikamp wrote:
> JFS does rely on the order of the I/O being written to disk.
>
> In practice, I don't know how likely you are to have a problem with it.
> Since the journal pages are written sequentially, I would expect that
> most of the time, the order will be preserved. Also dirty metadata
> pages are not written until all of the journal pages associated with the
> change have been written, but these pages are not usually written
> immediately, so in most cases the journal records will probably precede
> the actually metadata to disk.
It actually happens more likely than I thought. We are in a
process of evaluating journal file systems. What we do is to hook
up a programmable power supply to a machine and cycle the power
every 8 mins. Upon reboot, the system runs a program that creates
empty directories as fast as possible. With the first journal
file system we tried, with the write cache turned on, we managed
to corrupt the file system (i.e., cannot be fixed by replaying
journal log, but have to run full fsck) within an hour. That's
why I asked if jfs or other file systems can handle this correctly.
Wilson
>
> The most likely place for error is when we update the "sync point" in
> the journal. The sync point marks the most recent transaction that
> still has outstanding I/O. If we move the sync point past transactions
> that we believe have been completely written to disk, but are still in
> the disk's cache, fsck will not replay those transactions when the
> system is rebooted.
>
> Of course, I have no real data to back up my how likely problems are to
> occur. This is just off the top of my head.
>
> In short, JFS cannot guarantee the filesystem integrity when a disk's
> write cache is on.
>
> W. Wilson Ho wrote:
>
>> Hi all,
>>
>> Can anyone tell me if JFS correctly handle the hard disk's "write
>> cache" reordering? Basically, if the disk's write cache is on, and it
>> shuffles the order where the journal blocks are being flushed to the
>> disk, AND if the power is cut off in the middle of this, the journal
>> log will be out of sequence.
>>
>> Can JFS correctly handle this case and only replay that part of
>> the journal log that is valid?
>>
>> Thanks!
>>
>> Wilson Ho
>
>
>
_______________________________________________
Jfs-discussion mailing list
[EMAIL PROTECTED]
http://www-124.ibm.com/developerworks/oss/mailman/listinfo/jfs-discussion
