Re: 2.6.0-test11 data loss

[Con Kolivas]

On Thu, 25 Dec 2003 13:07, Eric D. Mudama wrote:
> On Thu, Dec 25 at  9:34, Con Kolivas wrote:
> >On Thu, 25 Dec 2003 09:22, Gergely Tamas wrote:
> >> I don't think this is a reiserfs bug. This was my first thought and
> >> after first hitting this bug, I've moved all my partitions from reiserfs
> >> to jfs. But I've also had this problem with it... Now I'm back to
> >> 2.4.23, and everything works fine.
> >
> >Because of the numerous reboots and hangs I've seen with experimental
> > patches I've also seen this, but it's not reiserFS fault. The problem is
> > that most drives have write caching enabled and not all of them are safe
> > with this. If you disable it with hdparm (hdparm -W 0 /dev/hd*) you'll
> > find that open files during a hard reset or power outage will prevent
> > those open files from being corrupted.
>
> Write cache off will not prevent a file from being corrupted, however,
> it should limit the corruption to a single disk operation.
>
> I don't see how the behavior you describe could be the drive's
> fault...
>
> The user stated that their system hard locked, then they went and
> rebooted it, and following the reboot they had corruption...  From
> this, there are a few possibilities:
>
> 1. The drive had been given the commands to write the data prior to the
> hang.
>
> If this was the case, the drive would happilly keep writing the data
> it had been given and was caching in the background, even while you
> continued to send (or stopped sending) data for a new command over the
> interface.  An IDE interface lockup or system lockup will not prevent
> the drive from flushing the remainder of its write cache.  (Only
> possible exception might be faulty handling of a hard reset, but all
> drives today will flush their cache when they see the reset, prior to
> processing it.) Unless the user yanked power within a few hundred
> milliseconds of the write command, I think it is unlikely that cached
> data already in the drive wasn't flushed properly.
>
> 2. The drive was in the middle of a command writing important data
> during the hang.
>
> In this case, yes, your file you were writing would probably be
> corrupt on the media, but nothing more.  Drives detect power loss, and
> immediately disable write-gate and park the actuator.  If they don't
> get the actuator parked before they run out of back-EMF from the
> momentum of the platter(s), the head will stick to the media and
> you'll probably need a chisel to get that drive to spin again.
>
> 3. The drive hadn't yet been issued the commands for the data that was
> eventually corrupted.
>
> I find this to be the most likely case, and is a situation where the
> filesystem thinks objects were moved but those updates were not
> correctly sent to the disk (due to the hang?), so it might think
> they're in the old location or something.  (I'm not a filesystem
> wizard so if I'm way off-base, my apologies)
>
> It seems to me that the problem occurred at a higher system level than
> the disk, and disabling the write cache on the drive (besides being a
> *HUGE* performance loser) will only make the window for failure
> smaller, not eliminate it entirely.
>
> Unless you are using *really* old hard drives, the write caching in
> today's drives is really quite good and definitely should be usable.
> Sure, it makes things less safe in power events, but system lockups
> shouldn't affect the drive's ability to flush its cache.  Note too
> that Gergely reported that the problem went away on his 2.4.23 system.
> I don't believe that to be a small data point.

I hardly said it was the correct solution; just what worked for me, as I had 
exactly the same issue going 2.4->2.6. I can't even recall if write caching 
was actually on in 2.4, and my write performance under video capture has not 
shown any detriment. The filesystem gods should comment. Merry Christmas.

Con

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/