WRITE_DMA errors on SATA drive under 5.3-RELEASE

[Anthony Atkielski]

I've gotten two messages like the ones below today on my production server
(5.3-RELEASE):

messages:Feb 27 14:48:17 freebie kernel: ad10: TIMEOUT - WRITE_DMA retrying (2 retries left) LBA=4848803
messages:Feb 27 14:48:17 freebie kernel: ad10: FAILURE - WRITE_DMA timed out

What do these messages mean? The referenced drive is one of two identical SATA
drives on the server; it holds /tmp and /var. I don't recall seeing
these messages before.

Is there a way to work backwards from the LBA to the filesystem so that
I can see which file was being referenced when this occurred?

--
Anthony

[cpghost@cordula.ws]

On Sun, Feb 27, 2005 at 03:53:30PM +0100, Anthony Atkielski wrote:

> messages:Feb 27 14:48:17 freebie kernel: ad10: TIMEOUT - WRITE_DMA retrying (2 retries left) LBA=4848803
> messages:Feb 27 14:48:17 freebie kernel: ad10: FAILURE - WRITE_DMA timed out

[...]

> Is there a way to work backwards from the LBA to the filesystem so that
> I can see which file was being referenced when this occurred?

Theoretically, one could use 'fsdb -r' in a scripted manner, to
generate a mapping of file names to blocks (relative to the partition
of the file system you are mapping). Once you have the blocks, you'll
need to do so artithmetics to map those blocks to LBA address ranges
(perhaps via GEOM or using data in disklabels). Finally, you'll have
to locate the range for a particular LBA address and work backwards
up to the inode #, and then to the filename(s) that link to that inode.

Perhaps there's already a system utility or port for this? It would be
really useful!

> Anthony

Cheers,
-cpghost.

--
Cordula's Web. [url]http://www.cordula.ws/[/url]

[Anthony Atkielski]

[email]cpghost@cordula.ws[/email] writes:

> Theoretically, one could use 'fsdb -r' in a scripted manner, to
> generate a mapping of file names to blocks (relative to the partition
> of the file system you are mapping). Once you have the blocks, you'll
> need to do so artithmetics to map those blocks to LBA address ranges
> (perhaps via GEOM or using data in disklabels). Finally, you'll have
> to locate the range for a particular LBA address and work backwards
> up to the inode #, and then to the filename(s) that link to that inode.

Sounds complicated. Surely I'm not the first person to wish for such a
utility ... in UNIXland, there seems to be a command for just about
every conceivable purpose (?).

> Perhaps there's already a system utility or port for this? It would be
> really useful!

I'm mainly worried about exactly what the system was trying to write at
the time. It's not clear from the message whether the write succeeded
or not.

--
Anthony

[cpghost@cordula.ws]

On Sun, Feb 27, 2005 at 05:19:32PM +0100, Anthony Atkielski wrote:

> [email]cpghost@cordula.ws[/email] writes:
>

> > Theoretically, one could use 'fsdb -r' in a scripted manner, to
> > generate a mapping of file names to blocks (relative to the partition
> > of the file system you are mapping). Once you have the blocks, you'll
> > need to do so artithmetics to map those blocks to LBA address ranges
> > (perhaps via GEOM or using data in disklabels). Finally, you'll have
> > to locate the range for a particular LBA address and work backwards
> > up to the inode #, and then to the filename(s) that link to that inode.

>
> Sounds complicated. Surely I'm not the first person to wish for such a
> utility ... in UNIXland, there seems to be a command for just about
> every conceivable purpose (?).

Or you could write the missing ones :-).

Actually, it's not that hard. You need three mappings:

1. (lba address, (filesystem, block #))
2. ((filesystem, block #), (filesystem, inode #))
3. ((filesystem, inode #), (list of filenames linking to inode #))

Each of those mappings could be done and displayed by a single
utility. Combining all three into a lba2filenames program would
then be trivial.

> > Perhaps there's already a system utility or port for this? It would be
> > really useful!

>
> I'm mainly worried about exactly what the system was trying to write at
> the time. It's not clear from the message whether the write succeeded
> or not.

Yes, that's exactly my concern too.

> --
> Anthony

-cpghost.

--
Cordula's Web. [url]http://www.cordula.ws/[/url]