lkml.org 
[lkml]   [2005]   [Jun]   [27]   [last100]   RSS Feed
Views: [more markup]  [less markup]  [headers]  [forward] 
 
Messages in this thread
/
DateMon, 27 Jun 2005 15:58:27 -0500
FromSteve Lord <>
SubjectRe: reiser4 plugins
Theodore Ts'o wrote:
> On Mon, Jun 27, 2005 at 03:18:30PM -0500, Steve Lord wrote:
> 
>>I presume Ted is referring to problems guaranteeing the integrity of
>>the journal at recovery time. I am coming into this without all the
>>available context, so I may be barking up the wrong tree.... In
>>particular, I am not sure how journaling whole blocks protects
>>you from this.
> 
> 
> Actually, I was talking about the problem what happens when power
> fails while DMA'ing to the disk, and memory, which is more sensitive
> to voltage drops than the rest of the system, starts sending garbage
> to the bus, which the disk then faithfully writes to the inode table.
> 
> As I recall, you were the one who told me about this problem, and how
> it was fixed in Irix by using a powerfail interrupt to abort DMA
> transfers, as well as giving me a program which tests for this
> condition (basically it writes known test pattern to the disk, and
> then you do an unclean shutdown, and you look to see if garbage is
> written to the disk instead of one of the known test patterns).  If it
> wasn't you, it must have been Jim Mostek --- but I could have sworn it
> was you.....
> 
> 						- Ted
> 

Your memory is better than mine, not sure about the test program, but
there was at one point a scenario like that on Irix, and I quite
probably did mention it to you. That was certainly not PC hardware,
but more likely a large Irix system with multiple power busses spread
around a large amount of hardware. I presume you are saying that the fact
that ext3 journals complete inode blocks rather that subsections of inodes
means that, should a similar trashing of metadata occur during power down,
a journal replay would fix it?

I see no way short of hardware fixes of avoiding the general problem of
a system failing in an ugly manner like this. Unless you write everything
to disk twice (i.e. journal all data), you can still end up with a
legitimate set of metadata, and the master copy of your employee
database full of nasty little bits of corruption.

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

\
 
 \ /
  Last update: 2005-06-27 23:09    [from the cache]
©2003-2008