lkml.org 
[lkml]   [2004]   [Jun]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: ide errors in 7-rc1-mm1 and later
Date
On Thursday 10 of June 2004 02:37, Andrew Morton wrote:
> Bartlomiej Zolnierkiewicz <B.Zolnierkiewicz@elka.pw.edu.pl> wrote:
> > On Thursday 10 of June 2004 01:50, Andrew Morton wrote:
> > > Bartlomiej Zolnierkiewicz <B.Zolnierkiewicz@elka.pw.edu.pl> wrote:
> > > > Does journal has checksum or some other protection against failure
> > > > during writing journal to a disk? If not than it still can be
> > > > screwed even with ordered writes if we are unfortunate enough. ;-)
> > >
> > > A transaction is written to disk as two synchronous operations: write
> > > all the data, wait on it, write the single commit block, wait on that.
> >
> > That is how it looks from fs side, from disk side it may look like this:
> >
> > write some data sectors (rest stays in cache)
> > write rest of data sectors (from cache)
> > write some commit sectors (rest stays in cache)
> > write rest of commit sectors (from cache)
> >
> > fs atomic operations != disk atomic operations
>
> JBD is careful about that. There is a single commit block (1, 2 or 4k) and
> the first eight bytes of that block contain a magic number and a sequence
> number. If they're not both valid then replay considers the entire
> transaction (data blocks + commit block) to be invalid.
>
> So all we care about is the atomicity of the first eight bytes of a single
> 512-byte sector. I see no problem with internal-to-commit-block write
> reordering.

OK, thanks for explaining this.

> The problem is that the commit block may hit disk prior to the preceding
> data blocks, which is why we need a full flush prior to submitting the
> commit block.

Yes, yes, this is really obvious for me.
I was also worried about write cache vs commit block write.

> > > If the commit block were to hit disk before the data then we have a
> > > window in which poweroff+recovery would replay garbage into the
> > > filesystem.
> >
> > Yes.
> >
> > The quoted part of my mail is about situation when poweroff happens
> > between 'write some commit sectors' and 'write rest of commit sectors
> > (from cache)' or during transferring commit sectors to a disk.
>
> There is just a single commit sector.

Only one 512-bytes sector? Good!

> > Sure. What's your opinion about doing blk_issue_flush() and ordinary
> > commit (pros+cons given in my previous mail)?
>
> I think we need:
>
> submit_data_sectors();
> blkdev_issue_flush();
> wait_on_data_sectors();
>
> /*
> * All of the transaction's data sectors are now on disk. Submit the
> * commit sector
> */
> mark_buffer_ordered(commit_bh);

Ordered write is not really needed because the next
'data cycle' will provide us with needed ordering.

submit_data_sectors();
blkdev_issue_flush();

^^^
flushes previous commit before the new one is submitted

wait_on_data_sectors();

> submit_bh(commit_bh);
> wait_on_buffer(commit_bh);
>
> Or something like that. Haven't really looked at the blkdev_issue_flush()
> design yet. It has this mysterious comment: "Caller must run
> wait_for_completion() on its own.". Wait for what completion??

-
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: 2005-03-22 14:03    [W:1.163 / U:0.020 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site