Re: dm thin: superblock may write succeed before other metadata blocks because of wirting metadata in async mode.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Wed, Jun 20 2018 at  1:03pm -0400,
monty <monty_pavel@xxxxxxxx> wrote:

> 
> On Tue, Jun 19, 2018 at 11:00:32AM -0400, Mike Snitzer wrote:
> > 
> > On Tue, Jun 19 2018 at 10:43am -0400,
> > Joe Thornber <thornber@xxxxxxxxxx> wrote:
> > 
> > > On Tue, Jun 19, 2018 at 09:11:06AM -0400, Mike Snitzer wrote:
> > > > On Mon, May 21 2018 at  8:53pm -0400,
> > > > Monty Pavel <monty_pavel@xxxxxxxx> wrote:
> > > > 
> > > > > 
> > > > > If dm_bufio_write_dirty_buffers func is called by __commit_transaction
> > > > > func and power loss happens during executing it, coincidencely
> > > > > superblock wrote correctly but some metadata blocks didn't. The reason
> > > > > is we write all metadata in async mode. We can guarantee that we send
> > > > > superblock after other blocks but we cannot guarantee that superblock
> > > > > write completely early than other blocks.
> > > > > So, We need to commit other metadata blocks before change superblock.
> > > > > 
> > > > > Signed-off-by: Monty Pavel <monty_pavel@xxxxxxxx>
> > > > > ---
> > > > >  drivers/md/dm-thin-metadata.c |    8 ++++++++
> > > > >  1 files changed, 8 insertions(+), 0 deletions(-)
> > > > > 
> > > > > diff --git a/drivers/md/dm-thin-metadata.c b/drivers/md/dm-thin-metadata.c
> > > > > index 36ef284..897d7d6 100644
> > > > > --- a/drivers/md/dm-thin-metadata.c
> > > > > +++ b/drivers/md/dm-thin-metadata.c
> > > > > @@ -813,6 +813,14 @@ static int __commit_transaction(struct dm_pool_metadata *pmd)
> > > > >  	if (r)
> > > > >  		return r;
> > > > >  
> > > > > +	r = dm_tm_commit(pmd->tm, sblock);
> > > > > +	if (r)
> > > > > +		return r;
> > > > > +
> > > > > +	r = superblock_lock(pmd, &sblock);
> > > > > +	if (r)
> > > > > +		return r;
> > > > > +
> > > > >  	disk_super = dm_block_data(sblock);
> > > > >  	disk_super->time = cpu_to_le32(pmd->time);
> > > > >  	disk_super->data_mapping_root = cpu_to_le64(pmd->root);
> > > 
> > > I don't believe you've tested this; sblock is passed to dm_tm_commit()
> > > uninitialised, and you didn't even bother to remove the later (and correct)
> > > call to dm_tm_commit().
> > 
> > I pointed out to Joe that the patch, in isolation, is decieving.  It
> > _looks_ like sblock may be uninitialized, etc.  But once the patch is
> > applied and you look at the entirety of __commit_transaction() it is
> > clear that you're reusing the existing superblock_lock() to safely
> > accomplish your additional call to dm_tm_commit().
> > 
> > > What is the issue that started you looking in this area?
> > 
> > Right, as my previous reply asked: please clarify if you _know_ your
> > patch fixes an actual problem you've experienced.  The more details the
> > better.
> > 
> > Thanks,
> > Mike
> > 
> Hi, Mike and Joe. Thanks for your reply. I read __commit_transaction
> many times and didn't find any problem of 2-phase commit. I use
> md-raid1(PCIe nvme and md-raid5) in write-behind mode to store dm-thin
> metadata.
> Test case:
> 1. I do copy-diff test on thin device and then reboot my machine.
> 2. Rebuild our device stack and exec "vgchang -ay".
> The thin-pool can not be established(details_root become a bitmap node
> and metadata's bitmap_root become a btree_node).

But are you saying your double commit in __commit_transaction() serves
as a workaround for the corruption you're seeing?

Is it just a case where raid5's writebehind mode is _not_ safe for your
storage config?  By "reboot" do you mean a clean shutdown?  Or a forced
powerfail scenario?

Mike

--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel



[Index of Archives]     [DM Crypt]     [Fedora Desktop]     [ATA RAID]     [Fedora Marketing]     [Fedora Packaging]     [Fedora SELinux]     [Yosemite Discussion]     [KDE Users]     [Fedora Docs]

  Powered by Linux