Hibernate/Resume and dm-cache, dm-writeboost...

Hibernate/Resume and dm-cache, dm-writeboost...

[Rolf Fokkens]

Hi all,

On the linux-bcache mailing list the potential issue was brought up 
about writes to backing or caching device after bringing up devices in 
initramdisk, but prior to an actual resume. The actual issue is that 
there should be no writes between hibernation and resume:

http://www.spinics.net/lists/linux-bcache/msg02692.html

If this is an issue for bcache, I can imagine it also is for other 
caching solutions like dm-cache and dm-writeboost. I'm not an expert on 
this, but from my user perspective it seems somewhat concerning (for 
both bcache and the DM caching solutions) and an expert view on this 
might be needed.

Thanks,

Rolf

Re: Hibernate/Resume and dm-cache, dm-writeboost...

[Joe Thornber]

On Wed, Dec 24, 2014 at 07:35:49PM +0100, Rolf Fokkens wrote:
> Hi all,
> 
> On the linux-bcache mailing list the potential issue was brought up
> about writes to backing or caching device after bringing up devices
> in initramdisk, but prior to an actual resume. The actual issue is
> that there should be no writes between hibernation and resume:
> 
> http://www.spinics.net/lists/linux-bcache/msg02692.html
> 
> If this is an issue for bcache, I can imagine it also is for other
> caching solutions like dm-cache and dm-writeboost. I'm not an expert
> on this, but from my user perspective it seems somewhat concerning
> (for both bcache and the DM caching solutions) and an expert view on
> this might be needed.

Thanks for bringing this to our attention.  I think dm-cache provides
the neccessary tools to do this; either setting the migration
threshold to zero, or switch to write through mode for the duration of
the hibernate.  But atm this would need to be triggered from userland.
I need to look at hibernate more.

- Joe

Re: Hibernate/Resume and dm-cache, dm-writeboost...

[Akira Hayakawa]

Hi, I am the developer of DM-Writeboost.

I am not sure what the problem is. Is the problem caching driver
may update the underlying devices before resume?

If so, stopping daemons will do by sending the driver these messages
- allow_writeback 0
- enable_writeback_modulator 0
- update_record_interval 0
The first two stops writeback and the third one stops periodical
update of the superblock sector.

btw, DM-Writeboost isn't in LVM family yet so this problem won't happen
at the moment however, I doesn't mean I can be unaware. Thanks.

- Akira