zfs discuss - Dec 2005 - zvol questions...

Hi All,

I have some implementation questions about the zvol capability. 
According to Eric''s picture at:

http://opensolaris.org/os/community/zfs/source

The zvol interfaces do not talk to the ZAP interfaces. The only thing I 
can find in the zvol code is that it uses the ZAP for size(volsize). Is 
this the only attributes it gets/sets from the ZAP?

Also, again from the picture and looking at the source, it doesn''t look
like zvol uses the ZIL interfaces in any way. So, if this is true, does 
that mean that data is written immediately to disk through the zvol 
interfaces? Well by immediately I mean they are never written to the 
intent log, it may be cached intermediately at some point, correct?.  
And, if you open the device with O_DSYNC I assume they will get written 
immediately to disk?

thanks,
sarah

On Thu, Dec 01, 2005 at 02:49:18PM -0700, Sarah Jelinek
wrote:
> Hi All,
> 
> I have some implementation questions about the zvol capability. 
> According to Eric''s picture at:
> 
> http://opensolaris.org/os/community/zfs/source
> 
> The zvol interfaces do not talk to the ZAP interfaces. The only thing I 
> can find in the zvol code is that it uses the ZAP for size(volsize). Is 
> this the only attributes it gets/sets from the ZAP?
Yes, that is correct.  It uses a private ZAP object to store the volume
size so that it stored when taking a snapshot.  There are a lot of lines
that could be drawn into that picture (such as the fact that the DSL
makes heavy use of the ZAP), but it would make it significantly more
complex, and we stuck with just the main consumers.

- Eric

--
Eric Schrock, Solaris Kernel Development       http://blogs.sun.com/eschrock

Sarah Jelinek wrote On 12/01/05 14:49,:
> Hi All,
>
> I have some implementation questions about the zvol capability.
> According to Eric''s picture at:
>
> http://opensolaris.org/os/community/zfs/source
>
> The zvol interfaces do not talk to the ZAP interfaces. The only
> thing I can find in the zvol code is that it uses the ZAP
> for size(volsize). Is this the only attributes it gets/sets
> from the ZAP?
Eric answered this one.
> Also, again from the picture and looking at the source, it
> doesn''t look like zvol uses the ZIL interfaces in any way.
Correct. The ZIL is only currently used for the POSIX transactions
of the ZPL, though is potentially useful for any client that wishes
to ensure transactions are on stable storage.
> So, if this is true, does
> that mean that data is written immediately to disk through the zvol
> interfaces? Well by immediately I mean they are never written to the
> intent log, it may be cached intermediately at some point, correct?.
> And, if you open the device with O_DSYNC I assume they will
> get written immediately to disk?
It looks like spec_write() will not set the the segmap_release()
flags to SM_ASYNC if FDSYNC (O_DSYNC) is set. Looking at
zvol_strategy() which segmap_release() eventually calls to
put any page, it ignores B_ASYNC. As soon as dmu_write() returns
the IO is marked as complete by biodone(). I think this is a bug.
It should wait until they are committed using txg_wait_synced()
(which would be very slow) or use the ZIL to log the page.
Fixing ZVOL to use the ZIL would take a while, but might be
a useful exercise to validate the ZIL architecture using
a different client.

Anyway I think a bug needs to be raised against ZVOL not
honoring O_DSYNC semantics. Thanks for bringing this to our
attention (and making me familiarise myself with ZVOL!).
>
> thanks,
> sarah
Neil