Re: [nfsv4] draft-eisler-nfsv4-enterprise-apps-00.txt

[Chuck Lever <[e-mail]>]

On Oct 14, 2010, at 12:09 PM, J. Bruce Fields wrote:

> On Wed, Oct 13, 2010 at 03:40:21PM -0700, Mike Eisler wrote:
>> 
>> (apologies if this a dup)
>> 
>> Margaret Susairaj of Oracle and I (of NetApp) posted this
>> internet-draft, which proposes extensions to NFSv4 for
>> better support of enterprise applications, such
>> as databases.
>> 
>> http://www.ietf.org/id/draft-eisler-nfsv4-enterprise-apps-00.txt
>> 
>> I have requested time at the IETF meeting in Beijing to discuss this.
> 
> Miscellaneous, possibly dumb questions:
> 
> 	- Can you define "enterprise application" for the purpose of
> 	  this draft?
> 
> 	- INITIALIZE:
> 		- I'm confused by the motivation.  So an application is
> 		  trying to detect whether an unused section of a file
> 		  has been corrupted?  Does it even care about
> 		  corruption of regions where it hasn't stored any data?
> 		  What does data corruption in an unallocated region
> 		  even mean?  If you care about data corruption, don't
> 		  you need additional application-level mechanisms (say,
> 		  checksums of some kind) that would render INITIALIZE
> 		  unnecessary?  I'm sure I'm just missing something;
> 		  pointers to literature welcomed.

The goal is to look for misplaced writes by applications.  In other words, it's a way to detect application software bugs, or inappropriate accesses to files by other applications.  On regular disks, you might also look for drive firmware bugs that cause a drive to write data into the wrong sectors.

> 		- Has anyone talked to e.g. linux filesystem people to
> 		  figure out support for interfaces necessary for an
> 		  application to perform INITIALIZE from a client,
> 		  and/or to let an nfs server perform it on a
> 		  filesystem?

I think these are not necessarily intended for a POSIX-style file system API with a VFS layer.  A VFS client implementation might implement these operations via ioctl.  But... it might be appropriate for the server to report to clients that it cannot support these features.

> 	- ADVISE ops: the types look almost like those in posix_fadvise,
> 	  but not exactly; out of curiosity, why the differences?
> 
> 	- SESSION_CTL: why is managing an additional sessions more
> 	  complicated than implementing SESSION_CTL?

And I wonder why the client can't simply terminate a session and negotiate a new one with the new parameters.  Increasing the size of the session's slot table while there are still slots in use might be kind of interesting to implement.

-- 
chuck[dot]lever[at]oracle[dot]com