uids-as-names, compound, reply cache, cabbages and kings

New Message Reply About this list Date view Thread view Subject view Author view Attachment view

From: Noveck, Dave (Dave.Noveck@netapp.com)
Date: 05/30/02-11:19:13 AM Z


Message-ID: <8C610D86AF6CD4119C9800B0D0499E336A8E28@red.nane.netapp.com>
From: "Noveck, Dave" <Dave.Noveck@netapp.com>
Subject: uids-as-names, compound, reply cache, cabbages and kings
Date: Thu, 30 May 2002 09:19:13 -0700

I had an interesting experience in implementing some 
elements of the v4 readdir in which there was some
unexpected interactions (positive as well as negative)
among protocol features and I thought I'd share it with
everybody (after all, you can hit delete), since it's
kind of amusing, at least if you can be amused by weird
protocol implementation issues (OK, so maybe everybody is
going to hit delete).  If you're going to be at the 
bake-off you might want to skip to the last paragraph,
instead of hitting delete. 

This is the kind of thing that might be an interesting 
section in that implementation RFC that we used to talk
about doing after the spec was done.  By the way Spencer, 
how is the spec coming?  I'm hoping that we will test the
*final* protocol at the August bake-off.

Anyway, the issue is that for reasons that are too 
complicated to go into here, I may post-process some of
the readdir output to deal with the fact that uid's and 
gid's need to be mapped to strings.  I maintain a cache
of this stuff and install the strings in the output if
I can find them in the cache.  In cases in which the 
name isn't in the cache, I leave some bread crumbs in 
what would otherwise be the readdir output and mark
the request for later post-processing, after the
request is finished, while creating the reply.

So here's the ugly case.  Somebody gives me a readdir
with a very small max output.  It might be big enough
to hold the output for a single file in the directory 
or it might not, depending on how long the string for
the uid turns out ot be.  So at post-processing, I
may install the actual string if it fits, or return
a NOSPC error.

But wait a minute.  I'm doing this after processing the
request.  This might have been a COMPOUND in which there
is some operation that followed the READDIR.  If I fail
the readdir, then I shouldn't have done the following 
operation, and if it had visible consequences then I
would have to undo that operation, which would be very
difficult or impossible to do.

So am I screwed here?  The answer is "No" because of some
interesting serendipity.  Suppose someone did a COMPOUND
consisting of READDIR followed by RENAME.  Or READ followed
by WRITE or by REMOVE.  How could this be accommodated by
the reply cache.  Not very easily.  The COMPOUND is clearly
not idempotent so the reply cache information has to be saved.
But that includes the idempotent operations before the 
non-idempotent including operations like READ and READDIR
that can generate lots of output.  Without COMPOUND, such
operations would never have to be stored in the reply cache,
and saving large amounts of data in a reply cache is highly
undesirable.

So because the above issues with the reply cache (which is
not part of the spec), we decided way back that although
such things as READDIR followed by RENAME are allowed by the
spec, the server may reject a compound with NFS4ERR_RESOURCE,
if it is just too complicated to handle for this or any other 
reason.  RFC3010 says:

   It is the client's responsibility for recovering from any 
   partially completed COMPOUND procedure.  Partially completed 
   COMPOUND procedures may occur at any point due to errors such 
   as NFS4ERR_RESOURCE and NFS4ERR_LONG_DELAY.  This may occur 
   even given an otherwise valid operation string.   ...   
   Therefore, the client should avoid overly complex COMPOUND 
   procedures in the event of the failure of an operation within 
   the procedure. 

Dealing with a non-idempotent op after READ, READLINK, READDIR seems
just about impossible, from the reply cache point of view.  Dealing
with READ, READLINK, READDIR after a non-idempotent operation is more
tractable but still difficult.  At the next bake-off, my server is 
going to return NFS4ERR_RESOURCE for either case.  I don't expect this
will be a problem for anyone, but you never know.  BTW, DAFS which
does define reply cache behavior as part of its architecture, specifies
that the first of these combinations is an instance of invalid chaining
(DAFS has chaining rather than COMPOUND).     
 


New Message Reply About this list Date view Thread view Subject view Author view Attachment view

This archive was generated by hypermail 2.1.2 : 03/04/05-01:49:45 AM Z CST