Re: [PATCH 2/2] nfsd41: mask out unsupported pnfs attributes

[Benny Halevy <bhalevy@panasas.com>]

On 2010-10-01 17:20, J. Bruce Fields wrote:
> On Fri, Oct 01, 2010 at 05:17:51PM +0200, Benny Halevy wrote:
>> On 2010-10-01 16:48,  J. Bruce Fields wrote:
>>> On Thu, Sep 30, 2010 at 08:47:58PM +0200, Benny Halevy wrote:
>>>> These attributes are valid in NFSv4.1, the just doesn't support them yet.
>>>
>>> The existing code handles unsupported attributes in the operations
>>> themselves.  Perhaps it makes sense to move those checks here, but if
>>> so, explain why, and let's do this for all unsupported attributes, not
>>> just these two.
>>
>> The client can run a DOS attack on the server by requesting invalid attributes
>> and tripping the BUG_ONs in nfsd4_encode_fattr.
> 
> How can they do that?  getattr and readdir, for example, both handle
> this.  But I may well be missing something!
> 

You're right.  I got confused by the pnfsd specific bug I fixed here:
http://marc.info/?l=linux-nfs&m=128587444310102&w=2

In this case LAYOUT_BLKSIZE was marked as supported but nothing
really supported it so we ended up with an invalid result sent
to the client.

So let's drop this patch. sorry.

Benny

>> We can/should also change the BUG_ONs to either report invalid
>> attribute or just silently ignore them, but the client is
>> perfectly entitled to get attrs we don't support :)
> 
> Sure.
> 
>>> Looking back at the spec....  I guess it's only on operations that set
>>> attributes that we return NFS4ERR_ATTRNOTSUPP, and otherwise we silently
>>> ignore them?
>>
>> For the GETATTR case, we just return the attrmask for the attrs we support.
>> IOW:
>>    The server returns an attribute bitmap that
>>    indicates the attribute values that it was able to return, which will
>>    include all attributes requested by the client that are attributes
>>    supported by the server for the target file system.
> 
> OK, makes sense.
> 
> --b.