[pnfs] Status meeting pNFS

[Benny Halevy]

Marc Eshel wrote:
> So we agreed that the interface will for the xdr routines will include a 
> pointer and length going in and length or return code going out. We also 
> agreed that we will make the changes after the bakeathon if it is OK with 
> Benny.
>   
Absolutely, no problem...
Benny
> Marc.
>
> pnfs-bounces at linux-nfs.org wrote on 08/31/2006 08:38:27 AM:
>
>   
>> marc at almaden.ibm.com said:
>>     
>>> Another alternative is to use *resp in the prototype and extract 'p' 
>>>       
> in 
>   
>>> the routine that we call. This would be cleaner allowing the routine 
>>>       
> to 
>   
>>> also update 'p' so the caller can tell where we are in the buffer 
>>>       
> without 
>   
>>> any extra parameter that return the length, and we can use the 
>>>       
> returned 
>   
>>> int for error code instead.
>>>
>>> Marc.
>>>       
>> bhalevy at panasas.com said:
>>     
>>> I also like this approach better, especially since it allows using
>>> semi-standard XDR macros RESERVE_SPACE, WRITE* or ADJUST_ARGS. If we 
>>>       
> want to
>   
>>> provide a solution which is independent from xdr I think we should 
>>>       
> also
>   
>>> provide a size_t parameter with the maximum allowed  buffer size. the
>>> declaration would look like layout_get(unsigned int **p, size_t 
>>>       
> maxlength,
>   
>>> void *layout); 
>>>       
>> i don't see any reason to pass in this whole structure. only 'p' 
>>     
> and'end' are 
>   
>> used.
>>
>> struct nfsd4_compoundres {
>>         /* scratch variables for XDR encode */
>>         u32 *                           p;
>>         u32 *                           end;
>>         struct xdr_buf *                xbuf;
>>         struct svc_rqst *               rqstp;
>>
>>         u32                             taglen;
>>         char *                          tag;
>>         u32                             opcnt;
>>         u32 *                           tagp; /* where to encode tag and 
>>     
>  
>   
>> opcount */
>> };
>>
>> the WRITE* macro's work on 'p' so resp is not needed to use them.
>>
>>
>> bhalevy at panasas.com said:
>>     
>>> Currently the layout needs to be encoded in XDR by the file system 
>>>       
> code
>   
>>> which means preceding the opaque layout with a 4 byte, XDR encoded 
>>>       
> length.
>   
>>> If you want to keep the file system ignorant about XDR then the 
>>>       
>> routing also
>>     
>>> needs to return the length of whatever it encoded opaquely  and the 
>>>       
> nfsd
>   
>>> should "frame" the layout bits with XDR length and realign  resp->p if
>>> needed. 
>>>       
>> it's not so much keeping the file system ignorant of the XDR as it is 
>>     
> keeping 
>   
>> the nfs layer ignorant of the file system opaque layout. either way, 
>>     
> benny's 
>   
>> suggestion will work.
>>
>> int layout_get(unsigned int **p, size_t  maxbytes, void *layout); 
>>
>> layout_get returns a positive length upon succes, or a negative error 
>>     
> code 
>   
>> upon failure.
>>
>> and maxbytes = (resp->end - resp->p) << 2; always on a 4 byte boundary
>>
>> for layouts that want to do XDR, the macro RESERVE_SPACE(nbytes) becomes 
>>     
>
>   
>> something like
>>
>> RESERVE_SPACE(nbutes, maxbytes)
>>    BUG_ON(XDR_QUADLEN(nbytes) > (maxbytes >> 2);            \
>>
>> nfsd4_encode_layoutget then 'frames' the layout bytes with the 
>> returned length 
>> and realign resp->p.
>>
>>
>> -->Andy
>>
>>
>>
>>
>> _______________________________________________
>> pNFS mailing list
>> pNFS at linux-nfs.org
>> http://linux-nfs.org/cgi-bin/mailman/listinfo/pnfs
>>