On 02/28/2012 11:02 AM, Pádraig Brady wrote:
> On 02/28/2012 08:50 AM, Paul Eggert wrote:
>> On 02/27/2012 05:09 PM, Pádraig Brady wrote:
>>> how
>>> to handle existing regular files with conv=trunc.
>>> I.E. seeking over existing possible non NUL data.
>>> It's too dangerous/inconsistent to do this for files I think.
>>
>> Why?  This is *dd* we're talking about here.
>> It's *supposed* to be used for tricky stuff like this.
>>
>> If one interprets conv=sparse to mean "write sparsely",
>> rather than "create a sparse file that exactly mimics
>> the input's sparseness", then everything should be clear,
>> no?
> 
> Well I wasn't considering reproducing sparseness.
> I was thinking it would be more beneficial to be able to
> update a file in place, but do so sparsely if possible.
> I.E. support something like:
> 
> truncate -r src.img backup.img &&
> dd if=src.img conv=notrunc,sparse of=backup.img
> 
> To do that, one could not simply skip over the output
> for NUL input. Doing that would be "dangerous" as I said above,
> or in other words, surprising to users to not update
> possibly non NUL data in the output file.
> 
> As for the "inconsistent" point I mentioned. The last patch
> currently just seeks the output for NUL input,
> _except_ for the last byte of the file if part of a NUL write.
> So that inconsistency means one couldn't use NUL input as
> a write mask so to speak, if one ever did want such an
> esoteric feature. I guess we could address the consistency aspect
> by using ftruncate() rather than the write(outfd, "\0", 1) technique
> (only doing that if we've extended the file).
> However I'm not sure such functionality is useful for files,
> and that we should try to align with the first operation?
> Note ftruncate() might be good to do anyway to avoid
> writes completely to a fully sparse file.
> 
> Also I'm just thought about oflag=append.
> That will cause writes to ignore the seek beyond end of file.
> I.E. with the last patch, this will create a file.test with just "ab"
> 
>   printf "a\000\000b" |
>   src/dd conv=sparse bs=1 count=10 oflag=append > file.test
> 
> I guess an ftruncate() would cater for that too, except in the case
> where another process is writing to the file, in which case
> you would might want to disallow the conv=sparse oflag=append combo.
> We can't really ftruncate as we go as that would be racy,
> so I guess we should disallow the combo?

Ok the attached should be pretty much complete.
It disables conv=sparse when used with
conv=notrunc or oflag=append.

Comments appreciated.

Roman are you around?

cheers,
Pádraig.