Re: [PATCH 1/2] fs: add SEEK_HOLE and SEEK_DATA flags

[Eric Blake]

On 04/22/2011 10:57 AM, Sunil Mushran wrote:
> On 04/22/2011 09:40 AM, Eric Blake wrote:
>> On 04/22/2011 10:28 AM, Sunil Mushran wrote:
>>> while(1) {
>>>      read(block);
>>>      if (block_all_zeroes)
>>>          lseek(SEEK_DATA);
>>> }
>>>
>>> What's wrong with the above? If this is the case, even SEEK_HOLE
>>> is not needed but should be added as it is already in Solaris.
> 
> Holes are an implementation detail.

Nobody's arguing that.  And on Solaris, a file system with no holes
support tells you that up front - lseek(fd, 0, SEEK_HOLE) returns the
end of the file (unless the file is 0 bytes, then it fails with ENXIO).

> 
> cp can read whatever blocksize it chooses. If that block contains
> zero, it would signal cp that maybe it should SEEK_DATA and skip
> reading all those blocks. That's all. We are not trying to achieve
> perfection. We are just trying to reduce cpu waste.
> 
> If the fs supports SEEK_*, then great. If it does not, then it is no
> worse than before.

But providing just SEEK_DATA _is_ worse than before if you don't provide
the correct SEEK_HOLE everywhere.  Because then your algorithm of trying
lseek(SEEK_DATA) after every run of zeros in the hopes of an
optimization is a wasted syscall, since it will just return your current
offset every time, so you end up with more syscalls than if you had used
the single lseek(SEEK_DATA) that returns the end of the file up front,
and known that the remainder of the file has no holes to even try
seeking past.

-- 
Eric Blake   eblake@xxxxxxxxxx    +1-801-349-2682
Libvirt virtualization library http://libvirt.org

Attachment: signature.asc
Description: OpenPGP digital signature