On Mon, Apr 18, 2011 at 10:28:10PM +0300, Kostik Belousov wrote:
> > + if (offset + len > vap->va_size) {
> > + VATTR_NULL(vap);
> > + vap->va_size = offset + len;
> > + error = VOP_SETATTR(vp, vap, td->td_ucred);
> > + if (error != 0)
> > + goto out;
> > + }
> I still do not see a reason to do VOP_SETATTR() there. VOP_WRITE() will
> do auto-extend as needed. Also, see below.
Yeah, also when we extend file size we could skip reading zeros.
> > + if (offset < vap->va_size) {
[...]
> > + } else {
> > + bzero(buf, cur);
> > + }
> Wouldn't VOP_SETATTR() at the start of the function mostly prevent
> this bzero from executing ?
Once we drop the vnode lock, the file size can change under us, no?
> I estimated what it would take to do the optimized implementation for UFS,
> and I think that the following change would allow to lessen the code
> duplication much.
>
> What if the vnode lock drop and looping be handled by the syscall, instead
> of the vop implementation ? In other words, allow the VOP_ALLOCATE()
> to allocate less then requested, and return the allocated amount to
> the caller. The loop would be centralized then, freeing fs from doing
> the dance. Also, if fs considers that suitable, it would do a whole
> allocation in one run.
I'd still go with SEEK_DATA/SEEK_HOLE loop as I suggested on arch@.
If you would like to spend time on it, having SEEK_DATA/SEEK_HOLE
support in UFS would be beneficial for other purposes too.
--
Pawel Jakub Dawidek http://www.wheelsystems.com
FreeBSD committer http://www.FreeBSD.org
Am I Evil? Yes, I Am! http://yomoli.com
pgp7uMUMXLSum.pgp
Description: PGP signature
