Re: [PATCH] tmpfs: fix regressions from wider use of ZERO_PAGE

[Hugh Dickins]

On Sat, 9 Apr 2022, Christoph Hellwig wrote:
> On Fri, Apr 08, 2022 at 01:38:41PM -0700, Hugh Dickins wrote:
> > +		} else if (iter_is_iovec(to)) {
> > +			/*
> > +			 * Copy to user tends to be so well optimized, but
> > +			 * clear_user() not so much, that it is noticeably
> > +			 * faster to copy the zero page instead of clearing.
> > +			 */
> > +			ret = copy_page_to_iter(ZERO_PAGE(0), offset, nr, to);
> 
> Is the offset and length guaranteed to be less than PAGE_SIZE here?

Almost :) The offset is guaranteed to be less than PAGE_SIZE here, and
the length is guaranteed to be less than or equal to PAGE_SIZE - offset.

> 
> Either way I'd rather do this optimization in iov_iter_zero rather
> than hiding it in tmpfs.

Let's see what others say.  I think we would all prefer clear_user() to be
enhanced, and hack around it neither here in tmpfs nor in iov_iter_zero().
But that careful work won't get done by magic, nor by me.

And iov_iter_zero() has to deal with a wider range of possibilities,
when pulling in cache lines of ZERO_PAGE(0) will be less advantageous,
than in tmpfs doing a large dd - the case I'm aiming not to regress here
(tmpfs has been copying ZERO_PAGE(0) like this for years).

Hugh