Re: [PATCH][RFC] splice support

[Jens Axboe]

On Thu, Mar 30 2006, Andrew Morton wrote:
> Jens Axboe <axboe@xxxxxxx> wrote:
> >
> > 
> > ...
> >
> > > - I think the `size_t left' in do_splice_to() can overflow if f_pos is
> > >   sufficiently different from i_size.
> > 
> > They're both loff_t.
> 
> Nope:
> 
> +static long do_splice_to(struct file *in, struct inode *pipe, size_t len,
> +			 unsigned long flags)
> +{
> +	if (in->f_op && in->f_op->splice_read) {
> +		loff_t isize = i_size_read(in->f_mapping->host);
> +		size_t left;
> +
> +		if (unlikely(in->f_pos >= isize))
> +			return 0;
> +	
> +		left = isize - in->f_pos;
> 
> It's doing
> 
> 		32bit = 64bit - 64bit;

My mistake, I looked at a different spot. Fixed.

> > 
> > > - In generic_file_splice_read():
> > > 
> > >   - nonatomic modification of f_pos.  Is i_mutex held?  (see
> > >     generic_file_llseek())
> > 
> > Fixed.
> 
> OK.  In some ways I agree with Nick that a pwrite/pread-like interface is
> nicer, so things are more stateless and threads don't have to fight over
> f_pos.  Dunno..

I can go either way, I tend to agree that passing in offset may be the
best solution.

> > >   - These pages can get truncated at any time they're unlocked.  Does
> > >     the code cope with all that?
> > 
> > I guess page_cache_pipe_buf_map() needs the same ->mapping check?
> 
> That would seem appropriate.
> 
> btw, that function might have a problem I think - it returns NULL with
> the page locked, but pipe_to_sendpage() and other callers don't appear to
> unlock it.

Will fix that up.

> > > - hm.  What happens if the pages which find_get_pages() returned are
> > > not contiguous in pagecache?  I think your `pages' array gets all
> > > jumbled up.
> >
> > Hmm please expand. 
> 
> find_get_pages() does "find me the next N pages above `index' which are
> presently in pagecache'.  So it can return an array of page*'s which do not
> represent contiguous pages in the file - there can be holes in there.
> 
> IOW: pages[n]->index !necessarily= pages[n+1]->index-1
> 
> Maybe the code handles that by making sure that all the pages in the range
> are already in pagecache - I didn't check.  But that would take some heroic
> locking.

It doesn't, I'm assuming that find_get_pages() returns consequtive pages
atm. Would seem like the sane interface :-)

We continue doing find_or_create_page() on the remaining, but using 'i'
as the 'index' addition. So if we had non-conseq pages, we'd be screwed.

Needs a little thinking, input welcome..

-- 
Jens Axboe

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/