Re: [RFC PATCH V2] iomap: add support to track dirty state of sub pages

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Tue, Aug 18, 2020 at 04:53:05PM +0100, Matthew Wilcox wrote:
> On Tue, Aug 18, 2020 at 09:46:18PM +0800, Yu Kuai wrote:
> > changes from v1:
> >  - separate set dirty and clear dirty functions
> >  - don't test uptodate bit in iomap_writepage_map()
> >  - use one bitmap array for uptodate and dirty.
> 
> This looks much better.
> 
> > +	spinlock_t		state_lock;
> > +	/*
> > +	 * The first half bits are used to track sub-page uptodate status,
> > +	 * the second half bits are for dirty status.
> > +	 */
> > +	DECLARE_BITMAP(state, PAGE_SIZE / 256);
> 
> It would be better to use the same wording as below:
> 
> > +	bitmap_zero(iop->state, PAGE_SIZE * 2 / SECTOR_SIZE);

ISTR there was some reason why '512' was hardcoded in here instead of
SECTOR_SIZE.  I /think/ it was so that iomap.h did not then have a hard
dependency on blkdev.h and everything else that requires...

https://lore.kernel.org/linux-xfs/20181215105155.GD1575@lst.de/

--D

> 
> [...]
> 
> > +static void
> > +iomap_iop_set_range_dirty(struct page *page, unsigned int off,
> > +		unsigned int len)
> > +{
> > +	struct iomap_page *iop = to_iomap_page(page);
> > +	struct inode *inode = page->mapping->host;
> > +	unsigned int total = PAGE_SIZE / SECTOR_SIZE;
> > +	unsigned int first = off >> inode->i_blkbits;
> > +	unsigned int last = (off + len - 1) >> inode->i_blkbits;
> > +	unsigned long flags;
> > +	unsigned int i;
> > +
> > +	spin_lock_irqsave(&iop->state_lock, flags);
> > +	for (i = first; i <= last; i++)
> > +		set_bit(i + total, iop->state);
> > +	spin_unlock_irqrestore(&iop->state_lock, flags);
> > +}
> 
> How about:
> 
> -	unsigned int total = PAGE_SIZE / SECTOR_SIZE;
> ...
> +	first += PAGE_SIZE / SECTOR_SIZE;
> +	last += PAGE_SIZE / SECTOR_SIZE;
> ...
> 	for (i = first; i <= last; i++)
> -		set_bit(i + total, iop->state);
> +		set_bit(i, iop->state);
> 
> We might want
> 
> #define	DIRTY_BITS(x)	((x) + PAGE_SIZE / SECTOR_SIZE)
> 
> and then we could do:
> 
> +	unsigned int last = DIRTY_BITS((off + len - 1) >> inode->i_blkbits);
> 
> That might be overthinking things a bit though.
> 
> > @@ -705,6 +767,7 @@ __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
> >  	if (unlikely(copied < len && !PageUptodate(page)))
> >  		return 0;
> >  	iomap_set_range_uptodate(page, offset_in_page(pos), len);
> > +	iomap_set_range_dirty(page, offset_in_page(pos), len);
> >  	iomap_set_page_dirty(page);
> 
> I would move the call to iomap_set_page_dirty() into
> iomap_set_range_dirty() to parallel iomap_set_range_uptodate more closely.
> We don't want a future change to add a call to iomap_set_range_dirty()
> and miss the call to iomap_set_page_dirty().
> 
> >  	return copied;
> >  }
> > @@ -1030,6 +1093,7 @@ iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
> >  		WARN_ON_ONCE(!PageUptodate(page));
> >  		iomap_page_create(inode, page);
> >  		set_page_dirty(page);
> > +		iomap_set_range_dirty(page, offset_in_page(pos), length);
> 
> I would move all this from the mkwrite_actor() to iomap_page_mkwrite()
> and call it once with (0, PAGE_SIZE) rather than calling it once for
> each extent in the page.
> 
> > @@ -1435,6 +1500,8 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
> >  		 */
> >  		set_page_writeback_keepwrite(page);
> >  	} else {
> > +		iomap_clear_range_dirty(page, 0,
> > +				end_offset - page_offset(page) + 1);
> >  		clear_page_dirty_for_io(page);
> >  		set_page_writeback(page);
> 
> I'm not sure it's worth doing this calculation.  Better to just clear
> the dirty bits on the entire page?  Opinions?