Re: [PATCHv11 8/8] iomap: Add per-block dirty state tracking to improve performance

[Ritesh Harjani (IBM) <ritesh.list@gmail.com>]

Matthew Wilcox <willy@infradead.org> writes:

Sorry for the delayed response. I am currently on travel.

> On Fri, Jul 07, 2023 at 08:16:17AM +1000, Dave Chinner wrote:
>> On Thu, Jul 06, 2023 at 06:42:36PM +0100, Matthew Wilcox wrote:
>> > On Thu, Jul 06, 2023 at 08:16:05PM +0530, Ritesh Harjani wrote:
>> > > > @@ -1645,6 +1766,11 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
>> > > >  	int error = 0, count = 0, i;
>> > > >  	LIST_HEAD(submit_list);
>> > > >  
>> > > > +	if (!ifs && nblocks > 1) {
>> > > > +		ifs = ifs_alloc(inode, folio, 0);
>> > > > +		iomap_set_range_dirty(folio, 0, folio_size(folio));
>> > > > +	}
>> > > > +
>> > > >  	WARN_ON_ONCE(ifs && atomic_read(&ifs->write_bytes_pending) != 0);
>> > > >  
>> > > >  	/*
>> > > > @@ -1653,7 +1779,7 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
>> > > >  	 * invalid, grab a new one.
>> > > >  	 */
>> > > >  	for (i = 0; i < nblocks && pos < end_pos; i++, pos += len) {
>> > > > -		if (ifs && !ifs_block_is_uptodate(ifs, i))
>> > > > +		if (ifs && !ifs_block_is_dirty(folio, ifs, i))
>> > > >  			continue;
>> > > >  
>> > > >  		error = wpc->ops->map_blocks(wpc, inode, pos);
>> > > > @@ -1697,6 +1823,7 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
>> > > >  		}
>> > > >  	}
>> > > >  
>> > > > +	iomap_clear_range_dirty(folio, 0, end_pos - folio_pos(folio));
>> > > >  	folio_start_writeback(folio);
>> > > >  	folio_unlock(folio);
>> > > >  
>> > > 
>> > > I think we should fold below change with this patch. 
>> > > end_pos is calculated in iomap_do_writepage() such that it is either
>> > > folio_pos(folio) + folio_size(folio), or if this value becomes more then
>> > > isize, than end_pos is made isize.
>> > > 
>> > > The current patch does not have a functional problem I guess. But in
>> > > some cases where truncate races with writeback, it will end up marking
>> > > more bits & later doesn't clear those. Hence I think we should correct
>> > > it using below diff.
>> > 
>> > I don't think this is the only place where we'll set dirty bits beyond
>> > EOF.  For example, if we mmap the last partial folio in a file,
>> > page_mkwrite will dirty the entire folio, but we won't write back
>> > blocks past EOF.  I think we'd be better off clearing all the dirty
>> > bits in the folio, even the ones past EOF.  What do you think?

Yup. I agree, it's better that way to clear all dirty bits in the folio.
Thanks for the suggestion & nice catch!! 

>> 
>> Clear the dirty bits beyond EOF where we zero the data range beyond
>> EOF in iomap_do_writepage() via folio_zero_segment()?
>
> That would work, but I think it's simpler to change:
>
> -	iomap_clear_range_dirty(folio, 0, end_pos - folio_pos(folio));
> +	iomap_clear_range_dirty(folio, 0, folio_size(folio));

Right. 

@Darrick,
IMO, we should fold below change with Patch-8. If you like I can send a v12
with this change. I re-tested 1k-blocksize fstests on x86 with
below changes included and didn't find any surprise. Also v11 series
including the below folded change is cleanly applicable on your
iomap-for-next branch.


diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index b6280e053d68..de212b6fe467 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -1766,9 +1766,11 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
        int error = 0, count = 0, i;
        LIST_HEAD(submit_list);

+       WARN_ON_ONCE(end_pos <= pos);
+
        if (!ifs && nblocks > 1) {
                ifs = ifs_alloc(inode, folio, 0);
-               iomap_set_range_dirty(folio, 0, folio_size(folio));
+               iomap_set_range_dirty(folio, 0, end_pos - pos);
        }

        WARN_ON_ONCE(ifs && atomic_read(&ifs->write_bytes_pending) != 0);
@@ -1823,7 +1825,12 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
                }
        }

-       iomap_clear_range_dirty(folio, 0, end_pos - folio_pos(folio));
+       /*
+        * We can have dirty bits set past end of file in page_mkwrite path
+        * while mapping the last partial folio. Hence it's better to clear
+        * all the dirty bits in the folio here.
+        */
+       iomap_clear_range_dirty(folio, 0, folio_size(folio));
        folio_start_writeback(folio);
        folio_unlock(folio);

--
2.30.2


-ritesh