[Cluster-devel] [PATCH 2/2] gfs2: writeout truncated pages

[Benjamin Marzinski]

On Thu, Jun 16, 2016 at 04:56:49PM +0100, Steven Whitehouse wrote:
> Hi,
> 
> On 14/06/16 23:02, Benjamin Marzinski wrote:
> >When gfs2 attempts to write a page to a file that is being truncated,
> >and notices that the page is completely outside of the file size, it
> >tries to invalidate it.  However, this may require a transaction for
> >journaled data files to revoke any buffers from the page on the active
> >items list. Unfortunately, this can happen inside a log flush, where a
> >transaction cannot be started. Also, gfs2 may need to be able to remove
> >the buffer from the ail1 list before it can finish the log flush.
> >
> >To deal with this, gfs2 now skips the check to see if the write is
> >outside the file size, and simply writes it anyway. This situation can
> >only occur when the truncate code still has the file locked exclusively,
> >and hasn't marked this block as free in the metadata (which happens
> >later in truc_dealloc).  After gfs2 writes this page out, the truncation
> >code will shortly invalidate it and write out any revokes if necessary.
> >
> >To do this, gfs2 now implements its own version of block_write_full_page
> >without the check, and calls the newly exported __block_write_full_page.
> >The check still exists in nobh_writepage, which is gfs2 calls in the
> >non-journaled case. But in this case the buffers will never be in the
> >journal. Thus, no revokes will need to be issued and the write can
> >safely be skipped without causing any possible journal replay issues.
> >
> >Signed-off-by: Benjamin Marzinski <bmarzins at redhat.com>
> >---
> >  fs/gfs2/aops.c | 37 +++++++++++++++++++++++++++----------
> >  1 file changed, 27 insertions(+), 10 deletions(-)
> >
> >diff --git a/fs/gfs2/aops.c b/fs/gfs2/aops.c
> >index 37b7bc1..d3a7301 100644
> >--- a/fs/gfs2/aops.c
> >+++ b/fs/gfs2/aops.c
> >@@ -100,20 +100,11 @@ static int gfs2_writepage_common(struct page *page,
> >  	struct inode *inode = page->mapping->host;
> >  	struct gfs2_inode *ip = GFS2_I(inode);
> >  	struct gfs2_sbd *sdp = GFS2_SB(inode);
> >-	loff_t i_size = i_size_read(inode);
> >-	pgoff_t end_index = i_size >> PAGE_SHIFT;
> >-	unsigned offset;
> >  	if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
> >  		goto out;
> >  	if (current->journal_info)
> >  		goto redirty;
> >-	/* Is the page fully outside i_size? (truncate in progress) */
> >-	offset = i_size & (PAGE_SIZE-1);
> >-	if (page->index > end_index || (page->index == end_index && !offset)) {
> >-		page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
> >-		goto out;
> >-	}
> >  	return 1;
> I wonder whether it would make more sense to simply stop the jdata writepage
> calling the writepage_common function altogether and move the remaining
> checks to the jdata writepage function. That way the existing writepage
> would continue to do the invalidation, or does this check not buy us
> anything in the ordered/writeback cases either?
> 
> Otherwise I think it all looks good,

Like I told Bob, we already run into cases where the file i_size is
shrunk after we call gfs2_writepage_common, and we rely on
nobh_writepage to do the check.  So, from a correctness point of view, I
don't think this changes much.  We are invalidating the page sooner, so
it's possible that there is some difference in performance, but I don't
know that it would be measureable.

If people would prefer, I'm fine following this suggestion, since it
means that we aren't changing code that has been working fine all along,
and doing it my way certainly won't improve our performance.

-Ben

> 
> Steve.
> 
> >  redirty:
> >  	redirty_page_for_writepage(wbc, page);
> >@@ -140,6 +131,32 @@ static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
> >  	return nobh_writepage(page, gfs2_get_block_noalloc, wbc);
> >  }
> >+/* This is the same as calling block_write_full_page, but it also
> >+ * writes pages outside of i_size
> >+ */
> >+int gfs2_write_full_page(struct page *page, get_block_t *get_block,
> >+			 struct writeback_control *wbc)
> >+{
> >+	struct inode * const inode = page->mapping->host;
> >+	loff_t i_size = i_size_read(inode);
> >+	const pgoff_t end_index = i_size >> PAGE_SHIFT;
> >+	unsigned offset;
> >+
> >+	/*
> >+	 * The page straddles i_size.  It must be zeroed out on each and every
> >+	 * writepage invocation because it may be mmapped.  "A file is mapped
> >+	 * in multiples of the page size.  For a file that is not a multiple of
> >+	 * the  page size, the remaining memory is zeroed when mapped, and
> >+	 * writes to that region are not written out to the file."
> >+	 */
> >+	offset = i_size & (PAGE_SIZE-1);
> >+	if (page->index == end_index && offset)
> >+		zero_user_segment(page, offset, PAGE_SIZE);
> >+
> >+	return __block_write_full_page(inode, page, get_block, wbc,
> >+				       end_buffer_async_write);
> >+}
> >+
> >  /**
> >   * __gfs2_jdata_writepage - The core of jdata writepage
> >   * @page: The page to write
> >@@ -165,7 +182,7 @@ static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *w
> >  		}
> >  		gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
> >  	}
> >-	return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
> >+	return gfs2_write_full_page(page, gfs2_get_block_noalloc, wbc);
> >  }
> >  /**