Re: [PATCH 07/14] iomap: Support THPs in readpage

From: Matthew Wilcox <willy@infradead.org>
To: "Darrick J. Wong" <darrick.wong@oracle.com>
Cc: linux-fsdevel@vger.kernel.org, linux-xfs@vger.kernel.org,
	linux-mm@kvack.org
Subject: Re: [PATCH 07/14] iomap: Support THPs in readpage
Date: Wed, 14 Oct 2020 18:35:00 +0100	[thread overview]
Message-ID: <20201014173500.GR20115@casper.infradead.org> (raw)
In-Reply-To: <20201014163936.GJ9832@magnolia>

On Wed, Oct 14, 2020 at 09:39:36AM -0700, Darrick J. Wong wrote:
> On Wed, Oct 14, 2020 at 04:03:50AM +0100, Matthew Wilcox (Oracle) wrote:
> > +/*
> > + * The page that was passed in has become Uptodate.  This may be due to
> > + * the storage being synchronous or due to a page split finding the page
> > + * is actually uptodate.  The page is still locked.
> > + * Lift this into the VFS at some point.
> > + */
> > +#define AOP_UPDATED_PAGE       (AOP_TRUNCATED_PAGE + 1)
> 
> Er... why not lift it now?
> 
> "Because touching fs.h causes the whole kernel to be rebuilt and that's
> annoying"? :D

Hah!  Because I already have a patch series out that does lift it to
the VFS:

https://lore.kernel.org/linux-fsdevel/20201009143104.22673-1-willy@infradead.org/

which I'm kind of hoping gets merged first.  Then I can adjust this patch.

> > +static int iomap_split_page(struct inode *inode, struct page *page)
> > +{
> > +	struct page *head = thp_head(page);
> > +	bool uptodate = iomap_range_uptodate(inode, head,
> > +				(page - head) * PAGE_SIZE, PAGE_SIZE);
> > +
> > +	iomap_page_release(head);
> > +	if (split_huge_page(page) < 0) {
> > +		unlock_page(page);
> > +		return AOP_TRUNCATED_PAGE;
> > +	}
> > +	if (!uptodate)
> > +		return 0;
> > +	SetPageUptodate(page);
> > +	return AOP_UPDATED_PAGE;
> > +}
> > +
> >  int
> >  iomap_readpage(struct page *page, const struct iomap_ops *ops)
> >  {
> >  	struct iomap_readpage_ctx ctx = { .cur_page = page };
> > -	struct inode *inode = page->mapping->host;
> > +	struct inode *inode = thp_head(page)->mapping->host;
> >  	unsigned poff;
> >  	loff_t ret;
> >  
> > -	trace_iomap_readpage(page->mapping->host, 1);
> > +	trace_iomap_readpage(inode, 1);
> > +
> > +	if (PageTransCompound(page)) {
> > +		int status = iomap_split_page(inode, page);
> > +		if (status == AOP_UPDATED_PAGE) {
> > +			unlock_page(page);
> 
> /me wonders why not do the unlock in iomap_split_page?

This is working around AOP_UPDATED_PAGE not existing in the VFS.  So
adjusting this patch becomes:

		int status = iomap_split_page(inode, page);
		if (status)
			return status;