Re: [PATCH] mm: Remove redundant test from find_get_pages_contig

[Matthew Wilcox]

On Mon, Jan 07, 2019 at 03:09:04PM -0800, Andrew Morton wrote:
> On Mon, 7 Jan 2019 14:39:35 -0800 Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> 
> > On Mon, Jan 07, 2019 at 02:33:19PM -0800, Andrew Morton wrote:
> > > On Mon,  7 Jan 2019 12:02:24 -0800 Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > > 
> > > > After we establish a reference on the page, we check the pointer continues
> > > > to be in the correct position in i_pages.  There's no need to check the
> > > > page->mapping or page->index afterwards; if those can change after we've
> > > > got the reference, they can change after we return the page to the caller.
> > > 
> > > But that isn't what the comment says.
> > 
> > Right.  That patch from Nick moved the check from before taking the
> > ref to after taking the ref.  It was racy to have it before.  But it's
> > unnecessary to have it afterwards -- pages can't move once there's a
> > ref on them.  Or if they can move, they can move after the ref is taken.
> 
> So Nick's patch was never necessary?  I wonder what inspired it.

It was necessary to not check before the pin; that was clearly correct.
Checking after the pin, even with the code the way it was in 2006, was
unnecessary.  Look with a bit more context:

-               if (page->mapping == NULL || page->index != index)
-                       break;
-
                if (!page_cache_get_speculative(page))
                        goto repeat;
 
                /* Has the page moved? */
                if (unlikely(page != *((void **)pages[i]))) {
                        page_cache_release(page);
                        goto repeat;
                }
 
+               /*
+                * must check mapping and index after taking the ref.
+                * otherwise we can get both false positives and false
+                * negatives, which is just confusing to the caller.
+                */
+               if (page->mapping == NULL || page->index != index) {
+                       page_cache_release(page);
+                       break;
+               }
+

It's not immediately obvious that those added lines merely re-check the
condition checked by the 'page != *((void **)pages[i])', but if you think
about it, if page->index changes, then page must necessarily move within
the radix tree / xarray.

> Would it be excessively cautious to put a WARN_ON_ONCE() in there for a
> while?

I think it would ... it'd get in the way of a subsequent patch to store
only head pages in the page cache.