On Tue, Apr 10, 2018 at 05:05:28AM -0700, Matthew Wilcox wrote:
> On Tue, Apr 10, 2018 at 10:26:43AM +0200, Michal Hocko wrote:
> > On Mon 09-04-18 12:40:44, Matthew Wilcox wrote:
> > > The problem is that the mapping gfp flags are used not only for allocating
> > > pages, but also for allocating the page cache data structures that hold
> > > the pages. F2FS is the only filesystem that set the __GFP_ZERO bit,
> > > so it's the first time anyone's noticed that the page cache passes the
> > > __GFP_ZERO bit through to the radix tree allocation routines, which
> > > causes the radix tree nodes to be zeroed instead of constructed.
> > >
> > > I think the right solution to this is:
> > This just hides the underlying problem that the node is not fully and
> > properly initialized. Relying on the previous released state is just too
> > subtle.
> That's the fundamental design of slab-with-constructors. The user provides
> a constructor, so all newly allocagted objects are initialised to a known
> state, then the user will restore the object to that state when it frees
> the object to slab.
Fully agreed. I'm surprised there is so much pushback to what Willy is
saying here, this stuff has been established for years.
> > Are you going to blacklist all potential gfp flags that come
> > from the mapping? This is just unmaintainable! If anything this should
> > be an explicit & with the allowed set of allowed flags.
> Oh, I agree that using the set of flags used to allocate the page
> in order to allocate the radix tree nodes is a pretty horrible idea.
> Your suggestion, then, is:
> - error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
> + error = radix_tree_preload(gfp_mask & GFP_RECLAIM_MASK);
That looks reasonable to me.