Re: [PATCH] dax: Fix missed PMD wakeups

[Matthew Wilcox]

On Wed, Jul 10, 2019 at 10:26:47PM +0200, Jan Kara wrote:
> On Wed 10-07-19 13:15:39, Matthew Wilcox wrote:
> > On Wed, Jul 10, 2019 at 09:02:04PM +0200, Jan Kara wrote:
> > > +#define DAX_ENTRY_CONFLICT dax_make_entry(pfn_to_pfn_t(1), DAX_EMPTY)
> > 
> > I was hoping to get rid of DAX_EMPTY ... it's almost unused now.  Once
> > we switch to having a single DAX_LOCK value instead of a single bit,
> > I think it can go away, freeing up two bits.
> > 
> > If you really want a special DAX_ENTRY_CONFLICT, I think we can make
> > one in the 2..4094 range.
> > 
> > That aside, this looks pretty similar to the previous patch I sent, so
> > if you're now happy with this, let's add
> > 
> > #define XA_DAX_CONFLICT_ENTRY xa_mk_internal(258)
> > 
> > to xarray.h and do it that way?
> 
> Yeah, that would work for me as well. The chosen value for DAX_ENTRY_CONFLICT
> was pretty arbitrary. Or we could possibly use:
> 
> #define DAX_ENTRY_CONFLICT XA_ZERO_ENTRY
> 
> so that we don't leak DAX-specific internal definition into xarray.h?

I don't want to use the ZERO entry as our conflict marker because that
could legitimately appear in an XArray.  Not the i_pages XArray today,
but I hold out hope for using that in place of the DAX_ZERO_PAGE bit too.
That's going to be a bit more tricky since we currently distinguish
between DAX_ZERO_PAGE and DAX_ZERO_PAGE | DAX_PMD.

However, the XA_RETRY_ENTRY might be a good choice.  It doesn't normally
appear in an XArray (it may appear if you're looking at a deleted node,
but since we're holding the lock, we can't see deleted nodes).