On Sun, Jan 31, 2016 at 10:27 AM, Matthew Wilcox <willy(a)linux.intel.com> wrote:
On Sun, Jan 31, 2016 at 10:18:46AM -0800, Dan Williams wrote:
> On Sun, Jan 31, 2016 at 10:07 AM, Matthew Wilcox <willy(a)linux.intel.com>
> > Yes. This is just the default implementation of dax_map_pfn() which works
> > for most situations. We can introduce more complex implementations of
> > dax_map_pfn() as necessary. You make another excellent point for why
> > we should store PFNs in the radix tree instead of kaddrs :-)
> How much complexity do we want to add in support of an fsync/msync
> mechanism that is not the recommended way to use DAX?
It actually makes the dax_io path much, much simpler. And it's not
primarily about fixing fsync/msync. It also makes the fault path cheaper
in the case where we're refaulting a page that's already been faulted
by another process (or was previously faulted by this process and now
needs to be faulted at a different address).
And it fixes the problem with filesystems that use multiple block_devices.
It also makes DAX much less reliant on buffer heads, which is good for
the problem that Jared raised where he doesn't have a block_device in
an embedded system.
Oh I thought we were talking about what goes in the radix. Sure,
de-emphasizing the usage of a block_device throughout the dax
implementation is interesting. It also has some synergy with the
LSF/MM topic I'm writing up "pmem as storage device vs pmem as