Re: [PATCH v10] mm: introduce MEMORY_DEVICE_FS_DAX and CONFIG_DEV_PAGEMAP_OPS

On Tue, 29 May 2018 13:26:33 -0700 Dan Williams <dan.j.williams(a)intel.com&gt; wrote: > On Wed, May 23, 2018 at 11:50 AM, Gerald Schaefer > <gerald.schaefer(a)de.ibm.com&gt; wrote: > > On Tue, 22 May 2018 08:28:06 +0200 > > Christoph Hellwig <hch(a)lst.de&gt; wrote: > > > >> On Mon, May 21, 2018 at 11:04:10AM +0200, Jan Kara wrote: > >> > We definitely do have customers using "execute in place" on s390x from > >> > dcssblk. I've got about two bug reports for it when customers were updating > >> > from old kernels using original XIP to kernels using DAX. So we need to > >> > keep that working. > >> > >> That is all good an fine, but I think time has come where s390 needs > >> to migrate to provide the pmem API so that we can get rid of these > >> special cases. Especially given that the old XIP/legacy DAX has all > >> kinds of known bugs at this point in time. > > > > I haven't yet looked at this patch series, but I can feel that this > > FS_DAX_LIMITED workaround is beginning to cause some headaches, apart > > from being quite ugly of course. > > > > Just to make sure I still understand the basic problem, which I thought > > was missing struct pages for the dcssblk memory, what exactly do you > > mean with "provide the pmem API", is there more to do? > > No, just 'struct page' is needed. > > What used to be the pmem API is now pushed down into to dax_operations > provided by the device driver. dcssblk is free to just redirect to the > generic implementations for copy_from_iter() and copy_to_iter(), and > be done. I.e. we've removed the "pmem API" requirement. > > > I do have a prototype patch lying around that adds struct pages, but > > didn't yet have time to fully test/complete it. Of course we initially > > introduced XIP as a mechanism to reduce memory consumption, and that > > is probably the use case for the remaining customer(s). Adding struct > > pages would somehow reduce that benefit, but as long as we can still > > "execute in place", I guess it will be OK. > > The pmem driver has the option to allocate the 'struct page' map out > of pmem directly. If the overhead of having the map in System RAM is > too high it could borrow the same approach, but that adds another > degree of configuration complexity freedom. > Thanks for clarifying, and mentioning the pmem altmap support, that looks interesting. I also noticed that I probably should enable CONFIG_ZONE_DEVICE for s390, and use devm_memremap_pages() to get the struct pages, rather than my homegrown solution so far. This will take some time however, so I hope you can live with the FS_DAX_LIMITED a little longer.