On Tue, 2016-09-06 at 14:52 -0700, Dan Williams wrote: > On Tue, Sep 6, 2016 at 1:17 PM, Andrew Morton <akpm@linux-foundation. > org> wrote: > > > > On Tue, 06 Sep 2016 09:49:41 -0700 Dan Williams <dan.j.williams@int > > el.com> wrote: > > > > > > > > track_pfn_insert() is marking dax mappings as uncacheable. > > > > > > It is used to keep mappings attributes consistent across a > > > remapped range. However, since dax regions are never registered > > > via track_pfn_remap(), the caching mode lookup for dax pfns > > > always returns _PAGE_CACHE_MODE_UC. We do not use > > > track_pfn_insert() in the dax-pte path, and we always want to use > > > the pgprot of the vma itself, so drop this call. > > > > > > Cc: Ross Zwisler <ross.zwisler@xxxxxxxxxxxxxxx> > > > Cc: Matthew Wilcox <mawilcox@xxxxxxxxxxxxx> > > > Cc: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx> > > > Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> > > > Cc: Nilesh Choudhury <nilesh.choudhury@xxxxxxxxxx> > > > Reported-by: Kai Zhang <kai.ka.zhang@xxxxxxxxxx> > > > Reported-by: Toshi Kani <toshi.kani@xxxxxxx> > > > Cc: <stable@xxxxxxxxxxxxxxx> > > > Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx> > > > > Changelog fails to explain the user-visible effects of the > > patch. The stable maintainer(s) will look at this and wonder "ytf > > was I sent this". > > True, I'll change it to this: > > track_pfn_insert() is marking dax mappings as uncacheable rendering > them impractical for application usage. DAX-pte mappings are cached > and the goal of establishing DAX-pmd mappings is to attain more > performance, not dramatically less (3 orders of magnitude). > > Deleting the call to track_pfn_insert() in vmf_insert_pfn_pmd() lets > the default pgprot (write-back cache enabled) from the vma be used > for the mapping which yields the expected performance improvement > over DAX-pte mappings. > > track_pfn_insert() is meant to keep the cache mode for a given range > synchronized across different users of remap_pfn_range() and > vm_insert_pfn_prot(). DAX uses neither of those mapping methods, and > the pmem driver is already marking its memory ranges as write-back > cache enabled. So, removing the call to track_pfn_insert() leaves > the kernel no worse off than the current situation where a user could > map the range via /dev/mem with an incompatible cache mode compared > to the driver. I think devm_memremap_pages() should call reserve_memtype() on x86 to keep it consistent with devm_memremap() on this regard. We may need an arch stub for reserve_memtype(), though. Then, track_pfn_insert() should have no issue in this case. Thanks, -Toshi��.n������g��� �a��� �&ޖ)���)��h���&������梷���� �Ǟ�m������)������^�����������v���O��zf������