On Mon, Feb 08, 2021 at 07:26:25PM -0400, Jason Gunthorpe wrote: > On Mon, Feb 08, 2021 at 05:02:59PM -0500, Peter Xu wrote: > > On Mon, Feb 08, 2021 at 02:51:33PM -0400, Jason Gunthorpe wrote: > > > On Fri, Feb 05, 2021 at 01:14:11PM -0500, Peter Xu wrote: > > > > > > > But I do have a question on why dax as the only user needs to pass in the > > > > notifier to follow_pte() for initialization. > > > > > > Not sure either, why does DAX opencode something very much like > > > page_mkclean() with dax_entry_mkclean()? > > > > > > Also it looks like DAX uses the wrong notifier, it calls > > > MMU_NOTIFY_CLEAR but page_mkclean_one() uses > > > MMU_NOTIFY_PROTECTION_PAGE for the same PTE modification sequence?? > > > > > > page_mkclean() has some technique to make the notifier have the right > > > size without becoming entangled in the PTL locks.. > > > > Right. I guess it's because dax doesn't have "struct page*" on the > > back, so it > > It doesn't? I thought DAX cases did? I'm not familiar with dax at all.. but it seems so: e.g. dax_iomap_pte_fault() looks like the general fault handler for dax mappings, in which there's calls to things like vmf_insert_mixed_mkwrite() trying to install ptes with raw pfn. Or I could also be missing something very important.. Thanks, -- Peter Xu
