Dan Williams <dan.j.williams@xxxxxxxxx> writes: > Typical pages take a reference at pte insertion and drop it at > zap_pte_range() time. That reference management is missing for DAX > leading to a situation where DAX pages are mapped in user page tables, > but are not referenced. > > Once fsdax decides it wants to unmap the page it can drop its reference, > but unlike typical pages it needs to maintain the association of the > page to the inode that arbitrated the access in the first instance. It > maintains that association until explicit truncate(), or the implicit > truncate() that occurs at inode death, truncate_inode_pages_final(). > > The zapped state tracks whether the fsdax has dropped its interest in a > page, but still allows the associated i_pages entry to live until > truncate. This facilitates inode lookup while awaiting any page pin > users to drop their pins. For example, if memory_failure() is triggered > on the page after it has been unmapped, but before it has been truncated > from the inode, memory_failure() can still associate the event with the > inode. > > Once truncate begins fsdax unmaps the page to prevent any new references > from being taken without calling back into fsdax core to reestablish > the mapping. The gup path should now check dax_is_zap()? Where do we prevent new referenced from being taken? -aneesh
