On Wed, Nov 16, 2016 at 12:01:01PM +0100, Jan Kara wrote: > On Wed 16-11-16 11:51:32, Peter Zijlstra wrote: > > Now, I'm entirely out of touch wrt DAX, so I've not idea what that > > needs/wants. > > Yeah, DAX does not have 'struct page' for its pages so it directly installs > PFNs in the page tables. As a result it needs to know about page tables and > stuff. Not convinced, a physical address should then be the equivalent of a struct page. You still don't need access to the actual pages tables. The VM core can then convert the physical address to a PFN and stuff it in the PTE entry. > Now I've abstracted knowledge about that into helper functions back > in mm/ but still we need to pass the information through the ->fault handler > into those helpers and vm_fault structure is simply natural for that. > So far we have tried to avoid that but the result was not pretty (special > return codes from DAX ->fault handlers essentially leaking information > about DAX internal locking into mm/ code to direct generic mm code to do > the right thing for DAX). Its probably the DAX locking bit I'm missing, because I cannot see why VM_FAULT_DAX_LOCKED is 'broken' -- also, I'd have called that VM_FAULT_PFN or similar and not have used the full entry but only the PFN bits from it. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>