On Mon, Jan 13, 2025 at 10:12:41PM -0800, Dan Williams wrote: > Alistair Popple wrote: > > Device DAX pages are currently not reference counted when mapped, > > instead relying on the devmap PTE bit to ensure mapping code will not > > get/put references. This requires special handling in various page > > table walkers, particularly GUP, to manage references on the > > underlying pgmap to ensure the pages remain valid. > > > > However there is no reason these pages can't be refcounted properly at > > map time. Doning so eliminates the need for the devmap PTE bit, > > freeing up a precious PTE bit. It also simplifies GUP as it no longer > > needs to manage the special pgmap references and can instead just > > treat the pages normally as defined by vm_normal_page(). > > > > Signed-off-by: Alistair Popple <apopple@xxxxxxxxxx> > > --- > > drivers/dax/device.c | 15 +++++++++------ > > mm/memremap.c | 13 ++++++------- > > 2 files changed, 15 insertions(+), 13 deletions(-) > > > > diff --git a/drivers/dax/device.c b/drivers/dax/device.c > > index 6d74e62..fd22dbf 100644 > > --- a/drivers/dax/device.c > > +++ b/drivers/dax/device.c > > @@ -126,11 +126,12 @@ static vm_fault_t __dev_dax_pte_fault(struct dev_dax *dev_dax, > > return VM_FAULT_SIGBUS; > > } > > > > - pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP); > > + pfn = phys_to_pfn_t(phys, 0); > > > > dax_set_mapping(vmf, pfn, fault_size); > > > > - return vmf_insert_mixed(vmf->vma, vmf->address, pfn); > > + return vmf_insert_page_mkwrite(vmf, pfn_t_to_page(pfn), > > + vmf->flags & FAULT_FLAG_WRITE); > > } > > > > static vm_fault_t __dev_dax_pmd_fault(struct dev_dax *dev_dax, > > @@ -169,11 +170,12 @@ static vm_fault_t __dev_dax_pmd_fault(struct dev_dax *dev_dax, > > return VM_FAULT_SIGBUS; > > } > > > > - pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP); > > + pfn = phys_to_pfn_t(phys, 0); > > > > dax_set_mapping(vmf, pfn, fault_size); > > > > - return vmf_insert_pfn_pmd(vmf, pfn, vmf->flags & FAULT_FLAG_WRITE); > > + return vmf_insert_folio_pmd(vmf, page_folio(pfn_t_to_page(pfn)), > > + vmf->flags & FAULT_FLAG_WRITE); > > This looks suspect without initializing the compound page metadata. I initially wondered about this too, however I think the compound page metadata should be initialised by memmap_init_zone_device(). That said I kind of get lost in all the namespace/CXL/PMEM/DAX drivers in the stack so maybe I've overlooked something. > This might be getting compound pages by default with > CONFIG_ARCH_WANT_OPTIMIZE_DAX_VMEMMAP. The device-dax unit tests are ok > so far, but that is not super comforting until I can think about this a > bit more... but not tonight.
