Hi all, thanks for taking the time to comment. Replies in-line.
On Wed, Apr 19, 2017 at 08:53:13AM -0700, Laura Abbott wrote: > On 04/18/2017 11:48 AM, Ard Biesheuvel wrote: > >On 18 April 2017 at 19:21, Mark Rutland <mark.rutl...@arm.com> wrote: > >>On Fri, Apr 14, 2017 at 03:01:58PM +0100, Andrea Reale wrote: [...] > >> > >> From a quick scan, I see that it's necessary to use pgtable_page_ctor() > >>for pages that will be used for userspace page tables, but it's not > >>clear to me if it's ever necessary for pages used for kernel page > >>tables. > >> > >>If it is, we appear to have a bug on arm64. > >> > >>Laura, Ard, thoughts? > >> > > > >The generic apply_to_page_range() will expect the PTE lock to be > >initialized for page table pages that are not part of init_mm. For > >arm64, that is precisely efi_mm as far as I am aware. For EFI, the > >locking is unnecessary but does no harm (the permissions are set once > >via apply_to_page_range() at boot), so I added this call when adding > >support for strict permissions in EFI rt services mappings. > > > >So I think it is appropriate for create_pgd_mapping() to be in charge > >of calling the ctor(). We simply have no destroy_pgd_mapping() > >counterpart that would be the place for the dtor() call, given that we > >never take down EFI rt services mappi > > >Whether it makes sense or not to lock/unlock in apply_to_page_range() > >is something I did not spend any brain cycles on at the time. > > > > Agreed there shouldn't be a problem right now. I do think the locking is > appropriate in apply_to_page_range given what other functions also get > locked. > > I really wish this were less asymmetrical though since it get hard > to reason about. It looks like hotplug_paging will call the ctor, > so is there an issue with calling hot-remove on memory that was once > hot-added or is that not a concern? > > Thanks, > Laura I think the confusion comes from the fact that, in hotplug_paging, we are passing pgd_pgtable_alloc as the page allocator for __create_pgd_mapping, which always calls the ctor. If I got things right (but, please, correct me if I am wrong), we don't need to get the pte_lock that the ctor gets since - in hotplug - we are adding to init_mm. Moreover, I am just realizing that calling the dtor while hot-removing might create problems when removing memory that *was not* previously hotplugged, as we are calling a dtor on something that was never ctor'ed. Is that what you were hinting at, Laura? Thanks and best regards, Andrea
