Messages in this thread | | | Subject | Re: [PATCH v5 4/4] mm: Defer ZONE_DEVICE page initialization to the point where we init pgmap | From | Alexander Duyck <> | Date | Wed, 10 Oct 2018 08:27:58 -0700 |
| |
On 10/10/2018 5:52 AM, Yi Zhang wrote: > On 2018-10-09 at 14:19:32 -0700, Dan Williams wrote: >> On Tue, Oct 9, 2018 at 1:34 PM Alexander Duyck >> <alexander.h.duyck@linux.intel.com> wrote: >>> >>> On 10/9/2018 11:04 AM, Dan Williams wrote: >>>> On Tue, Oct 9, 2018 at 3:21 AM Yi Zhang <yi.z.zhang@linux.intel.com> wrote: >> [..] >>>> That comment is incorrect, device-pages are never onlined. So I think >>>> we can just skip that call to __SetPageReserved() unless the memory >>>> range is MEMORY_DEVICE_{PRIVATE,PUBLIC}. >>>> >>> >>> When pages are "onlined" via __free_pages_boot_core they clear the >>> reserved bit, that is the reason for the comment. The reserved bit is >>> meant to indicate that the page cannot be swapped out or moved based on >>> the description of the bit. >> >> ...but ZONE_DEVICE pages are never onlined so I would expect >> memmap_init_zone_device() to know that detail. >> >>> I would think with that being the case we still probably need the call >>> to __SetPageReserved to set the bit with the expectation that it will >>> not be cleared for device-pages since the pages are not onlined. >>> Removing the call to __SetPageReserved would probably introduce a number >>> of regressions as there are multiple spots that use the reserved bit to >>> determine if a page can be swapped out to disk, mapped as system memory, >>> or migrated. > > Another things, it seems page_init/set_reserved already been done in the > move_pfn_range_to_zone > |-->memmap_init_zone > |-->for_each_page_in_pfn > |-->__init_single_page > |-->SetPageReserved > > Why we haven't remove these redundant initial in memmap_init_zone? > > Correct me if I missed something.
In this case it isn't redundant as only the vmmemmap pages are initialized in memmap_init_zone now. So all of the pages that are going to be used as device pages are not initialized until the call to memmap_init_zone_device. What I did is split the initialization of the pages into two parts in order to allow us to initialize the pages outside of the hotplug lock.
>> >> Right, this is what Yi is working on... the PageReserved flag is >> problematic for KVM. Auditing those locations it seems as long as we >> teach hibernation to avoid ZONE_DEVICE ranges we can safely not set >> the reserved flag for DAX pages. What I'm trying to avoid is a local >> KVM hack to check for DAX pages when the Reserved flag is not >> otherwise needed. > Thanks Dan. Provide the patch link. > > https://lore.kernel.org/lkml/cover.1536342881.git.yi.z.zhang@linux.intel.com
So it looks like your current logic is just working around the bit then since it just allows for reserved DAX pages.
| |