Re: [PATCH v2] mm, zone_device: replace {get, put}_zone_device_page() with a single reference

[Dan Williams <dan.j.williams@intel.com>]

On Fri, Apr 28, 2017 at 12:16 PM, Jerome Glisse <jglisse@redhat.com> wrote:
>> On Fri, Apr 28, 2017 at 11:00 AM, Jerome Glisse <jglisse@redhat.com> wrote:
>> >> On Fri, Apr 28, 2017 at 10:34 AM, Jerome Glisse <jglisse@redhat.com>
>> >> wrote:
>> >> >> Kirill points out that the calls to {get,put}_dev_pagemap() can be
>> >> >> removed from the mm fast path if we take a single get_dev_pagemap()
>> >> >> reference to signify that the page is alive and use the final put of
>> >> >> the
>> >> >> page to drop that reference.
>> >> >>
>> >> >> This does require some care to make sure that any waits for the
>> >> >> percpu_ref to drop to zero occur *after* devm_memremap_page_release(),
>> >> >> since it now maintains its own elevated reference.
>> >> >
>> >> > This is NAK from HMM point of view as i need those call. So if you
>> >> > remove
>> >> > them now i will need to add them back as part of HMM.
>> >>
>> >> I thought you only need them at page free time? You can still hook
>> >> __put_page().
>> >
>> > No, i need a hook when page refcount reach 1, not 0. That being said
>> > i don't care about put_dev_pagemap(page->pgmap); so that part of the
>> > patch is fine from HMM point of view but i definitly need to hook my-
>> > self in the general put_page() function.
>> >
>> > So i will have to undo part of this patch for HMM (put_page() will
>> > need to handle ZONE_DEVICE page differently).
>>
>> Ok, I'd rather this go in now since it fixes the existing use case,
>> and unblocks the get_user_pages_fast() conversion to generic code.
>> That also gives Kirill and -mm folks a chance to review what HMM wants
>> to do on top of the page_ref infrastructure.  The
>> {get,put}_zone_device_page interface went in in 4.5 right before
>> page_ref went in during 4.6, so it was just an oversight that
>> {get,put}_zone_device_page were not removed earlier.
>>
>
> I don't mind this going in, i am hopping people won't ignore HMM patchset
> once i repost after 4.12 merge window. Note that there is absolutely no way
> around me hooking up inside put_page(). The only other way to do it would
> be to modify virtualy all places that call that function to handle ZONE_DEVICE
> case.

Are you sure about needing to hook the 2 -> 1 transition? Could we
change ZONE_DEVICE pages to not have an elevated reference count when
they are created so you can keep the HMM references out of the mm hot
path?