Re: [PATCH 2/3] mm: Introduce subsection_dev_map

[David Hildenbrand]

> Am 13.11.2019 um 20:06 schrieb Dan Williams <dan.j.williams@xxxxxxxxx>:
> 
> ïOn Wed, Nov 13, 2019 at 10:51 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>> 
>>> On 08.11.19 20:13, Dan Williams wrote:
>>> On Thu, Nov 7, 2019 at 4:15 PM Toshiki Fukasawa
>>> <t-fukasawa@xxxxxxxxxxxxx> wrote:
>>>> 
>>>> Currently, there is no way to identify pfn on ZONE_DEVICE.
>>>> Identifying pfn on system memory can be done by using a
>>>> section-level flag. On the other hand, identifying pfn on
>>>> ZONE_DEVICE requires a subsection-level flag since ZONE_DEVICE
>>>> can be created in units of subsections.
>>>> 
>>>> This patch introduces a new bitmap subsection_dev_map so that
>>>> we can identify pfn on ZONE_DEVICE.
>>>> 
>>>> Also, subsection_dev_map is used to prove that struct pages
>>>> included in the subsection have been initialized since it is
>>>> set after memmap_init_zone_device(). We can avoid accessing
>>>> pages currently being initialized by checking subsection_dev_map.
>>>> 
>>>> Signed-off-by: Toshiki Fukasawa <t-fukasawa@xxxxxxxxxxxxx>
>>>> ---
>>>>  include/linux/mmzone.h | 19 +++++++++++++++++++
>>>>  mm/memremap.c          |  2 ++
>>>>  mm/sparse.c            | 32 ++++++++++++++++++++++++++++++++
>>>>  3 files changed, 53 insertions(+)
>>>> 
>>>> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
>>>> index bda2028..11376c4 100644
>>>> --- a/include/linux/mmzone.h
>>>> +++ b/include/linux/mmzone.h
>>>> @@ -1174,11 +1174,17 @@ static inline unsigned long section_nr_to_pfn(unsigned long sec)
>>>> 
>>>>  struct mem_section_usage {
>>>>         DECLARE_BITMAP(subsection_map, SUBSECTIONS_PER_SECTION);
>>>> +#ifdef CONFIG_ZONE_DEVICE
>>>> +       DECLARE_BITMAP(subsection_dev_map, SUBSECTIONS_PER_SECTION);
>>>> +#endif
>>> 
>>> Hi Toshiki,
>>> 
>>> There is currently an effort to remove the PageReserved() flag as some
>>> code is using that to detect ZONE_DEVICE. In reviewing those patches
>>> we realized that what many code paths want is to detect online memory.
>>> So instead of a subsection_dev_map add a subsection_online_map. That
>>> way pfn_to_online_page() can reliably avoid ZONE_DEVICE ranges. I
>>> otherwise question the use case for pfn_walkers to return pages for
>>> ZONE_DEVICE pages, I think the skip behavior when pfn_to_online_page()
>>> == false is the right behavior.
>> 
>> To be more precise, I recommended an subsection_active_map, to indicate
>> which memmaps were fully initialized and can safely be touched (e.g., to
>> read the zone/nid). This map would also be set when the devmem memmaps
>> were initialized (race between adding memory/growing the section and
>> initializing the memmap).
>> 
>> See
>> 
>> https://lkml.org/lkml/2019/10/10/87
>> 
>> and
>> 
>> https://www.spinics.net/lists/linux-driver-devel/msg130012.html
> 
> I'm still struggling to understand the motivation of distinguishing
> "active" as something distinct from "online". As long as the "online"
> granularity is improved from sections down to subsections then most
> code paths are good to go. The others can use get_devpagemap() to
> check for ZONE_DEVICE in a race free manner as they currently do.

I thought we wanted to unify access if we donât really care about the zone as in most pfn walkers - E.g., for zone shrinking. Anyhow, a subsection online map would be a good start, we can reuse that later for ZONE_DEVICE as well.

> 
>> I dislike a map that is specific to ZONE_DEVICE or (currently)
>> !ZONE_DEVICE. I rather want an indication "this memmap is safe to
>> touch". As discussed along the mentioned threads, we can combine this
>> later with RCU to handle some races that are currently possible.
> 
> The rcu protection is independent of the pfn_active vs pfn_online
> distinction afaics.

Itâs one part of the bigger picture IMHO.

>