[LSF/MM TOPIC] Page flags, can we free up space ?

[LSF/MM TOPIC] Page flags, can we free up space ?

[Jerome Glisse]

So lattely i have been looking at page flags and we are using 6 flags
for memory reclaim and compaction:

    PG_referenced
    PG_lru
    PG_active
    PG_workingset
    PG_reclaim
    PG_unevictable

On top of which you can add the page anonymous flag (anonymous or
share memory)
    PG_anon // does not exist, lower bit of page->mapping

And also the movable flag (which alias with KSM)
    PG_movable // does not exist, lower bit of page->mapping


So i would like to explore if there is a way to express the same amount
of information with less bits. My methodology is to exhaustively list
all the possible states (valid combination of above flags) and then to
see how we change from one state to another (what event trigger the change
like mlock(), page being referenced, ...) and under which rules (ie do we
hold the page lock, zone lock, ...).

My hope is that there might be someway to use less bits to express the
same thing. I am doing this because for my work on generic page write
protection (ie KSM for file back page) which i talk about last year and
want to talk about again ;) I will need to unalias the movable bit from
KSM bit.


Right now this is more a temptative ie i do not know if i will succeed,
in any case i can report on failure or success and discuss my finding to
get people opinions on the matter.


I think everyone interested in mm will be interested in this topic :)

Cheers,
Jérôme

Re: [LSF/MM TOPIC] Page flags, can we free up space ?

[Andi Kleen]

Jerome Glisse <jglisse@redhat.com> writes:
>
> Right now this is more a temptative ie i do not know if i will succeed,
> in any case i can report on failure or success and discuss my finding to
> get people opinions on the matter.

I would just stop putting node/zone number into the flags. These
could be all handled with a small perfect hash table, like the original
x86_64 port did, which should be quite cheap to look up.
Then there should be enough bits for everyone again.

-Andi

Re: [LSF/MM TOPIC] Page flags, can we free up space ?

[Jerome Glisse]

On Tue, Jan 22, 2019 at 01:44:15PM -0800, Andi Kleen wrote:
> Jerome Glisse <jglisse@redhat.com> writes:
> >
> > Right now this is more a temptative ie i do not know if i will succeed,
> > in any case i can report on failure or success and discuss my finding to
> > get people opinions on the matter.
> 
> I would just stop putting node/zone number into the flags. These
> could be all handled with a small perfect hash table, like the original
> x86_64 port did, which should be quite cheap to look up.
> Then there should be enough bits for everyone again.

Definitly something i will look into, i was scare to remove those.

Cheers,
Jérôme

Re: [LSF/MM TOPIC] Page flags, can we free up space ?

[Mike Kravetz]

On 1/22/19 12:17 PM, Jerome Glisse wrote:
> So lattely i have been looking at page flags and we are using 6 flags
> for memory reclaim and compaction:
> 
>     PG_referenced
>     PG_lru
>     PG_active
>     PG_workingset
>     PG_reclaim
>     PG_unevictable
> 
> On top of which you can add the page anonymous flag (anonymous or
> share memory)
>     PG_anon // does not exist, lower bit of page->mapping
> 
> And also the movable flag (which alias with KSM)
>     PG_movable // does not exist, lower bit of page->mapping
> 
> 
> So i would like to explore if there is a way to express the same amount
> of information with less bits. My methodology is to exhaustively list
> all the possible states (valid combination of above flags) and then to
> see how we change from one state to another (what event trigger the change
> like mlock(), page being referenced, ...) and under which rules (ie do we
> hold the page lock, zone lock, ...).
> 
> My hope is that there might be someway to use less bits to express the
> same thing. I am doing this because for my work on generic page write
> protection (ie KSM for file back page) which i talk about last year and
> want to talk about again ;) I will need to unalias the movable bit from
> KSM bit.
> 
> 
> Right now this is more a temptative ie i do not know if i will succeed,
> in any case i can report on failure or success and discuss my finding to
> get people opinions on the matter.
> 
> 
> I think everyone interested in mm will be interested in this topic :)

Explicitly adding Matthew on Cc as I am pretty sure he has been working
in this area.

-- 
Mike Kravetz

Re: [LSF/MM TOPIC] Page flags, can we free up space ?

[David Hildenbrand]

On 22.01.19 21:17, Jerome Glisse wrote:
> So lattely i have been looking at page flags and we are using 6 flags
> for memory reclaim and compaction:
> 
>     PG_referenced
>     PG_lru
>     PG_active
>     PG_workingset
>     PG_reclaim
>     PG_unevictable
> 
> On top of which you can add the page anonymous flag (anonymous or
> share memory)
>     PG_anon // does not exist, lower bit of page->mapping
> 
> And also the movable flag (which alias with KSM)
>     PG_movable // does not exist, lower bit of page->mapping


I would really like to see an easier way to spot if a page is movable.

__PageMovable() can produce way to many false positives.

movable will usually not be paired with other flags you mentioned as of now.

If many of these flags are not used in combination, we could merge some
of the flags into a number field. Valid combinations would get a number
assigned.

To keep it simple, only flags that are completely exclusive might be a
candidate. But not sure if we really have many of these.

> 
> 
> So i would like to explore if there is a way to express the same amount
> of information with less bits. My methodology is to exhaustively list
> all the possible states (valid combination of above flags) and then to
> see how we change from one state to another (what event trigger the change
> like mlock(), page being referenced, ...) and under which rules (ie do we
> hold the page lock, zone lock, ...).
> 
> My hope is that there might be someway to use less bits to express the
> same thing. I am doing this because for my work on generic page write
> protection (ie KSM for file back page) which i talk about last year and
> want to talk about again ;) I will need to unalias the movable bit from
> KSM bit.
> 
> 
> Right now this is more a temptative ie i do not know if i will succeed,
> in any case i can report on failure or success and discuss my finding to
> get people opinions on the matter.
> 
> 
> I think everyone interested in mm will be interested in this topic :)
> 
> Cheers,
> Jérôme
> 


-- 

Thanks,

David / dhildenb