How to reclaim inode pages on demand

How to reclaim inode pages on demand

[Mel Gorman]

Hi,

I am working on a direct reclaim strategy to free up large blocks of
contiguous pages. The part I have is working fine, but I am finding a
hundreds of pages that are being used for inodes that I need to reclaim. I
tried purging the inode lists using a variation of prune_icache() but it
is not working out.

Given a struct page, that one knows is an inode, can anyone suggest the
best way to find the inode using it and free it?

-- 
Mel Gorman
Part-time Phd Student                          Java Applications Developer
University of Limerick                         IBM Dublin Software Lab

Re: How to reclaim inode pages on demand

[Jörn Engel]

On Mon, 8 August 2005 16:52:52 +0100, Mel Gorman wrote:
> 
> I am working on a direct reclaim strategy to free up large blocks of
> contiguous pages. The part I have is working fine, but I am finding a
> hundreds of pages that are being used for inodes that I need to reclaim. I
> tried purging the inode lists using a variation of prune_icache() but it
> is not working out.
> 
> Given a struct page, that one knows is an inode, can anyone suggest the
> best way to find the inode using it and free it?

A struct page ain't an inode.  So I'm assuming you mean something like
"giving a struct page that is known to be part of the inode slab
cache".

In that case, you have to decide how intimate you want the slab code
and your page/inode reclaim code to be.  You could use slab structures
to find out the (ofs,len) tupel of contained structures, cast those
to a struct inode and then continue.  Whether you actually want that,
depends on your moral standards.

Jörn

-- 
I don't understand it. Nobody does.
-- Richard P. Feynman

Re: How to reclaim inode pages on demand

[Mel Gorman]

On Mon, 8 Aug 2005, Jörn Engel wrote:

> On Mon, 8 August 2005 16:52:52 +0100, Mel Gorman wrote:
> >
> > I am working on a direct reclaim strategy to free up large blocks of
> > contiguous pages. The part I have is working fine, but I am finding a
> > hundreds of pages that are being used for inodes that I need to reclaim. I
> > tried purging the inode lists using a variation of prune_icache() but it
> > is not working out.
> >
> > Given a struct page, that one knows is an inode, can anyone suggest the
> > best way to find the inode using it and free it?
>
> A struct page ain't an inode.  So I'm assuming you mean something like
> "giving a struct page that is known to be part of the inode slab
> cache".
>

I explained this badly.

Inodes have access to an address_space via inode->i_data. This
address_space is used to allocate pages from functions like
page_cache_alloc() or grab_cache_page(). It is these pages that are the
problem, not the slab pages holding the inode structures. The question is
how these pages can be directly reclaimed. Does anyone have suggestions?


-- 
Mel Gorman
Part-time Phd Student                          Java Applications Developer
University of Limerick                         IBM Dublin Software Lab

Re: How to reclaim inode pages on demand

[Andrew Morton]

Mel Gorman <mel@csn.ul.ie> wrote:
>
> Hi,
> 
> I am working on a direct reclaim strategy to free up large blocks of
> contiguous pages. The part I have is working fine, but I am finding a
> hundreds of pages that are being used for inodes that I need to reclaim. I
> tried purging the inode lists using a variation of prune_icache() but it
> is not working out.
> 
> Given a struct page, that one knows is an inode, can anyone suggest the
> best way to find the inode using it and free it?

Simple answer: invalidate_mapping_pages(page->mapping, start, end).

Problem: races.  If you pick a random page up off the LRU and start playing
with its mapping, what stops that mapping from getting freed under your
feet?

If the caller has a ref on the inode (say, you came in via a syscall
against an fd) then fine.

But if you picked the page up off the LRU then to avoid super-rare oopses
you'll need to pin the address_space somehow.  One way is to

a) lock the page, which pins the address_space.

b) check that page->mapping is still non-NULL.  Bale if it is.

c) bump page->mapping->host->i_count

d) unlock the page

e) diddle away with the address_space.

f) iput() the inode.

But beware that iput() can do a *ton* of stuff, so you need to be not
holding any locks and to not be called from any fancy contexts.

Re: How to reclaim inode pages on demand

[Mel Gorman]

On Mon, 8 Aug 2005, Andrew Morton wrote:

> Mel Gorman <mel@csn.ul.ie> wrote:
> >
> > Hi,
> >
> > I am working on a direct reclaim strategy to free up large blocks of
> > contiguous pages. The part I have is working fine, but I am finding a
> > hundreds of pages that are being used for inodes that I need to reclaim. I
> > tried purging the inode lists using a variation of prune_icache() but it
> > is not working out.
> >
> > Given a struct page, that one knows is an inode, can anyone suggest the
> > best way to find the inode using it and free it?
>
> Simple answer: invalidate_mapping_pages(page->mapping, start, end).
>

The majority of pages I am seeing no longer have page->mapping set. Does
this mean they are in the process of being cleared up?

> Problem: races.  If you pick a random page up off the LRU and start playing
> with its mapping, what stops that mapping from getting freed under your
> feet?
>

Nothing at all, which explains why I occasionally got oops when I was
trying to free pages with the mapping set. I am going to follow your
suggestions and see how I get on.

If it is still not working out, I will try reclaiming when the workload is
something other than a large number of kernel compiles to see if I am on
the right track at all.

Thanks a lot

Re: How to reclaim inode pages on demand

[Andrew Morton]

Mel Gorman <mel@csn.ul.ie> wrote:
>
> On Mon, 8 Aug 2005, Andrew Morton wrote:
> 
> > Mel Gorman <mel@csn.ul.ie> wrote:
> > >
> > > Hi,
> > >
> > > I am working on a direct reclaim strategy to free up large blocks of
> > > contiguous pages. The part I have is working fine, but I am finding a
> > > hundreds of pages that are being used for inodes that I need to reclaim. I
> > > tried purging the inode lists using a variation of prune_icache() but it
> > > is not working out.
> > >
> > > Given a struct page, that one knows is an inode, can anyone suggest the
> > > best way to find the inode using it and free it?
> >
> > Simple answer: invalidate_mapping_pages(page->mapping, start, end).
> >
> 
> The majority of pages I am seeing no longer have page->mapping set. Does
> this mean they are in the process of being cleared up?

They're just anonymous pages, aren't they?  But you said "pages that are
being used for inodes".  Confused.

Re: How to reclaim inode pages on demand

[Mel Gorman]

On Wed, 10 Aug 2005, Andrew Morton wrote:

> Mel Gorman <mel@csn.ul.ie> wrote:
> >
> > On Mon, 8 Aug 2005, Andrew Morton wrote:
> >
> > > Mel Gorman <mel@csn.ul.ie> wrote:
> > > >
> > > > Hi,
> > > >
> > > > I am working on a direct reclaim strategy to free up large blocks of
> > > > contiguous pages. The part I have is working fine, but I am finding a
> > > > hundreds of pages that are being used for inodes that I need to reclaim. I
> > > > tried purging the inode lists using a variation of prune_icache() but it
> > > > is not working out.
> > > >
> > > > Given a struct page, that one knows is an inode, can anyone suggest the
> > > > best way to find the inode using it and free it?
> > >
> > > Simple answer: invalidate_mapping_pages(page->mapping, start, end).
> > >
> >
> > The majority of pages I am seeing no longer have page->mapping set. Does
> > this mean they are in the process of being cleared up?
>
> They're just anonymous pages, aren't they?  But you said "pages that are
> being used for inodes".  Confused.
>

So am I, I'm missing something really stupid.

What I have is the following;

1. Add a new flag GFP_INODE to mark inode pages
2. Add a GFP_INODE to the flags passed to mapping_set_gfp_mask() in
   fs/inode.c#alloc_inode(). This means that the page allocator will now
   know when it is allocating pages for inodes
3. Added a PG_inode flag for page->flags which will flag all pages that
   were allocated for inodes

(Note, I don't intend to use this flags in the long term, I've added them
for investigation purposes).

I later linearly scan the mem_map looking for pages that can be freed up
(usually LRU pages). I was expecting any page with PG_inode set to have a
page->mapping but not all of them do. It is the pages without a ->mapping
that are confusing the hell out of me.

-- 
Mel Gorman
Part-time Phd Student                          Java Applications Developer
University of Limerick                         IBM Dublin Software Lab

Re: How to reclaim inode pages on demand

[Andrew Morton]

Mel Gorman <mel@csn.ul.ie> wrote:
>
> On Wed, 10 Aug 2005, Andrew Morton wrote:
> 
> > Mel Gorman <mel@csn.ul.ie> wrote:
> > >
> > > On Mon, 8 Aug 2005, Andrew Morton wrote:
> > >
> > > > Mel Gorman <mel@csn.ul.ie> wrote:
> > > > >
> > > > > Hi,
> > > > >
> > > > > I am working on a direct reclaim strategy to free up large blocks of
> > > > > contiguous pages. The part I have is working fine, but I am finding a
> > > > > hundreds of pages that are being used for inodes that I need to reclaim. I
> > > > > tried purging the inode lists using a variation of prune_icache() but it
> > > > > is not working out.
> > > > >
> > > > > Given a struct page, that one knows is an inode, can anyone suggest the
> > > > > best way to find the inode using it and free it?
> > > >
> > > > Simple answer: invalidate_mapping_pages(page->mapping, start, end).
> > > >
> > >
> > > The majority of pages I am seeing no longer have page->mapping set. Does
> > > this mean they are in the process of being cleared up?
> >
> > They're just anonymous pages, aren't they?  But you said "pages that are
> > being used for inodes".  Confused.
> >
> 
> So am I, I'm missing something really stupid.
> 
> What I have is the following;
> 
> 1. Add a new flag GFP_INODE to mark inode pages
> 2. Add a GFP_INODE to the flags passed to mapping_set_gfp_mask() in
>    fs/inode.c#alloc_inode(). This means that the page allocator will now
>    know when it is allocating pages for inodes
> 3. Added a PG_inode flag for page->flags which will flag all pages that
>    were allocated for inodes
> 
> (Note, I don't intend to use this flags in the long term, I've added them
> for investigation purposes).
> 
> I later linearly scan the mem_map looking for pages that can be freed up
> (usually LRU pages). I was expecting any page with PG_inode set to have a
> page->mapping but not all of them do. It is the pages without a ->mapping
> that are confusing the hell out of me.

Well there are conditions in which mmapped file pages can get converted to
anonymous pages due to truncate(), but I have a feeling that we stopped
that from happening.

Also there are situations in which truncate of a still-committing ext3
pagecache page can cause the page to remain ont he page LRUs - it's been
truncated from the file, but ext4 still has a hold of it for journalling
purposes.

You cold lock the pages then check ->mapping.

Re: How to reclaim inode pages on demand

[Mel Gorman]

On Wed, 10 Aug 2005, Andrew Morton wrote:

> Mel Gorman <mel@csn.ul.ie> wrote:
> >
> > On Wed, 10 Aug 2005, Andrew Morton wrote:
> >
> > > Mel Gorman <mel@csn.ul.ie> wrote:
> > > >
> > > > On Mon, 8 Aug 2005, Andrew Morton wrote:
> > > >
> > > > > Mel Gorman <mel@csn.ul.ie> wrote:
> > > > > >
> > > > > > Hi,
> > > > > >
> > > > > > I am working on a direct reclaim strategy to free up large blocks of
> > > > > > contiguous pages. The part I have is working fine, but I am finding a
> > > > > > hundreds of pages that are being used for inodes that I need to reclaim. I
> > > > > > tried purging the inode lists using a variation of prune_icache() but it
> > > > > > is not working out.
> > > > > >
> > > > > > Given a struct page, that one knows is an inode, can anyone suggest the
> > > > > > best way to find the inode using it and free it?
> > > > >
> > > > > Simple answer: invalidate_mapping_pages(page->mapping, start, end).
> > > > >
> > > >
> > > > The majority of pages I am seeing no longer have page->mapping set. Does
> > > > this mean they are in the process of being cleared up?
> > >
> > > They're just anonymous pages, aren't they?  But you said "pages that are
> > > being used for inodes".  Confused.
> > >
> >
> > So am I, I'm missing something really stupid.
> >
> > What I have is the following;
> >
> > 1. Add a new flag GFP_INODE to mark inode pages
> > 2. Add a GFP_INODE to the flags passed to mapping_set_gfp_mask() in
> >    fs/inode.c#alloc_inode(). This means that the page allocator will now
> >    know when it is allocating pages for inodes
> > 3. Added a PG_inode flag for page->flags which will flag all pages that
> >    were allocated for inodes
> >
> > (Note, I don't intend to use this flags in the long term, I've added them
> > for investigation purposes).
> >
> > I later linearly scan the mem_map looking for pages that can be freed up
> > (usually LRU pages). I was expecting any page with PG_inode set to have a
> > page->mapping but not all of them do. It is the pages without a ->mapping
> > that are confusing the hell out of me.
>
> Well there are conditions in which mmapped file pages can get converted to
> anonymous pages due to truncate(), but I have a feeling that we stopped
> that from happening.
>

Does that also apply to when a file is unlinked rather than truncated? As
the load is a lot of kernel compiles, there are a lot of files being
created and deleted shortly afterwards.

> Also there are situations in which truncate of a still-committing ext3
> pagecache page can cause the page to remain ont he page LRUs - it's been
> truncated from the file, but ext4 still has a hold of it for journalling
> purposes.
>

In case it is journalling-related, I am going to rerun the tests on an
ext2 filesystem over the weekend.

> You cold lock the pages then check ->mapping.
>

Will do.

-- 
Mel Gorman
Part-time Phd Student                          Java Applications Developer
University of Limerick                         IBM Dublin Software Lab

Re: How to reclaim inode pages on demand

[Andrew Morton]

Mel Gorman <mel@csn.ul.ie> wrote:
>
> > Well there are conditions in which mmapped file pages can get converted to
> > anonymous pages due to truncate(), but I have a feeling that we stopped
> > that from happening.
> >
> 
> Does that also apply to when a file is unlinked rather than truncated?

Yup.  unlink() does truncate if it unlinked the final link.

> In case it is journalling-related, I am going to rerun the tests on an
> ext2 filesystem over the weekend.

Good idea.

Re: How to reclaim inode pages on demand

[Dave Hansen]

On Wed, 2005-08-10 at 18:27 +0100, Mel Gorman wrote:
> I later linearly scan the mem_map looking for pages that can be freed up
> (usually LRU pages). I was expecting any page with PG_inode set to have a
> page->mapping but not all of them do. It is the pages without a ->mapping
> that are confusing the hell out of me.

How about putting a check for PG_inode and a periodic dump_stack()
wherever page->mapping is cleared?  That should at least let you catch
the culprit who is clearing them. __remove_from_page_cache() is the only
real place I see this being done.

Are you remembering to clean PG_inode when a page is freed?  Perhaps
it's allocated for the page cache, reclaimed, returned to the allocator,
and reallocated as anonymous memory where you're seeing the null
->mapping.  Might want to check for PG_inode in free_pages_check().

-- Dave

Re: How to reclaim inode pages on demand

[Mel Gorman]

On Wed, 10 Aug 2005, Dave Hansen wrote:

> On Wed, 2005-08-10 at 18:27 +0100, Mel Gorman wrote:
> > I later linearly scan the mem_map looking for pages that can be freed up
> > (usually LRU pages). I was expecting any page with PG_inode set to have a
> > page->mapping but not all of them do. It is the pages without a ->mapping
> > that are confusing the hell out of me.
>
> How about putting a check for PG_inode and a periodic dump_stack()
> wherever page->mapping is cleared?  That should at least let you catch
> the culprit who is clearing them. __remove_from_page_cache() is the only
> real place I see this being done.
>

Will do, it might shed some light on whats happening.

> Are you remembering to clean PG_inode when a page is freed?

Yes, I was using the allocator to count how many inode-related pages there
were in the system. The answer was lots, another reason why I'm thinking
that I picked the worst possible benchmark to test this stuff with.

> Perhaps
> it's allocated for the page cache, reclaimed, returned to the allocator,
> and reallocated as anonymous memory where you're seeing the null
> ->mapping.  Might want to check for PG_inode in free_pages_check().
>

I am 99.99% certain I make that check but I can't double check the code
until late in the weekend.

-- 
Mel Gorman
Part-time Phd Student                          Java Applications Developer
University of Limerick                         IBM Dublin Software Lab

Re: How to reclaim inode pages on demand

[Magnus Damm]

On 8/11/05, Andrew Morton <akpm@osdl.org> wrote:
> Mel Gorman <mel@csn.ul.ie> wrote:
> > The majority of pages I am seeing no longer have page->mapping set. Does
> > this mean they are in the process of being cleared up?
> 
> They're just anonymous pages, aren't they?  But you said "pages that are
> being used for inodes".  Confused.

I thought page->mapping was used by rmap for both inode-backed and
anonymous pages. And the PAGE_MAPPING_ANON bit was used to determine
that page->mapping points to struct anon_vma instead of struct
address_space.

When is page->mapping NULL?

/ magnus

Re: How to reclaim inode pages on demand

[Andrew Morton]

Magnus Damm <magnus.damm@gmail.com> wrote:
>
> On 8/11/05, Andrew Morton <akpm@osdl.org> wrote:
> > Mel Gorman <mel@csn.ul.ie> wrote:
> > > The majority of pages I am seeing no longer have page->mapping set. Does
> > > this mean they are in the process of being cleared up?
> > 
> > They're just anonymous pages, aren't they?  But you said "pages that are
> > being used for inodes".  Confused.
> 
> I thought page->mapping was used by rmap for both inode-backed and
> anonymous pages. And the PAGE_MAPPING_ANON bit was used to determine
> that page->mapping points to struct anon_vma instead of struct
> address_space.
> 
> When is page->mapping NULL?
> 

mm.h:page_mapping() handles all that.

Re: How to reclaim inode pages on demand

[Magnus Damm]

On 8/11/05, Andrew Morton <akpm@osdl.org> wrote:
> Magnus Damm <magnus.damm@gmail.com> wrote:
> >
> > On 8/11/05, Andrew Morton <akpm@osdl.org> wrote:
> > > Mel Gorman <mel@csn.ul.ie> wrote:
> > > > The majority of pages I am seeing no longer have page->mapping set. Does
> > > > this mean they are in the process of being cleared up?
> > >
> > > They're just anonymous pages, aren't they?  But you said "pages that are
> > > being used for inodes".  Confused.
> >
> > I thought page->mapping was used by rmap for both inode-backed and
> > anonymous pages. And the PAGE_MAPPING_ANON bit was used to determine
> > that page->mapping points to struct anon_vma instead of struct
> > address_space.
> >
> > When is page->mapping NULL?
> >
> 
> mm.h:page_mapping() handles all that.

Yes, thanks, but I believe that page->mapping is never NULL for
anonymous pages. Does not all anonymous pages have page->mapping
pointing to struct anon_vma?

/ magnus

Re: How to reclaim inode pages on demand

[Coywolf Qi Hunt]

On 8/11/05, Andrew Morton <akpm@osdl.org> wrote:
> Magnus Damm <magnus.damm@gmail.com> wrote:
> >
> > On 8/11/05, Andrew Morton <akpm@osdl.org> wrote:
> > > Mel Gorman <mel@csn.ul.ie> wrote:
> > > > The majority of pages I am seeing no longer have page->mapping set. Does
> > > > this mean they are in the process of being cleared up?
> > >
> > > They're just anonymous pages, aren't they?  But you said "pages that are
> > > being used for inodes".  Confused.
> >
> > I thought page->mapping was used by rmap for both inode-backed and
> > anonymous pages. And the PAGE_MAPPING_ANON bit was used to determine
> > that page->mapping points to struct anon_vma instead of struct
> > address_space.
> >
> > When is page->mapping NULL?
> >
> 
> mm.h:page_mapping() handles all that.

at http://sosdg.org/~coywolf/lxr/source/include/linux/mm.h#L561
Should the comment be s/page_mapped/page->mapping/ ?

-- 
Coywolf Qi Hunt
http://ahbl.org/~coywolf/

Re: How to reclaim inode pages on demand

[Hugh Dickins]

On Thu, 11 Aug 2005, Coywolf Qi Hunt wrote:
> 
> at http://sosdg.org/~coywolf/lxr/source/include/linux/mm.h#L561
> Should the comment be s/page_mapped/page->mapping/ ?

No.

Re: How to reclaim inode pages on demand

[Rik van Riel]

On Mon, 8 Aug 2005, Mel Gorman wrote:

> Given a struct page, that one knows is an inode, can anyone suggest
> the best way to find the inode using it and free it?

Note that you can only free the inodes that aren't currently
open files for any of the processes in the system.

-- 
All Rights Reversed