Re: Page Replacement in 2.6.10

[Rahul Iyer <rni@xxxxxxxxxxxxxx>]

Hi Rik,
Did as you suggested... Learnt a lot of VM and rmap in the process :) 
Also found this on rmap:
http://www.finux.org/Reprints/Reprint-McCracken-OLS2004.pdf
I must say, it is well written and explains current rmap very well

But i'm still a bit confused...
Here's what i understand
first eviction....
1. shrink_list does the final writing out of the pages
2. If it as an anonymous page, it is added to the swap cache if not 
already there
3. for each page, it calls try_to_unmap()
4. try_to_unmap() calls try_to_unmap_anon() or try_to_unmap_file() based 
on type of page
5. try_to_unmap_anon() calls try_to_unmap_one() for each page
6.try_to_unmap_one() stores the swp_entry_t  in the pte for anonymous pages
7. try_to_unmap_file() calls try_to_unmap_one() for each linear mapping 
and try_to_unmap_cluster() for each non linear one
8. in the former case, it does nothing to the pte because it is not an 
anonymous page, but in the latter, it stores the index within the mapped 
file in the pte

now, retrieval, ignoring COW details
1. On a page fault, the address id fetched from cr2
2. the vma corresponding to this address is found
3. if it an anonymous page
       if the page was never used before, the no_page function of the 
vma_ops is called and it ends here.
       if it was swapped out, the pte has the swp_entry_t and so we can 
fetch it from the swap space
4. if it is file backed
       the struct file associated with this vma is got.
       The address_space corresponding to the struct file is got.
        if the page was not accessed before
               the no_page function of vma_ops is called
       
       if it is a linear vma
               the struct file, the offset within the file is faulted 
address - start address of vma

      if it is a nonlinear vma
               the offset is in the pte

      the populate function of the vma_ops is used to read the data in
       
is this right?

If so, then i'm confused...
ok, i can store the swp_entry_t as an id for the anonymous pages... but 
what about file backed ones? there doesn't seem to be any id as there 
needs to be none. Am i right here?
A lot of reading... but this one still evades me.
Thanks
Rahul



Rik van Riel wrote:

> On Fri, 25 Mar 2005, Rahul Iyer wrote:
>
>  
>
> > I agree... pid is not a good idea. So i thought of something else... the
> > swp_entry_t. What i forgot was that this is only for anonymous pages... disk
> > buffers and mmapped pages are not put on swap (right?). So, I'm looking at
> > something else now... primarily the page->private field.
> >    
>
> You cannot use such an identifier, since the page will be 
> GONE after it is swapped out, meaning the identifier (which
> is your lookup key) will be gone too!
>
>  
>
> > I guess, what i need is an identifier to tag every page that leaves the 
> > cache... whether anonymous, mmapped or disk buffer. Unfortunately I 
> > cannot find such an id. Actually, this is all that holds me up... the 
> > rest of what i need to do is in place - mentally that is. :) So, I could 
> > really use a tip ;)
> >    
>
> OK, here's a tip:  how do the page cache code and the
> page fault code decide which page to bring into memory ?
>
>  

--
Kernelnewbies: Help each other learn about the Linux kernel.
Archive:       http://mail.nl.linux.org/kernelnewbies/
FAQ:           http://kernelnewbies.org/faq/