On Fri, 2002-12-06 at 04:45, Martin Maletinsky wrote:

Hello Joe,

Thank you for your reply. I have an additional question.

Joseph A Knapka wrote:

quoted

Martin Maletinsky wrote:

quoted

Hello,

I am looking at the swapping mechanism in Linux. I have read the

relevant chapter 16 in 'Understanding the Linux Kernel' from
Bovet&Cesati, and looked at the 2.2.18 kernel

quoted

quoted

source code. I still have the follwing question:

Function try_to_swap_out() [p. 481 in 'Understanding the Linux

Kernel']:

quoted

quoted

If the page in question already belongs to the swap cache, the

function performs no data transfer to the swap space on the disk (but
only marks the page as swapped out).

quoted

quoted

The corresponding comment in the try_to_swap_out() functions states

'Is the page already in the swap cache? If so, ..... - it is already
up-to-date on disk.

quoted

quoted

Understanding the Linux Kernel states on p. 482 'If the page belongs

to the swap cache .... no memory transfer is performed'.

quoted

quoted

Now my question is, couldn't the page have been modified since it

was added to the swap cache (and written to disk), and thus differ from
the data in the swap space? In

quoted

quoted

this case shouldn't the page be written to disk (again)?

If the page is in the swap cache, it's *effectively* up to date on

disk,

quoted

because the swap cache page is *the* authoritative image of the page.
If it's dirty it will get written out by page_launder() in short
order, because whomever dirtied it set the page_dirty bit in the
page struct. That issue is unimportant to the process doing the
swap_out, though - all it cares about is that the page is going
to be taken care of by the cache machinery.

Assume a page P that is marked as clean (i.e. PG_dirty bit not set), and
is in the page cache. Additionaly assume that P is mapped by a process
A. Now let A perform a store
operation into the page P, which will mark A's page table entry mapping
P as dirty (i.e. set the dirty bit).
Subsequently assume that try_to_swap_out() is called on A's page table
entry that maps P. try_to_swap_out() will see that P is in the swap
cache already,

quoted

quoted

quoted

For the first time P would never be found in the swap cache, infact

try_to_swap_out shall do following
a] Page is dirty (in page table entry), so set PG_DIRTY in struct page
b] Allocate a swap entry and add this page to swap cache
c] release the page, and add the modify page table entry to point it to
swap entry

Now We have 
a] Page table entry for P contains swap info
b] Page P in swap cache
c] PG_DIRTY _is_ set (infact for a page in swap cache this is always
true)

Do remember, along with the swap cache P may be party of inactive_dirty
list.

The actual swapping to backing store is done by page scanner.
It shall do following. Assume it has decided to _really_ free P
1] As page is dirty, call the page write back function. Thus here for
the first time page found its place in swap.
2] send P back home, to buddy allocator

If process A again access the page, then page fault handler shall do
following
1] allocate a swap cache page
2] read the page from swap.
3] Modify page table entry of A to point to this page 

Just give a little thought about all this, VM shall reveal herself to
you :)

bye
Amol



 and thus drop the pte.

This leads to a situation, where P is in the swap cache, marked as clear
(i.e. PG_dirty bit clear), while the disk image differs from the data
that is in the main memory page
frame.
I would have expected try_to_swap_out() to check the page table entries
dirty bit, and to mark the page dirty. However, I can't see any such
code in the function (I pasted the
relevant lines of code from linux 2.2.18 below).

static int try_to_swap_out(struct task_struct * tsk, struct
vm_area_struct* vma,
         unsigned long address, pte_t * page_table, int gfp_mask)
 {
         pte_t pte;
         unsigned long entry;
         unsigned long page;
         struct page * page_map;

         pte = *page_table;
         if (!pte_present(pte))
                 return 0;
         page = pte_page(pte);
         if (MAP_NR(page) >= max_mapnr)
                 return 0;
         page_map = mem_map + MAP_NR(page);

         if (pte_young(pte)) {
                 /*
                  * Transfer the "accessed" bit from the page
                  * tables to the global page map.
                  */
                 set_pte(page_table, pte_mkold(pte));
                 flush_tlb_page(vma, address);
                 set_bit(PG_referenced, &page_map->flags);
                 return 0;
         }

         if (PageReserved(page_map)
             || PageLocked(page_map)
             || ((gfp_mask & __GFP_DMA) && !PageDMA(page_map)))
                 return 0;

         /*
          * Is the page already in the swap cache? If so, then
          * we can just drop our reference to it without doing
          * any IO - it's already up-to-date on disk.
          *
          * Return 0, as we didn't actually free any real
          * memory, and we should just continue our scan.
          */
         if (PageSwapCache(page_map)) {
                 entry = page_map->offset;
                 swap_duplicate(entry);
                 set_pte(page_table, __pte(entry));
 drop_pte:
                 vma->vm_mm->rss--;
                 flush_tlb_page(vma, address);
                 __free_page(page_map);
                 return 0;
         }
............


Thanks again, with best regards
Martin

--
Supercomputing System AG          email: maletinsky@scs.ch
Martin Maletinsky                 phone: +41 (0)1 445 16 05
Technoparkstrasse 1               fax:   +41 (0)1 445 16 10
CH-8005 Zurich


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/