Re: Should a swapped out page be deleted from swap cache?

From: Sha Zhengju
Date: Tue Feb 19 2013 - 04:38:32 EST


On Tue, Feb 19, 2013 at 2:53 PM, Li Haifeng <omycle@xxxxxxxxx> wrote:
> 2013/2/19 Will Huck <will.huckk@xxxxxxxxx>:
>> On 02/19/2013 10:04 AM, Li Haifeng wrote:
>>>
>>> 2013/2/19 Hugh Dickins <hughd@xxxxxxxxxx>
>>>>
>>>> On Mon, 18 Feb 2013, Li Haifeng wrote:
>>>>
>>>>> For explain my question, the two points should be displayed as below.
>>>>>
>>>>> 1. If an anonymous page is swapped out, this page will be deleted
>>>>> from swap cache and be put back into buddy system.
>>>>
>>>> Yes, unless the page is referenced again before it comes to be
>>>> deleted from swap cache.
>>>>
>>>>> 2. When a page is swapped out, the sharing count of swap slot must not
>>>>> be zero. That is, page_swapcount(page) will not return zero.
>>>>
>>>> I would not say "must not": we just prefer not to waste time on swapping
>>>> a page out if its use count has already gone to 0. And its use count
>>>> might go down to 0 an instant after swap_writepage() makes that check.
>>>>
>>> Thanks for your reply and patience.
>>>
>>> If a anonymous page is swapped out and comes to be reclaimable,
>>> shrink_page_list() will call __remove_mapping() to delete the page
>>> swapped out from swap cache. Corresponding code lists as below.
>>
>>
>> I'm not sure if
>> if (PageAnon(page) && !PageSwapCache(page)) {
>> .................
>> }
>> will add the page to swap cache again.
>>
>
> Adding the page to swap cache is the first stage of memory reclaiming.
>
> When an anonymous page will be reclaimed, it should be swapped out. If
> it's not in the swap cache, it will insert into swap cache first and
> set the bit of PG_swapcache on page->flags. Then, it will be swapped
> out by try_to_unmap(). After it's swapped out, and no processes swap

Swapout(writing to swap disk) is not done by try_to_unmap() which only
tries to remove all page table mappings to a page. Before unmapping,
add_to_swap() will set the swap cache page dirty and it will be
written out by pageout()->swap_writepage().


Thanks,
Sha


> in this page, the page will be deleted from swap cache and be put into
> buddy system.
>
> Please correct me if my understanding is wrong.
>
> Regards,
> Haifeng Li.
>
>>>
>>> 765 static unsigned long shrink_page_list(struct list_head *page_list,
>>> 766 struct mem_cgroup_zone *mz,
>>> 767 struct scan_control *sc,
>>> 768 int priority,
>>> 769 unsigned long *ret_nr_dirty,
>>> 770 unsigned long
>>> *ret_nr_writeback)
>>> 771 {
>>> ...
>>> 971 if (!mapping || !__remove_mapping(mapping, page))
>>> 972 goto keep_locked;
>>> 973
>>> 974 /*
>>> 975 * At this point, we have no other references and
>>> there is
>>> 976 * no way to pick any more up (removed from LRU,
>>> removed
>>> 977 * from pagecache). Can use non-atomic bitops now
>>> (and
>>> 978 * we obviously don't have to worry about waking
>>> up a process
>>> 979 * waiting on the page lock, because there are no
>>> references.
>>> 980 */
>>> 981 __clear_page_locked(page);
>>> 982 free_it:
>>> 983 nr_reclaimed++;
>>> 984
>>> 985 /*
>>> 986 * Is there need to periodically free_page_list? It
>>> would
>>> 987 * appear not as the counts should be low
>>> 988 */
>>> 989 list_add(&page->lru, &free_pages);
>>> 990 continue;
>>>
>>> Please correct me if my understanding is wrong.
>>>
>>> Thanks.
>>>>>
>>>>> Are both of them above right?
>>>>>
>>>>> According the two points above, I was confused to the line 655 below.
>>>>> When a page is swapped out, the return value of page_swapcount(page)
>>>>> will not be zero. So, the page couldn't be deleted from swap cache.
>>>>
>>>> Yes, we cannot free the swap as long as its data might be needed again.
>>>>
>>>> But a swap cache page may linger in memory for an indefinite time,
>>>> in between being queued for write out, and actually being freed from
>>>> the end of the lru by memory pressure.
>>>>
>>>> At various points where we hold the page lock on a swap cache page,
>>>> it's worth checking whether it is still actually needed, or could
>>>> now be freed from swap cache, and the corresponding swap slot freed:
>>>> that's what try_to_free_swap() does.
>>>
>>> I do agree. Thanks again.
>>>>
>>>> Hugh
>>>>
>>>>> 644 * If swap is getting full, or if there are no more mappings of
>>>>> this page,
>>>>> 645 * then try_to_free_swap is called to free its swap space.
>>>>> 646 */
>>>>> 647 int try_to_free_swap(struct page *page)
>>>>> 648 {
>>>>> 649 VM_BUG_ON(!PageLocked(page));
>>>>> 650
>>>>> 651 if (!PageSwapCache(page))
>>>>> 652 return 0;
>>>>> 653 if (PageWriteback(page))
>>>>> 654 return 0;
>>>>> 655 if (page_swapcount(page))//Has referenced by other swap
>>>>> out
>>>>> page.
>>>>> 656 return 0;
>>>>> 657
>>>>> 658 /*
>>>>> 659 * Once hibernation has begun to create its image of
>>>>> memory,
>>>>> 660 * there's a danger that one of the calls to
>>>>> try_to_free_swap()
>>>>> 661 * - most probably a call from __try_to_reclaim_swap()
>>>>> while
>>>>> 662 * hibernation is allocating its own swap pages for the
>>>>> image,
>>>>> 663 * but conceivably even a call from memory reclaim - will
>>>>> free
>>>>> 664 * the swap from a page which has already been recorded in
>>>>> the
>>>>> 665 * image as a clean swapcache page, and then reuse its
>>>>> swap
>>>>> for
>>>>> 666 * another page of the image. On waking from hibernation,
>>>>> the
>>>>> 667 * original page might be freed under memory pressure,
>>>>> then
>>>>> 668 * later read back in from swap, now with the wrong data.
>>>>> 669 *
>>>>> 670 * Hibration suspends storage while it is writing the
>>>>> image
>>>>> 671 * to disk so check that here.
>>>>> 672 */
>>>>> 673 if (pm_suspended_storage())
>>>>> 674 return 0;
>>>>> 675
>>>>> 676 delete_from_swap_cache(page);
>>>>> 677 SetPageDirty(page);
>>>>> 678 return 1;
>>>>> 679 }
>>>>>
>>>>> Thanks.
>>>
>>> --
>>>
>>> To unsubscribe, send a message with 'unsubscribe linux-mm' in
>>> the body to majordomo@xxxxxxxxxx For more info on Linux MM,
>>> see: http://www.linux-mm.org/ .
>>> Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
>>
>>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/