linux-mm.kvack.org archive mirror
 help / color / mirror / Atom feed
From: Balbir Singh <balbir@linux.vnet.ibm.com>
To: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: linux-mm@kvack.org, akpm@linux-foundation.org, mel@csn.ul.ie,
	clameter@sgi.com, riel@redhat.com, andrea@suse.de,
	a.p.zijlstra@chello.nl, eric.whitney@hp.com, npiggin@suse.de
Subject: Re: [PATCH/RFC 6/14] Reclaim Scalability: "No Reclaim LRU Infrastructure"
Date: Wed, 19 Sep 2007 11:30:16 +0530	[thread overview]
Message-ID: <46F0BAF0.2020806@linux.vnet.ibm.com> (raw)
In-Reply-To: <20070914205438.6536.49500.sendpatchset@localhost>

Lee Schermerhorn wrote:
> PATCH/RFC 06/14 Reclaim Scalability: "No Reclaim LRU Infrastructure"
> 
> Against:  2.6.23-rc4-mm1
> 
> Infrastructure to manage pages excluded from reclaim--i.e., hidden
> from vmscan.  Based on a patch by Larry Woodman of Red Hat. Reworked
> to maintain "nonreclaimable" pages on a separate per-zone LRU list,
> to "hide" them from vmscan.  A separate noreclaim pagevec is provided
> for shrink_active_list() to move nonreclaimable pages to the noreclaim
> list without over burdening the zone lru_lock.
> 
> Pages on the noreclaim list have both PG_noreclaim and PG_lru set.
> Thus, PG_noreclaim is analogous to and mutually exclusive with
> PG_active--it specifies which LRU list the page is on.  
> 
> The noreclaim infrastructure is enabled by a new mm Kconfig option
> [CONFIG_]NORECLAIM.
> 

Could we use a different name. CONFIG_NORECLAIM could be misunderstood
to be that reclaim is disabled on the system all together.


> 
> 4.  TODO:  Memory Controllers maintain separate active and inactive lists.
>     Need to consider whether they should also maintain a noreclaim list.  
>     Also, convert to use Christoph's array of indexed lru variables?
> 
>     See //TODO note in mm/memcontrol.c re:  isolating non-reclaimable
>     pages. 
> 

Thanks, I'll look into exploiting this in the memory controller.

> Index: Linux/mm/swap.c
> ===================================================================
> --- Linux.orig/mm/swap.c	2007-09-14 10:21:45.000000000 -0400
> +++ Linux/mm/swap.c	2007-09-14 10:21:48.000000000 -0400
> @@ -116,14 +116,14 @@ int rotate_reclaimable_page(struct page 
>  		return 1;
>  	if (PageDirty(page))
>  		return 1;
> -	if (PageActive(page))
> +	if (PageActive(page) | PageNoreclaim(page))

Did you intend to make this bitwise or?

> -	if (PageLRU(page) && !PageActive(page)) {
> +	if (PageLRU(page) && !PageActive(page) && !PageNoreclaim(page)) {

Since we use this even below, does it make sense to wrap it into an
inline function and call it check_page_lru_inactive_reclaimable()?


>  void lru_add_drain(void)
> @@ -277,14 +312,18 @@ void release_pages(struct page **pages, 
> 
>  		if (PageLRU(page)) {
>  			struct zone *pagezone = page_zone(page);
> +			int is_lru_page;
> +
>  			if (pagezone != zone) {
>  				if (zone)
>  					spin_unlock_irq(&zone->lru_lock);
>  				zone = pagezone;
>  				spin_lock_irq(&zone->lru_lock);
>  			}
> -			VM_BUG_ON(!PageLRU(page));
> -			__ClearPageLRU(page);
> +			is_lru_page = PageLRU(page);
> +			VM_BUG_ON(!(is_lru_page));
> +			if (is_lru_page)

This is a little confusing, after asserting that the page
is indeed in LRU, why add the check for is_lru_page again?
Comments will be helpful here.


> +#ifdef CONFIG_NORECLAIM
> +void __pagevec_lru_add_noreclaim(struct pagevec *pvec)
> +{
> +	int i;
> +	struct zone *zone = NULL;
> +
> +	for (i = 0; i < pagevec_count(pvec); i++) {
> +		struct page *page = pvec->pages[i];
> +		struct zone *pagezone = page_zone(page);
> +
> +		if (pagezone != zone) {
> +			if (zone)
> +				spin_unlock_irq(&zone->lru_lock);
> +			zone = pagezone;
> +			spin_lock_irq(&zone->lru_lock);
> +		}
> +		VM_BUG_ON(PageLRU(page));
> +		SetPageLRU(page);

> +		VM_BUG_ON(PageActive(page) || PageNoreclaim(page));
> +		SetPageNoreclaim(page);
> +		add_page_to_noreclaim_list(zone, page);

These two calls seem to be the only difference between __pagevec_lru_add
and this routine, any chance we could refactor to reuse most of the
code? Something like __pagevec_lru_add_prepare(), do the stuff and
then call __pagevec_lru_add_finish()


> +/*
> + * move_to_lru() - place @page onto appropriate lru list
> + * based on preserved page flags:  active, noreclaim, none
> + */
>  static inline void move_to_lru(struct page *page)
>  {
> -	if (PageActive(page)) {
> +	if (PageNoreclaim(page)) {
> +		VM_BUG_ON(PageActive(page));
> +		ClearPageNoreclaim(page);
> +		lru_cache_add_noreclaim(page);

I know that lru_cache_add_noreclaim() does the right thing
by looking at PageNoReclaim(), but the sequence is a little
confusing to read.


> -int __isolate_lru_page(struct page *page, int mode)
> +int __isolate_lru_page(struct page *page, int mode, int take_nonreclaimable)
>  {
>  	int ret = -EINVAL;
> 
> @@ -652,12 +660,27 @@ int __isolate_lru_page(struct page *page
>  		return ret;
> 
>  	/*
> -	 * When checking the active state, we need to be sure we are
> -	 * dealing with comparible boolean values.  Take the logical not
> -	 * of each.
> +	 * Non-reclaimable pages shouldn't make it onto the inactive list,
> +	 * so if we encounter one, we should be scanning either the active
> +	 * list--e.g., after splicing noreclaim list to end of active list--
> +	 * or nearby pages [lumpy reclaim].  Take it only if scanning active
> +	 * list.
>  	 */
> -	if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
> -		return ret;
> +	if (PageNoreclaim(page)) {
> +		if (!take_nonreclaimable)
> +			return -EBUSY;	/* lumpy reclaim -- skip this page */
> +		/*
> +		 * else fall thru' and try to isolate
> +		 */

I think we need to distinguish between the types of nonreclaimable
pages. Is it the heavily mapped pages that you pass on further?
A casual reader like me finds it hard to understand how lumpy reclaim
might try to reclaim a non-reclaimable page :-)

-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

  parent reply	other threads:[~2007-09-19  6:00 UTC|newest]

Thread overview: 77+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2007-09-14 20:53 [PATCH/RFC 0/14] Page Reclaim Scalability Lee Schermerhorn
2007-09-14 20:54 ` [PATCH/RFC 1/14] Reclaim Scalability: Convert anon_vma lock to read/write lock Lee Schermerhorn
2007-09-17 11:02   ` Mel Gorman
2007-09-18  2:41     ` KAMEZAWA Hiroyuki
2007-09-18 11:01       ` Mel Gorman
2007-09-18 14:57         ` Rik van Riel
2007-09-18 15:37       ` Lee Schermerhorn
2007-09-18 20:17     ` Lee Schermerhorn
2007-09-20 10:19       ` Mel Gorman
2007-09-14 20:54 ` [PATCH/RFC 2/14] Reclaim Scalability: convert inode i_mmap_lock to reader/writer lock Lee Schermerhorn
2007-09-17 12:53   ` Mel Gorman
2007-09-20  1:24   ` Andrea Arcangeli
2007-09-20 14:10     ` Lee Schermerhorn
2007-09-20 14:16       ` Andrea Arcangeli
2007-09-14 20:54 ` [PATCH/RFC 3/14] Reclaim Scalability: move isolate_lru_page() to vmscan.c Lee Schermerhorn
2007-09-14 21:34   ` Peter Zijlstra
2007-09-15  1:55     ` Rik van Riel
2007-09-17 14:11     ` Lee Schermerhorn
2007-09-17  9:20   ` Balbir Singh
2007-09-17 19:19     ` Lee Schermerhorn
2007-09-14 20:54 ` [PATCH/RFC 4/14] Reclaim Scalability: Define page_anon() function Lee Schermerhorn
2007-09-15  2:00   ` Rik van Riel
2007-09-17 13:19   ` Mel Gorman
2007-09-18  1:58   ` KAMEZAWA Hiroyuki
2007-09-18  2:27     ` Rik van Riel
2007-09-18  2:40       ` KAMEZAWA Hiroyuki
2007-09-18 15:04     ` Lee Schermerhorn
2007-09-18 19:41       ` Christoph Lameter
2007-09-19  0:30       ` KAMEZAWA Hiroyuki
2007-09-19 16:58         ` Lee Schermerhorn
2007-09-20  0:56           ` KAMEZAWA Hiroyuki
2007-09-14 20:54 ` [PATCH/RFC 5/14] Reclaim Scalability: Use an indexed array for LRU variables Lee Schermerhorn
2007-09-17 13:40   ` Mel Gorman
2007-09-17 14:17     ` Lee Schermerhorn
2007-09-17 14:39       ` Lee Schermerhorn
2007-09-17 18:58   ` Balbir Singh
2007-09-17 19:12     ` Lee Schermerhorn
2007-09-17 19:36       ` Balbir Singh
2007-09-17 19:36     ` Rik van Riel
2007-09-17 20:21       ` Balbir Singh
2007-09-17 21:01         ` Rik van Riel
2007-09-14 20:54 ` [PATCH/RFC 6/14] Reclaim Scalability: "No Reclaim LRU Infrastructure" Lee Schermerhorn
2007-09-14 22:47   ` Christoph Lameter
2007-09-17 15:17     ` Lee Schermerhorn
2007-09-17 18:41       ` Christoph Lameter
2007-09-18  9:54         ` Mel Gorman
2007-09-18 19:45           ` Christoph Lameter
2007-09-19 11:11             ` Mel Gorman
2007-09-19 18:03               ` Christoph Lameter
2007-09-19  6:00   ` Balbir Singh [this message]
2007-09-19 14:47     ` Lee Schermerhorn
2007-09-14 20:54 ` [PATCH/RFC 7/14] Reclaim Scalability: Non-reclaimable page statistics Lee Schermerhorn
2007-09-17  1:56   ` Rik van Riel
2007-09-14 20:54 ` [PATCH/RFC 8/14] Reclaim Scalability: Ram Disk Pages are non-reclaimable Lee Schermerhorn
2007-09-17  1:57   ` Rik van Riel
2007-09-17 14:40     ` Lee Schermerhorn
2007-09-17 18:42       ` Christoph Lameter
2007-09-14 20:54 ` [PATCH/RFC 9/14] Reclaim Scalability: SHM_LOCKED pages are nonreclaimable Lee Schermerhorn
2007-09-17  2:18   ` Rik van Riel
2007-09-14 20:55 ` [PATCH/RFC 10/14] Reclaim Scalability: track anon_vma "related vmas" Lee Schermerhorn
2007-09-17  2:52   ` Rik van Riel
2007-09-17 15:52     ` Lee Schermerhorn
2007-09-14 20:55 ` [PATCH/RFC 11/14] Reclaim Scalability: swap backed pages are nonreclaimable when no swap space available Lee Schermerhorn
2007-09-17  2:53   ` Rik van Riel
2007-09-18 17:46     ` Lee Schermerhorn
2007-09-18 20:01       ` Rik van Riel
2007-09-19 14:55         ` Lee Schermerhorn
2007-09-18  2:59   ` KAMEZAWA Hiroyuki
2007-09-18 15:47     ` Lee Schermerhorn
2007-09-14 20:55 ` [PATCH/RFC 12/14] Reclaim Scalability: Non-reclaimable Mlock'ed pages Lee Schermerhorn
2007-09-14 20:55 ` [PATCH/RFC 13/14] Reclaim Scalability: Handle Mlock'ed pages during map/unmap and truncate Lee Schermerhorn
2007-09-14 20:55 ` [PATCH/RFC 14/14] Reclaim Scalability: cull non-reclaimable anon pages in fault path Lee Schermerhorn
2007-09-14 21:11 ` [PATCH/RFC 0/14] Page Reclaim Scalability Peter Zijlstra
2007-09-14 21:42   ` Linus Torvalds
2007-09-14 22:02     ` Peter Zijlstra
2007-09-15  0:07       ` Linus Torvalds
2007-09-17  6:44 ` Balbir Singh

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=46F0BAF0.2020806@linux.vnet.ibm.com \
    --to=balbir@linux.vnet.ibm.com \
    --cc=a.p.zijlstra@chello.nl \
    --cc=akpm@linux-foundation.org \
    --cc=andrea@suse.de \
    --cc=clameter@sgi.com \
    --cc=eric.whitney@hp.com \
    --cc=lee.schermerhorn@hp.com \
    --cc=linux-mm@kvack.org \
    --cc=mel@csn.ul.ie \
    --cc=npiggin@suse.de \
    --cc=riel@redhat.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox